diff --git a/parse/train/1Kof-nkmQB8/1Kof-nkmQB8.md b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8.md new file mode 100644 index 0000000000000000000000000000000000000000..a5c8910808362b968ec2bc35624176aa0e16ba1c --- /dev/null +++ b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8.md @@ -0,0 +1,272 @@ +# Collaborating with Humans without Human Data + +DJ Strouse⇤, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett⇤ DeepMind {strouse, kevinrmckee, botvinick, edwardhughes, reverett}@deepmind.com + +# Abstract + +Collaborating with humans requires rapidly adapting to their individual strengths, weaknesses, and preferences. Unfortunately, most standard multi-agent reinforcement learning techniques, such as self-play (SP) or population play (PP), produce agents that overfit to their training partners and do not generalize well to humans. Alternatively, researchers can collect human data, train a human model using behavioral cloning, and then use that model to train “human-aware” agents (“behavioral cloning play”, or BCP). While such an approach can improve the generalization of agents to new human co-players, it involves the onerous and expensive step of collecting large amounts of human data first. Here, we study the problem of how to train agents that collaborate well with human partners without using human data. We argue that the crux of the problem is to produce a diverse set of training partners. Drawing inspiration from successful multi-agent approaches in competitive domains, we find that a surprisingly simple approach is highly effective. We train our agent partner as the best response to a population of self-play agents and their past checkpoints taken throughout training, a method we call Fictitious Co-Play (FCP). Our experiments focus on a two-player collaborative cooking simulator that has recently been proposed as a challenge problem for coordination with humans. We find that FCP agents score significantly higher than SP, PP, and BCP when paired with novel agent and human partners. Furthermore, humans also report a strong subjective preference to partnering with FCP agents over all baselines. + +# 1 Introduction + +Generating agents which collaborate with novel partners is a longstanding challenge for Artificial Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31, $\boxed { 6 6 }$ is especially important in situations where an AI must generalize to novel human partners [6, 61]. Many successful approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly [12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach humanlevel using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64]. This begs the question: Can model-free RL without human data generate agents that can collaborate with novel humans? + +We seek an answer to this question in the space of common-payoff games, where all agents work towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from repeated games played against copies of itself, does not produce agents that generalize well to novel co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to coordinate with themselves, and so make for brittle and stubborn collaborators with new partners who act differently. Population play (PP) trains a population of agents, all of whom interact with each other $\pmb { \| 3 9 \| }$ . While PP can generate agents capable of cooperation with humans in competitive team games $\pmb { \Vert 3 4 \Vert }$ , it still fails to produce robust partners for novel humans in pure common-payoff settings [12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing strategic diversity and producing agents not so different from self-play $\dot { \lVert 2 4 \rVert }$ . + +![](images/70b2b8c1e9478aed36abef766805e05c03561ba29533d4a7de7ad061fa2d261f.jpg) +Figure 1: In this work, we evaluate a variety of agent training methods (Section 2) in zero-shot coordination with agents (Section 4). We then run a human-agent collaborative study designed to elicit human preferences over agents (Section 5) + +Our approach starts with the intuition that the key to producing robust agent collaborators is exposure to diverse training partners. We find that a surprisingly simple strategy is effective in generating sufficient diversity. We train $N$ self-play agents varying only their random seed for neural network initialization. Periodically during training, we save agent “checkpoints” representing their strategy at that point in time. Then, we train an agent partner as the best-response to both the fully-trained agents and their past checkpoints. The different checkpoints simulate different skill levels, and the different random seeds simulate breaking symmetries in different ways. We refer to this agent training procedure as Fictitious $\mathbf { C o }$ -Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69]. + +We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator. Based on the game Overcooked $\mathbb { \left. \boldsymbol { \ 2 5 } \right. }$ , it has recently been proposed as a coordination challenge for AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel humans was achieved in $[ \mathbb { 1 2 } ]$ via behavioral cloning (BC) of human data. More precisely, BC was used to produce models that can stand in as human proxies during training in simulation, a method we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to both novel agent and human partners, and that humans express a significant preference for partnering with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human data for training, while achieving better outcomes for humans at test time. + +We summarize the novel contributions of this paper as follows: + +1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with humans (Section 2.1). +2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot coordination with a variety of held-out agents (Section 4.2). +3. We propose a rigorous human-agent interaction study with behavioral analysis and participant feedback (Section $5 . 1 )$ . +4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task score and in human partner preference (Section $\underline { { \overline { { 5 . 2 } } } } )$ . + +# 2 Methods + +# 2.1 Fictitious Co-Play (FCP) + +Diverse training conditions have been shown to make agents more robust, from environmental variations (i.e. domain randomization $\textcircled { 1 5 4 } , \textcircled { 5 6 } , \textcircled { 6 7 } \textcircled { 1 }$ to heterogeneity in training partners $\mathbb { \left. \boldsymbol { \mathfrak { G } } \boldsymbol { \mathfrak { g } } \right. }$ . We seek to train agents that are robust partners for humans in common-payoff games, and so extend this line of work to that setting. + +One important challenge in collaborating with novel partners is dealing with symmetries $\pmb { \mathbb { B } } \mathbf { \mathbb { 1 } }$ . For example, two agents A and B facing each other may move past each other by A going left and B going right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between these conventions if a human clearly prefers one over the other. A second important challenge is dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled partners, as well as partners who are still learning. + +![](images/2372b76b2e16cbfb2db5439982c244e9b3e60b30aff99699aa5043d431598f6f.jpg) +Figure 2: The four agent training methods we evaluate in this work. Self-play (SP) where an agent learns with itself, population-play (PP) where a population of agents are co-trained together, and behavioral cloning play (BCP) where data from human games is used to create a behaviorally cloned agent with which an RL agent is then trained. In our method, Fictitious Co-Play (FCP), $N$ self-play agents are trained independently and checkpointed throughout training. An agent is then trained to best respond to the entire population of SP agents and their checkpoints. + +Fictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both of these challenges (Figure $2 ,$ right). In the first stage, we train a diverse pool of partners. To allow the pool to represent different symmetry breaking conventions, we train $N$ partner agents in self-play. Since these partners are trained independently, they can arrive at different arbitrary conventions for breaking symmetries. To allow the pool to represent different skill levels, we use multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost. + +In the second stage, we train an FCP agent as the best response to the pool of diverse partners created in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the lead of human partners, and learn a general policy across a range of strategies and skills. We call our method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69]. + +# 2.2 Baselines and ablations + +We compare FCP agents to the three baseline training methods listed below, each varying only in their set of training partners, with the RL algorithm and architecture consistent across all agents: + +1. Self-play (SP), where agents learn solely through interaction with themselves. +2. Population-play (PP), where a population of agents are co-trained through random pairings. +3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12]. + +We also evaluate three variations on FCP to better understand the conditions for its success: + +1. To test the importance of including past checkpoints in training, we evaluate an ablation of FCP in which agents are trained only with the converged checkpoints of their partners $\mathrm { F C P } _ { - T }$ for “FCP minus time”). 2. To test whether FCP would benefit from additional diversity in its partner population, we evaluate an augmentation of FCP in which the population of SP partners varies not just in random seed, but also in architecture $\operatorname { F C P } _ { + A }$ for “FCP plus architectural variation”). 3. To test whether architectural variation can serve as a full replacement for playing with past checkpoints, we evaluate the combination of both modifications $( \mathrm { F C P } _ { - T , + A } )$ . + +# 2.3 Environment + +Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment in Carroll et al. [12]. For full details, see Appendix A. + +In this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential high-level actions to which both players can contribute: collecting tomatoes, depositing them into cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering it. Upon a successful delivery, both players are rewarded equally. + +To effectively complete the task, players must learn to navigate the kitchen and interact with objects in the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them. This environment therefore presents the challenges of both movement and strategic coordination. + +Each player observes an egocentric RGB view of the world, and at every step can perform one of six actions: stand still, move {up, down, left, right}, interact. The behavior ofLÈ¡áyҏįŘįºµò‹µÒ¡y® interact varies based on the cell which the player is facing (e.g. place tomato on counter). + +![](images/48e8af75423e0fa5e5c23e2088c49716c4c6044ee3579f26cdd9911dbb4b4925.jpg) +Figure 3: The Overcooked environment: a two-player common-payoff game in which players must coordinate to cook and deliver soup. + +![](images/2b3253fcaac0677dbb15cc1ed5aa0663e561ff52605ba15bd61605f96c716872.jpg) +Figure 4: Layouts: the kitchens which agents and humans play in, each emphasizing different coordination strategies. Highlighted in bold are the terms used to refer to each in the rest of this paper. + +# 2.4 Implementation details + +Here we highlight several key implementation details for our training methods. For full details, including the architectures, hyperparameters, and compute used, please see Appendix B. + +For our reinforcement learning agents, we use the V-MPO $\boldsymbol { \left[ \left[ 6 5 \right] \right] }$ algorithm along with a ResNet [26] plus LSTM $\mathbb { \left| \bigstar \bigstar \right\| }$ architecture which we found led to optimal behavior across all layouts. Agents are trained using a distributed set of environments running in parallel $\textcircled { 1 1 7 }$ , each sampling two agents from the training population to play together every episode. + +Both PP and FCP are trained with a population size of $N = 3 2$ agents which are sampled uniformly. For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden: (1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert agent), and (3) at the middle of training, defined as when the agent reaches $50 \%$ of its final reward (i.e. an average-skilled agent). When varying architecture for the training partners of the $\operatorname { F C P } _ { + A }$ and $\mathrm { F C P } _ { - T , + A }$ variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4 combinations, leaving the total population size of $N = 3 2$ unchanged, ensuring a fair comparison. + +To train agents via behavioral cloning $\left[ \left[ 5 8 \right] \right]$ , we use the open-source Acme $\pmb { \mathbb { B } } 0 \|$ to learn a policy from human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for agent-agent evaluation. Following Carroll et al. $\bar { \lVert 1 2 \rVert }$ , we use a set of feature-based observations for the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts (asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced). + +# 3 Related work + +Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play $ { \mathbb { B } } , { \mathbb { G } } 6 { \mathbb { I } }$ , also known as zero-shot coordination $\bar { \mathbb { B } } \mathbb { I }$ . Prior work based in game-theoretic settings has suggested the benefits of planning $\pmb { \mathbb { Z } 1 }$ , online learning $\mathbb { \left| \left. \sum \right. { ] } \right| }$ , and novel solution concepts $\bar { \mathbb { D } }$ , to name a few examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale to more complex gridworld or continuous control settings, leading to work on hierarchical social planning $\widehat { \left\| 3 6 \right\| }$ , adapting to existing social conventions $\checkmark$ , trajectory diversity $\lVert \rVert \dot { \boldsymbol { \mathrm { E } } } \rVert$ , and theory of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization to held-out agent partners as a proxy for human co-players. + +Collaborative play with novel humans has been evaluated more actively in the context of training agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the stateof-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich gridworld environment, as measured by the challenge tasks in Carroll et al. [12]. + +Diversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents that train with behaviorally diverse populations of game partners tend to demonstrate stronger performance than their self-play counterparts. For example, across a range of multi-agent games, generalization to held-out populations can be improved by training larger and more diverse populations [13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be optimized through selective matchmaking between increasingly diverse agents $[ \overbrace { 2 4 } , \overbrace { 3 9 } , \overbrace { 6 9 } ]$ . + +Despite the increased focus on improving multi-agent performance, evaluation has typically been constrained to agent-agent settings. High-performing agents have infrequently been evaluated with humans, particularly in non-competitive domains $\bar { \lVert 1 6 rVert }$ . We add to this growing literature, showing that training with diversity is a powerful approach for effective human-agent collaboration. + +Human-agent interaction In recent years, increased attention has been directed toward designing machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also $\boxed { \boxed { 1 6 } }$ for a broader review on Cooperative AI). Tylkin et al. $\lVert \rVert$ is particularly notable in also demonstrating that partially trained agents can be useful learning targets for human helpers, although in a different domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to our agents’ performance (Tables 1 and 2 and Figure 7b). + +A key preceding entry in this research area is Carroll et al. $\pmb { \mathbb { I } }$ , who similarly investigated humanagent coordination in Overcooked. We use their method (BCP) as a baseline throughout our experiments (Section $\boxed { 2 . 2 }$ . Relative to BCP, our approach removes the need for the expensive step of human data collection for agent training. Furthermore, through our novel human-agent experimental design, we go beyond objective performance metrics to compare the subjective preferences that agents generate. For a detailed comparison of methods and results, see Appendix E. + +# 4 Zero-shot coordination with agents + +In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents. + +# 4.1 Evaluation method: collaborative evaluation with agent partners + +Our primary concern in this work is generalization to novel human partners (as investigated in Section $\textcircled{5}$ . However, just as collecting human-human data for behavioral cloning is expensive, so too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent partners as a cheap proxy of performance with humans. This is then used to guide our model selection process, allowing us to be more targeted with the agents we select for our human-agent evaluations. + +We evaluate with three held-out populations: + +1. A BC model trained on human data, $H _ { \mathrm { p r o x y } }$ , intended as a proxy of generalization to humans, as done by Carroll et al. $[ \overbrace { | 1 2 | }$ . 2. A set of self-play agents varying in seed, architecture, and training time (specifically, heldout seeds of the $N = 3 2$ partners trained for the $\mathrm { F C P } _ { + A }$ agent; see Section 2.4). These are intended to test generalization to a diverse yet still skillful population. 3. Randomly initialized agents intended to test generalization to low-skill partners. + +For all results, we report the average number of deliveries made by both players within an episode, aggregated across the 5 different layouts from Figure $^ 4$ (with the per-layout results reported in Appendix $\underline { { \overline { { ( \mathrm { C . 2 } ) } } } }$ . We estimate mean and standard deviation across 5 random seeds. For each seed, we evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair. + +# 4.2 Results + +# Finding 1: FCP significantly outperforms all baselines + +To begin, we compare our FCP agent and the baselines when partnered with the three held-out populations introduced above. As can be seen in Figure $\boxed { 5 }$ FCP significantly outperforms all baselines when partnered with all three held-out populations. Notably, it performs better than BCP with $H _ { \mathrm { p r o x y } }$ even though BCP trains with such a model and FCP does not. Similar to Carroll et al. $\mathbb { \lVert 1 2 \rVert }$ , we find that BCP significantly outscores SP. + +When paired with a randomly initialized partner which behaves suboptimally, we see an even greater difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of such agents, it may not be surprising that it does so well with partners that behave poorly. However, what is surprising is how brittle the other training methods are. This suggests that they may not perform well with humans who are not highly skilled players, which we will see in Section 5. + +![](images/8083dcac415b6ef9624e1aa4f428a0a33c3e9d47ce8903df3798f3e80ef0427c.jpg) +Figure 5: Agent-agent collaborative evaluation: Performance of each agent when partnered with each of the held-out populations (Section $4 . 1 )$ in episodes of length $T = 5 4 0$ . Importantly, FCP scores higher than all baselines with a variety of test partners. Error bars represent standard deviation over five random training seeds. Plots aggregate data across kitchen layouts; results calculated by individual layout can be found in Appendix C.2. + +Finding 2: Training with past checkpoints is the most beneficial variation for performance Next, we investigate how the different training partner variations influence FCP’s performance. In particular, we separately ablate the past checkpoints $( T )$ and architecture $( A )$ variations, evaluating them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1. Comparing the FCP and $\mathrm { F C P } _ { - T }$ columns, we see that removing past checkpoints from training significantly reduces performance. Comparing the FCP and $\operatorname { F C P } _ { + A }$ columns, we see that adding architectural variation to the training population offers no improvement over training with past + +
PartnerFCPFCP-TFCP+AFCP-T,+A
Hproxy10.6± 0.54.7± 0.49.9±0.67.0±0.8
Diverse SP11.2 ± 0.16.9 ± 0.111.1 ± 0.48.6 ± 0.4
Random8.6± 0.21.0 ± 0.18.4±0.43.2 ± 0.5
+ +Table 1: Ablation results: Performance of each variation of FCP – training with past partner checkpoints $T$ for time) and adding partner variation in architecture $( A )$ . Scores are mean deliveries with standard deviation over 5 random seeds. Notably, we find that the inclusion of past checkpoints is essential for strong performance $( \mathrm { F C P } > \mathrm { F C P } _ { - T }$ ), and additionally including architectural variation does not improve performance $( \mathrm { F C P } \approx \mathrm { F C P } _ { + A . }$ ). However, architectural variation is better than no variation, improving performance when past checkpoints are not available $( \mathrm { F C P } _ { - T , + A } > \mathrm { F C P } _ { - T , }$ ). + +checkpoints. However, comparing the $\mathrm { F C P } _ { - T }$ and $\mathrm { F C P } _ { - T , + A }$ columns, we see that without training with past checkpoints, architectural variation in the population does improve performance. + +# 5 Zero-shot coordination with humans + +Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative play with human partners. + +![](images/5856cbdd90eeef2a7a6093668472063862a759746c339aa0324e78efb5d67b22.jpg) +Figure 6: Human-agent collaborative study: For our human-agent collaboration study, we recruited participants online to play games with FCP and baseline agents. Participants played a randomized sequence of episodes with different agent partners and kitchen layouts. After every two episodes, participants reported the direction and strength of their preference between their last two partners. + +# 5.1 Evaluation method: collaborative evaluation with human participants + +To test how effectively FCP’s performance generalizes to human partners, we recruited participants from Prolific $\mathbb { 1 1 8 } , \lvert 5 5 \rvert$ for an online collaboration study $N = 1 1 4$ ; $3 7 . 7 \%$ female, $5 9 . 6 \%$ male, $1 . 8 \%$ nonbinary; median age between 25–34 years). We used a within-participant design for the study: each participant played with a full cohort of agents (i.e. generated through every training method). This design allowed us to evaluate both objective performance as well as subjective preferences. + +Participants first read game instructions and played a short tutorial episode guiding them through the dish preparation sequence (see Appendix $\dot { \mathrm { \bf D } } . 1 . 1$ for instruction text and study screenshots). Participants then played 20 episodes with a randomized sequence of agent partners and kitchen layouts. Episodes lasted $T = 3 0 0$ steps (1 minute) each. After every two episodes, participants reported their preference over the agent partners from those episodes on a five-point Likert-type scale. After playing all 20 episodes, participants completed a debrief questionnaire collecting standard demographic information and open-ended feedback on the study. Our statistical analysis below primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D for additional details of our study design and analysis, including independent ethical review. + +# 5.2 Results + +Finding 1: FCP coordinates best with humans, achieving the highest score across maps To begin, we compare the objective team performance supported by our FCP and baseline agents. The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration: + +the FCP-human teams significantly outperform all other agent-human teams, achieving the highest average scores across maps, every $p < 0 . 0 0 1$ (Figure $\mathrm { 7 a ) }$ , while performing as well as or better than the other teams on each individual map (see Appendix $\mathbf { D } . 3 )$ . Echoing the results from our agent-agent ablation experiments (Table $\perp )$ , the inclusion of past checkpoints in training proves critical to FCP’s strong performance, $p < 0 . 0 { \overline { { 0 1 } } }$ (Figure $\textcircled { 7 6 }$ . Similar to Carroll et al. $[ \left[ 1 2 \right] ]$ , we find that BCP outscores SP when collaborating with human players, $p < 0 . 0 0 1$ . + +# Finding 2: Participants prefer FCP over all baselines + +FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences. Participants expressed a significant preference for FCP partners over all other agents, including BCP, with every $p < 0 . 0 5$ (Figure $\dot { \bigtriangledown } \dot { \mathbf { c } } { \big \rVert }$ . Notably, while human-BCP and human-PP teams did not significantly differ in their completed deliveries, participants reported significantly preferring BCP over PP, $p = 0 . 0 0 3$ , highlighting the informativeness of our subjective analysis. + +![](images/14e048153d10325b3617f94b393393ed3ab6cf3e38198f8791acf0fc83b051df.jpg) +Figure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from humanagent play in episodes of length $T = 3 0 0$ . Error bars represents $9 5 \%$ confidence intervals, calculated over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can be found in Appendix D.3. + +# 5.3 Exploratory behavioral analysis + +To better understand how the human-agent scores and preferences may have arisen, here we analyze the resulting action trajectories of each human and agent player in our experiment. + +![](images/65959c0ab23a18c0f8de72bc451d199e7d9d9cc2698e7ca2ac4f320e26dac829.jpg) + +Figure 8: Behavioral analysis: (a) FCP is able to move most frequently $3 5 \%$ of the time), corresponding to the best movement coordination with human partners. (b) FCP exhibits the most equal preferences over cooking pots (0.11 difference), aligning with human preferences. Values are calculated as the absolute difference in preferences between the two pots; 1 indicates that the player only uses one of the two available pots, while 0 indicates that the player uses both pots equally. + +# Finding 1: FCP exhibits the best movement coordination with humans + +First, we investigate how much each player moves in an episode (Figure $\textcircled { 8 \mathrm { a } }$ , where moving in a higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner. Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP. + +(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its movement strategy with its partner. This result was also reported by human participants, for example: “I noticed that some of my partners seemed to know they needed to move around me, while others seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples). + +# Finding 2: FCP’s preferences over cooking pots aligns best with that of humans + +Next, we investigate whether there was a preference for a specific cooking pot in the layouts which included two cooking pots (Figure $\textcircled { 8 6 }$ . To do this, we calculate the difference in the number of times each pot was used by each player, where a high value indicates a strong preference for one pot and a low value indicates more equal preference for the two pots. + +As can be seen in the FCP column, our agent typically has the most aligned preferences with that of humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers one cooking pot over the other $5 5 . 5 \%$ of the time (i.e. a 0.11 point difference). In contrast, all other agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes poorly to typical human behavior of using both pots, leading to worse performance. + +# 6 Discussion + +Summary In this work, we investigated the challenging problem of zero-shot collaboration with humans without using human data in the training pipeline. To accomplish this, we introduced Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set of training partners. We found that FCP agents scored significantly higher than all baselines when partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent experimental design, we also found that humans reported a strong subjective preference to partnering with FCP agents over all baselines. + +Limitations and future work Our method currently relies on the manual process of initially training and selecting a diverse set of partners. This is not only time consuming, but also prone to researcher biases that may negatively influence the behavior of the created agents. Additionally, while we found FCP with a partner population size of $N = 3 2$ sufficient here, for more complex games, FCP may require an unrealistically large partner population size to represent sufficiently diverse strategies. To address these concerns, methods for automatically generating partner diversity for common-payoff games may be important. Possibilities include adaptive population matchmaking as been used in competitive zero-sum games $\mathbb { \lVert 6 9 \rVert }$ , as well as auxiliary objectives that explicitly encourage behavioral diversity [19, 45, 46]. + +Our method requires a known and fixed reward function. We also focus on one domain in order to compare with prior work which has argued that human-in-the-loop training is necessary. Consequently, the resulting agents are only designed to adaptively collaborate on a single task, and not to infer human preferences in general $\textcircled { 1 1 } \textcircled { 3 3 } \textcircled { 5 9 }$ . Moreover, if a task’s reward function is poorly aligned with how humans approach the task, our method may well produce subpar partners, as would any method without access to human data. Thus, additional domains and tasks should be studied to better understand how our method generalizes. Targeted experiments to test specific forms of generalization may be especially helpful in this regard $\overline { { [ 3 8 ] } }$ , as could approaches that procedurally generate environment layouts requiring diverse solutions $\pmb { \mathbb { Z } } 2 \mathbf { l }$ + +Finally, it may be possible to produce even stronger agent assistants by combining the strengths of FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al. $\textcircled { \lvert 3 8 \rvert }$ recently demonstrated that modifying BCP to train with multiple BC partners produces more robust collaboration with held-out agents, a finding that would be interesting to test with human partners. + +Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner expresses preferences or intentions to harm others). This could therefore produce negative societal effects either if training leads to poor alignment or if agents are optimized for harmful metrics. + +One potential strategy for mitigating these risks is the use of human preference data [15]. Such data could be used to fine-tune and filter trained agents before deployment, encouraging better alignment with human values. A key question in this line of research is how human preference data should be aggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human beliefs and perceptions of AI $\lVert \overline { { 4 8 } } \rVert$ , and how they steer human-agent interaction, would help inform agent design for positive societal impact. For instance, developers could incorporate specific priors into agents to reinforce tendencies for fair outcomes $\pm \mathbb { Z } 0 . \pm \mathbb { B } 2 \mathbb { I }$ . + +Conclusion We proposed a method which is both effective at collaborating with humans and simple to implement. We also presented a rigorous and general methodology for evaluating with humans and eliciting their preferences. Together, these establish a strong foundation for future research on the important challenge of human-agent collaboration for benefiting society. + +# Acknowledgements + +The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions. + +# Funding disclosure + +This work was funded solely by DeepMind. The authors declare no competing interests. + +# References + +[1] J. Abramson, A. Ahuja, I. Barr, A. Brussee, F. Carnevale, M. Cassin, R. Chhaparia, S. Clark, B. Damoc, A. Dudzik, et al. Imitating interactive intelligence. arXiv preprint arXiv:2012.05672, 2020. +[2] S. V. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method for ad hoc coordination in multiagent systems. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2013. +[3] B. Baker. Emergent reciprocity and team formation from randomized uncertain social preferences. In Neural Information Processing Systems (NeurIPS), 2020. +[4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin, S. Moitra, E. Hughes, et al. The Hanabi challenge: A new frontier for AI research. Artificial Intelligence, 280:103216, 2020. +[5] S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with new teammates. Artificial Intelligence, 242:132–171, 2017. +[6] A. Bauer, D. Wollherr, and M. Buss. Human–robot collaboration: A survey. International Journal of Humanoid Robotics, 5(01):47–66, 2008. +[7] G. W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and allocation, 13(1):374–376, 1951. +[8] N. Brown and T. Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science, 359(6374):418–424, 2018. +[9] N. Brown and T. Sandholm. Superhuman AI for multiplayer poker. Science, 365(6456):885–890, 2019. +[10] K. Bullard, F. Meier, D. Kiela, J. Pineau, and J. Foerster. Exploring zero-shot emergent communication in embodied multi-agent populations. arXiv preprint arXiv:2010.15896, 2020. +[11] K. Bullard, D. Kiela, J. Pineau, and J. Foerster. Quasi-equivalence discovery for zero-shot emergent communication. arXiv preprint arXiv:2103.08067, 2021. +[12] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the utility of learning about humans for human-AI coordination. In Neural Information Processing Systems (NeurIPS), 2019. +[13] R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Investigating partner diversification methods in cooperative multi-agent deep reinforcement learning. In International Conference on Neural Information Processing, 2020. +[14] R. Choudhury, G. Swamy, D. Hadfield-Menell, and A. D. Dragan. On the utility of model learning in HRI. In ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2019. +[15] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 4302–4310, 2017. +[16] A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and T. Graepel. Open problems in cooperative AI. arXiv preprint arXiv:2012.08630, 2020. +[17] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu, T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable distributed deep-RL with importance weighted actor-learner architectures. In International Conference on Machine Learning (ICML), 2018. +[18] P. Eyal, R. David, G. Andrew, E. Zak, and D. Ekaterina. Data quality of platforms and panels for online behavioral research. Behavior Research Methods, pages 1–20, 2021. +[19] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations (ICLR), 2019. +[20] E. Fehr and K. M. Schmidt. A theory of fairness, competition, and cooperation. The Quarterly Journal of Economics, 114(3):817–868, 1999. +[21] J. Foerster, F. Song, E. Hughes, N. Burch, I. Dunning, S. Whiteson, M. Botvinick, and M. Bowling. Bayesian action decoder for deep multi-agent reinforcement learning. In International Conference on Machine Learning (ICML), 2019. +[22] M. Fontaine, Y.-C. Hsu, Y. Zhang, B. Tjanaka, and S. Nikolaidis. On the importance of environments in human-robot coordination. In Robotics: Science and Systems (RSS), 2021. +[23] I. Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30(3):411–437, 2020. +[24] M. Garnelo, W. M. Czarnecki, S. Liu, D. Tirumala, J. Oh, G. Gidel, H. van Hasselt, and D. Balduzzi. Pick your battles: Interaction graphs as population-level objectives for strategic diversity. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. +[25] Ghost Town Games. Overcooked, 2016. +[26] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. +[27] J. Heinrich and D. Silver. Deep reinforcement learning from self-play in imperfect-information games. In NIPS Deep Reinforcement Learning (DRL) Workshop, 2016. +[28] J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In International Conference on Machine Learning (ICML), 2015. +[29] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8): 1735–1780, 1997. +[30] M. Hoffman, B. Shahriari, J. Aslanides, G. Barth-Maron, F. Behbahani, T. Norman, A. Abdolmaleki, A. Cassirer, F. Yang, K. Baumli, S. Henderson, A. Novikov, S. G. Colmenarejo, S. Cabi, C. Gulcehre, T. L. Paine, A. Cowie, Z. Wang, B. Piot, and N. de Freitas. Acme: A research framework for distributed reinforcement learning. arXiv preprint arXiv:2006.00979, 2020. +[31] H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “Other-Play” for zero-shot coordination. In International Conference on Machine Learning (ICML), 2020. +[32] E. Hughes, J. Z. Leibo, M. G. Phillips, K. Tuyls, E. A. Duéñez-Guzmán, A. G. Castañeda, I. Dunning, T. Zhu, K. R. McKee, R. Koster, H. Roff, and T. Graepel. Inequity aversion improves cooperation in intertemporal social dilemmas. In Neural Information Processing Systems (NeurIPS), 2018. +[33] B. Ibarz, J. Leike, T. Pohlen, G. Irving, S. Legg, and D. Amodei. Reward learning from human preferences and demonstrations in Atari. In Neural Information Processing Systems (NeurIPS), 2018. +[34] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie, N. C. Rabinowitz, A. S. Morcos, A. Ruderman, et al. Human-level performance in 3D multiplayer games with population-based reinforcement learning. Science, 364(6443):859–865, 2019. +[35] S. Javdani, S. S. Srinivasa, and J. A. Bagnell. Shared autonomy via hindsight optimization. Robotics Science and Systems, 2015. +[36] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, and J. B. Tenenbaum. Coordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In Conference of the Cognitive Science Society (CogSci), 2016. +[37] G. Klien, D. D. Woods, J. M. Bradshaw, R. R. Hoffman, and P. J. Feltovich. Ten challenges for making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19 (6):91–95, 2004. +[38] P. Knott, M. Carroll, S. Devlin, K. Ciosek, K. Hofmann, A. Dragan, and R. Shah. Evaluating the robustness of collaborative agents. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. +[39] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and T. Graepel. A unified game-theoretic approach to multiagent reinforcement learning. In Neural Information Processing Systems (NIPS), 2017. +[40] A. Lerer and A. Peysakhovich. Learning existing social conventions via observationally augmented self-play. In AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society (AIES), 2019. +[41] E. Lockhart, N. Burch, N. Bard, S. Borgeaud, T. Eccles, L. Smaira, and R. Smith. Human-agent cooperation in bridge bidding. arXiv preprint arXiv:2011.14124, 2020. +[42] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Neural Information Processing Systems (NIPS), 2017. +[43] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. Learning to learn to communicate. In ICML Adaptive & Multitask Learning Workshop, 2019. +[44] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. On the interaction between supervision and self-play in emergent communication. In International Conference on Learning Representations (ICLR), 2020. +[45] A. Lupu, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. +[46] A. Mahajan, T. Rashid, M. Samvelyan, and S. Whiteson. Maven: Multi-agent variational exploration. In Neural Information Processing Systems (NeurIPS), 2019. +[47] K. R. McKee, I. Gemp, B. McWilliams, E. A. Duèñez-Guzmán, E. Hughes, and J. Z. Leibo. Social diversity and social preferences in mixed-motive reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020. +[48] K. R. McKee, X. Bai, and S. Fiske. Understanding human impressions of artificial intelligence. PsyArXiv, 2021. +[49] K. R. McKee, E. Hughes, T. O. Zhu, M. J. Chadwick, R. Koster, A. G. Castaneda, C. Beattie, T. Graepel, M. Botvinick, and J. Z. Leibo. Deep reinforcement learning models the emergent dynamics of human cooperation. arXiv preprint arXiv:2103.04982, 2021. +[50] K. R. McKee, J. Z. Leibo, C. Beattie, and R. Everett. Quantifying environment and population diversity in multi-agent reinforcement learning. arXiv preprint arXiv:2102.08370, 2021. +[51] F. Melo and A. Sardinha. Ad hoc teamwork by learning teammates’ task. International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2015. +[52] B. Mutlu, A. Terrell, and C.-M. Huang. Coordination mechanisms in human-robot collaboration. In HRI Collaborative Manipulation Workshop, 2013. +[53] S. Nikolaidis and J. Shah. Human-robot cross-training: Computational formulation, modeling and evaluation of a human team training strategy. In ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2013. +[54] OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng, Q. Yuan, W. Zaremba, and L. Zhang. Solving Rubik’s Cube with a robot hand. arXiv preprint arXiv:1910.07113, 2019. +[55] E. Peer, L. Brandimarte, S. Samat, and A. Acquisti. Beyond the Turk: Alternative platforms for crowdsourcing behavioral research. Journal of Experimental Social Psychology, 70:153–163, 2017. +[56] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. In IEEE International Conference on Robotics and Automation (ICRA), 2018. +[57] P. M. Pilarski, A. Butcher, M. Johanson, M. M. Botvinick, A. Bolt, and A. S. Parker. Learned human-agent decision-making, communication and joint action in a virtual reality environment. In Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM), 2019. +[58] D. A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation. Neural Computation, 3(1):88–97, 1991. +[59] S. Russell. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin, 2019. +[60] D. Sadigh, N. Landolfi, S. S. Sastry, S. A. Seshia, and A. D. Dragan. Planning for cars that coordinate with people: leveraging effects on human actions for planning and active information gathering over human internal state. Autonomous Robots, 42(7):1405–1426, 2018. +[61] N. Schurr, J. Marecki, M. Tambe, and P. Scerri. Towards flexible coordination of human-agent teams. Multiagent and Grid Systems, 1(1):3–16, 2005. +[62] A. Shih, A. Sawhney, J. Kondic, S. Ermon, and D. Sadigh. On the critical role of conventions in adaptive human-AI collaboration. In International Conference on Learning Representations (ICLR), 2021. +[63] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and D. Hassabis. Mastering the game of go without human knowledge. Nature, 550(7676):354–359, 2017. +[64] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419): 1140–1144, 2018. +[65] H. F. Song, A. Abdolmaleki, J. T. Springenberg, A. Clark, H. Soyer, J. W. Rae, S. Noury, A. Ahuja, S. Liu, D. Tirumala, N. Heess, D. Belov, M. Riedmiller, and M. M. Botvinick. VMPO: On-policy maximum a posteriori policy optimization for discrete and continuous control. In International Conference on Learning Representations (ICLR), 2020. +[66] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Collaboration without pre-coordination. In AAAI Conference on Artificial Intelligence, 2010. +[67] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017. +[68] P. Tylkin, G. Radanovic, and D. C. Parkes. Learning robust helpful behaviors in two-player cooperative Atari environments. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. +[69] O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi, R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre, T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V. Dalibard, D. Budden, Y. Sulsky, J. Molloy, T. L. Paine, C. Gulcehre, Z. Wang, T. Pfaff, Y. Wu, R. Ring, D. Yogatama, D. Wünsch, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, K. Kavukcuoglu, D. Hassabis, C. Apps, and D. Silver. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):350–354, 2019. +[70] R. E. Wang, S. A. Wu, J. A. Evans, J. B. Tenenbaum, D. C. Parkes, and M. Kleiman-Weiner. Too many cooks: Coordinating multi-agent collaboration through inverse planning. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020. +[71] F. Wu, S. Zilberstein, and X. Chen. Online planning for ad hoc autonomous agent teams. In International Joint Conference on Artificial Intelligence (IJCAI), 2011. +[72] S. Zheng, A. Trott, S. Srinivasa, N. Naik, M. Gruesbeck, D. C. Parkes, and R. Socher. The AI economist: Improving equality and productivity with AI-driven tax policies. arXiv preprint arXiv:2004.13332, 2020. \ No newline at end of file diff --git a/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_content_list.json b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e8dbd858ceda63b66ada229b7b3f12658bd8ce12 --- /dev/null +++ b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_content_list.json @@ -0,0 +1,1192 @@ +[ + { + "type": "text", + "text": "Collaborating with Humans without Human Data ", + "text_level": 1, + "bbox": [ + 197, + 122, + 802, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DJ Strouse⇤, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett⇤ DeepMind {strouse, kevinrmckee, botvinick, edwardhughes, reverett}@deepmind.com ", + "bbox": [ + 199, + 200, + 799, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 279, + 535, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Collaborating with humans requires rapidly adapting to their individual strengths, weaknesses, and preferences. Unfortunately, most standard multi-agent reinforcement learning techniques, such as self-play (SP) or population play (PP), produce agents that overfit to their training partners and do not generalize well to humans. Alternatively, researchers can collect human data, train a human model using behavioral cloning, and then use that model to train “human-aware” agents (“behavioral cloning play”, or BCP). While such an approach can improve the generalization of agents to new human co-players, it involves the onerous and expensive step of collecting large amounts of human data first. Here, we study the problem of how to train agents that collaborate well with human partners without using human data. We argue that the crux of the problem is to produce a diverse set of training partners. Drawing inspiration from successful multi-agent approaches in competitive domains, we find that a surprisingly simple approach is highly effective. We train our agent partner as the best response to a population of self-play agents and their past checkpoints taken throughout training, a method we call Fictitious Co-Play (FCP). Our experiments focus on a two-player collaborative cooking simulator that has recently been proposed as a challenge problem for coordination with humans. We find that FCP agents score significantly higher than SP, PP, and BCP when paired with novel agent and human partners. Furthermore, humans also report a strong subjective preference to partnering with FCP agents over all baselines. ", + "bbox": [ + 232, + 310, + 766, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 176, + 614, + 310, + 631 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating agents which collaborate with novel partners is a longstanding challenge for Artificial Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31, $\\boxed { 6 6 }$ is especially important in situations where an AI must generalize to novel human partners [6, 61]. Many successful approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly [12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach humanlevel using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64]. This begs the question: Can model-free RL without human data generate agents that can collaborate with novel humans? ", + "bbox": [ + 174, + 646, + 825, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We seek an answer to this question in the space of common-payoff games, where all agents work towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from repeated games played against copies of itself, does not produce agents that generalize well to novel co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to coordinate with themselves, and so make for brittle and stubborn collaborators with new partners who act differently. Population play (PP) trains a population of agents, all of whom interact with each other $\\pmb { \\| 3 9 \\| }$ . While PP can generate agents capable of cooperation with humans in competitive team games $\\pmb { \\Vert 3 4 \\Vert }$ , it still fails to produce robust partners for novel humans in pure common-payoff settings [12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing strategic diversity and producing agents not so different from self-play $\\dot { \\lVert 2 4 \\rVert }$ . ", + "bbox": [ + 174, + 763, + 825, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/70b2b8c1e9478aed36abef766805e05c03561ba29533d4a7de7ad061fa2d261f.jpg", + "image_caption": [ + "Figure 1: In this work, we evaluate a variety of agent training methods (Section 2) in zero-shot coordination with agents (Section 4). We then run a human-agent collaborative study designed to elicit human preferences over agents (Section 5) " + ], + "image_footnote": [], + "bbox": [ + 191, + 88, + 805, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 301, + 821, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach starts with the intuition that the key to producing robust agent collaborators is exposure to diverse training partners. We find that a surprisingly simple strategy is effective in generating sufficient diversity. We train $N$ self-play agents varying only their random seed for neural network initialization. Periodically during training, we save agent “checkpoints” representing their strategy at that point in time. Then, we train an agent partner as the best-response to both the fully-trained agents and their past checkpoints. The different checkpoints simulate different skill levels, and the different random seeds simulate breaking symmetries in different ways. We refer to this agent training procedure as Fictitious $\\mathbf { C o }$ -Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69]. ", + "bbox": [ + 174, + 335, + 825, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator. Based on the game Overcooked $\\mathbb { \\left. \\boldsymbol { \\ 2 5 } \\right. }$ , it has recently been proposed as a coordination challenge for AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel humans was achieved in $[ \\mathbb { 1 2 } ]$ via behavioral cloning (BC) of human data. More precisely, BC was used to produce models that can stand in as human proxies during training in simulation, a method we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to both novel agent and human partners, and that humans express a significant preference for partnering with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human data for training, while achieving better outcomes for humans at test time. ", + "bbox": [ + 173, + 453, + 825, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize the novel contributions of this paper as follows: ", + "bbox": [ + 174, + 583, + 584, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with humans (Section 2.1). \n2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot coordination with a variety of held-out agents (Section 4.2). \n3. We propose a rigorous human-agent interaction study with behavioral analysis and participant feedback (Section $5 . 1 )$ . \n4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task score and in human partner preference (Section $\\underline { { \\overline { { 5 . 2 } } } } )$ . ", + "bbox": [ + 210, + 609, + 825, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Methods ", + "text_level": 1, + "bbox": [ + 174, + 751, + 279, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Fictitious Co-Play (FCP) ", + "text_level": 1, + "bbox": [ + 174, + 781, + 385, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diverse training conditions have been shown to make agents more robust, from environmental variations (i.e. domain randomization $\\textcircled { 1 5 4 } , \\textcircled { 5 6 } , \\textcircled { 6 7 } \\textcircled { 1 }$ to heterogeneity in training partners $\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { g } } \\right. }$ . We seek to train agents that are robust partners for humans in common-payoff games, and so extend this line of work to that setting. ", + "bbox": [ + 174, + 808, + 825, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One important challenge in collaborating with novel partners is dealing with symmetries $\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }$ . For example, two agents A and B facing each other may move past each other by A going left and B going right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between these conventions if a human clearly prefers one over the other. A second important challenge is dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled partners, as well as partners who are still learning. ", + "bbox": [ + 176, + 869, + 825, + 911 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2372b76b2e16cbfb2db5439982c244e9b3e60b30aff99699aa5043d431598f6f.jpg", + "image_caption": [ + "Figure 2: The four agent training methods we evaluate in this work. Self-play (SP) where an agent learns with itself, population-play (PP) where a population of agents are co-trained together, and behavioral cloning play (BCP) where data from human games is used to create a behaviorally cloned agent with which an RL agent is then trained. In our method, Fictitious Co-Play (FCP), $N$ self-play agents are trained independently and checkpointed throughout training. An agent is then trained to best respond to the entire population of SP agents and their checkpoints. " + ], + "image_footnote": [], + "bbox": [ + 192, + 88, + 808, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 378, + 821, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both of these challenges (Figure $2 ,$ right). In the first stage, we train a diverse pool of partners. To allow the pool to represent different symmetry breaking conventions, we train $N$ partner agents in self-play. Since these partners are trained independently, they can arrive at different arbitrary conventions for breaking symmetries. To allow the pool to represent different skill levels, we use multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost. ", + "bbox": [ + 174, + 426, + 825, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second stage, we train an FCP agent as the best response to the pool of diverse partners created in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the lead of human partners, and learn a general policy across a range of strategies and skills. We call our method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69]. ", + "bbox": [ + 174, + 544, + 825, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Baselines and ablations ", + "text_level": 1, + "bbox": [ + 174, + 643, + 375, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We compare FCP agents to the three baseline training methods listed below, each varying only in their set of training partners, with the RL algorithm and architecture consistent across all agents: ", + "bbox": [ + 176, + 667, + 825, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. Self-play (SP), where agents learn solely through interaction with themselves. \n2. Population-play (PP), where a population of agents are co-trained through random pairings. \n3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12]. ", + "bbox": [ + 210, + 707, + 826, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also evaluate three variations on FCP to better understand the conditions for its success: ", + "bbox": [ + 179, + 767, + 772, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "1. To test the importance of including past checkpoints in training, we evaluate an ablation of FCP in which agents are trained only with the converged checkpoints of their partners $\\mathrm { F C P } _ { - T }$ for “FCP minus time”). 2. To test whether FCP would benefit from additional diversity in its partner population, we evaluate an augmentation of FCP in which the population of SP partners varies not just in random seed, but also in architecture $\\operatorname { F C P } _ { + A }$ for “FCP plus architectural variation”). 3. To test whether architectural variation can serve as a full replacement for playing with past checkpoints, we evaluate the combination of both modifications $( \\mathrm { F C P } _ { - T , + A } )$ . ", + "bbox": [ + 212, + 791, + 825, + 912 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Environment ", + "text_level": 1, + "bbox": [ + 174, + 92, + 303, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment in Carroll et al. [12]. For full details, see Appendix A. ", + "bbox": [ + 174, + 117, + 825, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential high-level actions to which both players can contribute: collecting tomatoes, depositing them into cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering it. Upon a successful delivery, both players are rewarded equally. ", + "bbox": [ + 173, + 165, + 825, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively complete the task, players must learn to navigate the kitchen and interact with objects in the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them. This environment therefore presents the challenges of both movement and strategic coordination. ", + "bbox": [ + 174, + 241, + 825, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each player observes an egocentric RGB view of the world, and at every step can perform one of six actions: stand still, move {up, down, left, right}, interact. The behavior ofLÈ¡áyҏįŘį\u000eºµò‹µÒ¡y® interact varies based on the cell which the player is facing (e.g. place tomato on counter). ", + "bbox": [ + 176, + 289, + 823, + 332 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/48e8af75423e0fa5e5c23e2088c49716c4c6044ee3579f26cdd9911dbb4b4925.jpg", + "image_caption": [ + "Figure 3: The Overcooked environment: a two-player common-payoff game in which players must coordinate to cook and deliver soup. " + ], + "image_footnote": [], + "bbox": [ + 192, + 344, + 803, + 529 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b3253fcaac0677dbb15cc1ed5aa0663e561ff52605ba15bd61605f96c716872.jpg", + "image_caption": [ + "Figure 4: Layouts: the kitchens which agents and humans play in, each emphasizing different coordination strategies. Highlighted in bold are the terms used to refer to each in the rest of this paper. " + ], + "image_footnote": [], + "bbox": [ + 196, + 579, + 803, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 Implementation details ", + "text_level": 1, + "bbox": [ + 174, + 719, + 374, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we highlight several key implementation details for our training methods. For full details, including the architectures, hyperparameters, and compute used, please see Appendix B. ", + "bbox": [ + 174, + 746, + 825, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For our reinforcement learning agents, we use the V-MPO $\\boldsymbol { \\left[ \\left[ 6 5 \\right] \\right] }$ algorithm along with a ResNet [26] plus LSTM $\\mathbb { \\left| \\bigstar \\bigstar \\right\\| }$ architecture which we found led to optimal behavior across all layouts. Agents are trained using a distributed set of environments running in parallel $\\textcircled { 1 1 7 }$ , each sampling two agents from the training population to play together every episode. ", + "bbox": [ + 174, + 779, + 825, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Both PP and FCP are trained with a population size of $N = 3 2$ agents which are sampled uniformly. For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden: (1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert agent), and (3) at the middle of training, defined as when the agent reaches $50 \\%$ of its final reward (i.e. an average-skilled agent). When varying architecture for the training partners of the $\\operatorname { F C P } _ { + A }$ and $\\mathrm { F C P } _ { - T , + A }$ variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4 combinations, leaving the total population size of $N = 3 2$ unchanged, ensuring a fair comparison. ", + "bbox": [ + 173, + 842, + 825, + 911 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 92, + 825, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train agents via behavioral cloning $\\left[ \\left[ 5 8 \\right] \\right]$ , we use the open-source Acme $\\pmb { \\mathbb { B } } 0 \\|$ to learn a policy from human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for agent-agent evaluation. Following Carroll et al. $\\bar { \\lVert 1 2 \\rVert }$ , we use a set of feature-based observations for the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts (asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced). ", + "bbox": [ + 174, + 138, + 825, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3 Related work ", + "text_level": 1, + "bbox": [ + 174, + 260, + 316, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play $ { \\mathbb { B } } , { \\mathbb { G } } 6 { \\mathbb { I } }$ , also known as zero-shot coordination $\\bar { \\mathbb { B } } \\mathbb { I }$ . Prior work based in game-theoretic settings has suggested the benefits of planning $\\pmb { \\mathbb { Z } 1 }$ , online learning $\\mathbb { \\left| \\left. \\sum \\right. { ] } \\right| }$ , and novel solution concepts $\\bar { \\mathbb { D } }$ , to name a few examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale to more complex gridworld or continuous control settings, leading to work on hierarchical social planning $\\widehat { \\left\\| 3 6 \\right\\| }$ , adapting to existing social conventions $\\checkmark$ , trajectory diversity $\\lVert \\rVert \\dot { \\boldsymbol { \\mathrm { E } } } \\rVert$ , and theory of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization to held-out agent partners as a proxy for human co-players. ", + "bbox": [ + 173, + 294, + 825, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collaborative play with novel humans has been evaluated more actively in the context of training agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the stateof-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich gridworld environment, as measured by the challenge tasks in Carroll et al. [12]. ", + "bbox": [ + 174, + 425, + 825, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents that train with behaviorally diverse populations of game partners tend to demonstrate stronger performance than their self-play counterparts. For example, across a range of multi-agent games, generalization to held-out populations can be improved by training larger and more diverse populations [13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be optimized through selective matchmaking between increasingly diverse agents $[ \\overbrace { 2 4 } , \\overbrace { 3 9 } , \\overbrace { 6 9 } ]$ . ", + "bbox": [ + 174, + 487, + 825, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Despite the increased focus on improving multi-agent performance, evaluation has typically been constrained to agent-agent settings. High-performing agents have infrequently been evaluated with humans, particularly in non-competitive domains $\\bar { \\lVert 1 6 rVert }$ . We add to this growing literature, showing that training with diversity is a powerful approach for effective human-agent collaboration. ", + "bbox": [ + 174, + 590, + 825, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human-agent interaction In recent years, increased attention has been directed toward designing machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also $\\boxed { \\boxed { 1 6 } }$ for a broader review on Cooperative AI). Tylkin et al. $\\lVert \\rVert$ is particularly notable in also demonstrating that partially trained agents can be useful learning targets for human helpers, although in a different domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to our agents’ performance (Tables 1 and 2 and Figure 7b). ", + "bbox": [ + 173, + 652, + 825, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key preceding entry in this research area is Carroll et al. $\\pmb { \\mathbb { I } }$ , who similarly investigated humanagent coordination in Overcooked. We use their method (BCP) as a baseline throughout our experiments (Section $\\boxed { 2 . 2 }$ . Relative to BCP, our approach removes the need for the expensive step of human data collection for agent training. Furthermore, through our novel human-agent experimental design, we go beyond objective performance metrics to compare the subjective preferences that agents generate. For a detailed comparison of methods and results, see Appendix E. ", + "bbox": [ + 174, + 756, + 825, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 Zero-shot coordination with agents ", + "text_level": 1, + "bbox": [ + 174, + 863, + 498, + 880 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents. ", + "bbox": [ + 169, + 896, + 795, + 911 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 Evaluation method: collaborative evaluation with agent partners ", + "text_level": 1, + "bbox": [ + 174, + 90, + 663, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our primary concern in this work is generalization to novel human partners (as investigated in Section $\\textcircled{5}$ . However, just as collecting human-human data for behavioral cloning is expensive, so too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent partners as a cheap proxy of performance with humans. This is then used to guide our model selection process, allowing us to be more targeted with the agents we select for our human-agent evaluations. ", + "bbox": [ + 174, + 116, + 825, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate with three held-out populations: ", + "bbox": [ + 176, + 193, + 467, + 207 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1. A BC model trained on human data, $H _ { \\mathrm { p r o x y } }$ , intended as a proxy of generalization to humans, as done by Carroll et al. $[ \\overbrace { | 1 2 | }$ . 2. A set of self-play agents varying in seed, architecture, and training time (specifically, heldout seeds of the $N = 3 2$ partners trained for the $\\mathrm { F C P } _ { + A }$ agent; see Section 2.4). These are intended to test generalization to a diverse yet still skillful population. 3. Randomly initialized agents intended to test generalization to low-skill partners. ", + "bbox": [ + 210, + 217, + 826, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For all results, we report the average number of deliveries made by both players within an episode, aggregated across the 5 different layouts from Figure $^ 4$ (with the per-layout results reported in Appendix $\\underline { { \\overline { { ( \\mathrm { C . 2 } ) } } } }$ . We estimate mean and standard deviation across 5 random seeds. For each seed, we evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair. ", + "bbox": [ + 174, + 320, + 826, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 Results ", + "text_level": 1, + "bbox": [ + 174, + 391, + 264, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finding 1: FCP significantly outperforms all baselines ", + "text_level": 1, + "bbox": [ + 176, + 416, + 549, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To begin, we compare our FCP agent and the baselines when partnered with the three held-out populations introduced above. As can be seen in Figure $\\boxed { 5 }$ FCP significantly outperforms all baselines when partnered with all three held-out populations. Notably, it performs better than BCP with $H _ { \\mathrm { p r o x y } }$ even though BCP trains with such a model and FCP does not. Similar to Carroll et al. $\\mathbb { \\lVert 1 2 \\rVert }$ , we find that BCP significantly outscores SP. ", + "bbox": [ + 174, + 433, + 825, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When paired with a randomly initialized partner which behaves suboptimally, we see an even greater difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of such agents, it may not be surprising that it does so well with partners that behave poorly. However, what is surprising is how brittle the other training methods are. This suggests that they may not perform well with humans who are not highly skilled players, which we will see in Section 5. ", + "bbox": [ + 173, + 508, + 825, + 579 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8083dcac415b6ef9624e1aa4f428a0a33c3e9d47ce8903df3798f3e80ef0427c.jpg", + "image_caption": [ + "Figure 5: Agent-agent collaborative evaluation: Performance of each agent when partnered with each of the held-out populations (Section $4 . 1 )$ in episodes of length $T = 5 4 0$ . Importantly, FCP scores higher than all baselines with a variety of test partners. Error bars represent standard deviation over five random training seeds. Plots aggregate data across kitchen layouts; results calculated by individual layout can be found in Appendix C.2. " + ], + "image_footnote": [], + "bbox": [ + 165, + 593, + 808, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finding 2: Training with past checkpoints is the most beneficial variation for performance Next, we investigate how the different training partner variations influence FCP’s performance. In particular, we separately ablate the past checkpoints $( T )$ and architecture $( A )$ variations, evaluating them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1. Comparing the FCP and $\\mathrm { F C P } _ { - T }$ columns, we see that removing past checkpoints from training significantly reduces performance. Comparing the FCP and $\\operatorname { F C P } _ { + A }$ columns, we see that adding architectural variation to the training population offers no improvement over training with past ", + "bbox": [ + 173, + 804, + 826, + 911 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/56b2734b6d3a9e5266ca9b0dd8d3ba9f6d27df29190a11ac23347bb9e9fb3c37.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PartnerFCPFCP-TFCP+AFCP-T,+A
Hproxy10.6± 0.54.7± 0.49.9±0.67.0±0.8
Diverse SP11.2 ± 0.16.9 ± 0.111.1 ± 0.48.6 ± 0.4
Random8.6± 0.21.0 ± 0.18.4±0.43.2 ± 0.5
", + "bbox": [ + 272, + 88, + 720, + 147 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1: Ablation results: Performance of each variation of FCP – training with past partner checkpoints $T$ for time) and adding partner variation in architecture $( A )$ . Scores are mean deliveries with standard deviation over 5 random seeds. Notably, we find that the inclusion of past checkpoints is essential for strong performance $( \\mathrm { F C P } > \\mathrm { F C P } _ { - T }$ ), and additionally including architectural variation does not improve performance $( \\mathrm { F C P } \\approx \\mathrm { F C P } _ { + A . }$ ). However, architectural variation is better than no variation, improving performance when past checkpoints are not available $( \\mathrm { F C P } _ { - T , + A } > \\mathrm { F C P } _ { - T , }$ ). ", + "bbox": [ + 173, + 155, + 825, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "checkpoints. However, comparing the $\\mathrm { F C P } _ { - T }$ and $\\mathrm { F C P } _ { - T , + A }$ columns, we see that without training with past checkpoints, architectural variation in the population does improve performance. ", + "bbox": [ + 174, + 262, + 825, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 Zero-shot coordination with humans ", + "text_level": 1, + "bbox": [ + 174, + 310, + 511, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative play with human partners. ", + "bbox": [ + 173, + 342, + 826, + 383 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5856cbdd90eeef2a7a6093668472063862a759746c339aa0324e78efb5d67b22.jpg", + "image_caption": [ + "Figure 6: Human-agent collaborative study: For our human-agent collaboration study, we recruited participants online to play games with FCP and baseline agents. Participants played a randomized sequence of episodes with different agent partners and kitchen layouts. After every two episodes, participants reported the direction and strength of their preference between their last two partners. " + ], + "image_footnote": [], + "bbox": [ + 148, + 397, + 851, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 Evaluation method: collaborative evaluation with human participants ", + "text_level": 1, + "bbox": [ + 173, + 598, + 697, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test how effectively FCP’s performance generalizes to human partners, we recruited participants from Prolific $\\mathbb { 1 1 8 } , \\lvert 5 5 \\rvert$ for an online collaboration study $N = 1 1 4$ ; $3 7 . 7 \\%$ female, $5 9 . 6 \\%$ male, $1 . 8 \\%$ nonbinary; median age between 25–34 years). We used a within-participant design for the study: each participant played with a full cohort of agents (i.e. generated through every training method). This design allowed us to evaluate both objective performance as well as subjective preferences. ", + "bbox": [ + 174, + 625, + 826, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Participants first read game instructions and played a short tutorial episode guiding them through the dish preparation sequence (see Appendix $\\dot { \\mathrm { \\bf D } } . 1 . 1$ for instruction text and study screenshots). Participants then played 20 episodes with a randomized sequence of agent partners and kitchen layouts. Episodes lasted $T = 3 0 0$ steps (1 minute) each. After every two episodes, participants reported their preference over the agent partners from those episodes on a five-point Likert-type scale. After playing all 20 episodes, participants completed a debrief questionnaire collecting standard demographic information and open-ended feedback on the study. Our statistical analysis below primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D for additional details of our study design and analysis, including independent ethical review. ", + "bbox": [ + 174, + 699, + 825, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 Results ", + "text_level": 1, + "bbox": [ + 174, + 842, + 264, + 856 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finding 1: FCP coordinates best with humans, achieving the highest score across maps To begin, we compare the objective team performance supported by our FCP and baseline agents. The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration: ", + "bbox": [ + 171, + 867, + 826, + 911 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the FCP-human teams significantly outperform all other agent-human teams, achieving the highest average scores across maps, every $p < 0 . 0 0 1$ (Figure $\\mathrm { 7 a ) }$ , while performing as well as or better than the other teams on each individual map (see Appendix $\\mathbf { D } . 3 )$ . Echoing the results from our agent-agent ablation experiments (Table $\\perp )$ , the inclusion of past checkpoints in training proves critical to FCP’s strong performance, $p < 0 . 0 { \\overline { { 0 1 } } }$ (Figure $\\textcircled { 7 6 }$ . Similar to Carroll et al. $[ \\left[ 1 2 \\right] ]$ , we find that BCP outscores SP when collaborating with human players, $p < 0 . 0 0 1$ . ", + "bbox": [ + 173, + 90, + 825, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding 2: Participants prefer FCP over all baselines ", + "text_level": 1, + "bbox": [ + 173, + 189, + 544, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences. Participants expressed a significant preference for FCP partners over all other agents, including BCP, with every $p < 0 . 0 5$ (Figure $\\dot { \\bigtriangledown } \\dot { \\mathbf { c } } { \\big \\rVert }$ . Notably, while human-BCP and human-PP teams did not significantly differ in their completed deliveries, participants reported significantly preferring BCP over PP, $p = 0 . 0 0 3$ , highlighting the informativeness of our subjective analysis. ", + "bbox": [ + 173, + 205, + 826, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14e048153d10325b3617f94b393393ed3ab6cf3e38198f8791acf0fc83b051df.jpg", + "image_caption": [ + "Figure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from humanagent play in episodes of length $T = 3 0 0$ . Error bars represents $9 5 \\%$ confidence intervals, calculated over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can be found in Appendix D.3. " + ], + "image_footnote": [], + "bbox": [ + 174, + 290, + 823, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 Exploratory behavioral analysis ", + "text_level": 1, + "bbox": [ + 174, + 531, + 434, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To better understand how the human-agent scores and preferences may have arisen, here we analyze the resulting action trajectories of each human and agent player in our experiment. ", + "bbox": [ + 173, + 556, + 825, + 585 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65959c0ab23a18c0f8de72bc451d199e7d9d9cc2698e7ca2ac4f320e26dac829.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 603, + 774, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 8: Behavioral analysis: (a) FCP is able to move most frequently $3 5 \\%$ of the time), corresponding to the best movement coordination with human partners. (b) FCP exhibits the most equal preferences over cooking pots (0.11 difference), aligning with human preferences. Values are calculated as the absolute difference in preferences between the two pots; 1 indicates that the player only uses one of the two available pots, while 0 indicates that the player uses both pots equally. ", + "bbox": [ + 173, + 758, + 826, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finding 1: FCP exhibits the best movement coordination with humans ", + "text_level": 1, + "bbox": [ + 174, + 839, + 665, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "First, we investigate how much each player moves in an episode (Figure $\\textcircled { 8 \\mathrm { a } }$ , where moving in a higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner. Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP. ", + "bbox": [ + 174, + 856, + 826, + 911 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its movement strategy with its partner. This result was also reported by human participants, for example: “I noticed that some of my partners seemed to know they needed to move around me, while others seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples). ", + "bbox": [ + 173, + 92, + 825, + 147 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finding 2: FCP’s preferences over cooking pots aligns best with that of humans ", + "text_level": 1, + "bbox": [ + 173, + 162, + 725, + 178 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Next, we investigate whether there was a preference for a specific cooking pot in the layouts which included two cooking pots (Figure $\\textcircled { 8 6 }$ . To do this, we calculate the difference in the number of times each pot was used by each player, where a high value indicates a strong preference for one pot and a low value indicates more equal preference for the two pots. ", + "bbox": [ + 174, + 180, + 825, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "As can be seen in the FCP column, our agent typically has the most aligned preferences with that of humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers one cooking pot over the other $5 5 . 5 \\%$ of the time (i.e. a 0.11 point difference). In contrast, all other agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes poorly to typical human behavior of using both pots, leading to worse performance. ", + "bbox": [ + 174, + 242, + 825, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Discussion ", + "text_level": 1, + "bbox": [ + 174, + 327, + 294, + 344 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Summary In this work, we investigated the challenging problem of zero-shot collaboration with humans without using human data in the training pipeline. To accomplish this, we introduced Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set of training partners. We found that FCP agents scored significantly higher than all baselines when partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent experimental design, we also found that humans reported a strong subjective preference to partnering with FCP agents over all baselines. ", + "bbox": [ + 173, + 352, + 825, + 449 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Limitations and future work Our method currently relies on the manual process of initially training and selecting a diverse set of partners. This is not only time consuming, but also prone to researcher biases that may negatively influence the behavior of the created agents. Additionally, while we found FCP with a partner population size of $N = 3 2$ sufficient here, for more complex games, FCP may require an unrealistically large partner population size to represent sufficiently diverse strategies. To address these concerns, methods for automatically generating partner diversity for common-payoff games may be important. Possibilities include adaptive population matchmaking as been used in competitive zero-sum games $\\mathbb { \\lVert 6 9 \\rVert }$ , as well as auxiliary objectives that explicitly encourage behavioral diversity [19, 45, 46]. ", + "bbox": [ + 174, + 455, + 825, + 580 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our method requires a known and fixed reward function. We also focus on one domain in order to compare with prior work which has argued that human-in-the-loop training is necessary. Consequently, the resulting agents are only designed to adaptively collaborate on a single task, and not to infer human preferences in general $\\textcircled { 1 1 } \\textcircled { 3 3 } \\textcircled { 5 9 }$ . Moreover, if a task’s reward function is poorly aligned with how humans approach the task, our method may well produce subpar partners, as would any method without access to human data. Thus, additional domains and tasks should be studied to better understand how our method generalizes. Targeted experiments to test specific forms of generalization may be especially helpful in this regard $\\overline { { [ 3 8 ] } }$ , as could approaches that procedurally generate environment layouts requiring diverse solutions $\\pmb { \\mathbb { Z } } 2 \\mathbf { l }$ ", + "bbox": [ + 174, + 587, + 825, + 712 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Finally, it may be possible to produce even stronger agent assistants by combining the strengths of FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al. $\\textcircled { \\lvert 3 8 \\rvert }$ recently demonstrated that modifying BCP to train with multiple BC partners produces more robust collaboration with held-out agents, a finding that would be interesting to test with human partners. ", + "bbox": [ + 174, + 718, + 825, + 773 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner expresses preferences or intentions to harm others). This could therefore produce negative societal effects either if training leads to poor alignment or if agents are optimized for harmful metrics. ", + "bbox": [ + 174, + 779, + 825, + 863 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "One potential strategy for mitigating these risks is the use of human preference data [15]. Such data could be used to fine-tune and filter trained agents before deployment, encouraging better alignment with human values. A key question in this line of research is how human preference data should be aggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human beliefs and perceptions of AI $\\lVert \\overline { { 4 8 } } \\rVert$ , and how they steer human-agent interaction, would help inform agent design for positive societal impact. For instance, developers could incorporate specific priors into agents to reinforce tendencies for fair outcomes $\\pm \\mathbb { Z } 0 . \\pm \\mathbb { B } 2 \\mathbb { I }$ . ", + "bbox": [ + 176, + 869, + 823, + 911 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 92, + 825, + 160 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Conclusion We proposed a method which is both effective at collaborating with humans and simple to implement. We also presented a rigorous and general methodology for evaluating with humans and eliciting their preferences. Together, these establish a strong foundation for future research on the important challenge of human-agent collaboration for benefiting society. ", + "bbox": [ + 174, + 166, + 825, + 223 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Acknowledgements ", + "text_level": 1, + "bbox": [ + 176, + 242, + 338, + 258 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions. ", + "bbox": [ + 174, + 272, + 825, + 328 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Funding disclosure ", + "text_level": 1, + "bbox": [ + 174, + 345, + 334, + 363 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "This work was funded solely by DeepMind. The authors declare no competing interests. ", + "bbox": [ + 176, + 376, + 750, + 391 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References ", + "text_level": 1, + "bbox": [ + 174, + 410, + 266, + 426 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "[1] J. Abramson, A. Ahuja, I. Barr, A. Brussee, F. Carnevale, M. Cassin, R. Chhaparia, S. Clark, B. Damoc, A. Dudzik, et al. Imitating interactive intelligence. arXiv preprint arXiv:2012.05672, 2020. \n[2] S. V. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method for ad hoc coordination in multiagent systems. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2013. \n[3] B. Baker. Emergent reciprocity and team formation from randomized uncertain social preferences. In Neural Information Processing Systems (NeurIPS), 2020. \n[4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin, S. Moitra, E. Hughes, et al. The Hanabi challenge: A new frontier for AI research. Artificial Intelligence, 280:103216, 2020. \n[5] S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with new teammates. Artificial Intelligence, 242:132–171, 2017. \n[6] A. Bauer, D. Wollherr, and M. Buss. Human–robot collaboration: A survey. International Journal of Humanoid Robotics, 5(01):47–66, 2008. \n[7] G. W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and allocation, 13(1):374–376, 1951. \n[8] N. Brown and T. Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top professionals. Science, 359(6374):418–424, 2018. \n[9] N. Brown and T. Sandholm. Superhuman AI for multiplayer poker. Science, 365(6456):885–890, 2019. \n[10] K. Bullard, F. Meier, D. Kiela, J. Pineau, and J. Foerster. Exploring zero-shot emergent communication in embodied multi-agent populations. arXiv preprint arXiv:2010.15896, 2020. \n[11] K. Bullard, D. Kiela, J. Pineau, and J. Foerster. Quasi-equivalence discovery for zero-shot emergent communication. arXiv preprint arXiv:2103.08067, 2021. \n[12] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the utility of learning about humans for human-AI coordination. In Neural Information Processing Systems (NeurIPS), 2019. \n[13] R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Investigating partner diversification methods in cooperative multi-agent deep reinforcement learning. In International Conference on Neural Information Processing, 2020. \n[14] R. Choudhury, G. Swamy, D. Hadfield-Menell, and A. D. Dragan. On the utility of model learning in HRI. In ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2019. \n[15] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pages 4302–4310, 2017. \n[16] A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and T. Graepel. Open problems in cooperative AI. arXiv preprint arXiv:2012.08630, 2020. \n[17] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu, T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable distributed deep-RL with importance weighted actor-learner architectures. In International Conference on Machine Learning (ICML), 2018. \n[18] P. Eyal, R. David, G. Andrew, E. Zak, and D. Ekaterina. Data quality of platforms and panels for online behavioral research. Behavior Research Methods, pages 1–20, 2021. \n[19] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills without a reward function. In International Conference on Learning Representations (ICLR), 2019. \n[20] E. Fehr and K. M. Schmidt. A theory of fairness, competition, and cooperation. The Quarterly Journal of Economics, 114(3):817–868, 1999. \n[21] J. Foerster, F. Song, E. Hughes, N. Burch, I. Dunning, S. Whiteson, M. Botvinick, and M. Bowling. Bayesian action decoder for deep multi-agent reinforcement learning. In International Conference on Machine Learning (ICML), 2019. \n[22] M. Fontaine, Y.-C. Hsu, Y. Zhang, B. Tjanaka, and S. Nikolaidis. On the importance of environments in human-robot coordination. In Robotics: Science and Systems (RSS), 2021. \n[23] I. Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30(3):411–437, 2020. \n[24] M. Garnelo, W. M. Czarnecki, S. Liu, D. Tirumala, J. Oh, G. Gidel, H. van Hasselt, and D. Balduzzi. Pick your battles: Interaction graphs as population-level objectives for strategic diversity. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. \n[25] Ghost Town Games. Overcooked, 2016. \n[26] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. \n[27] J. Heinrich and D. Silver. Deep reinforcement learning from self-play in imperfect-information games. In NIPS Deep Reinforcement Learning (DRL) Workshop, 2016. \n[28] J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In International Conference on Machine Learning (ICML), 2015. \n[29] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8): 1735–1780, 1997. \n[30] M. Hoffman, B. Shahriari, J. Aslanides, G. Barth-Maron, F. Behbahani, T. Norman, A. Abdolmaleki, A. Cassirer, F. Yang, K. Baumli, S. Henderson, A. Novikov, S. G. Colmenarejo, S. Cabi, C. Gulcehre, T. L. Paine, A. Cowie, Z. Wang, B. Piot, and N. de Freitas. Acme: A research framework for distributed reinforcement learning. arXiv preprint arXiv:2006.00979, 2020. \n[31] H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “Other-Play” for zero-shot coordination. In International Conference on Machine Learning (ICML), 2020. \n[32] E. Hughes, J. Z. Leibo, M. G. Phillips, K. Tuyls, E. A. Duéñez-Guzmán, A. G. Castañeda, I. Dunning, T. Zhu, K. R. McKee, R. Koster, H. Roff, and T. Graepel. Inequity aversion improves cooperation in intertemporal social dilemmas. In Neural Information Processing Systems (NeurIPS), 2018. \n[33] B. Ibarz, J. Leike, T. Pohlen, G. Irving, S. Legg, and D. Amodei. Reward learning from human preferences and demonstrations in Atari. In Neural Information Processing Systems (NeurIPS), 2018. \n[34] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie, N. C. Rabinowitz, A. S. Morcos, A. Ruderman, et al. Human-level performance in 3D multiplayer games with population-based reinforcement learning. Science, 364(6443):859–865, 2019. \n[35] S. Javdani, S. S. Srinivasa, and J. A. Bagnell. Shared autonomy via hindsight optimization. Robotics Science and Systems, 2015. \n[36] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, and J. B. Tenenbaum. Coordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In Conference of the Cognitive Science Society (CogSci), 2016. \n[37] G. Klien, D. D. Woods, J. M. Bradshaw, R. R. Hoffman, and P. J. Feltovich. Ten challenges for making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19 (6):91–95, 2004. \n[38] P. Knott, M. Carroll, S. Devlin, K. Ciosek, K. Hofmann, A. Dragan, and R. Shah. Evaluating the robustness of collaborative agents. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. \n[39] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and T. Graepel. A unified game-theoretic approach to multiagent reinforcement learning. In Neural Information Processing Systems (NIPS), 2017. \n[40] A. Lerer and A. Peysakhovich. Learning existing social conventions via observationally augmented self-play. In AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society (AIES), 2019. \n[41] E. Lockhart, N. Burch, N. Bard, S. Borgeaud, T. Eccles, L. Smaira, and R. Smith. Human-agent cooperation in bridge bidding. arXiv preprint arXiv:2011.14124, 2020. \n[42] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. In Neural Information Processing Systems (NIPS), 2017. \n[43] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. Learning to learn to communicate. In ICML Adaptive & Multitask Learning Workshop, 2019. \n[44] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. On the interaction between supervision and self-play in emergent communication. In International Conference on Learning Representations (ICLR), 2020. \n[45] A. Lupu, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. \n[46] A. Mahajan, T. Rashid, M. Samvelyan, and S. Whiteson. Maven: Multi-agent variational exploration. In Neural Information Processing Systems (NeurIPS), 2019. \n[47] K. R. McKee, I. Gemp, B. McWilliams, E. A. Duèñez-Guzmán, E. Hughes, and J. Z. Leibo. Social diversity and social preferences in mixed-motive reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020. \n[48] K. R. McKee, X. Bai, and S. Fiske. Understanding human impressions of artificial intelligence. PsyArXiv, 2021. \n[49] K. R. McKee, E. Hughes, T. O. Zhu, M. J. Chadwick, R. Koster, A. G. Castaneda, C. Beattie, T. Graepel, M. Botvinick, and J. Z. Leibo. Deep reinforcement learning models the emergent dynamics of human cooperation. arXiv preprint arXiv:2103.04982, 2021. \n[50] K. R. McKee, J. Z. Leibo, C. Beattie, and R. Everett. Quantifying environment and population diversity in multi-agent reinforcement learning. arXiv preprint arXiv:2102.08370, 2021. \n[51] F. Melo and A. Sardinha. Ad hoc teamwork by learning teammates’ task. International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2015. \n[52] B. Mutlu, A. Terrell, and C.-M. Huang. Coordination mechanisms in human-robot collaboration. In HRI Collaborative Manipulation Workshop, 2013. \n[53] S. Nikolaidis and J. Shah. Human-robot cross-training: Computational formulation, modeling and evaluation of a human team training strategy. In ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2013. \n[54] OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino, M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng, Q. Yuan, W. Zaremba, and L. Zhang. Solving Rubik’s Cube with a robot hand. arXiv preprint arXiv:1910.07113, 2019. \n[55] E. Peer, L. Brandimarte, S. Samat, and A. Acquisti. Beyond the Turk: Alternative platforms for crowdsourcing behavioral research. Journal of Experimental Social Psychology, 70:153–163, 2017. \n[56] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic control with dynamics randomization. In IEEE International Conference on Robotics and Automation (ICRA), 2018. \n[57] P. M. Pilarski, A. Butcher, M. Johanson, M. M. Botvinick, A. Bolt, and A. S. Parker. Learned human-agent decision-making, communication and joint action in a virtual reality environment. In Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM), 2019. \n[58] D. A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation. Neural Computation, 3(1):88–97, 1991. \n[59] S. Russell. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin, 2019. \n[60] D. Sadigh, N. Landolfi, S. S. Sastry, S. A. Seshia, and A. D. Dragan. Planning for cars that coordinate with people: leveraging effects on human actions for planning and active information gathering over human internal state. Autonomous Robots, 42(7):1405–1426, 2018. \n[61] N. Schurr, J. Marecki, M. Tambe, and P. Scerri. Towards flexible coordination of human-agent teams. Multiagent and Grid Systems, 1(1):3–16, 2005. \n[62] A. Shih, A. Sawhney, J. Kondic, S. Ermon, and D. Sadigh. On the critical role of conventions in adaptive human-AI collaboration. In International Conference on Learning Representations (ICLR), 2021. \n[63] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and D. Hassabis. Mastering the game of go without human knowledge. Nature, 550(7676):354–359, 2017. \n[64] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419): 1140–1144, 2018. \n[65] H. F. Song, A. Abdolmaleki, J. T. Springenberg, A. Clark, H. Soyer, J. W. Rae, S. Noury, A. Ahuja, S. Liu, D. Tirumala, N. Heess, D. Belov, M. Riedmiller, and M. M. Botvinick. VMPO: On-policy maximum a posteriori policy optimization for discrete and continuous control. In International Conference on Learning Representations (ICLR), 2020. \n[66] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Collaboration without pre-coordination. In AAAI Conference on Artificial Intelligence, 2010. \n[67] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017. \n[68] P. Tylkin, G. Radanovic, and D. C. Parkes. Learning robust helpful behaviors in two-player cooperative Atari environments. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. \n[69] O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi, R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre, T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V. Dalibard, D. Budden, Y. Sulsky, J. Molloy, T. L. Paine, C. Gulcehre, Z. Wang, T. Pfaff, Y. Wu, R. Ring, D. Yogatama, D. Wünsch, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, K. Kavukcuoglu, D. Hassabis, C. Apps, and D. Silver. Grandmaster level in StarCraft II using multi-agent reinforcement learning. Nature, 575(7782):350–354, 2019. \n[70] R. E. Wang, S. A. Wu, J. A. Evans, J. B. Tenenbaum, D. C. Parkes, and M. Kleiman-Weiner. Too many cooks: Coordinating multi-agent collaboration through inverse planning. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020. \n[71] F. Wu, S. Zilberstein, and X. Chen. Online planning for ad hoc autonomous agent teams. In International Joint Conference on Artificial Intelligence (IJCAI), 2011. \n[72] S. Zheng, A. Trott, S. Srinivasa, N. Naik, M. Gruesbeck, D. C. Parkes, and R. Socher. The AI economist: Improving equality and productivity with AI-driven tax policies. arXiv preprint arXiv:2004.13332, 2020. ", + "bbox": [ + 178, + 426, + 826, + 912 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 50, + 828, + 917 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 74, + 828, + 915 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 56, + 828, + 912 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 90, + 826, + 213 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_middle.json b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..d0c37ffa39e6e24e61446805fbaab48e10e740c8 --- /dev/null +++ b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_middle.json @@ -0,0 +1,34736 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 121, + 97, + 491, + 117 + ], + "lines": [ + { + "bbox": [ + 121, + 97, + 491, + 119 + ], + "spans": [ + { + "bbox": [ + 121, + 97, + 491, + 119 + ], + "score": 1.0, + "content": "Collaborating with Humans without Human Data", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 122, + 159, + 489, + 193 + ], + "lines": [ + { + "bbox": [ + 135, + 159, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 135, + 159, + 480, + 172 + ], + "score": 1.0, + "content": "DJ Strouse⇤, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett⇤", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 282, + 170, + 329, + 182 + ], + "spans": [ + { + "bbox": [ + 282, + 170, + 329, + 182 + ], + "score": 1.0, + "content": "DeepMind", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 180, + 491, + 195 + ], + "spans": [ + { + "bbox": [ + 120, + 180, + 491, + 195 + ], + "score": 1.0, + "content": "{strouse, kevinrmckee, botvinick, edwardhughes, reverett}@deepmind.com", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 283, + 221, + 328, + 234 + ], + "lines": [ + { + "bbox": [ + 281, + 220, + 330, + 235 + ], + "spans": [ + { + "bbox": [ + 281, + 220, + 330, + 235 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 142, + 246, + 469, + 465 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "score": 1.0, + "content": "Collaborating with humans requires rapidly adapting to their individual strengths,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "score": 1.0, + "content": "weaknesses, and preferences. Unfortunately, most standard multi-agent reinforce-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "ment learning techniques, such as self-play (SP) or population play (PP), produce", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 280, + 471, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 471, + 292 + ], + "score": 1.0, + "content": "agents that overfit to their training partners and do not generalize well to humans.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 290, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 302 + ], + "score": 1.0, + "content": "Alternatively, researchers can collect human data, train a human model using behav-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 302, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 313 + ], + "score": 1.0, + "content": "ioral cloning, and then use that model to train “human-aware” agents (“behavioral", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "score": 1.0, + "content": "cloning play”, or BCP). While such an approach can improve the generalization", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 322, + 470, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 336 + ], + "score": 1.0, + "content": "of agents to new human co-players, it involves the onerous and expensive step of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "collecting large amounts of human data first. Here, we study the problem of how", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 345, + 471, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 471, + 357 + ], + "score": 1.0, + "content": "to train agents that collaborate well with human partners without using human data.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 355, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 470, + 369 + ], + "score": 1.0, + "content": "We argue that the crux of the problem is to produce a diverse set of training part-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 367, + 470, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 380 + ], + "score": 1.0, + "content": "ners. Drawing inspiration from successful multi-agent approaches in competitive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "score": 1.0, + "content": "domains, we find that a surprisingly simple approach is highly effective. We train", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "score": 1.0, + "content": "our agent partner as the best response to a population of self-play agents and their", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 399, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 412 + ], + "score": 1.0, + "content": "past checkpoints taken throughout training, a method we call Fictitious Co-Play", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "score": 1.0, + "content": "(FCP). Our experiments focus on a two-player collaborative cooking simulator that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 422, + 471, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 422, + 471, + 434 + ], + "score": 1.0, + "content": "has recently been proposed as a challenge problem for coordination with humans.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "score": 1.0, + "content": "We find that FCP agents score significantly higher than SP, PP, and BCP when", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "score": 1.0, + "content": "paired with novel agent and human partners. Furthermore, humans also report a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 455, + 453, + 466 + ], + "spans": [ + { + "bbox": [ + 141, + 455, + 453, + 466 + ], + "score": 1.0, + "content": "strong subjective preference to partnering with FCP agents over all baselines.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 487, + 190, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 192, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 192, + 502 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "Generating agents which collaborate with novel partners is a longstanding challenge for Artificial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 434, + 536 + ], + "score": 1.0, + "content": "Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31,", + "type": "text" + }, + { + "bbox": [ + 434, + 523, + 451, + 534 + ], + "score": 0.26, + "content": "\\boxed { 6 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "is especially", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "important in situations where an AI must generalize to novel human partners [6, 61]. Many successful", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 108, + 556, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 108, + 556, + 506, + 567 + ], + "score": 1.0, + "content": "[12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach human-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "level using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64].", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "This begs the question: Can model-free RL without human data generate agents that can collaborate", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 589, + 188, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 188, + 600 + ], + "score": 1.0, + "content": "with novel humans?", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "We seek an answer to this question in the space of common-payoff games, where all agents work", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "repeated games played against copies of itself, does not produce agents that generalize well to novel", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "coordinate with themselves, and so make for brittle and stubborn collaborators with new partners", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "who act differently. Population play (PP) trains a population of agents, all of whom interact with each", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 130, + 684 + ], + "score": 1.0, + "content": "other", + "type": "text" + }, + { + "bbox": [ + 130, + 670, + 147, + 682 + ], + "score": 0.65, + "content": "\\pmb { \\| 3 9 \\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 669, + 506, + 684 + ], + "score": 1.0, + "content": ". While PP can generate agents capable of cooperation with humans in competitive team", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 134, + 693 + ], + "score": 1.0, + "content": "games", + "type": "text" + }, + { + "bbox": [ + 134, + 682, + 152, + 693 + ], + "score": 0.43, + "content": "\\pmb { \\Vert 3 4 \\Vert }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 681, + 505, + 695 + ], + "score": 1.0, + "content": ", it still fails to produce robust partners for novel humans in pure common-payoff settings", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 37.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 105, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 702, + 192, + 712 + ], + "lines": [ + { + "bbox": [ + 119, + 700, + 194, + 713 + ], + "spans": [ + { + "bbox": [ + 119, + 700, + 194, + 713 + ], + "score": 1.0, + "content": "⇤Equal contribution.", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 121, + 97, + 491, + 117 + ], + "lines": [ + { + "bbox": [ + 121, + 97, + 491, + 119 + ], + "spans": [ + { + "bbox": [ + 121, + 97, + 491, + 119 + ], + "score": 1.0, + "content": "Collaborating with Humans without Human Data", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 122, + 159, + 489, + 193 + ], + "lines": [ + { + "bbox": [ + 135, + 159, + 480, + 172 + ], + "spans": [ + { + "bbox": [ + 135, + 159, + 480, + 172 + ], + "score": 1.0, + "content": "DJ Strouse⇤, Kevin R. McKee, Matt Botvinick, Edward Hughes, Richard Everett⇤", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 282, + 170, + 329, + 182 + ], + "spans": [ + { + "bbox": [ + 282, + 170, + 329, + 182 + ], + "score": 1.0, + "content": "DeepMind", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 180, + 491, + 195 + ], + "spans": [ + { + "bbox": [ + 120, + 180, + 491, + 195 + ], + "score": 1.0, + "content": "{strouse, kevinrmckee, botvinick, edwardhughes, reverett}@deepmind.com", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 120, + 159, + 491, + 195 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 221, + 328, + 234 + ], + "lines": [ + { + "bbox": [ + 281, + 220, + 330, + 235 + ], + "spans": [ + { + "bbox": [ + 281, + 220, + 330, + 235 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 142, + 246, + 469, + 465 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 471, + 260 + ], + "score": 1.0, + "content": "Collaborating with humans requires rapidly adapting to their individual strengths,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "score": 1.0, + "content": "weaknesses, and preferences. Unfortunately, most standard multi-agent reinforce-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "ment learning techniques, such as self-play (SP) or population play (PP), produce", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 280, + 471, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 471, + 292 + ], + "score": 1.0, + "content": "agents that overfit to their training partners and do not generalize well to humans.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 290, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 302 + ], + "score": 1.0, + "content": "Alternatively, researchers can collect human data, train a human model using behav-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 302, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 313 + ], + "score": 1.0, + "content": "ioral cloning, and then use that model to train “human-aware” agents (“behavioral", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 469, + 325 + ], + "score": 1.0, + "content": "cloning play”, or BCP). While such an approach can improve the generalization", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 322, + 470, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 470, + 336 + ], + "score": 1.0, + "content": "of agents to new human co-players, it involves the onerous and expensive step of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "spans": [ + { + "bbox": [ + 142, + 335, + 469, + 347 + ], + "score": 1.0, + "content": "collecting large amounts of human data first. Here, we study the problem of how", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 345, + 471, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 471, + 357 + ], + "score": 1.0, + "content": "to train agents that collaborate well with human partners without using human data.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 355, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 355, + 470, + 369 + ], + "score": 1.0, + "content": "We argue that the crux of the problem is to produce a diverse set of training part-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 367, + 470, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 380 + ], + "score": 1.0, + "content": "ners. Drawing inspiration from successful multi-agent approaches in competitive", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 390 + ], + "score": 1.0, + "content": "domains, we find that a surprisingly simple approach is highly effective. We train", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 401 + ], + "score": 1.0, + "content": "our agent partner as the best response to a population of self-play agents and their", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 399, + 470, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 399, + 470, + 412 + ], + "score": 1.0, + "content": "past checkpoints taken throughout training, a method we call Fictitious Co-Play", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 410, + 470, + 423 + ], + "score": 1.0, + "content": "(FCP). Our experiments focus on a two-player collaborative cooking simulator that", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 422, + 471, + 434 + ], + "spans": [ + { + "bbox": [ + 141, + 422, + 471, + 434 + ], + "score": 1.0, + "content": "has recently been proposed as a challenge problem for coordination with humans.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "spans": [ + { + "bbox": [ + 141, + 432, + 469, + 444 + ], + "score": 1.0, + "content": "We find that FCP agents score significantly higher than SP, PP, and BCP when", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 444, + 470, + 456 + ], + "score": 1.0, + "content": "paired with novel agent and human partners. Furthermore, humans also report a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 455, + 453, + 466 + ], + "spans": [ + { + "bbox": [ + 141, + 455, + 453, + 466 + ], + "score": 1.0, + "content": "strong subjective preference to partnering with FCP agents over all baselines.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 14.5, + "bbox_fs": [ + 141, + 246, + 471, + 466 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 487, + 190, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 192, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 192, + 502 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 600 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 524 + ], + "score": 1.0, + "content": "Generating agents which collaborate with novel partners is a longstanding challenge for Artificial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 522, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 434, + 536 + ], + "score": 1.0, + "content": "Intelligence (AI) [4, 16, 37, 52]. Achieving ad-hoc, zero-shot coordination [31,", + "type": "text" + }, + { + "bbox": [ + 434, + 523, + 451, + 534 + ], + "score": 0.26, + "content": "\\boxed { 6 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 522, + 505, + 536 + ], + "score": 1.0, + "content": "is especially", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 547 + ], + "score": 1.0, + "content": "important in situations where an AI must generalize to novel human partners [6, 61]. Many successful", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "approaches have employed human models, either constructed explicitly [14, 35, 53] or learnt implicitly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 108, + 556, + 506, + 567 + ], + "spans": [ + { + "bbox": [ + 108, + 556, + 506, + 567 + ], + "score": 1.0, + "content": "[12, 60]. By contrast, recent work in competitive domains has shown that it is possible to reach human-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 505, + 579 + ], + "score": 1.0, + "content": "level using model-free reinforcement learning (RL) without human data, via self-play [8, 9, 63, 64].", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "This begs the question: Can model-free RL without human data generate agents that can collaborate", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 589, + 188, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 188, + 600 + ], + "score": 1.0, + "content": "with novel humans?", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 512, + 506, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 605, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "We seek an answer to this question in the space of common-payoff games, where all agents work", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "towards a shared goal and receive the same reward. Self-play (SP), in which an agent learns from", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "repeated games played against copies of itself, does not produce agents that generalize well to novel", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "co-players [10, 11, 21, 44]. Intuitively, this is because agents trained in self-play only ever need to", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "coordinate with themselves, and so make for brittle and stubborn collaborators with new partners", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 505, + 672 + ], + "score": 1.0, + "content": "who act differently. Population play (PP) trains a population of agents, all of whom interact with each", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 130, + 684 + ], + "score": 1.0, + "content": "other", + "type": "text" + }, + { + "bbox": [ + 130, + 670, + 147, + 682 + ], + "score": 0.65, + "content": "\\pmb { \\| 3 9 \\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 669, + 506, + 684 + ], + "score": 1.0, + "content": ". While PP can generate agents capable of cooperation with humans in competitive team", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 681, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 134, + 693 + ], + "score": 1.0, + "content": "games", + "type": "text" + }, + { + "bbox": [ + 134, + 682, + 152, + 693 + ], + "score": 0.43, + "content": "\\pmb { \\Vert 3 4 \\Vert }", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 681, + 505, + 695 + ], + "score": 1.0, + "content": ", it still fails to produce robust partners for novel humans in pure common-payoff settings", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "[12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 249, + 410, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 389, + 262 + ], + "score": 1.0, + "content": "strategic diversity and producing agents not so different from self-play", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 390, + 249, + 407, + 261 + ], + "score": 0.63, + "content": "\\dot { \\lVert 2 4 \\rVert }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 408, + 250, + 410, + 262 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 605, + 506, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 70, + 493, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 70, + 493, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 70, + 493, + 187 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 493, + 187 + ], + "score": 0.97, + "type": "image", + "image_path": "70b2b8c1e9478aed36abef766805e05c03561ba29533d4a7de7ad061fa2d261f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 70, + 493, + 109.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 109.0, + 493, + 148.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 148.0, + 493, + 187.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 194, + 505, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "Figure 1: In this work, we evaluate a variety of agent training methods (Section 2) in zero-shot", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "coordination with agents (Section 4). We then run a human-agent collaborative study designed to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 217, + 300, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 300, + 229 + ], + "score": 1.0, + "content": "elicit human preferences over agents (Section 5)", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 239, + 503, + 262 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "[12]. PP in common-payoff settings naturally encourages agents to play the same way, reducing", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 249, + 410, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 389, + 262 + ], + "score": 1.0, + "content": "strategic diversity and producing agents not so different from self-play", + "type": "text" + }, + { + "bbox": [ + 390, + 249, + 407, + 261 + ], + "score": 0.63, + "content": "\\dot { \\lVert 2 4 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 250, + 410, + 262 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 279 + ], + "score": 1.0, + "content": "Our approach starts with the intuition that the key to producing robust agent collaborators is exposure", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "to diverse training partners. We find that a surprisingly simple strategy is effective in generating", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 222, + 301 + ], + "score": 1.0, + "content": "sufficient diversity. We train", + "type": "text" + }, + { + "bbox": [ + 223, + 289, + 233, + 298 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 288, + 506, + 301 + ], + "score": 1.0, + "content": "self-play agents varying only their random seed for neural network", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "score": 1.0, + "content": "initialization. Periodically during training, we save agent “checkpoints” representing their strategy", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "at that point in time. Then, we train an agent partner as the best-response to both the fully-trained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "agents and their past checkpoints. The different checkpoints simulate different skill levels, and the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "different random seeds simulate breaking symmetries in different ways. We refer to this agent training", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 342, + 487, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 202, + 355 + ], + "score": 1.0, + "content": "procedure as Fictitious", + "type": "text" + }, + { + "bbox": [ + 202, + 343, + 216, + 353 + ], + "score": 0.26, + "content": "\\mathbf { C o }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 342, + 487, + 355 + ], + "score": 1.0, + "content": "-Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69].", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 359, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "score": 1.0, + "content": "We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 237, + 383 + ], + "score": 1.0, + "content": "Based on the game Overcooked", + "type": "text" + }, + { + "bbox": [ + 237, + 369, + 254, + 381 + ], + "score": 0.63, + "content": "\\mathbb { \\left. \\boldsymbol { \\ 2 5 } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 369, + 506, + 383 + ], + "score": 1.0, + "content": ", it has recently been proposed as a coordination challenge for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "score": 1.0, + "content": "AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 207, + 405 + ], + "score": 1.0, + "content": "humans was achieved in", + "type": "text" + }, + { + "bbox": [ + 207, + 391, + 225, + 403 + ], + "score": 0.66, + "content": "[ \\mathbb { 1 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "via behavioral cloning (BC) of human data. More precisely, BC was", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "used to produce models that can stand in as human proxies during training in simulation, a method", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 411, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 411, + 505, + 427 + ], + "score": 1.0, + "content": "we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "score": 1.0, + "content": "both novel agent and human partners, and that humans express a significant preference for partnering", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 446, + 403, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 403, + 458 + ], + "score": 1.0, + "content": "data for training, while achieving better outcomes for humans at test time.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 358, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 360, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 360, + 475 + ], + "score": 1.0, + "content": "We summarize the novel contributions of this paper as follows:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 129, + 483, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 129, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 129, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 494, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 141, + 494, + 230, + 506 + ], + "score": 1.0, + "content": "humans (Section 2.1).", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 128, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 128, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 519, + 381, + 532 + ], + "spans": [ + { + "bbox": [ + 141, + 519, + 381, + 532 + ], + "score": 1.0, + "content": "coordination with a variety of held-out agents (Section 4.2).", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 128, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 128, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "3. We propose a rigorous human-agent interaction study with behavioral analysis and partici-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 543, + 255, + 556 + ], + "spans": [ + { + "bbox": [ + 141, + 545, + 235, + 556 + ], + "score": 1.0, + "content": "pant feedback (Section", + "type": "text" + }, + { + "bbox": [ + 235, + 543, + 252, + 556 + ], + "score": 0.32, + "content": "5 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 545, + 255, + 556 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 128, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 128, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 568, + 352, + 581 + ], + "spans": [ + { + "bbox": [ + 141, + 569, + 332, + 581 + ], + "score": 1.0, + "content": "score and in human partner preference (Section", + "type": "text" + }, + { + "bbox": [ + 332, + 568, + 350, + 581 + ], + "score": 0.57, + "content": "\\underline { { \\overline { { 5 . 2 } } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 569, + 352, + 581 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 171, + 609 + ], + "lines": [ + { + "bbox": [ + 104, + 592, + 172, + 611 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 172, + 611 + ], + "score": 1.0, + "content": "2 Methods", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 107, + 619, + 236, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 237, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 237, + 634 + ], + "score": 1.0, + "content": "2.1 Fictitious Co-Play (FCP)", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "Diverse training conditions have been shown to make agents more robust, from environmental", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 264, + 663 + ], + "score": 1.0, + "content": "variations (i.e. domain randomization", + "type": "text" + }, + { + "bbox": [ + 264, + 650, + 313, + 662 + ], + "score": 0.43, + "content": "\\textcircled { 1 5 4 } , \\textcircled { 5 6 } , \\textcircled { 6 7 } \\textcircled { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 650, + 466, + 663 + ], + "score": 1.0, + "content": "to heterogeneity in training partners", + "type": "text" + }, + { + "bbox": [ + 466, + 650, + 484, + 662 + ], + "score": 0.42, + "content": "\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { g } } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 650, + 506, + 663 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "score": 1.0, + "content": "seek to train agents that are robust partners for humans in common-payoff games, and so extend this", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 671, + 217, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 217, + 685 + ], + "score": 1.0, + "content": "line of work to that setting.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 466, + 702 + ], + "score": 1.0, + "content": "One important challenge in collaborating with novel partners is dealing with symmetries", + "type": "text" + }, + { + "bbox": [ + 466, + 688, + 484, + 700 + ], + "score": 0.33, + "content": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 688, + 506, + 702 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "example, two agents A and B facing each other may move past each other by A going left and B going", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 8 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 70, + 493, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 70, + 493, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 70, + 493, + 187 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 493, + 187 + ], + "score": 0.97, + "type": "image", + "image_path": "70b2b8c1e9478aed36abef766805e05c03561ba29533d4a7de7ad061fa2d261f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 70, + 493, + 109.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 109.0, + 493, + 148.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 148.0, + 493, + 187.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 194, + 505, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "Figure 1: In this work, we evaluate a variety of agent training methods (Section 2) in zero-shot", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "coordination with agents (Section 4). We then run a human-agent collaborative study designed to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 217, + 300, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 300, + 229 + ], + "score": 1.0, + "content": "elicit human preferences over agents (Section 5)", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 108, + 239, + 503, + 262 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 106, + 237, + 505, + 262 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 505, + 279 + ], + "score": 1.0, + "content": "Our approach starts with the intuition that the key to producing robust agent collaborators is exposure", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 277, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "to diverse training partners. We find that a surprisingly simple strategy is effective in generating", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 222, + 301 + ], + "score": 1.0, + "content": "sufficient diversity. We train", + "type": "text" + }, + { + "bbox": [ + 223, + 289, + 233, + 298 + ], + "score": 0.78, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 288, + 506, + 301 + ], + "score": 1.0, + "content": "self-play agents varying only their random seed for neural network", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "score": 1.0, + "content": "initialization. Periodically during training, we save agent “checkpoints” representing their strategy", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 322 + ], + "score": 1.0, + "content": "at that point in time. Then, we train an agent partner as the best-response to both the fully-trained", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 333 + ], + "score": 1.0, + "content": "agents and their past checkpoints. The different checkpoints simulate different skill levels, and the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "different random seeds simulate breaking symmetries in different ways. We refer to this agent training", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 342, + 487, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 202, + 355 + ], + "score": 1.0, + "content": "procedure as Fictitious", + "type": "text" + }, + { + "bbox": [ + 202, + 343, + 216, + 353 + ], + "score": 0.26, + "content": "\\mathbf { C o }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 342, + 487, + 355 + ], + "score": 1.0, + "content": "-Play (FCP) for its relationship to fictitious self-play [7, 27, 28, 69].", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 265, + 506, + 355 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 359, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 506, + 371 + ], + "score": 1.0, + "content": "We evaluate FCP in a fully-observable two-player common-payoff collaborative cooking simulator.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 369, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 237, + 383 + ], + "score": 1.0, + "content": "Based on the game Overcooked", + "type": "text" + }, + { + "bbox": [ + 237, + 369, + 254, + 381 + ], + "score": 0.63, + "content": "\\mathbb { \\left. \\boldsymbol { \\ 2 5 } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 369, + 506, + 383 + ], + "score": 1.0, + "content": ", it has recently been proposed as a coordination challenge for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 394 + ], + "score": 1.0, + "content": "AI [12, 50, 70]. State-of-the-art performance in producing agents capable of generalization to novel", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 207, + 405 + ], + "score": 1.0, + "content": "humans was achieved in", + "type": "text" + }, + { + "bbox": [ + 207, + 391, + 225, + 403 + ], + "score": 0.66, + "content": "[ \\mathbb { 1 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "via behavioral cloning (BC) of human data. More precisely, BC was", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 416 + ], + "score": 1.0, + "content": "used to produce models that can stand in as human proxies during training in simulation, a method", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 411, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 104, + 411, + 505, + 427 + ], + "score": 1.0, + "content": "we call behavioral cloning play (BCP). We demonstrate that FCP outperforms BCP in generalizing to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "score": 1.0, + "content": "both novel agent and human partners, and that humans express a significant preference for partnering", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "with FCP over BCP. Our method avoids the cost and potential privacy concerns of collecting human", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 446, + 403, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 403, + 458 + ], + "score": 1.0, + "content": "data for training, while achieving better outcomes for humans at test time.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 358, + 506, + 458 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 462, + 358, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 360, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 360, + 475 + ], + "score": 1.0, + "content": "We summarize the novel contributions of this paper as follows:", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 461, + 360, + 475 + ] + }, + { + "type": "list", + "bbox": [ + 129, + 483, + 505, + 581 + ], + "lines": [ + { + "bbox": [ + 129, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 129, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "1. We propose Fictitious Co-Play (FCP) to train agents capable of zero-shot coordination with", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 494, + 230, + 506 + ], + "spans": [ + { + "bbox": [ + 141, + 494, + 230, + 506 + ], + "score": 1.0, + "content": "humans (Section 2.1).", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 128, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "2. We demonstrate that FCP agents generalize better than SP, PP, and BCP in zero-shot", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 519, + 381, + 532 + ], + "spans": [ + { + "bbox": [ + 141, + 519, + 381, + 532 + ], + "score": 1.0, + "content": "coordination with a variety of held-out agents (Section 4.2).", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 532, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 128, + 532, + 506, + 545 + ], + "score": 1.0, + "content": "3. We propose a rigorous human-agent interaction study with behavioral analysis and partici-", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 543, + 255, + 556 + ], + "spans": [ + { + "bbox": [ + 141, + 545, + 235, + 556 + ], + "score": 1.0, + "content": "pant feedback (Section", + "type": "text" + }, + { + "bbox": [ + 235, + 543, + 252, + 556 + ], + "score": 0.32, + "content": "5 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 545, + 255, + 556 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 128, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "4. We demonstrate that FCP significantly outperforms the BCP state-of-the-art, both in task", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 568, + 352, + 581 + ], + "spans": [ + { + "bbox": [ + 141, + 569, + 332, + 581 + ], + "score": 1.0, + "content": "score and in human partner preference (Section", + "type": "text" + }, + { + "bbox": [ + 332, + 568, + 350, + 581 + ], + "score": 0.57, + "content": "\\underline { { \\overline { { 5 . 2 } } } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 569, + 352, + 581 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + } + ], + "index": 29.5, + "bbox_fs": [ + 128, + 483, + 506, + 581 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 171, + 609 + ], + "lines": [ + { + "bbox": [ + 104, + 592, + 172, + 611 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 172, + 611 + ], + "score": 1.0, + "content": "2 Methods", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 107, + 619, + 236, + 631 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 237, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 237, + 634 + ], + "score": 1.0, + "content": "2.1 Fictitious Co-Play (FCP)", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 640, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "Diverse training conditions have been shown to make agents more robust, from environmental", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 650, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 264, + 663 + ], + "score": 1.0, + "content": "variations (i.e. domain randomization", + "type": "text" + }, + { + "bbox": [ + 264, + 650, + 313, + 662 + ], + "score": 0.43, + "content": "\\textcircled { 1 5 4 } , \\textcircled { 5 6 } , \\textcircled { 6 7 } \\textcircled { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 650, + 466, + 663 + ], + "score": 1.0, + "content": "to heterogeneity in training partners", + "type": "text" + }, + { + "bbox": [ + 466, + 650, + 484, + 662 + ], + "score": 0.42, + "content": "\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { g } } \\right. }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 650, + 506, + 663 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "score": 1.0, + "content": "seek to train agents that are robust partners for humans in common-payoff games, and so extend this", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 671, + 217, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 217, + 685 + ], + "score": 1.0, + "content": "line of work to that setting.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 640, + 506, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 466, + 702 + ], + "score": 1.0, + "content": "One important challenge in collaborating with novel partners is dealing with symmetries", + "type": "text" + }, + { + "bbox": [ + 466, + 688, + 484, + 700 + ], + "score": 0.33, + "content": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 688, + 506, + 702 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "example, two agents A and B facing each other may move past each other by A going left and B going", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "right, or vice versa. Both are valid solutions, but a good agent partner will adaptively switch between", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "these conventions if a human clearly prefers one over the other. A second important challenge is", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 322, + 308, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 308, + 336 + ], + "score": 1.0, + "content": "partners, as well as partners who are still learning.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 688, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 70, + 495, + 219 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 70, + 495, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 70, + 495, + 219 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 495, + 219 + ], + "score": 0.97, + "type": "image", + "image_path": "2372b76b2e16cbfb2db5439982c244e9b3e60b30aff99699aa5043d431598f6f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 70, + 495, + 119.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 119.66666666666666, + 495, + 169.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 169.33333333333331, + 495, + 218.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 225, + 505, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "Figure 2: The four agent training methods we evaluate in this work. Self-play (SP) where an agent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "learns with itself, population-play (PP) where a population of agents are co-trained together, and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 246, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 260 + ], + "score": 1.0, + "content": "behavioral cloning play (BCP) where data from human games is used to create a behaviorally", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 258, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 493, + 271 + ], + "score": 1.0, + "content": "cloned agent with which an RL agent is then trained. In our method, Fictitious Co-Play (FCP),", + "type": "text" + }, + { + "bbox": [ + 493, + 258, + 504, + 268 + ], + "score": 0.68, + "content": "N", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "self-play agents are trained independently and checkpointed throughout training. An agent is then", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 280, + 438, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 438, + 292 + ], + "score": 1.0, + "content": "trained to best respond to the entire population of SP agents and their checkpoints.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 108, + 300, + 503, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 313 + ], + "score": 1.0, + "content": "these conventions if a human clearly prefers one over the other. A second important challenge is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "dealing with variations in skill level. Good agent partners should be able to assist both highly-skilled", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 322, + 308, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 322, + 308, + 336 + ], + "score": 1.0, + "content": "partners, as well as partners who are still learning.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 352 + ], + "score": 1.0, + "content": "Fictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 221, + 362 + ], + "score": 1.0, + "content": "of these challenges (Figure", + "type": "text" + }, + { + "bbox": [ + 221, + 349, + 231, + 362 + ], + "score": 0.45, + "content": "2 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "right). In the first stage, we train a diverse pool of partners. To", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 432, + 373 + ], + "score": 1.0, + "content": "allow the pool to represent different symmetry breaking conventions, we train", + "type": "text" + }, + { + "bbox": [ + 432, + 361, + 443, + 370 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 360, + 505, + 373 + ], + "score": 1.0, + "content": "partner agents", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "in self-play. Since these partners are trained independently, they can arrive at different arbitrary", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "conventions for breaking symmetries. To allow the pool to represent different skill levels, we use", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 415, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 506, + 427 + ], + "score": 1.0, + "content": "using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "score": 1.0, + "content": "In the second stage, we train an FCP agent as the best response to the pool of diverse partners created", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 452, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 506, + 466 + ], + "score": 1.0, + "content": "partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "lead of human partners, and learn a general policy across a range of strategies and skills. We call our", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 485, + 456, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 456, + 498 + ], + "score": 1.0, + "content": "trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69].", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 107, + 510, + 230, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 230, + 522 + ], + "score": 1.0, + "content": "2.2 Baselines and ablations", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 529, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "We compare FCP agents to the three baseline training methods listed below, each varying only in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 540, + 493, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 493, + 553 + ], + "score": 1.0, + "content": "their set of training partners, with the RL algorithm and architecture consistent across all agents:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 129, + 560, + 506, + 601 + ], + "lines": [ + { + "bbox": [ + 129, + 560, + 455, + 573 + ], + "spans": [ + { + "bbox": [ + 129, + 560, + 455, + 573 + ], + "score": 1.0, + "content": "1. Self-play (SP), where agents learn solely through interaction with themselves.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 128, + 573, + 507, + 588 + ], + "spans": [ + { + "bbox": [ + 128, + 573, + 507, + 588 + ], + "score": 1.0, + "content": "2. Population-play (PP), where a population of agents are co-trained through random pairings.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 128, + 587, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 128, + 587, + 506, + 602 + ], + "score": 1.0, + "content": "3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12].", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 110, + 608, + 473, + 620 + ], + "lines": [ + { + "bbox": [ + 107, + 607, + 474, + 621 + ], + "spans": [ + { + "bbox": [ + 107, + 607, + 474, + 621 + ], + "score": 1.0, + "content": "We also evaluate three variations on FCP to better understand the conditions for its success:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 130, + 627, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 129, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 129, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "1. To test the importance of including past checkpoints in training, we evaluate an ablation", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 141, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "of FCP in which agents are trained only with the converged checkpoints of their partners", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 142, + 650, + 273, + 662 + ], + "spans": [ + { + "bbox": [ + 142, + 650, + 145, + 661 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 145, + 650, + 177, + 662 + ], + "score": 0.84, + "content": "\\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 650, + 273, + 661 + ], + "score": 1.0, + "content": "for “FCP minus time”).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 129, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 129, + 664, + 506, + 677 + ], + "score": 1.0, + "content": "2. To test whether FCP would benefit from additional diversity in its partner population, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 142, + 676, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 142, + 676, + 505, + 687 + ], + "score": 1.0, + "content": "evaluate an augmentation of FCP in which the population of SP partners varies not just in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 686, + 485, + 698 + ], + "spans": [ + { + "bbox": [ + 141, + 686, + 294, + 697 + ], + "score": 1.0, + "content": "random seed, but also in architecture", + "type": "text" + }, + { + "bbox": [ + 295, + 686, + 326, + 698 + ], + "score": 0.88, + "content": "\\operatorname { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 686, + 485, + 697 + ], + "score": 1.0, + "content": "for “FCP plus architectural variation”).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 128, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 128, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "3. To test whether architectural variation can serve as a full replacement for playing with past", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 710, + 455, + 724 + ], + "spans": [ + { + "bbox": [ + 141, + 710, + 399, + 724 + ], + "score": 1.0, + "content": "checkpoints, we evaluate the combination of both modifications", + "type": "text" + }, + { + "bbox": [ + 399, + 711, + 450, + 723 + ], + "score": 0.87, + "content": "( \\mathrm { F C P } _ { - T , + A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 710, + 455, + 724 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 70, + 495, + 219 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 70, + 495, + 219 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 70, + 495, + 219 + ], + "spans": [ + { + "bbox": [ + 118, + 70, + 495, + 219 + ], + "score": 0.97, + "type": "image", + "image_path": "2372b76b2e16cbfb2db5439982c244e9b3e60b30aff99699aa5043d431598f6f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 70, + 495, + 119.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 119.66666666666666, + 495, + 169.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 169.33333333333331, + 495, + 218.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 225, + 505, + 291 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "Figure 2: The four agent training methods we evaluate in this work. Self-play (SP) where an agent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 506, + 249 + ], + "score": 1.0, + "content": "learns with itself, population-play (PP) where a population of agents are co-trained together, and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 246, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 246, + 505, + 260 + ], + "score": 1.0, + "content": "behavioral cloning play (BCP) where data from human games is used to create a behaviorally", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 258, + 504, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 493, + 271 + ], + "score": 1.0, + "content": "cloned agent with which an RL agent is then trained. In our method, Fictitious Co-Play (FCP),", + "type": "text" + }, + { + "bbox": [ + 493, + 258, + 504, + 268 + ], + "score": 0.68, + "content": "N", + "type": "inline_equation" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 505, + 282 + ], + "score": 1.0, + "content": "self-play agents are trained independently and checkpointed throughout training. An agent is then", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 280, + 438, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 438, + 292 + ], + "score": 1.0, + "content": "trained to best respond to the entire population of SP agents and their checkpoints.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 108, + 300, + 503, + 334 + ], + "lines": [], + "index": 10, + "bbox_fs": [ + 106, + 299, + 505, + 336 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 506, + 352 + ], + "score": 1.0, + "content": "Fictitious co-play (FCP) is a simple two-stage approach for training agents that overcomes both", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 349, + 505, + 362 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 221, + 362 + ], + "score": 1.0, + "content": "of these challenges (Figure", + "type": "text" + }, + { + "bbox": [ + 221, + 349, + 231, + 362 + ], + "score": 0.45, + "content": "2 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 349, + 505, + 362 + ], + "score": 1.0, + "content": "right). In the first stage, we train a diverse pool of partners. To", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 360, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 106, + 360, + 432, + 373 + ], + "score": 1.0, + "content": "allow the pool to represent different symmetry breaking conventions, we train", + "type": "text" + }, + { + "bbox": [ + 432, + 361, + 443, + 370 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 360, + 505, + 373 + ], + "score": 1.0, + "content": "partner agents", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "in self-play. Since these partners are trained independently, they can arrive at different arbitrary", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 382, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 106, + 382, + 505, + 394 + ], + "score": 1.0, + "content": "conventions for breaking symmetries. To allow the pool to represent different skill levels, we use", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 393, + 506, + 406 + ], + "score": 1.0, + "content": "multiple checkpoints of each self-play partner throughout training. The final checkpoint represents a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "fully-trained “skillful” partner, while earlier checkpoints represent less skilled partners. Notably, by", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 415, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 506, + 427 + ], + "score": 1.0, + "content": "using multiple checkpoints per partner, this additional diversity in skill incurs no extra training cost.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 338, + 506, + 427 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 431, + 505, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 431, + 506, + 444 + ], + "score": 1.0, + "content": "In the second stage, we train an FCP agent as the best response to the pool of diverse partners created", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 506, + 455 + ], + "score": 1.0, + "content": "in the first stage. Importantly, the partner parameters are frozen and thus FCP must learn to adapt to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 452, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 506, + 466 + ], + "score": 1.0, + "content": "partners, rather than expect partners to adapt to it. In this way, FCP agents are prepared to follow the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 477 + ], + "score": 1.0, + "content": "lead of human partners, and learn a general policy across a range of strategies and skills. We call our", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "method “fictitious” co-play for its relationship to fictitious self-play in which competitive agents are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 485, + 456, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 456, + 498 + ], + "score": 1.0, + "content": "trained with past checkpoints (in that case, to avoid strategy cycling) [7, 27, 28, 39, 69].", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 431, + 506, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 510, + 230, + 521 + ], + "lines": [ + { + "bbox": [ + 106, + 509, + 230, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 509, + 230, + 522 + ], + "score": 1.0, + "content": "2.2 Baselines and ablations", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 108, + 529, + 505, + 552 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "We compare FCP agents to the three baseline training methods listed below, each varying only in", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 540, + 493, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 493, + 553 + ], + "score": 1.0, + "content": "their set of training partners, with the RL algorithm and architecture consistent across all agents:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 106, + 529, + 505, + 553 + ] + }, + { + "type": "index", + "bbox": [ + 129, + 560, + 506, + 601 + ], + "lines": [ + { + "bbox": [ + 129, + 560, + 455, + 573 + ], + "spans": [ + { + "bbox": [ + 129, + 560, + 455, + 573 + ], + "score": 1.0, + "content": "1. Self-play (SP), where agents learn solely through interaction with themselves.", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 573, + 507, + 588 + ], + "spans": [ + { + "bbox": [ + 128, + 573, + 507, + 588 + ], + "score": 1.0, + "content": "2. Population-play (PP), where a population of agents are co-trained through random pairings.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 587, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 128, + 587, + 506, + 602 + ], + "score": 1.0, + "content": "3. Behavioral cloning play (BCP), where an agent is trained with a BC model of a human [12].", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + } + ], + "index": 30, + "bbox_fs": [ + 128, + 560, + 507, + 602 + ] + }, + { + "type": "text", + "bbox": [ + 110, + 608, + 473, + 620 + ], + "lines": [ + { + "bbox": [ + 107, + 607, + 474, + 621 + ], + "spans": [ + { + "bbox": [ + 107, + 607, + 474, + 621 + ], + "score": 1.0, + "content": "We also evaluate three variations on FCP to better understand the conditions for its success:", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32, + "bbox_fs": [ + 107, + 607, + 474, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 130, + 627, + 505, + 723 + ], + "lines": [ + { + "bbox": [ + 129, + 627, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 129, + 627, + 505, + 641 + ], + "score": 1.0, + "content": "1. To test the importance of including past checkpoints in training, we evaluate an ablation", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 639, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 141, + 639, + 505, + 651 + ], + "score": 1.0, + "content": "of FCP in which agents are trained only with the converged checkpoints of their partners", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 142, + 650, + 273, + 662 + ], + "spans": [ + { + "bbox": [ + 142, + 650, + 145, + 661 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 145, + 650, + 177, + 662 + ], + "score": 0.84, + "content": "\\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 650, + 273, + 661 + ], + "score": 1.0, + "content": "for “FCP minus time”).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 129, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 129, + 664, + 506, + 677 + ], + "score": 1.0, + "content": "2. To test whether FCP would benefit from additional diversity in its partner population, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 142, + 676, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 142, + 676, + 505, + 687 + ], + "score": 1.0, + "content": "evaluate an augmentation of FCP in which the population of SP partners varies not just in", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 141, + 686, + 485, + 698 + ], + "spans": [ + { + "bbox": [ + 141, + 686, + 294, + 697 + ], + "score": 1.0, + "content": "random seed, but also in architecture", + "type": "text" + }, + { + "bbox": [ + 295, + 686, + 326, + 698 + ], + "score": 0.88, + "content": "\\operatorname { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 686, + 485, + 697 + ], + "score": 1.0, + "content": "for “FCP plus architectural variation”).", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 128, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 128, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "3. To test whether architectural variation can serve as a full replacement for playing with past", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 141, + 710, + 455, + 724 + ], + "spans": [ + { + "bbox": [ + 141, + 710, + 399, + 724 + ], + "score": 1.0, + "content": "checkpoints, we evaluate the combination of both modifications", + "type": "text" + }, + { + "bbox": [ + 399, + 711, + 450, + 723 + ], + "score": 0.87, + "content": "( \\mathrm { F C P } _ { - T , + A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 710, + 455, + 724 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 36.5, + "bbox_fs": [ + 128, + 627, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 73, + 186, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 188, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 188, + 86 + ], + "score": 1.0, + "content": "2.3 Environment", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 93, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "score": 1.0, + "content": "environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 322, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 322, + 127 + ], + "score": 1.0, + "content": "in Carroll et al. [12]. For full details, see Appendix A.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 130, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 130, + 506, + 144 + ], + "score": 1.0, + "content": "In this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 141, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 504, + 153 + ], + "score": 1.0, + "content": "many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "high-level actions to which both players can contribute: collecting tomatoes, depositing them into", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 161, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 505, + 179 + ], + "score": 1.0, + "content": "cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 368, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 368, + 188 + ], + "score": 1.0, + "content": "it. Upon a successful delivery, both players are rewarded equally.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "To effectively complete the task, players must learn to navigate the kitchen and interact with objects in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "score": 1.0, + "content": "the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 495, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 495, + 225 + ], + "score": 1.0, + "content": "This environment therefore presents the challenges of both movement and strategic coordination.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 229, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 506, + 242 + ], + "score": 1.0, + "content": "Each player observes an egocentric RGB view of the world, and at every step can perform one", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 240, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 506, + 252 + ], + "score": 1.0, + "content": "of six actions: stand still, move {up, down, left, right}, interact. The behavior ofLÈ¡áyҏįŘį\u000eºµò‹µÒ¡y®", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 476, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 476, + 263 + ], + "score": 1.0, + "content": "interact varies based on the cell which the player is facing (e.g. place tomato on counter).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 118, + 273, + 492, + 419 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 273, + 492, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 273, + 492, + 419 + ], + "spans": [ + { + "bbox": [ + 118, + 273, + 492, + 419 + ], + "score": 0.969, + "type": "image", + "image_path": "48e8af75423e0fa5e5c23e2088c49716c4c6044ee3579f26cdd9911dbb4b4925.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 118, + 273, + 492, + 321.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 118, + 321.6666666666667, + 492, + 370.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 118, + 370.33333333333337, + 492, + 419.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 426, + 503, + 448 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 425, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 504, + 438 + ], + "score": 1.0, + "content": "Figure 3: The Overcooked environment: a two-player common-payoff game in which players must", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 436, + 254, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 254, + 450 + ], + "score": 1.0, + "content": "coordinate to cook and deliver soup.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + } + ], + "index": 17.25 + }, + { + "type": "image", + "bbox": [ + 120, + 459, + 492, + 521 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 459, + 492, + 521 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 120, + 459, + 492, + 521 + ], + "spans": [ + { + "bbox": [ + 120, + 459, + 492, + 521 + ], + "score": 0.963, + "type": "image", + "image_path": "2b3253fcaac0677dbb15cc1ed5aa0663e561ff52605ba15bd61605f96c716872.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 120, + 459, + 492, + 479.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 120, + 479.6666666666667, + 492, + 500.33333333333337 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 120, + 500.33333333333337, + 492, + 521.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 527, + 504, + 550 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "Figure 4: Layouts: the kitchens which agents and humans play in, each emphasizing different", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "coordination strategies. Highlighted in bold are the terms used to refer to each in the rest of this paper.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + } + ], + "index": 22.25 + }, + { + "type": "title", + "bbox": [ + 107, + 570, + 229, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 230, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 230, + 584 + ], + "score": 1.0, + "content": "2.4 Implementation details", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 507, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 507, + 603 + ], + "score": 1.0, + "content": "Here we highlight several key implementation details for our training methods. For full details,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "score": 1.0, + "content": "including the architectures, hyperparameters, and compute used, please see Appendix B.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 341, + 630 + ], + "score": 1.0, + "content": "For our reinforcement learning agents, we use the V-MPO", + "type": "text" + }, + { + "bbox": [ + 342, + 617, + 359, + 629 + ], + "score": 0.45, + "content": "\\boldsymbol { \\left[ \\left[ 6 5 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 617, + 504, + 630 + ], + "score": 1.0, + "content": "algorithm along with a ResNet [26]", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 155, + 641 + ], + "score": 1.0, + "content": "plus LSTM", + "type": "text" + }, + { + "bbox": [ + 155, + 628, + 173, + 640 + ], + "score": 0.7, + "content": "\\mathbb { \\left| \\bigstar \\bigstar \\right\\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "architecture which we found led to optimal behavior across all layouts. Agents are", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 375, + 653 + ], + "score": 1.0, + "content": "trained using a distributed set of environments running in parallel", + "type": "text" + }, + { + "bbox": [ + 376, + 639, + 393, + 651 + ], + "score": 0.69, + "content": "\\textcircled { 1 1 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 639, + 505, + 653 + ], + "score": 1.0, + "content": ", each sampling two agents", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 650, + 346, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 346, + 663 + ], + "score": 1.0, + "content": "from the training population to play together every episode.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 507, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 324, + 680 + ], + "score": 1.0, + "content": "Both PP and FCP are trained with a population size of", + "type": "text" + }, + { + "bbox": [ + 324, + 667, + 357, + 677 + ], + "score": 0.89, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 667, + 507, + 680 + ], + "score": 1.0, + "content": "agents which are sampled uniformly.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "(1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 410, + 713 + ], + "score": 1.0, + "content": "agent), and (3) at the middle of training, defined as when the agent reaches", + "type": "text" + }, + { + "bbox": [ + 410, + 700, + 430, + 711 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "of its final reward", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 455, + 723 + ], + "score": 1.0, + "content": "(i.e. an average-skilled agent). When varying architecture for the training partners of the", + "type": "text" + }, + { + "bbox": [ + 456, + 711, + 487, + 722 + ], + "score": 0.9, + "content": "\\operatorname { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 73, + 186, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 188, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 188, + 86 + ], + "score": 1.0, + "content": "2.3 Environment", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 93, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 505, + 105 + ], + "score": 1.0, + "content": "Following prior work on zero-shot coordination in human-agent interaction, we study the Overcooked", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 117 + ], + "score": 1.0, + "content": "environment (see Figure 3) [12, 13, 38, 50, 70]. We draw particular inspiration from the environment", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 322, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 322, + 127 + ], + "score": 1.0, + "content": "in Carroll et al. [12]. For full details, see Appendix A.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 92, + 506, + 127 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 131, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 130, + 506, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 130, + 506, + 144 + ], + "score": 1.0, + "content": "In this environment, players are placed into a gridworld kitchen as chefs and tasked with delivering as", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 141, + 504, + 153 + ], + "spans": [ + { + "bbox": [ + 106, + 141, + 504, + 153 + ], + "score": 1.0, + "content": "many cooked dishes of tomato soup as possible within an episode. This involves a series of sequential", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "high-level actions to which both players can contribute: collecting tomatoes, depositing them into", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 161, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 505, + 179 + ], + "score": 1.0, + "content": "cooking pots, letting the tomatoes cook into soup, collecting a dish, getting the soup, and delivering", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 174, + 368, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 368, + 188 + ], + "score": 1.0, + "content": "it. Upon a successful delivery, both players are rewarded equally.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 130, + 506, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 505, + 224 + ], + "lines": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "To effectively complete the task, players must learn to navigate the kitchen and interact with objects in", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 506, + 214 + ], + "score": 1.0, + "content": "the correct order, all while maintaining awareness of their partner’s behavior to coordinate with them.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 495, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 495, + 225 + ], + "score": 1.0, + "content": "This environment therefore presents the challenges of both movement and strategic coordination.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 191, + 506, + 225 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 229, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 506, + 242 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 506, + 242 + ], + "score": 1.0, + "content": "Each player observes an egocentric RGB view of the world, and at every step can perform one", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 240, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 506, + 252 + ], + "score": 1.0, + "content": "of six actions: stand still, move {up, down, left, right}, interact. The behavior ofLÈ¡áyҏįŘį\u000eºµò‹µÒ¡y®", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 250, + 476, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 476, + 263 + ], + "score": 1.0, + "content": "interact varies based on the cell which the player is facing (e.g. place tomato on counter).", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 228, + 506, + 263 + ] + }, + { + "type": "image", + "bbox": [ + 118, + 273, + 492, + 419 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 273, + 492, + 419 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 273, + 492, + 419 + ], + "spans": [ + { + "bbox": [ + 118, + 273, + 492, + 419 + ], + "score": 0.969, + "type": "image", + "image_path": "48e8af75423e0fa5e5c23e2088c49716c4c6044ee3579f26cdd9911dbb4b4925.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 118, + 273, + 492, + 321.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 118, + 321.6666666666667, + 492, + 370.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 118, + 370.33333333333337, + 492, + 419.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 426, + 503, + 448 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 425, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 504, + 438 + ], + "score": 1.0, + "content": "Figure 3: The Overcooked environment: a two-player common-payoff game in which players must", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 436, + 254, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 436, + 254, + 450 + ], + "score": 1.0, + "content": "coordinate to cook and deliver soup.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + } + ], + "index": 17.25 + }, + { + "type": "image", + "bbox": [ + 120, + 459, + 492, + 521 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 120, + 459, + 492, + 521 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 120, + 459, + 492, + 521 + ], + "spans": [ + { + "bbox": [ + 120, + 459, + 492, + 521 + ], + "score": 0.963, + "type": "image", + "image_path": "2b3253fcaac0677dbb15cc1ed5aa0663e561ff52605ba15bd61605f96c716872.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 120, + 459, + 492, + 479.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 120, + 479.6666666666667, + 492, + 500.33333333333337 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 120, + 500.33333333333337, + 492, + 521.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 527, + 504, + 550 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "Figure 4: Layouts: the kitchens which agents and humans play in, each emphasizing different", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "coordination strategies. Highlighted in bold are the terms used to refer to each in the rest of this paper.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + } + ], + "index": 22.25 + }, + { + "type": "title", + "bbox": [ + 107, + 570, + 229, + 582 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 230, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 230, + 584 + ], + "score": 1.0, + "content": "2.4 Implementation details", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 591, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 591, + 507, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 507, + 603 + ], + "score": 1.0, + "content": "Here we highlight several key implementation details for our training methods. For full details,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "score": 1.0, + "content": "including the architectures, hyperparameters, and compute used, please see Appendix B.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 591, + 507, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 504, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 341, + 630 + ], + "score": 1.0, + "content": "For our reinforcement learning agents, we use the V-MPO", + "type": "text" + }, + { + "bbox": [ + 342, + 617, + 359, + 629 + ], + "score": 0.45, + "content": "\\boldsymbol { \\left[ \\left[ 6 5 \\right] \\right] }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 617, + 504, + 630 + ], + "score": 1.0, + "content": "algorithm along with a ResNet [26]", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 155, + 641 + ], + "score": 1.0, + "content": "plus LSTM", + "type": "text" + }, + { + "bbox": [ + 155, + 628, + 173, + 640 + ], + "score": 0.7, + "content": "\\mathbb { \\left| \\bigstar \\bigstar \\right\\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "architecture which we found led to optimal behavior across all layouts. Agents are", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 639, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 375, + 653 + ], + "score": 1.0, + "content": "trained using a distributed set of environments running in parallel", + "type": "text" + }, + { + "bbox": [ + 376, + 639, + 393, + 651 + ], + "score": 0.69, + "content": "\\textcircled { 1 1 7 }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 639, + 505, + 653 + ], + "score": 1.0, + "content": ", each sampling two agents", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 650, + 346, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 650, + 346, + 663 + ], + "score": 1.0, + "content": "from the training population to play together every episode.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 617, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 507, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 324, + 680 + ], + "score": 1.0, + "content": "Both PP and FCP are trained with a population size of", + "type": "text" + }, + { + "bbox": [ + 324, + 667, + 357, + 677 + ], + "score": 0.89, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 667, + 507, + 680 + ], + "score": 1.0, + "content": "agents which are sampled uniformly.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "For FCP, we use 3 checkpoints for each agent, therefore incurring no additional training burden:", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "(1) at initialization (i.e. a low-skilled agent), (2) at the end of training (i.e. a fully-trained expert", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 410, + 713 + ], + "score": 1.0, + "content": "agent), and (3) at the middle of training, defined as when the agent reaches", + "type": "text" + }, + { + "bbox": [ + 410, + 700, + 430, + 711 + ], + "score": 0.87, + "content": "50 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "of its final reward", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 455, + 723 + ], + "score": 1.0, + "content": "(i.e. an average-skilled agent). When varying architecture for the training partners of the", + "type": "text" + }, + { + "bbox": [ + 456, + 711, + 487, + 722 + ], + "score": 0.9, + "content": "\\operatorname { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 152, + 85 + ], + "score": 0.85, + "content": "\\mathrm { F C P } _ { - T , + A }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 153, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 499, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 306, + 107 + ], + "score": 1.0, + "content": "combinations, leaving the total population size of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 306, + 95, + 339, + 105 + ], + "score": 0.89, + "content": "N = 3 2", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 340, + 95, + 499, + 107 + ], + "score": 1.0, + "content": "unchanged, ensuring a fair comparison.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 667, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 152, + 85 + ], + "score": 0.85, + "content": "\\mathrm { F C P } _ { - T , + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 72, + 505, + 86 + ], + "score": 1.0, + "content": "variants, we vary whether the partners use memory (i.e. LSTM vs not) and the width", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "of their policy and value networks (i.e. 16 vs 256). In total, we train 8 agents for each of the 4", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 499, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 306, + 107 + ], + "score": 1.0, + "content": "combinations, leaving the total population size of", + "type": "text" + }, + { + "bbox": [ + 306, + 95, + 339, + 105 + ], + "score": 0.89, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 95, + 499, + 107 + ], + "score": 1.0, + "content": "unchanged, ensuring a fair comparison.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 257, + 124 + ], + "score": 1.0, + "content": "To train agents via behavioral cloning", + "type": "text" + }, + { + "bbox": [ + 257, + 110, + 274, + 122 + ], + "score": 0.79, + "content": "\\left[ \\left[ 5 8 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 110, + 398, + 124 + ], + "score": 1.0, + "content": ", we use the open-source Acme", + "type": "text" + }, + { + "bbox": [ + 399, + 110, + 416, + 122 + ], + "score": 0.77, + "content": "\\pmb { \\mathbb { B } } 0 \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 110, + 505, + 124 + ], + "score": 1.0, + "content": "to learn a policy from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "score": 1.0, + "content": "human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 298, + 167 + ], + "score": 1.0, + "content": "agent-agent evaluation. Following Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 299, + 154, + 316, + 166 + ], + "score": 0.61, + "content": "\\bar { \\lVert 1 2 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 154, + 505, + 167 + ], + "score": 1.0, + "content": ", we use a set of feature-based observations for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 177, + 452, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 452, + 189 + ], + "score": 1.0, + "content": "(asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 107, + 206, + 194, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 205, + 196, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 196, + 221 + ], + "score": 1.0, + "content": "3 Related work", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 456, + 245 + ], + "score": 1.0, + "content": "Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play", + "type": "text" + }, + { + "bbox": [ + 456, + 232, + 483, + 245 + ], + "score": 0.73, + "content": " { \\mathbb { B } } , { \\mathbb { G } } 6 { \\mathbb { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 233, + 505, + 245 + ], + "score": 1.0, + "content": ", also", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 243, + 257 + ], + "score": 1.0, + "content": "known as zero-shot coordination", + "type": "text" + }, + { + "bbox": [ + 243, + 244, + 260, + 256 + ], + "score": 0.49, + "content": "\\bar { \\mathbb { B } } \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 244, + 505, + 257 + ], + "score": 1.0, + "content": ". Prior work based in game-theoretic settings has suggested", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 205, + 267 + ], + "score": 1.0, + "content": "the benefits of planning", + "type": "text" + }, + { + "bbox": [ + 205, + 255, + 223, + 267 + ], + "score": 0.64, + "content": "\\pmb { \\mathbb { Z } 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 255, + 291, + 267 + ], + "score": 1.0, + "content": ", online learning", + "type": "text" + }, + { + "bbox": [ + 291, + 255, + 309, + 267 + ], + "score": 0.35, + "content": "\\mathbb { \\left| \\left. \\sum \\right. { ] } \\right| }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 255, + 428, + 267 + ], + "score": 1.0, + "content": ", and novel solution concepts", + "type": "text" + }, + { + "bbox": [ + 429, + 255, + 441, + 266 + ], + "score": 0.78, + "content": "\\bar { \\mathbb { D } }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 255, + 505, + 267 + ], + "score": 1.0, + "content": ", to name a few", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "to more complex gridworld or continuous control settings, leading to work on hierarchical social", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 144, + 302 + ], + "score": 1.0, + "content": "planning", + "type": "text" + }, + { + "bbox": [ + 144, + 288, + 162, + 299 + ], + "score": 0.76, + "content": "\\widehat { \\left\\| 3 6 \\right\\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 286, + 324, + 302 + ], + "score": 1.0, + "content": ", adapting to existing social conventions", + "type": "text" + }, + { + "bbox": [ + 324, + 288, + 356, + 299 + ], + "score": 0.34, + "content": "\\checkmark", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 286, + 438, + 302 + ], + "score": 1.0, + "content": ", trajectory diversity", + "type": "text" + }, + { + "bbox": [ + 439, + 288, + 456, + 299 + ], + "score": 0.6, + "content": "\\lVert \\rVert \\dot { \\boldsymbol { \\mathrm { E } } } \\rVert", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 286, + 505, + 302 + ], + "score": 1.0, + "content": ", and theory", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 321, + 345, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 345, + 333 + ], + "score": 1.0, + "content": "to held-out agent partners as a proxy for human co-players.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 505, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 350 + ], + "score": 1.0, + "content": "Collaborative play with novel humans has been evaluated more actively in the context of training", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 347, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 347, + 507, + 361 + ], + "score": 1.0, + "content": "agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the state-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "of-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 370, + 426, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 426, + 381 + ], + "score": 1.0, + "content": "gridworld environment, as measured by the challenge tasks in Carroll et al. [12].", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 386, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "Diversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "that train with behaviorally diverse populations of game partners tend to demonstrate stronger", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "performance than their self-play counterparts. For example, across a range of multi-agent games,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "generalization to held-out populations can be improved by training larger and more diverse populations", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "[13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 472, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 422, + 464 + ], + "score": 1.0, + "content": "optimized through selective matchmaking between increasingly diverse agents", + "type": "text" + }, + { + "bbox": [ + 423, + 451, + 469, + 463 + ], + "score": 0.6, + "content": "[ \\overbrace { 2 4 } , \\overbrace { 3 9 } , \\overbrace { 6 9 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 451, + 472, + 464 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "Despite the increased focus on improving multi-agent performance, evaluation has typically been", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "constrained to agent-agent settings. High-performing agents have infrequently been evaluated with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 308, + 504 + ], + "score": 1.0, + "content": "humans, particularly in non-competitive domains", + "type": "text" + }, + { + "bbox": [ + 308, + 489, + 326, + 501 + ], + "score": 0.66, + "content": "\\bar { \\lVert 1 6 rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 489, + 505, + 504 + ], + "score": 1.0, + "content": ". We add to this growing literature, showing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 501, + 471, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 471, + 513 + ], + "score": 1.0, + "content": "that training with diversity is a powerful approach for effective human-agent collaboration.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "score": 1.0, + "content": "Human-agent interaction In recent years, increased attention has been directed toward designing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 465, + 541 + ], + "score": 1.0, + "content": "machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also", + "type": "text" + }, + { + "bbox": [ + 465, + 528, + 482, + 540 + ], + "score": 0.76, + "content": "\\boxed { \\boxed { 1 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 304, + 552 + ], + "score": 1.0, + "content": "broader review on Cooperative AI). Tylkin et al.", + "type": "text" + }, + { + "bbox": [ + 305, + 539, + 322, + 550 + ], + "score": 0.64, + "content": "\\lVert \\rVert", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "is particularly notable in also demonstrating", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "that partially trained agents can be useful learning targets for human helpers, although in a different", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 572, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 583 + ], + "score": 1.0, + "content": "multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 583, + 331, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 331, + 595 + ], + "score": 1.0, + "content": "our agents’ performance (Tables 1 and 2 and Figure 7b).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 343, + 612 + ], + "score": 1.0, + "content": "A key preceding entry in this research area is Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 343, + 599, + 361, + 610 + ], + "score": 0.58, + "content": "\\pmb { \\mathbb { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 599, + 506, + 612 + ], + "score": 1.0, + "content": ", who similarly investigated human-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 609, + 507, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 507, + 623 + ], + "score": 1.0, + "content": "agent coordination in Overcooked. We use their method (BCP) as a baseline throughout our exper-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 172, + 633 + ], + "score": 1.0, + "content": "iments (Section", + "type": "text" + }, + { + "bbox": [ + 172, + 620, + 191, + 633 + ], + "score": 0.44, + "content": "\\boxed { 2 . 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 621, + 506, + 633 + ], + "score": 1.0, + "content": ". Relative to BCP, our approach removes the need for the expensive step of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "human data collection for agent training. Furthermore, through our novel human-agent experimental", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 506, + 655 + ], + "score": 1.0, + "content": "design, we go beyond objective performance metrics to compare the subjective preferences that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 654, + 441, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 441, + 666 + ], + "score": 1.0, + "content": "agents generate. For a detailed comparison of methods and results, see Appendix E.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5 + }, + { + "type": "title", + "bbox": [ + 107, + 684, + 305, + 697 + ], + "lines": [ + { + "bbox": [ + 104, + 682, + 306, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 306, + 700 + ], + "score": 1.0, + "content": "4 Zero-shot coordination with agents", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 104, + 710, + 487, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 489, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 489, + 723 + ], + "score": 1.0, + "content": "In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 106 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 72, + 505, + 107 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 188 + ], + "lines": [ + { + "bbox": [ + 105, + 110, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 257, + 124 + ], + "score": 1.0, + "content": "To train agents via behavioral cloning", + "type": "text" + }, + { + "bbox": [ + 257, + 110, + 274, + 122 + ], + "score": 0.79, + "content": "\\left[ \\left[ 5 8 \\right] \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 110, + 398, + 124 + ], + "score": 1.0, + "content": ", we use the open-source Acme", + "type": "text" + }, + { + "bbox": [ + 399, + 110, + 416, + 122 + ], + "score": 0.77, + "content": "\\pmb { \\mathbb { B } } 0 \\|", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 110, + 505, + 124 + ], + "score": 1.0, + "content": "to learn a policy from", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 505, + 134 + ], + "score": 1.0, + "content": "human gameplay data. Specifically, we collected 5 human-human trajectories of length 1200 time", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "steps for each of the 5 layouts, resulting in 60k total environment steps. We divide this data in half", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "and train two BC agents: (1) a partner for training a BCP agent, and (2) a “human proxy” partner for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 154, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 298, + 167 + ], + "score": 1.0, + "content": "agent-agent evaluation. Following Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 299, + 154, + 316, + 166 + ], + "score": 0.61, + "content": "\\bar { \\lVert 1 2 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 154, + 505, + 167 + ], + "score": 1.0, + "content": ", we use a set of feature-based observations for", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "the agents (as opposed to RGB) and generate comparable results: performance is higher on 3 layouts", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 177, + 452, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 452, + 189 + ], + "score": 1.0, + "content": "(asymmetric, cramped, and ring) but poorer on the other 2 (circuit and forced).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 110, + 506, + 189 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 206, + 194, + 219 + ], + "lines": [ + { + "bbox": [ + 105, + 205, + 196, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 196, + 221 + ], + "score": 1.0, + "content": "3 Related work", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 233, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 456, + 245 + ], + "score": 1.0, + "content": "Ad-hoc team play There is a large and diverse body of literature on ad-hoc team-play", + "type": "text" + }, + { + "bbox": [ + 456, + 232, + 483, + 245 + ], + "score": 0.73, + "content": " { \\mathbb { B } } , { \\mathbb { G } } 6 { \\mathbb { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 233, + 505, + 245 + ], + "score": 1.0, + "content": ", also", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 243, + 257 + ], + "score": 1.0, + "content": "known as zero-shot coordination", + "type": "text" + }, + { + "bbox": [ + 243, + 244, + 260, + 256 + ], + "score": 0.49, + "content": "\\bar { \\mathbb { B } } \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 244, + 505, + 257 + ], + "score": 1.0, + "content": ". Prior work based in game-theoretic settings has suggested", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 255, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 205, + 267 + ], + "score": 1.0, + "content": "the benefits of planning", + "type": "text" + }, + { + "bbox": [ + 205, + 255, + 223, + 267 + ], + "score": 0.64, + "content": "\\pmb { \\mathbb { Z } 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 255, + 291, + 267 + ], + "score": 1.0, + "content": ", online learning", + "type": "text" + }, + { + "bbox": [ + 291, + 255, + 309, + 267 + ], + "score": 0.35, + "content": "\\mathbb { \\left| \\left. \\sum \\right. { ] } \\right| }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 255, + 428, + 267 + ], + "score": 1.0, + "content": ", and novel solution concepts", + "type": "text" + }, + { + "bbox": [ + 429, + 255, + 441, + 266 + ], + "score": 0.78, + "content": "\\bar { \\mathbb { D } }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 255, + 505, + 267 + ], + "score": 1.0, + "content": ", to name a few", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "examples. More recently, multi-agent deep reinforcement learning has provided the tools to scale", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 506, + 290 + ], + "score": 1.0, + "content": "to more complex gridworld or continuous control settings, leading to work on hierarchical social", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 144, + 302 + ], + "score": 1.0, + "content": "planning", + "type": "text" + }, + { + "bbox": [ + 144, + 288, + 162, + 299 + ], + "score": 0.76, + "content": "\\widehat { \\left\\| 3 6 \\right\\| }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 286, + 324, + 302 + ], + "score": 1.0, + "content": ", adapting to existing social conventions", + "type": "text" + }, + { + "bbox": [ + 324, + 288, + 356, + 299 + ], + "score": 0.34, + "content": "\\checkmark", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 286, + 438, + 302 + ], + "score": 1.0, + "content": ", trajectory diversity", + "type": "text" + }, + { + "bbox": [ + 439, + 288, + 456, + 299 + ], + "score": 0.6, + "content": "\\lVert \\rVert \\dot { \\boldsymbol { \\mathrm { E } } } \\rVert", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 286, + 505, + 302 + ], + "score": 1.0, + "content": ", and theory", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "of mind [14]. Ad-hoc team-play among novel agent partners is also an object of active study in the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "emergent communication literature [10, 11, 43]. This prior work has tended to focus on generalization", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 321, + 345, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 345, + 333 + ], + "score": 1.0, + "content": "to held-out agent partners as a proxy for human co-players.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 232, + 506, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 505, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 350 + ], + "score": 1.0, + "content": "Collaborative play with novel humans has been evaluated more actively in the context of training", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 347, + 507, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 347, + 507, + 361 + ], + "score": 1.0, + "content": "agent assistants; see for instance [57, 68]. To our knowledge, our FCP agents represent the state-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "of-the-art in coordinating with novel human partners on an equal footing of capabilities in a rich", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 370, + 426, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 426, + 381 + ], + "score": 1.0, + "content": "gridworld environment, as measured by the challenge tasks in Carroll et al. [12].", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 335, + 507, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 386, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "Diversity in multi-agent reinforcement learning In multi-agent reinforcement learning, agents", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "that train with behaviorally diverse populations of game partners tend to demonstrate stronger", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 407, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 506, + 421 + ], + "score": 1.0, + "content": "performance than their self-play counterparts. For example, across a range of multi-agent games,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "generalization to held-out populations can be improved by training larger and more diverse populations", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "[13, 42, 50]. In mixed-motive settings, cooperation among agents can be encouraged through social", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 441, + 505, + 453 + ], + "score": 1.0, + "content": "diversity, such as in player preferences and rewards [3, 47, 49]. Similarly, competitiveness can be", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 451, + 472, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 422, + 464 + ], + "score": 1.0, + "content": "optimized through selective matchmaking between increasingly diverse agents", + "type": "text" + }, + { + "bbox": [ + 423, + 451, + 469, + 463 + ], + "score": 0.6, + "content": "[ \\overbrace { 2 4 } , \\overbrace { 3 9 } , \\overbrace { 6 9 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 451, + 472, + 464 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 384, + 506, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 468, + 505, + 512 + ], + "lines": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 505, + 480 + ], + "score": 1.0, + "content": "Despite the increased focus on improving multi-agent performance, evaluation has typically been", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "constrained to agent-agent settings. High-performing agents have infrequently been evaluated with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 308, + 504 + ], + "score": 1.0, + "content": "humans, particularly in non-competitive domains", + "type": "text" + }, + { + "bbox": [ + 308, + 489, + 326, + 501 + ], + "score": 0.66, + "content": "\\bar { \\lVert 1 6 rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 489, + 505, + 504 + ], + "score": 1.0, + "content": ". We add to this growing literature, showing", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 501, + 471, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 471, + 513 + ], + "score": 1.0, + "content": "that training with diversity is a powerful approach for effective human-agent collaboration.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 468, + 505, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 531 + ], + "score": 1.0, + "content": "Human-agent interaction In recent years, increased attention has been directed toward designing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 465, + 541 + ], + "score": 1.0, + "content": "machine learning agents capable of collaborating with humans [41, 57, 68, 72] (see also", + "type": "text" + }, + { + "bbox": [ + 465, + 528, + 482, + 540 + ], + "score": 0.76, + "content": "\\boxed { \\boxed { 1 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "for a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 304, + 552 + ], + "score": 1.0, + "content": "broader review on Cooperative AI). Tylkin et al.", + "type": "text" + }, + { + "bbox": [ + 305, + 539, + 322, + 550 + ], + "score": 0.64, + "content": "\\lVert \\rVert", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 537, + 505, + 552 + ], + "score": 1.0, + "content": "is particularly notable in also demonstrating", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "that partially trained agents can be useful learning targets for human helpers, although in a different", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "domain (cooperative Atari). Our method, FCP, can be seen as extending theirs by training with", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 572, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 583 + ], + "score": 1.0, + "content": "multiple “skill levels” and random seeds, rather than just one, which we demonstrate to be crucial to", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 583, + 331, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 331, + 595 + ], + "score": 1.0, + "content": "our agents’ performance (Tables 1 and 2 and Figure 7b).", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 515, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 599, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 343, + 612 + ], + "score": 1.0, + "content": "A key preceding entry in this research area is Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 343, + 599, + 361, + 610 + ], + "score": 0.58, + "content": "\\pmb { \\mathbb { I } }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 599, + 506, + 612 + ], + "score": 1.0, + "content": ", who similarly investigated human-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 609, + 507, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 507, + 623 + ], + "score": 1.0, + "content": "agent coordination in Overcooked. We use their method (BCP) as a baseline throughout our exper-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 172, + 633 + ], + "score": 1.0, + "content": "iments (Section", + "type": "text" + }, + { + "bbox": [ + 172, + 620, + 191, + 633 + ], + "score": 0.44, + "content": "\\boxed { 2 . 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 191, + 621, + 506, + 633 + ], + "score": 1.0, + "content": ". Relative to BCP, our approach removes the need for the expensive step of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 506, + 644 + ], + "score": 1.0, + "content": "human data collection for agent training. Furthermore, through our novel human-agent experimental", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 642, + 506, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 642, + 506, + 655 + ], + "score": 1.0, + "content": "design, we go beyond objective performance metrics to compare the subjective preferences that", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 654, + 441, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 441, + 666 + ], + "score": 1.0, + "content": "agents generate. For a detailed comparison of methods and results, see Appendix E.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 599, + 507, + 666 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 684, + 305, + 697 + ], + "lines": [ + { + "bbox": [ + 104, + 682, + 306, + 700 + ], + "spans": [ + { + "bbox": [ + 104, + 682, + 306, + 700 + ], + "score": 1.0, + "content": "4 Zero-shot coordination with agents", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "text", + "bbox": [ + 104, + 710, + 487, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 489, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 489, + 723 + ], + "score": 1.0, + "content": "In this section, we evaluate our FCP agent, its ablations, and the baselines with held-out agents.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49, + "bbox_fs": [ + 105, + 709, + 489, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 406, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 407, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 407, + 87 + ], + "score": 1.0, + "content": "4.1 Evaluation method: collaborative evaluation with agent partners", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 92, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "Our primary concern in this work is generalization to novel human partners (as investigated in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 138, + 116 + ], + "score": 1.0, + "content": "Section", + "type": "text" + }, + { + "bbox": [ + 138, + 103, + 148, + 116 + ], + "score": 0.35, + "content": "\\textcircled{5}", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 103, + 506, + 116 + ], + "score": 1.0, + "content": ". However, just as collecting human-human data for behavioral cloning is expensive, so", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "score": 1.0, + "content": "partners as a cheap proxy of performance with humans. This is then used to guide our model selection", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "process, allowing us to be more targeted with the agents we select for our human-agent evaluations.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 108, + 153, + 286, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 286, + 166 + ], + "score": 1.0, + "content": "We evaluate with three held-out populations:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 129, + 172, + 506, + 246 + ], + "lines": [ + { + "bbox": [ + 129, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 129, + 171, + 283, + 185 + ], + "score": 1.0, + "content": "1. A BC model trained on human data,", + "type": "text" + }, + { + "bbox": [ + 284, + 173, + 312, + 185 + ], + "score": 0.91, + "content": "H _ { \\mathrm { p r o x y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 171, + 506, + 185 + ], + "score": 1.0, + "content": ", intended as a proxy of generalization to humans,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 182, + 261, + 196 + ], + "spans": [ + { + "bbox": [ + 141, + 182, + 240, + 196 + ], + "score": 1.0, + "content": "as done by Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 240, + 183, + 258, + 195 + ], + "score": 0.29, + "content": "[ \\overbrace { | 1 2 | }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 182, + 261, + 196 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 196, + 507, + 211 + ], + "spans": [ + { + "bbox": [ + 128, + 196, + 507, + 211 + ], + "score": 1.0, + "content": "2. A set of self-play agents varying in seed, architecture, and training time (specifically, held-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 207, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 208, + 207, + 222 + ], + "score": 1.0, + "content": "out seeds of the", + "type": "text" + }, + { + "bbox": [ + 207, + 209, + 241, + 219 + ], + "score": 0.89, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 208, + 335, + 222 + ], + "score": 1.0, + "content": "partners trained for the", + "type": "text" + }, + { + "bbox": [ + 335, + 209, + 367, + 220 + ], + "score": 0.9, + "content": "\\mathrm { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 208, + 460, + 222 + ], + "score": 1.0, + "content": "agent; see Section 2.4).", + "type": "text" + }, + { + "bbox": [ + 462, + 207, + 505, + 221 + ], + "score": 1.0, + "content": "These are", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 219, + 423, + 232 + ], + "spans": [ + { + "bbox": [ + 142, + 219, + 423, + 232 + ], + "score": 1.0, + "content": "intended to test generalization to a diverse yet still skillful population.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 234, + 464, + 246 + ], + "spans": [ + { + "bbox": [ + 129, + 234, + 464, + 246 + ], + "score": 1.0, + "content": "3. Randomly initialized agents intended to test generalization to low-skill partners.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 254, + 506, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "For all results, we report the average number of deliveries made by both players within an episode,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 330, + 277 + ], + "score": 1.0, + "content": "aggregated across the 5 different layouts from Figure", + "type": "text" + }, + { + "bbox": [ + 331, + 264, + 341, + 277 + ], + "score": 0.74, + "content": "^ 4", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "(with the per-layout results reported in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 146, + 288 + ], + "score": 1.0, + "content": "Appendix", + "type": "text" + }, + { + "bbox": [ + 146, + 275, + 166, + 288 + ], + "score": 0.51, + "content": "\\underline { { \\overline { { ( \\mathrm { C . 2 } ) } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 276, + 167, + 288 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 168, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "We estimate mean and standard deviation across 5 random seeds. For each seed, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "score": 1.0, + "content": "evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 310, + 162, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 163, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 163, + 323 + ], + "score": 1.0, + "content": "4.2 Results", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 330, + 336, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 338, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 338, + 344 + ], + "score": 1.0, + "content": "Finding 1: FCP significantly outperforms all baselines", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "To begin, we compare our FCP agent and the baselines when partnered with the three held-out", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 323, + 367 + ], + "score": 1.0, + "content": "populations introduced above. As can be seen in Figure", + "type": "text" + }, + { + "bbox": [ + 324, + 353, + 334, + 367 + ], + "score": 0.61, + "content": "\\boxed { 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "FCP significantly outperforms all baselines", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 103, + 361, + 502, + 381 + ], + "spans": [ + { + "bbox": [ + 103, + 361, + 473, + 381 + ], + "score": 1.0, + "content": "when partnered with all three held-out populations. Notably, it performs better than BCP with", + "type": "text" + }, + { + "bbox": [ + 474, + 365, + 502, + 377 + ], + "score": 0.88, + "content": "H _ { \\mathrm { p r o x y } }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 451, + 389 + ], + "score": 1.0, + "content": "even though BCP trains with such a model and FCP does not. Similar to Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 451, + 375, + 469, + 387 + ], + "score": 0.28, + "content": "\\mathbb { \\lVert 1 2 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 375, + 506, + 389 + ], + "score": 1.0, + "content": ", we find", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "score": 1.0, + "content": "that BCP significantly outscores SP.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 506, + 415 + ], + "score": 1.0, + "content": "When paired with a randomly initialized partner which behaves suboptimally, we see an even greater", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "score": 1.0, + "content": "such agents, it may not be surprising that it does so well with partners that behave poorly. However,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "what is surprising is how brittle the other training methods are. This suggests that they may not", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 480, + 460 + ], + "score": 1.0, + "content": "perform well with humans who are not highly skilled players, which we will see in Section 5.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 101, + 470, + 495, + 559 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 101, + 470, + 495, + 559 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 101, + 470, + 495, + 559 + ], + "spans": [ + { + "bbox": [ + 101, + 470, + 495, + 559 + ], + "score": 0.969, + "type": "image", + "image_path": "8083dcac415b6ef9624e1aa4f428a0a33c3e9d47ce8903df3798f3e80ef0427c.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 101, + 470, + 495, + 499.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 101, + 499.6666666666667, + 495, + 529.3333333333334 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 101, + 529.3333333333334, + 495, + 559.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 565, + 506, + 622 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Figure 5: Agent-agent collaborative evaluation: Performance of each agent when partnered with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 277, + 588 + ], + "score": 1.0, + "content": "each of the held-out populations (Section", + "type": "text" + }, + { + "bbox": [ + 278, + 575, + 296, + 588 + ], + "score": 0.62, + "content": "4 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 576, + 387, + 588 + ], + "score": 1.0, + "content": "in episodes of length", + "type": "text" + }, + { + "bbox": [ + 387, + 576, + 426, + 586 + ], + "score": 0.9, + "content": "T = 5 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 576, + 505, + 588 + ], + "score": 1.0, + "content": ". Importantly, FCP", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "scores higher than all baselines with a variety of test partners. Error bars represent standard deviation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "over five random training seeds. Plots aggregate data across kitchen layouts; results calculated by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 607, + 302, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 302, + 623 + ], + "score": 1.0, + "content": "individual layout can be found in Appendix C.2.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + } + ], + "index": 32.0 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 491, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 491, + 652 + ], + "score": 1.0, + "content": "Finding 2: Training with past checkpoints is the most beneficial variation for performance", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "Next, we investigate how the different training partner variations influence FCP’s performance. In", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 660, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 317, + 676 + ], + "score": 1.0, + "content": "particular, we separately ablate the past checkpoints", + "type": "text" + }, + { + "bbox": [ + 318, + 662, + 332, + 673 + ], + "score": 0.72, + "content": "( T )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 660, + 401, + 676 + ], + "score": 1.0, + "content": "and architecture", + "type": "text" + }, + { + "bbox": [ + 401, + 662, + 416, + 673 + ], + "score": 0.68, + "content": "( A )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 660, + 506, + 676 + ], + "score": 1.0, + "content": "variations, evaluating", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 671, + 498, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 498, + 687 + ], + "score": 1.0, + "content": "them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 210, + 703 + ], + "score": 1.0, + "content": "Comparing the FCP and", + "type": "text" + }, + { + "bbox": [ + 210, + 689, + 242, + 700 + ], + "score": 0.83, + "content": "\\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "columns, we see that removing past checkpoints from training", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 699, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 353, + 714 + ], + "score": 1.0, + "content": "significantly reduces performance. Comparing the FCP and", + "type": "text" + }, + { + "bbox": [ + 353, + 700, + 385, + 712 + ], + "score": 0.9, + "content": "\\operatorname { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 699, + 505, + 714 + ], + "score": 1.0, + "content": "columns, we see that adding", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "architectural variation to the training population offers no improvement over training with past", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 72, + 406, + 84 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 407, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 407, + 87 + ], + "score": 1.0, + "content": "4.1 Evaluation method: collaborative evaluation with agent partners", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 92, + 505, + 148 + ], + "lines": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "Our primary concern in this work is generalization to novel human partners (as investigated in", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 103, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 138, + 116 + ], + "score": 1.0, + "content": "Section", + "type": "text" + }, + { + "bbox": [ + 138, + 103, + 148, + 116 + ], + "score": 0.35, + "content": "\\textcircled{5}", + "type": "inline_equation" + }, + { + "bbox": [ + 148, + 103, + 506, + 116 + ], + "score": 1.0, + "content": ". However, just as collecting human-human data for behavioral cloning is expensive, so", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 114, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 506, + 128 + ], + "score": 1.0, + "content": "too is evaluating agents with humans. Consequently, we instead use generalization to held-out agent", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 138 + ], + "score": 1.0, + "content": "partners as a cheap proxy of performance with humans. This is then used to guide our model selection", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "process, allowing us to be more targeted with the agents we select for our human-agent evaluations.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 93, + 506, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 153, + 286, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 286, + 166 + ], + "score": 1.0, + "content": "We evaluate with three held-out populations:", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6, + "bbox_fs": [ + 106, + 151, + 286, + 166 + ] + }, + { + "type": "text", + "bbox": [ + 129, + 172, + 506, + 246 + ], + "lines": [ + { + "bbox": [ + 129, + 171, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 129, + 171, + 283, + 185 + ], + "score": 1.0, + "content": "1. A BC model trained on human data,", + "type": "text" + }, + { + "bbox": [ + 284, + 173, + 312, + 185 + ], + "score": 0.91, + "content": "H _ { \\mathrm { p r o x y } }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 171, + 506, + 185 + ], + "score": 1.0, + "content": ", intended as a proxy of generalization to humans,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 182, + 261, + 196 + ], + "spans": [ + { + "bbox": [ + 141, + 182, + 240, + 196 + ], + "score": 1.0, + "content": "as done by Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 240, + 183, + 258, + 195 + ], + "score": 0.29, + "content": "[ \\overbrace { | 1 2 | }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 182, + 261, + 196 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 128, + 196, + 507, + 211 + ], + "spans": [ + { + "bbox": [ + 128, + 196, + 507, + 211 + ], + "score": 1.0, + "content": "2. A set of self-play agents varying in seed, architecture, and training time (specifically, held-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 207, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 208, + 207, + 222 + ], + "score": 1.0, + "content": "out seeds of the", + "type": "text" + }, + { + "bbox": [ + 207, + 209, + 241, + 219 + ], + "score": 0.89, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 208, + 335, + 222 + ], + "score": 1.0, + "content": "partners trained for the", + "type": "text" + }, + { + "bbox": [ + 335, + 209, + 367, + 220 + ], + "score": 0.9, + "content": "\\mathrm { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 208, + 460, + 222 + ], + "score": 1.0, + "content": "agent; see Section 2.4).", + "type": "text" + }, + { + "bbox": [ + 462, + 207, + 505, + 221 + ], + "score": 1.0, + "content": "These are", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 219, + 423, + 232 + ], + "spans": [ + { + "bbox": [ + 142, + 219, + 423, + 232 + ], + "score": 1.0, + "content": "intended to test generalization to a diverse yet still skillful population.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 129, + 234, + 464, + 246 + ], + "spans": [ + { + "bbox": [ + 129, + 234, + 464, + 246 + ], + "score": 1.0, + "content": "3. Randomly initialized agents intended to test generalization to low-skill partners.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 128, + 171, + 507, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 254, + 506, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 506, + 266 + ], + "score": 1.0, + "content": "For all results, we report the average number of deliveries made by both players within an episode,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 330, + 277 + ], + "score": 1.0, + "content": "aggregated across the 5 different layouts from Figure", + "type": "text" + }, + { + "bbox": [ + 331, + 264, + 341, + 277 + ], + "score": 0.74, + "content": "^ 4", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "(with the per-layout results reported in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 146, + 288 + ], + "score": 1.0, + "content": "Appendix", + "type": "text" + }, + { + "bbox": [ + 146, + 275, + 166, + 288 + ], + "score": 0.51, + "content": "\\underline { { \\overline { { ( \\mathrm { C . 2 } ) } } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 276, + 167, + 288 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 168, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "We estimate mean and standard deviation across 5 random seeds. For each seed, we", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 507, + 299 + ], + "score": 1.0, + "content": "evaluate the agent with all members of the held-out population for 10 episodes per agent-partner pair.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 253, + 507, + 299 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 310, + 162, + 322 + ], + "lines": [ + { + "bbox": [ + 105, + 309, + 163, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 163, + 323 + ], + "score": 1.0, + "content": "4.2 Results", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 330, + 336, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 330, + 338, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 338, + 344 + ], + "score": 1.0, + "content": "Finding 1: FCP significantly outperforms all baselines", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "To begin, we compare our FCP agent and the baselines when partnered with the three held-out", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 323, + 367 + ], + "score": 1.0, + "content": "populations introduced above. As can be seen in Figure", + "type": "text" + }, + { + "bbox": [ + 324, + 353, + 334, + 367 + ], + "score": 0.61, + "content": "\\boxed { 5 }", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "FCP significantly outperforms all baselines", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 103, + 361, + 502, + 381 + ], + "spans": [ + { + "bbox": [ + 103, + 361, + 473, + 381 + ], + "score": 1.0, + "content": "when partnered with all three held-out populations. Notably, it performs better than BCP with", + "type": "text" + }, + { + "bbox": [ + 474, + 365, + 502, + 377 + ], + "score": 0.88, + "content": "H _ { \\mathrm { p r o x y } }", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 451, + 389 + ], + "score": 1.0, + "content": "even though BCP trains with such a model and FCP does not. Similar to Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 451, + 375, + 469, + 387 + ], + "score": 0.28, + "content": "\\mathbb { \\lVert 1 2 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 375, + 506, + 389 + ], + "score": 1.0, + "content": ", we find", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 252, + 399 + ], + "score": 1.0, + "content": "that BCP significantly outscores SP.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 103, + 343, + 506, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 506, + 415 + ], + "score": 1.0, + "content": "When paired with a randomly initialized partner which behaves suboptimally, we see an even greater", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "difference between FCP and the baselines. Given that FCP is trained with non-held-out versions of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 439 + ], + "score": 1.0, + "content": "such agents, it may not be surprising that it does so well with partners that behave poorly. However,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 436, + 506, + 448 + ], + "score": 1.0, + "content": "what is surprising is how brittle the other training methods are. This suggests that they may not", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 446, + 480, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 480, + 460 + ], + "score": 1.0, + "content": "perform well with humans who are not highly skilled players, which we will see in Section 5.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 403, + 506, + 460 + ] + }, + { + "type": "image", + "bbox": [ + 101, + 470, + 495, + 559 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 101, + 470, + 495, + 559 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 101, + 470, + 495, + 559 + ], + "spans": [ + { + "bbox": [ + 101, + 470, + 495, + 559 + ], + "score": 0.969, + "type": "image", + "image_path": "8083dcac415b6ef9624e1aa4f428a0a33c3e9d47ce8903df3798f3e80ef0427c.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 101, + 470, + 495, + 499.6666666666667 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 101, + 499.6666666666667, + 495, + 529.3333333333334 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 101, + 529.3333333333334, + 495, + 559.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 565, + 506, + 622 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 505, + 577 + ], + "score": 1.0, + "content": "Figure 5: Agent-agent collaborative evaluation: Performance of each agent when partnered with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 575, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 277, + 588 + ], + "score": 1.0, + "content": "each of the held-out populations (Section", + "type": "text" + }, + { + "bbox": [ + 278, + 575, + 296, + 588 + ], + "score": 0.62, + "content": "4 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 576, + 387, + 588 + ], + "score": 1.0, + "content": "in episodes of length", + "type": "text" + }, + { + "bbox": [ + 387, + 576, + 426, + 586 + ], + "score": 0.9, + "content": "T = 5 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 576, + 505, + 588 + ], + "score": 1.0, + "content": ". Importantly, FCP", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 599 + ], + "score": 1.0, + "content": "scores higher than all baselines with a variety of test partners. Error bars represent standard deviation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "over five random training seeds. Plots aggregate data across kitchen layouts; results calculated by", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 607, + 302, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 302, + 623 + ], + "score": 1.0, + "content": "individual layout can be found in Appendix C.2.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34 + } + ], + "index": 32.0 + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 636, + 491, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 491, + 652 + ], + "score": 1.0, + "content": "Finding 2: Training with past checkpoints is the most beneficial variation for performance", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "Next, we investigate how the different training partner variations influence FCP’s performance. In", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 660, + 506, + 676 + ], + "spans": [ + { + "bbox": [ + 104, + 660, + 317, + 676 + ], + "score": 1.0, + "content": "particular, we separately ablate the past checkpoints", + "type": "text" + }, + { + "bbox": [ + 318, + 662, + 332, + 673 + ], + "score": 0.72, + "content": "( T )", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 660, + 401, + 676 + ], + "score": 1.0, + "content": "and architecture", + "type": "text" + }, + { + "bbox": [ + 401, + 662, + 416, + 673 + ], + "score": 0.68, + "content": "( A )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 660, + 506, + 676 + ], + "score": 1.0, + "content": "variations, evaluating", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 671, + 498, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 498, + 687 + ], + "score": 1.0, + "content": "them with the same partners as in Figure 5. The results of this evaluation are presented in Table 1.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 687, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 210, + 703 + ], + "score": 1.0, + "content": "Comparing the FCP and", + "type": "text" + }, + { + "bbox": [ + 210, + 689, + 242, + 700 + ], + "score": 0.83, + "content": "\\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 687, + 506, + 703 + ], + "score": 1.0, + "content": "columns, we see that removing past checkpoints from training", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 699, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 353, + 714 + ], + "score": 1.0, + "content": "significantly reduces performance. Comparing the FCP and", + "type": "text" + }, + { + "bbox": [ + 353, + 700, + 385, + 712 + ], + "score": 0.9, + "content": "\\operatorname { F C P } _ { + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 699, + 505, + 714 + ], + "score": 1.0, + "content": "columns, we see that adding", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "architectural variation to the training population offers no improvement over training with past", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40, + "bbox_fs": [ + 104, + 636, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 70, + 441, + 117 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 167, + 70, + 441, + 117 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 70, + 441, + 117 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 441, + 117 + ], + "score": 0.974, + "html": "
PartnerFCPFCP-TFCP+AFCP-T,+A
Hproxy10.6± 0.54.7± 0.49.9±0.67.0±0.8
Diverse SP11.2 ± 0.16.9 ± 0.111.1 ± 0.48.6 ± 0.4
Random8.6± 0.21.0 ± 0.18.4±0.43.2 ± 0.5
", + "type": "table", + "image_path": "56b2734b6d3a9e5266ca9b0dd8d3ba9f6d27df29190a11ac23347bb9e9fb3c37.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 167, + 70, + 441, + 85.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 167, + 85.66666666666667, + 441, + 101.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 167, + 101.33333333333334, + 441, + 117.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 505, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Table 1: Ablation results: Performance of each variation of FCP – training with past partner", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 107, + 135, + 159, + 146 + ], + "score": 1.0, + "content": "checkpoints", + "type": "text" + }, + { + "bbox": [ + 159, + 135, + 168, + 145 + ], + "score": 0.63, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 135, + 379, + 146 + ], + "score": 1.0, + "content": "for time) and adding partner variation in architecture", + "type": "text" + }, + { + "bbox": [ + 379, + 135, + 393, + 146 + ], + "score": 0.49, + "content": "( A )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 135, + 505, + 146 + ], + "score": 1.0, + "content": ". Scores are mean deliveries", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "with standard deviation over 5 random seeds. Notably, we find that the inclusion of past checkpoints", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 244, + 169 + ], + "score": 1.0, + "content": "is essential for strong performance", + "type": "text" + }, + { + "bbox": [ + 245, + 156, + 306, + 168 + ], + "score": 0.89, + "content": "( \\mathrm { F C P } > \\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "), and additionally including architectural variation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 234, + 180 + ], + "score": 1.0, + "content": "does not improve performance", + "type": "text" + }, + { + "bbox": [ + 234, + 168, + 299, + 179 + ], + "score": 0.9, + "content": "( \\mathrm { F C P } \\approx \\mathrm { F C P } _ { + A . }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "). However, architectural variation is better than no", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 178, + 501, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 405, + 191 + ], + "score": 1.0, + "content": "variation, improving performance when past checkpoints are not available", + "type": "text" + }, + { + "bbox": [ + 405, + 178, + 495, + 190 + ], + "score": 0.91, + "content": "( \\mathrm { F C P } _ { - T , + A } > \\mathrm { F C P } _ { - T , }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 178, + 501, + 191 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 207, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 258, + 222 + ], + "score": 1.0, + "content": "checkpoints. However, comparing the", + "type": "text" + }, + { + "bbox": [ + 258, + 209, + 290, + 220 + ], + "score": 0.87, + "content": "\\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 207, + 308, + 222 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 209, + 353, + 221 + ], + "score": 0.88, + "content": "\\mathrm { F C P } _ { - T , + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 207, + 506, + 222 + ], + "score": 1.0, + "content": "columns, we see that without training", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 219, + 469, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 469, + 233 + ], + "score": 1.0, + "content": "with past checkpoints, architectural variation in the population does improve performance.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 246, + 313, + 259 + ], + "lines": [ + { + "bbox": [ + 104, + 245, + 314, + 262 + ], + "spans": [ + { + "bbox": [ + 104, + 245, + 314, + 262 + ], + "score": 1.0, + "content": "5 Zero-shot coordination with humans", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 506, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "score": 1.0, + "content": "Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 293, + 213, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 213, + 305 + ], + "score": 1.0, + "content": "play with human partners.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 91, + 315, + 521, + 401 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 91, + 315, + 521, + 401 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 91, + 315, + 521, + 401 + ], + "spans": [ + { + "bbox": [ + 91, + 315, + 521, + 401 + ], + "score": 0.814, + "type": "image", + "image_path": "5856cbdd90eeef2a7a6093668472063862a759746c339aa0324e78efb5d67b22.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 91, + 315, + 521, + 343.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 91, + 343.6666666666667, + 521, + 372.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 91, + 372.33333333333337, + 521, + 401.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 410, + 506, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "Figure 6: Human-agent collaborative study: For our human-agent collaboration study, we recruited", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 433 + ], + "score": 1.0, + "content": "participants online to play games with FCP and baseline agents. Participants played a randomized", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 432, + 507, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 444 + ], + "score": 1.0, + "content": "sequence of episodes with different agent partners and kitchen layouts. After every two episodes,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 442, + 498, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 498, + 455 + ], + "score": 1.0, + "content": "participants reported the direction and strength of their preference between their last two partners.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 17.75 + }, + { + "type": "title", + "bbox": [ + 106, + 474, + 427, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 429, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 429, + 489 + ], + "score": 1.0, + "content": "5.1 Evaluation method: collaborative evaluation with human participants", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 506, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 508 + ], + "score": 1.0, + "content": "To test how effectively FCP’s performance generalizes to human partners, we recruited participants", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 159, + 518 + ], + "score": 1.0, + "content": "from Prolific", + "type": "text" + }, + { + "bbox": [ + 160, + 505, + 192, + 517 + ], + "score": 0.31, + "content": "\\mathbb { 1 1 8 } , \\lvert 5 5 \\rvert", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 505, + 328, + 518 + ], + "score": 1.0, + "content": "for an online collaboration study", + "type": "text" + }, + { + "bbox": [ + 329, + 506, + 367, + 516 + ], + "score": 0.87, + "content": "N = 1 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 505, + 371, + 518 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 371, + 506, + 398, + 516 + ], + "score": 0.83, + "content": "3 7 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 505, + 430, + 518 + ], + "score": 1.0, + "content": "female,", + "type": "text" + }, + { + "bbox": [ + 431, + 506, + 457, + 516 + ], + "score": 0.86, + "content": "5 9 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 505, + 483, + 518 + ], + "score": 1.0, + "content": "male,", + "type": "text" + }, + { + "bbox": [ + 483, + 506, + 505, + 516 + ], + "score": 0.85, + "content": "1 . 8 \\%", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "nonbinary; median age between 25–34 years). We used a within-participant design for the study:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "each participant played with a full cohort of agents (i.e. generated through every training method).", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 537, + 492, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 492, + 553 + ], + "score": 1.0, + "content": "This design allowed us to evaluate both objective performance as well as subjective preferences.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "Participants first read game instructions and played a short tutorial episode guiding them through", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 297, + 578 + ], + "score": 1.0, + "content": "the dish preparation sequence (see Appendix", + "type": "text" + }, + { + "bbox": [ + 297, + 565, + 325, + 579 + ], + "score": 0.78, + "content": "\\dot { \\mathrm { \\bf D } } . 1 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "for instruction text and study screenshots).", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "Participants then played 20 episodes with a randomized sequence of agent partners and kitchen", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 209, + 600 + ], + "score": 1.0, + "content": "layouts. Episodes lasted", + "type": "text" + }, + { + "bbox": [ + 210, + 588, + 249, + 598 + ], + "score": 0.89, + "content": "T = 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "steps (1 minute) each. After every two episodes, participants", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "score": 1.0, + "content": "reported their preference over the agent partners from those episodes on a five-point Likert-type scale.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 609, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 623 + ], + "score": 1.0, + "content": "After playing all 20 episodes, participants completed a debrief questionnaire collecting standard", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 506, + 633 + ], + "score": 1.0, + "content": "demographic information and open-ended feedback on the study. Our statistical analysis below", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 641, + 476, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 476, + 655 + ], + "score": 1.0, + "content": "for additional details of our study design and analysis, including independent ethical review.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 107, + 667, + 162, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 163, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 163, + 680 + ], + "score": 1.0, + "content": "5.2 Results", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 687, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 477, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 477, + 700 + ], + "score": 1.0, + "content": "Finding 1: FCP coordinates best with humans, achieving the highest score across maps", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "To begin, we compare the objective team performance supported by our FCP and baseline agents.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "score": 1.0, + "content": "The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 753 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 753 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 167, + 70, + 441, + 117 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 167, + 70, + 441, + 117 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 167, + 70, + 441, + 117 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 441, + 117 + ], + "score": 0.974, + "html": "
PartnerFCPFCP-TFCP+AFCP-T,+A
Hproxy10.6± 0.54.7± 0.49.9±0.67.0±0.8
Diverse SP11.2 ± 0.16.9 ± 0.111.1 ± 0.48.6 ± 0.4
Random8.6± 0.21.0 ± 0.18.4±0.43.2 ± 0.5
", + "type": "table", + "image_path": "56b2734b6d3a9e5266ca9b0dd8d3ba9f6d27df29190a11ac23347bb9e9fb3c37.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 167, + 70, + 441, + 85.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 167, + 85.66666666666667, + 441, + 101.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 167, + 101.33333333333334, + 441, + 117.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 123, + 505, + 190 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 136 + ], + "score": 1.0, + "content": "Table 1: Ablation results: Performance of each variation of FCP – training with past partner", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 135, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 107, + 135, + 159, + 146 + ], + "score": 1.0, + "content": "checkpoints", + "type": "text" + }, + { + "bbox": [ + 159, + 135, + 168, + 145 + ], + "score": 0.63, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 135, + 379, + 146 + ], + "score": 1.0, + "content": "for time) and adding partner variation in architecture", + "type": "text" + }, + { + "bbox": [ + 379, + 135, + 393, + 146 + ], + "score": 0.49, + "content": "( A )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 135, + 505, + 146 + ], + "score": 1.0, + "content": ". Scores are mean deliveries", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "with standard deviation over 5 random seeds. Notably, we find that the inclusion of past checkpoints", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 156, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 156, + 244, + 169 + ], + "score": 1.0, + "content": "is essential for strong performance", + "type": "text" + }, + { + "bbox": [ + 245, + 156, + 306, + 168 + ], + "score": 0.89, + "content": "( \\mathrm { F C P } > \\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 156, + 505, + 169 + ], + "score": 1.0, + "content": "), and additionally including architectural variation", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 234, + 180 + ], + "score": 1.0, + "content": "does not improve performance", + "type": "text" + }, + { + "bbox": [ + 234, + 168, + 299, + 179 + ], + "score": 0.9, + "content": "( \\mathrm { F C P } \\approx \\mathrm { F C P } _ { + A . }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "). However, architectural variation is better than no", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 178, + 501, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 405, + 191 + ], + "score": 1.0, + "content": "variation, improving performance when past checkpoints are not available", + "type": "text" + }, + { + "bbox": [ + 405, + 178, + 495, + 190 + ], + "score": 0.91, + "content": "( \\mathrm { F C P } _ { - T , + A } > \\mathrm { F C P } _ { - T , }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 178, + 501, + 191 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 122, + 506, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 105, + 207, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 258, + 222 + ], + "score": 1.0, + "content": "checkpoints. However, comparing the", + "type": "text" + }, + { + "bbox": [ + 258, + 209, + 290, + 220 + ], + "score": 0.87, + "content": "\\mathrm { F C P } _ { - T }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 207, + 308, + 222 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 308, + 209, + 353, + 221 + ], + "score": 0.88, + "content": "\\mathrm { F C P } _ { - T , + A }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 207, + 506, + 222 + ], + "score": 1.0, + "content": "columns, we see that without training", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 219, + 469, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 469, + 233 + ], + "score": 1.0, + "content": "with past checkpoints, architectural variation in the population does improve performance.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 207, + 506, + 233 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 246, + 313, + 259 + ], + "lines": [ + { + "bbox": [ + 104, + 245, + 314, + 262 + ], + "spans": [ + { + "bbox": [ + 104, + 245, + 314, + 262 + ], + "score": 1.0, + "content": "5 Zero-shot coordination with humans", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 271, + 506, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 284 + ], + "score": 1.0, + "content": "Ultimately, our goal is to develop agents capable of coordinating with novel human partners. In this", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 506, + 295 + ], + "score": 1.0, + "content": "section, we run an online study to evaluate our FCP agent and the baseline agents in collaborative", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 293, + 213, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 213, + 305 + ], + "score": 1.0, + "content": "play with human partners.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 271, + 506, + 305 + ] + }, + { + "type": "image", + "bbox": [ + 91, + 315, + 521, + 401 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 91, + 315, + 521, + 401 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 91, + 315, + 521, + 401 + ], + "spans": [ + { + "bbox": [ + 91, + 315, + 521, + 401 + ], + "score": 0.814, + "type": "image", + "image_path": "5856cbdd90eeef2a7a6093668472063862a759746c339aa0324e78efb5d67b22.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 91, + 315, + 521, + 343.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 91, + 343.6666666666667, + 521, + 372.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 91, + 372.33333333333337, + 521, + 401.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 410, + 506, + 454 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 410, + 506, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 506, + 423 + ], + "score": 1.0, + "content": "Figure 6: Human-agent collaborative study: For our human-agent collaboration study, we recruited", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 420, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 505, + 433 + ], + "score": 1.0, + "content": "participants online to play games with FCP and baseline agents. Participants played a randomized", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 432, + 507, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 444 + ], + "score": 1.0, + "content": "sequence of episodes with different agent partners and kitchen layouts. After every two episodes,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 442, + 498, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 498, + 455 + ], + "score": 1.0, + "content": "participants reported the direction and strength of their preference between their last two partners.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 17.75 + }, + { + "type": "title", + "bbox": [ + 106, + 474, + 427, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 429, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 429, + 489 + ], + "score": 1.0, + "content": "5.1 Evaluation method: collaborative evaluation with human participants", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 495, + 506, + 550 + ], + "lines": [ + { + "bbox": [ + 105, + 493, + 506, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 508 + ], + "score": 1.0, + "content": "To test how effectively FCP’s performance generalizes to human partners, we recruited participants", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 159, + 518 + ], + "score": 1.0, + "content": "from Prolific", + "type": "text" + }, + { + "bbox": [ + 160, + 505, + 192, + 517 + ], + "score": 0.31, + "content": "\\mathbb { 1 1 8 } , \\lvert 5 5 \\rvert", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 505, + 328, + 518 + ], + "score": 1.0, + "content": "for an online collaboration study", + "type": "text" + }, + { + "bbox": [ + 329, + 506, + 367, + 516 + ], + "score": 0.87, + "content": "N = 1 1 4", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 505, + 371, + 518 + ], + "score": 1.0, + "content": ";", + "type": "text" + }, + { + "bbox": [ + 371, + 506, + 398, + 516 + ], + "score": 0.83, + "content": "3 7 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 505, + 430, + 518 + ], + "score": 1.0, + "content": "female,", + "type": "text" + }, + { + "bbox": [ + 431, + 506, + 457, + 516 + ], + "score": 0.86, + "content": "5 9 . 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 505, + 483, + 518 + ], + "score": 1.0, + "content": "male,", + "type": "text" + }, + { + "bbox": [ + 483, + 506, + 505, + 516 + ], + "score": 0.85, + "content": "1 . 8 \\%", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 506, + 529 + ], + "score": 1.0, + "content": "nonbinary; median age between 25–34 years). We used a within-participant design for the study:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "each participant played with a full cohort of agents (i.e. generated through every training method).", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 537, + 492, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 492, + 553 + ], + "score": 1.0, + "content": "This design allowed us to evaluate both objective performance as well as subjective preferences.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 493, + 506, + 553 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 505, + 654 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "Participants first read game instructions and played a short tutorial episode guiding them through", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 565, + 506, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 565, + 297, + 578 + ], + "score": 1.0, + "content": "the dish preparation sequence (see Appendix", + "type": "text" + }, + { + "bbox": [ + 297, + 565, + 325, + 579 + ], + "score": 0.78, + "content": "\\dot { \\mathrm { \\bf D } } . 1 . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 565, + 506, + 578 + ], + "score": 1.0, + "content": "for instruction text and study screenshots).", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 505, + 590 + ], + "score": 1.0, + "content": "Participants then played 20 episodes with a randomized sequence of agent partners and kitchen", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 209, + 600 + ], + "score": 1.0, + "content": "layouts. Episodes lasted", + "type": "text" + }, + { + "bbox": [ + 210, + 588, + 249, + 598 + ], + "score": 0.89, + "content": "T = 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "steps (1 minute) each. After every two episodes, participants", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 506, + 612 + ], + "score": 1.0, + "content": "reported their preference over the agent partners from those episodes on a five-point Likert-type scale.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 609, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 506, + 623 + ], + "score": 1.0, + "content": "After playing all 20 episodes, participants completed a debrief questionnaire collecting standard", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 620, + 506, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 620, + 506, + 633 + ], + "score": 1.0, + "content": "demographic information and open-ended feedback on the study. Our statistical analysis below", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "primarily relies upon the repeated-measures analysis of variance (ANOVA) method. See Appendix D", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 641, + 476, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 476, + 655 + ], + "score": 1.0, + "content": "for additional details of our study design and analysis, including independent ethical review.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 554, + 506, + 655 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 667, + 162, + 678 + ], + "lines": [ + { + "bbox": [ + 105, + 666, + 163, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 163, + 680 + ], + "score": 1.0, + "content": "5.2 Results", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 105, + 687, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 477, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 477, + 700 + ], + "score": 1.0, + "content": "Finding 1: FCP coordinates best with humans, achieving the highest score across maps", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "To begin, we compare the objective team performance supported by our FCP and baseline agents.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 507, + 724 + ], + "score": 1.0, + "content": "The strong FCP performance observed in agent-agent play generalizes to human-agent collaboration:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 686, + 507, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 139 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 84 + ], + "score": 1.0, + "content": "the FCP-human teams significantly outperform all other agent-human teams, achieving the highest", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 244, + 97 + ], + "score": 1.0, + "content": "average scores across maps, every", + "type": "text" + }, + { + "bbox": [ + 244, + 84, + 286, + 95 + ], + "score": 0.9, + "content": "p < 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 83, + 318, + 97 + ], + "score": 1.0, + "content": "(Figure", + "type": "text" + }, + { + "bbox": [ + 318, + 83, + 333, + 95 + ], + "score": 0.81, + "content": "\\mathrm { 7 a ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 83, + 506, + 97 + ], + "score": 1.0, + "content": ", while performing as well as or better than", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 319, + 107 + ], + "score": 1.0, + "content": "the other teams on each individual map (see Appendix", + "type": "text" + }, + { + "bbox": [ + 319, + 95, + 339, + 107 + ], + "score": 0.6, + "content": "\\mathbf { D } . 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 94, + 505, + 107 + ], + "score": 1.0, + "content": ". Echoing the results from our agent-agent", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 218, + 119 + ], + "score": 1.0, + "content": "ablation experiments (Table", + "type": "text" + }, + { + "bbox": [ + 218, + 105, + 228, + 117 + ], + "score": 0.63, + "content": "\\perp )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 105, + 506, + 119 + ], + "score": 1.0, + "content": ", the inclusion of past checkpoints in training proves critical to FCP’s", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 187, + 128 + ], + "score": 1.0, + "content": "strong performance,", + "type": "text" + }, + { + "bbox": [ + 188, + 117, + 230, + 128 + ], + "score": 0.88, + "content": "p < 0 . 0 { \\overline { { 0 1 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 116, + 261, + 128 + ], + "score": 1.0, + "content": "(Figure", + "type": "text" + }, + { + "bbox": [ + 261, + 116, + 276, + 129 + ], + "score": 0.82, + "content": "\\textcircled { 7 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 116, + 374, + 128 + ], + "score": 1.0, + "content": ". Similar to Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 375, + 116, + 392, + 128 + ], + "score": 0.58, + "content": "[ \\left[ 1 2 \\right] ]", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 116, + 506, + 128 + ], + "score": 1.0, + "content": ", we find that BCP outscores", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 329, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 282, + 140 + ], + "score": 1.0, + "content": "SP when collaborating with human players,", + "type": "text" + }, + { + "bbox": [ + 283, + 128, + 325, + 139 + ], + "score": 0.9, + "content": "p < 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 126, + 329, + 140 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 106, + 150, + 333, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 333, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 333, + 164 + ], + "score": 1.0, + "content": "Finding 2: Participants prefer FCP over all baselines", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 506, + 218 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 507, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 507, + 177 + ], + "score": 1.0, + "content": "FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 173, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 187 + ], + "score": 1.0, + "content": "Participants expressed a significant preference for FCP partners over all other agents, including", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 176, + 198 + ], + "score": 1.0, + "content": "BCP, with every", + "type": "text" + }, + { + "bbox": [ + 176, + 185, + 215, + 196 + ], + "score": 0.9, + "content": "p < 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 183, + 247, + 198 + ], + "score": 1.0, + "content": "(Figure", + "type": "text" + }, + { + "bbox": [ + 248, + 184, + 263, + 197 + ], + "score": 0.83, + "content": "\\dot { \\bigtriangledown } \\dot { \\mathbf { c } } { \\big \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 183, + 506, + 198 + ], + "score": 1.0, + "content": ". Notably, while human-BCP and human-PP teams did not", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "score": 1.0, + "content": "significantly differ in their completed deliveries, participants reported significantly preferring BCP", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 206, + 425, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 141, + 219 + ], + "score": 1.0, + "content": "over PP,", + "type": "text" + }, + { + "bbox": [ + 141, + 207, + 183, + 218 + ], + "score": 0.9, + "content": "p = 0 . 0 0 3", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 206, + 425, + 219 + ], + "score": 1.0, + "content": ", highlighting the informativeness of our subjective analysis.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 107, + 230, + 504, + 352 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 230, + 504, + 352 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 230, + 504, + 352 + ], + "spans": [ + { + "bbox": [ + 107, + 230, + 504, + 352 + ], + "score": 0.927, + "type": "image", + "image_path": "14e048153d10325b3617f94b393393ed3ab6cf3e38198f8791acf0fc83b051df.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 107, + 230, + 504, + 270.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 107, + 270.6666666666667, + 504, + 311.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 107, + 311.33333333333337, + 504, + 352.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 357, + 506, + 403 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "score": 1.0, + "content": "Figure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from human-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 368, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 233, + 379 + ], + "score": 1.0, + "content": "agent play in episodes of length", + "type": "text" + }, + { + "bbox": [ + 234, + 368, + 270, + 379 + ], + "score": 0.9, + "content": "T = 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 368, + 358, + 379 + ], + "score": 1.0, + "content": ". Error bars represents", + "type": "text" + }, + { + "bbox": [ + 359, + 368, + 378, + 379 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 368, + 505, + 379 + ], + "score": 1.0, + "content": "confidence intervals, calculated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 389, + 216, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 216, + 405 + ], + "score": 1.0, + "content": "be found in Appendix D.3.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + } + ], + "index": 14.75 + }, + { + "type": "title", + "bbox": [ + 107, + 421, + 266, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 267, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 267, + 435 + ], + "score": 1.0, + "content": "5.3 Exploratory behavioral analysis", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "To better understand how the human-agent scores and preferences may have arisen, here we analyze", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 452, + 437, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 437, + 465 + ], + "score": 1.0, + "content": "the resulting action trajectories of each human and agent player in our experiment.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "image", + "bbox": [ + 138, + 478, + 474, + 597 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 138, + 478, + 474, + 597 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 138, + 478, + 474, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 474, + 597 + ], + "score": 0.519, + "type": "image", + "image_path": "65959c0ab23a18c0f8de72bc451d199e7d9d9cc2698e7ca2ac4f320e26dac829.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 138, + 478, + 474, + 517.6666666666666 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 138, + 517.6666666666666, + 474, + 557.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 138, + 557.3333333333333, + 474, + 596.9999999999999 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 506, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 507, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 412, + 614 + ], + "score": 1.0, + "content": "Figure 8: Behavioral analysis: (a) FCP is able to move most frequently", + "type": "text" + }, + { + "bbox": [ + 412, + 601, + 432, + 613 + ], + "score": 0.85, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 601, + 507, + 614 + ], + "score": 1.0, + "content": "of the time), cor-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "responding to the best movement coordination with human partners. (b) FCP exhibits the most", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "equal preferences over cooking pots (0.11 difference), aligning with human preferences. Values are", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 633, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 649 + ], + "score": 1.0, + "content": "calculated as the absolute difference in preferences between the two pots; 1 indicates that the player", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 645, + 487, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 487, + 659 + ], + "score": 1.0, + "content": "only uses one of the two available pots, while 0 indicates that the player uses both pots equally.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 665, + 407, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 407, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 407, + 678 + ], + "score": 1.0, + "content": "Finding 1: FCP exhibits the best movement coordination with humans", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 405, + 691 + ], + "score": 1.0, + "content": "First, we investigate how much each player moves in an episode (Figure", + "type": "text" + }, + { + "bbox": [ + 406, + 677, + 421, + 691 + ], + "score": 0.27, + "content": "\\textcircled { 8 \\mathrm { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 677, + 506, + 691 + ], + "score": 1.0, + "content": ", where moving in a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 688, + 507, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 507, + 701 + ], + "score": 1.0, + "content": "higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 711, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 722 + ], + "score": 1.0, + "content": "for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 300, + 740, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 300, + 740, + 309, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 139 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 84 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 84 + ], + "score": 1.0, + "content": "the FCP-human teams significantly outperform all other agent-human teams, achieving the highest", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 244, + 97 + ], + "score": 1.0, + "content": "average scores across maps, every", + "type": "text" + }, + { + "bbox": [ + 244, + 84, + 286, + 95 + ], + "score": 0.9, + "content": "p < 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 83, + 318, + 97 + ], + "score": 1.0, + "content": "(Figure", + "type": "text" + }, + { + "bbox": [ + 318, + 83, + 333, + 95 + ], + "score": 0.81, + "content": "\\mathrm { 7 a ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 83, + 506, + 97 + ], + "score": 1.0, + "content": ", while performing as well as or better than", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 94, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 319, + 107 + ], + "score": 1.0, + "content": "the other teams on each individual map (see Appendix", + "type": "text" + }, + { + "bbox": [ + 319, + 95, + 339, + 107 + ], + "score": 0.6, + "content": "\\mathbf { D } . 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 339, + 94, + 505, + 107 + ], + "score": 1.0, + "content": ". Echoing the results from our agent-agent", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 105, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 218, + 119 + ], + "score": 1.0, + "content": "ablation experiments (Table", + "type": "text" + }, + { + "bbox": [ + 218, + 105, + 228, + 117 + ], + "score": 0.63, + "content": "\\perp )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 105, + 506, + 119 + ], + "score": 1.0, + "content": ", the inclusion of past checkpoints in training proves critical to FCP’s", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 187, + 128 + ], + "score": 1.0, + "content": "strong performance,", + "type": "text" + }, + { + "bbox": [ + 188, + 117, + 230, + 128 + ], + "score": 0.88, + "content": "p < 0 . 0 { \\overline { { 0 1 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 116, + 261, + 128 + ], + "score": 1.0, + "content": "(Figure", + "type": "text" + }, + { + "bbox": [ + 261, + 116, + 276, + 129 + ], + "score": 0.82, + "content": "\\textcircled { 7 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 116, + 374, + 128 + ], + "score": 1.0, + "content": ". Similar to Carroll et al.", + "type": "text" + }, + { + "bbox": [ + 375, + 116, + 392, + 128 + ], + "score": 0.58, + "content": "[ \\left[ 1 2 \\right] ]", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 116, + 506, + 128 + ], + "score": 1.0, + "content": ", we find that BCP outscores", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 126, + 329, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 282, + 140 + ], + "score": 1.0, + "content": "SP when collaborating with human players,", + "type": "text" + }, + { + "bbox": [ + 283, + 128, + 325, + 139 + ], + "score": 0.9, + "content": "p < 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 126, + 329, + 140 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 73, + 506, + 140 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 150, + 333, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 333, + 164 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 333, + 164 + ], + "score": 1.0, + "content": "Finding 2: Participants prefer FCP over all baselines", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 506, + 218 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 507, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 507, + 177 + ], + "score": 1.0, + "content": "FCP’s strong collaborative performance carries over to our participants’ subjective partner preferences.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 173, + 506, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 506, + 187 + ], + "score": 1.0, + "content": "Participants expressed a significant preference for FCP partners over all other agents, including", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 183, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 176, + 198 + ], + "score": 1.0, + "content": "BCP, with every", + "type": "text" + }, + { + "bbox": [ + 176, + 185, + 215, + 196 + ], + "score": 0.9, + "content": "p < 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 183, + 247, + 198 + ], + "score": 1.0, + "content": "(Figure", + "type": "text" + }, + { + "bbox": [ + 248, + 184, + 263, + 197 + ], + "score": 0.83, + "content": "\\dot { \\bigtriangledown } \\dot { \\mathbf { c } } { \\big \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 183, + 506, + 198 + ], + "score": 1.0, + "content": ". Notably, while human-BCP and human-PP teams did not", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 506, + 209 + ], + "score": 1.0, + "content": "significantly differ in their completed deliveries, participants reported significantly preferring BCP", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 206, + 425, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 206, + 141, + 219 + ], + "score": 1.0, + "content": "over PP,", + "type": "text" + }, + { + "bbox": [ + 141, + 207, + 183, + 218 + ], + "score": 0.9, + "content": "p = 0 . 0 0 3", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 206, + 425, + 219 + ], + "score": 1.0, + "content": ", highlighting the informativeness of our subjective analysis.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 162, + 507, + 219 + ] + }, + { + "type": "image", + "bbox": [ + 107, + 230, + 504, + 352 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 230, + 504, + 352 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 230, + 504, + 352 + ], + "spans": [ + { + "bbox": [ + 107, + 230, + 504, + 352 + ], + "score": 0.927, + "type": "image", + "image_path": "14e048153d10325b3617f94b393393ed3ab6cf3e38198f8791acf0fc83b051df.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 107, + 230, + 504, + 270.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 107, + 270.6666666666667, + 504, + 311.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 107, + 311.33333333333337, + 504, + 352.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 357, + 506, + 403 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 507, + 370 + ], + "score": 1.0, + "content": "Figure 7: Human-agent collaborative evaluation: Evaluation and preference metrics from human-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 368, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 233, + 379 + ], + "score": 1.0, + "content": "agent play in episodes of length", + "type": "text" + }, + { + "bbox": [ + 234, + 368, + 270, + 379 + ], + "score": 0.9, + "content": "T = 3 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 368, + 358, + 379 + ], + "score": 1.0, + "content": ". Error bars represents", + "type": "text" + }, + { + "bbox": [ + 359, + 368, + 378, + 379 + ], + "score": 0.87, + "content": "9 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 368, + 505, + 379 + ], + "score": 1.0, + "content": "confidence intervals, calculated", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "over episodes. Plots aggregate data across kitchen layouts; results calculated by individual layout can", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 389, + 216, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 216, + 405 + ], + "score": 1.0, + "content": "be found in Appendix D.3.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + } + ], + "index": 14.75 + }, + { + "type": "title", + "bbox": [ + 107, + 421, + 266, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 420, + 267, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 420, + 267, + 435 + ], + "score": 1.0, + "content": "5.3 Exploratory behavioral analysis", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 441, + 505, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "To better understand how the human-agent scores and preferences may have arisen, here we analyze", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 452, + 437, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 437, + 465 + ], + "score": 1.0, + "content": "the resulting action trajectories of each human and agent player in our experiment.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 440, + 505, + 465 + ] + }, + { + "type": "image", + "bbox": [ + 138, + 478, + 474, + 597 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 138, + 478, + 474, + 597 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 138, + 478, + 474, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 478, + 474, + 597 + ], + "score": 0.519, + "type": "image", + "image_path": "65959c0ab23a18c0f8de72bc451d199e7d9d9cc2698e7ca2ac4f320e26dac829.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 138, + 478, + 474, + 517.6666666666666 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 138, + 517.6666666666666, + 474, + 557.3333333333333 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 138, + 557.3333333333333, + 474, + 596.9999999999999 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 601, + 506, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 507, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 412, + 614 + ], + "score": 1.0, + "content": "Figure 8: Behavioral analysis: (a) FCP is able to move most frequently", + "type": "text" + }, + { + "bbox": [ + 412, + 601, + 432, + 613 + ], + "score": 0.85, + "content": "3 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 601, + 507, + 614 + ], + "score": 1.0, + "content": "of the time), cor-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "responding to the best movement coordination with human partners. (b) FCP exhibits the most", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 636 + ], + "score": 1.0, + "content": "equal preferences over cooking pots (0.11 difference), aligning with human preferences. Values are", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 633, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 649 + ], + "score": 1.0, + "content": "calculated as the absolute difference in preferences between the two pots; 1 indicates that the player", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 645, + 487, + 659 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 487, + 659 + ], + "score": 1.0, + "content": "only uses one of the two available pots, while 0 indicates that the player uses both pots equally.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 601, + 507, + 659 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 665, + 407, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 407, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 407, + 678 + ], + "score": 1.0, + "content": "Finding 1: FCP exhibits the best movement coordination with humans", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 678, + 506, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 405, + 691 + ], + "score": 1.0, + "content": "First, we investigate how much each player moves in an episode (Figure", + "type": "text" + }, + { + "bbox": [ + 406, + 677, + 421, + 691 + ], + "score": 0.27, + "content": "\\textcircled { 8 \\mathrm { a } }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 677, + 506, + 691 + ], + "score": 1.0, + "content": ", where moving in a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 688, + 507, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 507, + 701 + ], + "score": 1.0, + "content": "higher fraction of timesteps may suggest fewer collisions and thus better coordination with a partner.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "Notably, we observe two results: (1) humans rarely move, a behavior which is out-of-distribution", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 711, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 722 + ], + "score": 1.0, + "content": "for typical training methods (e.g. SP, PP) but is seen in the training distribution for BCP and FCP.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 677, + 507, + 722 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "movement strategy with its partner. This result was also reported by human participants, for example:", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "“I noticed that some of my partners seemed to know they needed to move around me, while others", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 473, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 473, + 119 + ], + "score": 1.0, + "content": "seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 106, + 129, + 444, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 446, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 446, + 143 + ], + "score": 1.0, + "content": "Finding 2: FCP’s preferences over cooking pots aligns best with that of humans", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 155 + ], + "score": 1.0, + "content": "Next, we investigate whether there was a preference for a specific cooking pot in the layouts which", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 243, + 167 + ], + "score": 1.0, + "content": "included two cooking pots (Figure", + "type": "text" + }, + { + "bbox": [ + 243, + 154, + 258, + 167 + ], + "score": 0.55, + "content": "\\textcircled { 8 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 153, + 506, + 167 + ], + "score": 1.0, + "content": ". To do this, we calculate the difference in the number of times", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "each pot was used by each player, where a high value indicates a strong preference for one pot and a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 344, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 344, + 188 + ], + "score": 1.0, + "content": "low value indicates more equal preference for the two pots.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 192, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "As can be seen in the FCP column, our agent typically has the most aligned preferences with that of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 505, + 216 + ], + "score": 1.0, + "content": "humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 230, + 227 + ], + "score": 1.0, + "content": "one cooking pot over the other", + "type": "text" + }, + { + "bbox": [ + 231, + 214, + 258, + 225 + ], + "score": 0.88, + "content": "5 5 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "of the time (i.e. a 0.11 point difference). In contrast, all other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 442, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 442, + 249 + ], + "score": 1.0, + "content": "poorly to typical human behavior of using both pots, leading to worse performance.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 259, + 180, + 273 + ], + "lines": [ + { + "bbox": [ + 104, + 258, + 181, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 258, + 181, + 275 + ], + "score": 1.0, + "content": "6 Discussion", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "Summary In this work, we investigated the challenging problem of zero-shot collaboration with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "humans without using human data in the training pipeline. To accomplish this, we introduced", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "of training partners. We found that FCP agents scored significantly higher than all baselines when", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "experimental design, we also found that humans reported a strong subjective preference to partnering", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 345, + 249, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 249, + 358 + ], + "score": 1.0, + "content": "with FCP agents over all baselines.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Limitations and future work Our method currently relies on the manual process of initially", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "training and selecting a diverse set of partners. This is not only time consuming, but also prone to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "researcher biases that may negatively influence the behavior of the created agents. Additionally, while", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 393, + 507, + 408 + ], + "spans": [ + { + "bbox": [ + 104, + 393, + 302, + 408 + ], + "score": 1.0, + "content": "we found FCP with a partner population size of", + "type": "text" + }, + { + "bbox": [ + 302, + 394, + 336, + 405 + ], + "score": 0.9, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 393, + 507, + 408 + ], + "score": 1.0, + "content": "sufficient here, for more complex games,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "FCP may require an unrealistically large partner population size to represent sufficiently diverse", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 416, + 504, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 504, + 428 + ], + "score": 1.0, + "content": "strategies. To address these concerns, methods for automatically generating partner diversity for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "common-payoff games may be important. Possibilities include adaptive population matchmaking", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 295, + 451 + ], + "score": 1.0, + "content": "as been used in competitive zero-sum games", + "type": "text" + }, + { + "bbox": [ + 295, + 437, + 313, + 449 + ], + "score": 0.8, + "content": "\\mathbb { \\lVert 6 9 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 437, + 505, + 451 + ], + "score": 1.0, + "content": ", as well as auxiliary objectives that explicitly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 281, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 281, + 461 + ], + "score": 1.0, + "content": "encourage behavioral diversity [19, 45, 46].", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 107, + 466, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 107, + 466, + 505, + 476 + ], + "score": 1.0, + "content": "Our method requires a known and fixed reward function. We also focus on one domain in order to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "compare with prior work which has argued that human-in-the-loop training is necessary. Consequently,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "score": 1.0, + "content": "the resulting agents are only designed to adaptively collaborate on a single task, and not to infer", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 230, + 511 + ], + "score": 1.0, + "content": "human preferences in general", + "type": "text" + }, + { + "bbox": [ + 231, + 497, + 273, + 509 + ], + "score": 0.46, + "content": "\\textcircled { 1 1 } \\textcircled { 3 3 } \\textcircled { 5 9 }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 497, + 506, + 511 + ], + "score": 1.0, + "content": ". Moreover, if a task’s reward function is poorly aligned", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "with how humans approach the task, our method may well produce subpar partners, as would", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 520, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 531 + ], + "score": 1.0, + "content": "any method without access to human data. Thus, additional domains and tasks should be studied", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "to better understand how our method generalizes. Targeted experiments to test specific forms of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 541, + 504, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 328, + 554 + ], + "score": 1.0, + "content": "generalization may be especially helpful in this regard", + "type": "text" + }, + { + "bbox": [ + 329, + 541, + 346, + 552 + ], + "score": 0.76, + "content": "\\overline { { [ 3 8 ] } }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 542, + 504, + 554 + ], + "score": 1.0, + "content": ", as could approaches that procedurally", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 552, + 352, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 335, + 565 + ], + "score": 1.0, + "content": "generate environment layouts requiring diverse solutions", + "type": "text" + }, + { + "bbox": [ + 335, + 552, + 352, + 564 + ], + "score": 0.39, + "content": "\\pmb { \\mathbb { Z } } 2 \\mathbf { l }", + "type": "inline_equation" + } + ], + "index": 39 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "Finally, it may be possible to produce even stronger agent assistants by combining the strengths of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 395, + 592 + ], + "score": 1.0, + "content": "FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al.", + "type": "text" + }, + { + "bbox": [ + 396, + 579, + 413, + 591 + ], + "score": 0.5, + "content": "\\textcircled { \\lvert 3 8 \\rvert }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "recently demonstrated", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "that modifying BCP to train with multiple BC partners produces more robust collaboration with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 602, + 425, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 425, + 614 + ], + "score": 1.0, + "content": "held-out agents, a finding that would be interesting to test with human partners.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "expresses preferences or intentions to harm others). This could therefore produce negative societal", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 672, + 486, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 486, + 685 + ], + "score": 1.0, + "content": "effects either if training leads to poor alignment or if agents are optimized for harmful metrics.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "One potential strategy for mitigating these risks is the use of human preference data [15]. Such data", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "could be used to fine-tune and filter trained agents before deployment, encouraging better alignment", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "with human values. A key question in this line of research is how human preference data should be", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 505, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 86 + ], + "score": 1.0, + "content": "(2) FCP moves the most on all layouts other than Forced, suggesting it is better at coordinating its", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "movement strategy with its partner. This result was also reported by human participants, for example:", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 104, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "“I noticed that some of my partners seemed to know they needed to move around me, while others", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 106, + 473, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 473, + 119 + ], + "score": 1.0, + "content": "seemed to get ‘stuck’ until I moved out of their way” (see Appendix D for more examples).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 104, + 73, + 506, + 119 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 129, + 444, + 141 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 446, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 446, + 143 + ], + "score": 1.0, + "content": "Finding 2: FCP’s preferences over cooking pots aligns best with that of humans", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 143, + 505, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 143, + 505, + 155 + ], + "score": 1.0, + "content": "Next, we investigate whether there was a preference for a specific cooking pot in the layouts which", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 243, + 167 + ], + "score": 1.0, + "content": "included two cooking pots (Figure", + "type": "text" + }, + { + "bbox": [ + 243, + 154, + 258, + 167 + ], + "score": 0.55, + "content": "\\textcircled { 8 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 153, + 506, + 167 + ], + "score": 1.0, + "content": ". To do this, we calculate the difference in the number of times", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "each pot was used by each player, where a high value indicates a strong preference for one pot and a", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 344, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 344, + 188 + ], + "score": 1.0, + "content": "low value indicates more equal preference for the two pots.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 143, + 506, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 192, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "As can be seen in the FCP column, our agent typically has the most aligned preferences with that of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 202, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 505, + 216 + ], + "score": 1.0, + "content": "humans (0.11 for FCP to 0.14 for humans). Behaviorally speaking, this means that our agent prefers", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 213, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 230, + 227 + ], + "score": 1.0, + "content": "one cooking pot over the other", + "type": "text" + }, + { + "bbox": [ + 231, + 214, + 258, + 225 + ], + "score": 0.88, + "content": "5 5 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 213, + 506, + 227 + ], + "score": 1.0, + "content": "of the time (i.e. a 0.11 point difference). In contrast, all other", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "agents have a strong preference for a single pot. This is a non-adaptive strategy which generalizes", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 442, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 442, + 249 + ], + "score": 1.0, + "content": "poorly to typical human behavior of using both pots, leading to worse performance.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 191, + 506, + 249 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 259, + 180, + 273 + ], + "lines": [ + { + "bbox": [ + 104, + 258, + 181, + 275 + ], + "spans": [ + { + "bbox": [ + 104, + 258, + 181, + 275 + ], + "score": 1.0, + "content": "6 Discussion", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 280, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 292 + ], + "score": 1.0, + "content": "Summary In this work, we investigated the challenging problem of zero-shot collaboration with", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "humans without using human data in the training pipeline. To accomplish this, we introduced", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 315 + ], + "score": 1.0, + "content": "Fictitious Co-Play (FCP) – a surprisingly simple yet effective method based on creating a diverse set", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "of training partners. We found that FCP agents scored significantly higher than all baselines when", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 336 + ], + "score": 1.0, + "content": "partnered with both novel agent and human partners. Furthermore, through a rigorous human-agent", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 348 + ], + "score": 1.0, + "content": "experimental design, we also found that humans reported a strong subjective preference to partnering", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 345, + 249, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 249, + 358 + ], + "score": 1.0, + "content": "with FCP agents over all baselines.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 280, + 506, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "Limitations and future work Our method currently relies on the manual process of initially", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "training and selecting a diverse set of partners. This is not only time consuming, but also prone to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 505, + 396 + ], + "score": 1.0, + "content": "researcher biases that may negatively influence the behavior of the created agents. Additionally, while", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 393, + 507, + 408 + ], + "spans": [ + { + "bbox": [ + 104, + 393, + 302, + 408 + ], + "score": 1.0, + "content": "we found FCP with a partner population size of", + "type": "text" + }, + { + "bbox": [ + 302, + 394, + 336, + 405 + ], + "score": 0.9, + "content": "N = 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 393, + 507, + 408 + ], + "score": 1.0, + "content": "sufficient here, for more complex games,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 405, + 505, + 418 + ], + "score": 1.0, + "content": "FCP may require an unrealistically large partner population size to represent sufficiently diverse", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 416, + 504, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 504, + 428 + ], + "score": 1.0, + "content": "strategies. To address these concerns, methods for automatically generating partner diversity for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 441 + ], + "score": 1.0, + "content": "common-payoff games may be important. Possibilities include adaptive population matchmaking", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 295, + 451 + ], + "score": 1.0, + "content": "as been used in competitive zero-sum games", + "type": "text" + }, + { + "bbox": [ + 295, + 437, + 313, + 449 + ], + "score": 0.8, + "content": "\\mathbb { \\lVert 6 9 \\rVert }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 437, + 505, + 451 + ], + "score": 1.0, + "content": ", as well as auxiliary objectives that explicitly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 448, + 281, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 281, + 461 + ], + "score": 1.0, + "content": "encourage behavioral diversity [19, 45, 46].", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 104, + 361, + 507, + 461 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 465, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 107, + 466, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 107, + 466, + 505, + 476 + ], + "score": 1.0, + "content": "Our method requires a known and fixed reward function. We also focus on one domain in order to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 489 + ], + "score": 1.0, + "content": "compare with prior work which has argued that human-in-the-loop training is necessary. Consequently,", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "score": 1.0, + "content": "the resulting agents are only designed to adaptively collaborate on a single task, and not to infer", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 230, + 511 + ], + "score": 1.0, + "content": "human preferences in general", + "type": "text" + }, + { + "bbox": [ + 231, + 497, + 273, + 509 + ], + "score": 0.46, + "content": "\\textcircled { 1 1 } \\textcircled { 3 3 } \\textcircled { 5 9 }", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 497, + 506, + 511 + ], + "score": 1.0, + "content": ". Moreover, if a task’s reward function is poorly aligned", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "with how humans approach the task, our method may well produce subpar partners, as would", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 520, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 531 + ], + "score": 1.0, + "content": "any method without access to human data. Thus, additional domains and tasks should be studied", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "to better understand how our method generalizes. Targeted experiments to test specific forms of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 541, + 504, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 328, + 554 + ], + "score": 1.0, + "content": "generalization may be especially helpful in this regard", + "type": "text" + }, + { + "bbox": [ + 329, + 541, + 346, + 552 + ], + "score": 0.76, + "content": "\\overline { { [ 3 8 ] } }", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 542, + 504, + 554 + ], + "score": 1.0, + "content": ", as could approaches that procedurally", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 552, + 352, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 335, + 565 + ], + "score": 1.0, + "content": "generate environment layouts requiring diverse solutions", + "type": "text" + }, + { + "bbox": [ + 335, + 552, + 352, + 564 + ], + "score": 0.39, + "content": "\\pmb { \\mathbb { Z } } 2 \\mathbf { l }", + "type": "inline_equation" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 466, + 506, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "Finally, it may be possible to produce even stronger agent assistants by combining the strengths of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 395, + 592 + ], + "score": 1.0, + "content": "FCP (i.e. diversity) and BCP (i.e. human-like play). Indeed, Knott et al.", + "type": "text" + }, + { + "bbox": [ + 396, + 579, + 413, + 591 + ], + "score": 0.5, + "content": "\\textcircled { \\lvert 3 8 \\rvert }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "recently demonstrated", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 604 + ], + "score": 1.0, + "content": "that modifying BCP to train with multiple BC partners produces more robust collaboration with", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 602, + 425, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 425, + 614 + ], + "score": 1.0, + "content": "held-out agents, a finding that would be interesting to test with human partners.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 569, + 506, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 630 + ], + "score": 1.0, + "content": "Societal impact A challenge for this line of work is ensuring agent behavior is aligned with human", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "values (i.e. the AI value alignment problem [23, 59]). Our method has no guarantees that the resulting", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 640, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 505, + 652 + ], + "score": 1.0, + "content": "policy aligns with the preferences, intentions, or welfare of its potential partners. It likewise does", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 505, + 663 + ], + "score": 1.0, + "content": "not exclude the possibility that the target being optimized for is harmful (e.g. if the agent’s partner", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 674 + ], + "score": 1.0, + "content": "expresses preferences or intentions to harm others). This could therefore produce negative societal", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 672, + 486, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 486, + 685 + ], + "score": 1.0, + "content": "effects either if training leads to poor alignment or if agents are optimized for harmful metrics.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 105, + 617, + 506, + 685 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "One potential strategy for mitigating these risks is the use of human preference data [15]. Such data", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "could be used to fine-tune and filter trained agents before deployment, encouraging better alignment", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 711, + 505, + 723 + ], + "score": 1.0, + "content": "with human values. A key question in this line of research is how human preference data should be", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "aggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 97 + ], + "score": 1.0, + "content": "agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 225, + 107 + ], + "score": 1.0, + "content": "beliefs and perceptions of AI", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 226, + 94, + 243, + 106 + ], + "score": 0.47, + "content": "\\lVert \\overline { { 4 8 } } \\rVert", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 244, + 93, + 505, + 107 + ], + "score": 1.0, + "content": ", and how they steer human-agent interaction, would help inform", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 104, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 506, + 120 + ], + "score": 1.0, + "content": "agent design for positive societal impact. For instance, developers could incorporate specific priors", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 351, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 316, + 128 + ], + "score": 1.0, + "content": "into agents to reinforce tendencies for fair outcomes", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 317, + 116, + 349, + 128 + ], + "score": 0.28, + "content": "\\pm \\mathbb { Z } 0 . \\pm \\mathbb { B } 2 \\mathbb { I }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 349, + 115, + 351, + 128 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 4 + } + ], + "index": 51, + "bbox_fs": [ + 106, + 689, + 505, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "aggregated—or selected, in the case of expert preferences—when our aim is to create socially aligned", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 505, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 97 + ], + "score": 1.0, + "content": "agents (i.e. agents that are sufficiently aligned for everyone). Relatedly, targeted research on human", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 225, + 107 + ], + "score": 1.0, + "content": "beliefs and perceptions of AI", + "type": "text" + }, + { + "bbox": [ + 226, + 94, + 243, + 106 + ], + "score": 0.47, + "content": "\\lVert \\overline { { 4 8 } } \\rVert", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 93, + 505, + 107 + ], + "score": 1.0, + "content": ", and how they steer human-agent interaction, would help inform", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 104, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 104, + 506, + 120 + ], + "score": 1.0, + "content": "agent design for positive societal impact. For instance, developers could incorporate specific priors", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 351, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 316, + 128 + ], + "score": 1.0, + "content": "into agents to reinforce tendencies for fair outcomes", + "type": "text" + }, + { + "bbox": [ + 317, + 116, + 349, + 128 + ], + "score": 0.28, + "content": "\\pm \\mathbb { Z } 0 . \\pm \\mathbb { B } 2 \\mathbb { I }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 115, + 351, + 128 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "Conclusion We proposed a method which is both effective at collaborating with humans and simple", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "to implement. We also presented a rigorous and general methodology for evaluating with humans", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "and eliciting their preferences. Together, these establish a strong foundation for future research on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 164, + 413, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 413, + 179 + ], + "score": 1.0, + "content": "the important challenge of human-agent collaboration for benefiting society.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "title", + "bbox": [ + 108, + 192, + 207, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 208, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 208, + 208 + ], + "score": 1.0, + "content": "Acknowledgements", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 420, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 420, + 261 + ], + "score": 1.0, + "content": "Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 107, + 274, + 205, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 207, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 207, + 291 + ], + "score": 1.0, + "content": "Funding disclosure", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 298, + 459, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 462, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 462, + 313 + ], + "score": 1.0, + "content": "This work was funded solely by DeepMind. The authors declare no competing interests.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 107, + 325, + 163, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 165, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 165, + 339 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 109, + 338, + 506, + 723 + ], + "lines": [ + { + "bbox": [ + 110, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 110, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "[1] J. Abramson, A. Ahuja, I. Barr, A. Brussee, F. Carnevale, M. Cassin, R. Chhaparia, S. Clark,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 126, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "B. Damoc, A. Dudzik, et al. Imitating interactive intelligence. arXiv preprint arXiv:2012.05672,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 365, + 153, + 378 + ], + "spans": [ + { + "bbox": [ + 126, + 365, + 153, + 378 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 110, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 110, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "[2] S. V. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 127, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 127, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "for ad hoc coordination in multiagent systems. In International Conference on Autonomous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 402, + 323, + 414 + ], + "spans": [ + { + "bbox": [ + 126, + 402, + 323, + 414 + ], + "score": 1.0, + "content": "Agents and Multiagent Systems (AAMAS), 2013.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 111, + 416, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 111, + 416, + 506, + 429 + ], + "score": 1.0, + "content": "[3] B. Baker. Emergent reciprocity and team formation from randomized uncertain social prefer-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 426, + 398, + 441 + ], + "spans": [ + { + "bbox": [ + 127, + 426, + 398, + 441 + ], + "score": 1.0, + "content": "ences. In Neural Information Processing Systems (NeurIPS), 2020.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 111, + 441, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 111, + 441, + 506, + 453 + ], + "score": 1.0, + "content": "[4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 127, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 127, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "S. Moitra, E. Hughes, et al. The Hanabi challenge: A new frontier for AI research. Artificial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 462, + 257, + 475 + ], + "spans": [ + { + "bbox": [ + 127, + 462, + 257, + 475 + ], + "score": 1.0, + "content": "Intelligence, 280:103216, 2020.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 110, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 110, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "[5] S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 488, + 369, + 500 + ], + "spans": [ + { + "bbox": [ + 126, + 488, + 369, + 500 + ], + "score": 1.0, + "content": "new teammates. Artificial Intelligence, 242:132–171, 2017.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 111, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 111, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "[6] A. Bauer, D. Wollherr, and M. Buss. Human–robot collaboration: A survey. International", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 126, + 513, + 335, + 525 + ], + "spans": [ + { + "bbox": [ + 126, + 513, + 335, + 525 + ], + "score": 1.0, + "content": "Journal of Humanoid Robotics, 5(01):47–66, 2008.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 110, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 110, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "[7] G. W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 539, + 262, + 550 + ], + "spans": [ + { + "bbox": [ + 127, + 539, + 262, + 550 + ], + "score": 1.0, + "content": "allocation, 13(1):374–376, 1951.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 111, + 550, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 111, + 550, + 505, + 566 + ], + "score": 1.0, + "content": "[8] N. Brown and T. Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 127, + 563, + 331, + 576 + ], + "spans": [ + { + "bbox": [ + 127, + 563, + 331, + 576 + ], + "score": 1.0, + "content": "professionals. Science, 359(6374):418–424, 2018.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 110, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 110, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "[9] N. Brown and T. Sandholm. Superhuman AI for multiplayer poker. Science, 365(6456):885–890,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 127, + 589, + 153, + 600 + ], + "spans": [ + { + "bbox": [ + 127, + 589, + 153, + 600 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "score": 1.0, + "content": "[10] K. Bullard, F. Meier, D. Kiela, J. Pineau, and J. Foerster. Exploring zero-shot emergent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 127, + 614, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 127, + 614, + 506, + 626 + ], + "score": 1.0, + "content": "communication in embodied multi-agent populations. arXiv preprint arXiv:2010.15896, 2020.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "[11] K. Bullard, D. Kiela, J. Pineau, and J. Foerster. Quasi-equivalence discovery for zero-shot", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 126, + 639, + 397, + 651 + ], + "spans": [ + { + "bbox": [ + 126, + 639, + 397, + 651 + ], + "score": 1.0, + "content": "emergent communication. arXiv preprint arXiv:2103.08067, 2021.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "score": 1.0, + "content": "[12] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 126, + 663, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 126, + 663, + 505, + 678 + ], + "score": 1.0, + "content": "utility of learning about humans for human-AI coordination. In Neural Information Processing", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 676, + 233, + 687 + ], + "spans": [ + { + "bbox": [ + 127, + 676, + 233, + 687 + ], + "score": 1.0, + "content": "Systems (NeurIPS), 2019.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "[13] R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Investigating partner diversification", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 127, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "methods in cooperative multi-agent deep reinforcement learning. In International Conference", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 711, + 294, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 294, + 723 + ], + "score": 1.0, + "content": "on Neural Information Processing, 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 32 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 73, + 505, + 127 + ], + "lines": [], + "index": 2, + "bbox_fs": [ + 104, + 73, + 506, + 128 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 505, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "Conclusion We proposed a method which is both effective at collaborating with humans and simple", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 505, + 156 + ], + "score": 1.0, + "content": "to implement. We also presented a rigorous and general methodology for evaluating with humans", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "and eliciting their preferences. Together, these establish a strong foundation for future research on", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 164, + 413, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 413, + 179 + ], + "score": 1.0, + "content": "the important challenge of human-agent collaboration for benefiting society.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 132, + 505, + 179 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 192, + 207, + 205 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 208, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 208, + 208 + ], + "score": 1.0, + "content": "Acknowledgements", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 216, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 215, + 505, + 228 + ], + "score": 1.0, + "content": "The authors would like to thank Mary Cassin for creating the game sprite art; Rohin Shah, Thore", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "Graepel, and Iason Gabriel for feedback on the draft; Lucy Campbell-Gillingham, Tina Zhu, and", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "Saffron Huang for support in evaluating agents with humans; and Max Kleiman-Weiner, Natasha", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 420, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 420, + 261 + ], + "score": 1.0, + "content": "Jaques, Marc Lanctot, Mike Bowling, and Dan Roberts for useful discussions.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 215, + 505, + 261 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 274, + 205, + 288 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 207, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 207, + 291 + ], + "score": 1.0, + "content": "Funding disclosure", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 108, + 298, + 459, + 310 + ], + "lines": [ + { + "bbox": [ + 105, + 296, + 462, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 462, + 313 + ], + "score": 1.0, + "content": "This work was funded solely by DeepMind. The authors declare no competing interests.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 296, + 462, + 313 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 325, + 163, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 165, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 165, + 339 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "list", + "bbox": [ + 109, + 338, + 506, + 723 + ], + "lines": [ + { + "bbox": [ + 110, + 343, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 110, + 343, + 506, + 356 + ], + "score": 1.0, + "content": "[1] J. Abramson, A. Ahuja, I. Barr, A. Brussee, F. Carnevale, M. Cassin, R. Chhaparia, S. Clark,", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 354, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 126, + 354, + 506, + 367 + ], + "score": 1.0, + "content": "B. Damoc, A. Dudzik, et al. Imitating interactive intelligence. arXiv preprint arXiv:2012.05672,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 365, + 153, + 378 + ], + "spans": [ + { + "bbox": [ + 126, + 365, + 153, + 378 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 110, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "[2] S. V. Albrecht and S. Ramamoorthy. A game-theoretic model and best-response learning method", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 127, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "for ad hoc coordination in multiagent systems. In International Conference on Autonomous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 402, + 323, + 414 + ], + "spans": [ + { + "bbox": [ + 126, + 402, + 323, + 414 + ], + "score": 1.0, + "content": "Agents and Multiagent Systems (AAMAS), 2013.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 416, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 111, + 416, + 506, + 429 + ], + "score": 1.0, + "content": "[3] B. Baker. Emergent reciprocity and team formation from randomized uncertain social prefer-", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 426, + 398, + 441 + ], + "spans": [ + { + "bbox": [ + 127, + 426, + 398, + 441 + ], + "score": 1.0, + "content": "ences. In Neural Information Processing Systems (NeurIPS), 2020.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 441, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 111, + 441, + 506, + 453 + ], + "score": 1.0, + "content": "[4] N. Bard, J. N. Foerster, S. Chandar, N. Burch, M. Lanctot, H. F. Song, E. Parisotto, V. Dumoulin,", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 127, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "S. Moitra, E. Hughes, et al. The Hanabi challenge: A new frontier for AI research. Artificial", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 127, + 462, + 257, + 475 + ], + "spans": [ + { + "bbox": [ + 127, + 462, + 257, + 475 + ], + "score": 1.0, + "content": "Intelligence, 280:103216, 2020.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 110, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "[5] S. Barrett, A. Rosenfeld, S. Kraus, and P. Stone. Making friends on the fly: Cooperating with", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 488, + 369, + 500 + ], + "spans": [ + { + "bbox": [ + 126, + 488, + 369, + 500 + ], + "score": 1.0, + "content": "new teammates. Artificial Intelligence, 242:132–171, 2017.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 502, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 111, + 502, + 505, + 515 + ], + "score": 1.0, + "content": "[6] A. Bauer, D. Wollherr, and M. Buss. Human–robot collaboration: A survey. International", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 513, + 335, + 525 + ], + "spans": [ + { + "bbox": [ + 126, + 513, + 335, + 525 + ], + "score": 1.0, + "content": "Journal of Humanoid Robotics, 5(01):47–66, 2008.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 110, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "[7] G. W. Brown. Iterative solution of games by fictitious play. Activity analysis of production and", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 539, + 262, + 550 + ], + "spans": [ + { + "bbox": [ + 127, + 539, + 262, + 550 + ], + "score": 1.0, + "content": "allocation, 13(1):374–376, 1951.", + "type": "text" + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 550, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 111, + 550, + 505, + 566 + ], + "score": 1.0, + "content": "[8] N. Brown and T. Sandholm. Superhuman AI for heads-up no-limit poker: Libratus beats top", + "type": "text" + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 563, + 331, + 576 + ], + "spans": [ + { + "bbox": [ + 127, + 563, + 331, + 576 + ], + "score": 1.0, + "content": "professionals. Science, 359(6374):418–424, 2018.", + "type": "text" + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 110, + 576, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 110, + 576, + 506, + 590 + ], + "score": 1.0, + "content": "[9] N. Brown and T. Sandholm. Superhuman AI for multiplayer poker. Science, 365(6456):885–890,", + "type": "text" + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 589, + 153, + 600 + ], + "spans": [ + { + "bbox": [ + 127, + 589, + 153, + 600 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 616 + ], + "score": 1.0, + "content": "[10] K. Bullard, F. Meier, D. Kiela, J. Pineau, and J. Foerster. Exploring zero-shot emergent", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 614, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 127, + 614, + 506, + 626 + ], + "score": 1.0, + "content": "communication in embodied multi-agent populations. arXiv preprint arXiv:2010.15896, 2020.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 628, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 506, + 641 + ], + "score": 1.0, + "content": "[11] K. Bullard, D. Kiela, J. Pineau, and J. Foerster. Quasi-equivalence discovery for zero-shot", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 639, + 397, + 651 + ], + "spans": [ + { + "bbox": [ + 126, + 639, + 397, + 651 + ], + "score": 1.0, + "content": "emergent communication. arXiv preprint arXiv:2103.08067, 2021.", + "type": "text" + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 505, + 666 + ], + "score": 1.0, + "content": "[12] M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P. Abbeel, and A. Dragan. On the", + "type": "text" + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 663, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 126, + 663, + 505, + 678 + ], + "score": 1.0, + "content": "utility of learning about humans for human-AI coordination. In Neural Information Processing", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 676, + 233, + 687 + ], + "spans": [ + { + "bbox": [ + 127, + 676, + 233, + 687 + ], + "score": 1.0, + "content": "Systems (NeurIPS), 2019.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 702 + ], + "score": 1.0, + "content": "[13] R. Charakorn, P. Manoonpong, and N. Dilokthanakul. Investigating partner diversification", + "type": "text" + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "methods in cooperative multi-agent deep reinforcement learning. In International Conference", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 711, + 294, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 294, + 723 + ], + "score": 1.0, + "content": "on Neural Information Processing, 2020.", + "type": "text" + } + ], + "index": 47, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "[14] R. Choudhury, G. Swamy, D. Hadfield-Menell, and A. D. Dragan. On the utility of model", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 82, + 507, + 98 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 507, + 98 + ], + "score": 1.0, + "content": "learning in HRI. In ACM/IEEE International Conference on Human-Robot Interaction (HRI),", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 92, + 155, + 109 + ], + "spans": [ + { + "bbox": [ + 126, + 92, + 155, + 109 + ], + "score": 1.0, + "content": "2019.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 109, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 123 + ], + "score": 1.0, + "content": "[15] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "learning from human preferences. In Proceedings of the 31st International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 131, + 392, + 145 + ], + "spans": [ + { + "bbox": [ + 126, + 131, + 392, + 145 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pages 4302–4310, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "[16] A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 157, + 477, + 172 + ], + "spans": [ + { + "bbox": [ + 126, + 157, + 477, + 172 + ], + "score": 1.0, + "content": "T. Graepel. Open problems in cooperative AI. arXiv preprint arXiv:2012.08630, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 172, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 507, + 187 + ], + "score": 1.0, + "content": "[17] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu,", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 182, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 125, + 182, + 507, + 198 + ], + "score": 1.0, + "content": "T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable distributed deep-RL", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 506, + 209 + ], + "score": 1.0, + "content": "with importance weighted actor-learner architectures. In International Conference on Machine", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 206, + 226, + 218 + ], + "spans": [ + { + "bbox": [ + 127, + 206, + 226, + 218 + ], + "score": 1.0, + "content": "Learning (ICML), 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "[18] P. Eyal, R. David, G. Andrew, E. Zak, and D. Ekaterina. Data quality of platforms and panels", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 231, + 447, + 245 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 447, + 245 + ], + "score": 1.0, + "content": "for online behavioral research. Behavior Research Methods, pages 1–20, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "[19] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 258, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 126, + 258, + 507, + 272 + ], + "score": 1.0, + "content": "without a reward function. In International Conference on Learning Representations (ICLR),", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 267, + 155, + 283 + ], + "spans": [ + { + "bbox": [ + 127, + 267, + 155, + 283 + ], + "score": 1.0, + "content": "2019.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 282, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 298 + ], + "score": 1.0, + "content": "[20] E. Fehr and K. M. Schmidt. A theory of fairness, competition, and cooperation. The Quarterly", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 294, + 314, + 309 + ], + "spans": [ + { + "bbox": [ + 126, + 294, + 314, + 309 + ], + "score": 1.0, + "content": "Journal of Economics, 114(3):817–868, 1999.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 309, + 507, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 507, + 324 + ], + "score": 1.0, + "content": "[21] J. Foerster, F. Song, E. Hughes, N. Burch, I. Dunning, S. Whiteson, M. Botvinick, and M. Bowl-", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 126, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "ing. Bayesian action decoder for deep multi-agent reinforcement learning. In International", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 331, + 325, + 345 + ], + "spans": [ + { + "bbox": [ + 126, + 331, + 325, + 345 + ], + "score": 1.0, + "content": "Conference on Machine Learning (ICML), 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 347, + 507, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 507, + 360 + ], + "score": 1.0, + "content": "[22] M. Fontaine, Y.-C. Hsu, Y. Zhang, B. Tjanaka, and S. Nikolaidis. On the importance of", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 358, + 496, + 370 + ], + "spans": [ + { + "bbox": [ + 128, + 358, + 496, + 370 + ], + "score": 1.0, + "content": "environments in human-robot coordination. In Robotics: Science and Systems (RSS), 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 374, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 506, + 385 + ], + "score": 1.0, + "content": "[23] I. Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30(3):411–437,", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 384, + 154, + 397 + ], + "spans": [ + { + "bbox": [ + 127, + 384, + 154, + 397 + ], + "score": 1.0, + "content": "2020.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 398, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 506, + 413 + ], + "score": 1.0, + "content": "[24] M. Garnelo, W. M. Czarnecki, S. Liu, D. Tirumala, J. Oh, G. Gidel, H. van Hasselt, and", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 410, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 126, + 410, + 505, + 424 + ], + "score": 1.0, + "content": "D. Balduzzi. Pick your battles: Interaction graphs as population-level objectives for strategic", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 419, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 126, + 419, + 507, + 435 + ], + "score": 1.0, + "content": "diversity. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS),", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 431, + 155, + 446 + ], + "spans": [ + { + "bbox": [ + 126, + 431, + 155, + 446 + ], + "score": 1.0, + "content": "2021.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 447, + 291, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 291, + 460 + ], + "score": 1.0, + "content": "[25] Ghost Town Games. Overcooked, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "[26] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 473, + 419, + 487 + ], + "spans": [ + { + "bbox": [ + 126, + 473, + 419, + 487 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "[27] J. Heinrich and D. Silver. Deep reinforcement learning from self-play in imperfect-information", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 499, + 415, + 512 + ], + "spans": [ + { + "bbox": [ + 126, + 499, + 415, + 512 + ], + "score": 1.0, + "content": "games. In NIPS Deep Reinforcement Learning (DRL) Workshop, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "[28] J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 525, + 380, + 538 + ], + "spans": [ + { + "bbox": [ + 126, + 525, + 380, + 538 + ], + "score": 1.0, + "content": "International Conference on Machine Learning (ICML), 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "[29] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 552, + 202, + 564 + ], + "spans": [ + { + "bbox": [ + 127, + 552, + 202, + 564 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 566, + 507, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 507, + 580 + ], + "score": 1.0, + "content": "[30] M. Hoffman, B. Shahriari, J. Aslanides, G. Barth-Maron, F. Behbahani, T. Norman, A. Abdol-", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 577, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 507, + 591 + ], + "score": 1.0, + "content": "maleki, A. Cassirer, F. Yang, K. Baumli, S. Henderson, A. Novikov, S. G. Colmenarejo, S. Cabi,", + "type": "text", + "cross_page": true + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 588, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 127, + 588, + 506, + 603 + ], + "score": 1.0, + "content": "C. Gulcehre, T. L. Paine, A. Cowie, Z. Wang, B. Piot, and N. de Freitas. Acme: A research", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 599, + 492, + 613 + ], + "spans": [ + { + "bbox": [ + 126, + 599, + 492, + 613 + ], + "score": 1.0, + "content": "framework for distributed reinforcement learning. arXiv preprint arXiv:2006.00979, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "[31] H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “Other-Play” for zero-shot coordination. In", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 625, + 381, + 639 + ], + "spans": [ + { + "bbox": [ + 126, + 625, + 381, + 639 + ], + "score": 1.0, + "content": "International Conference on Machine Learning (ICML), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "[32] E. Hughes, J. Z. Leibo, M. G. Phillips, K. Tuyls, E. A. Duéñez-Guzmán, A. G. Castañeda,", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 651, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 126, + 651, + 506, + 666 + ], + "score": 1.0, + "content": "I. Dunning, T. Zhu, K. R. McKee, R. Koster, H. Roff, and T. Graepel. Inequity aversion", + "type": "text", + "cross_page": true + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 127, + 663, + 506, + 677 + ], + "score": 1.0, + "content": "improves cooperation in intertemporal social dilemmas. In Neural Information Processing", + "type": "text", + "cross_page": true + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 674, + 232, + 686 + ], + "spans": [ + { + "bbox": [ + 127, + 674, + 232, + 686 + ], + "score": 1.0, + "content": "Systems (NeurIPS), 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 688, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 703 + ], + "score": 1.0, + "content": "[33] B. Ibarz, J. Leike, T. Pohlen, G. Irving, S. Legg, and D. Amodei. Reward learning from human", + "type": "text", + "cross_page": true + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "preferences and demonstrations in Atari. In Neural Information Processing Systems (NeurIPS),", + "type": "text", + "cross_page": true + } + ], + "index": 50 + }, + { + "bbox": [ + 126, + 709, + 155, + 725 + ], + "spans": [ + { + "bbox": [ + 126, + 709, + 155, + 725 + ], + "score": 1.0, + "content": "2018.", + "type": "text", + "cross_page": true + } + ], + "index": 51, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[34] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie,", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "score": 1.0, + "content": "N. C. Rabinowitz, A. S. Morcos, A. Ruderman, et al. Human-level performance in 3D mul-", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 95, + 507, + 109 + ], + "spans": [ + { + "bbox": [ + 127, + 95, + 507, + 109 + ], + "score": 1.0, + "content": "tiplayer games with population-based reinforcement learning. Science, 364(6443):859–865,", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 104, + 154, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 104, + 154, + 119 + ], + "score": 1.0, + "content": "2019.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 120, + 507, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 507, + 134 + ], + "score": 1.0, + "content": "[35] S. Javdani, S. S. Srinivasa, and J. A. Bagnell. Shared autonomy via hindsight optimization.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 131, + 277, + 145 + ], + "spans": [ + { + "bbox": [ + 127, + 131, + 277, + 145 + ], + "score": 1.0, + "content": "Robotics Science and Systems, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 147, + 507, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 507, + 159 + ], + "score": 1.0, + "content": "[36] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, and J. B. Tenenbaum. Co-", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 157, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 127, + 157, + 505, + 171 + ], + "score": 1.0, + "content": "ordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 168, + 371, + 182 + ], + "spans": [ + { + "bbox": [ + 127, + 168, + 371, + 182 + ], + "score": 1.0, + "content": "Conference of the Cognitive Science Society (CogSci), 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 183, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 506, + 197 + ], + "score": 1.0, + "content": "[37] G. Klien, D. D. Woods, J. M. Bradshaw, R. R. Hoffman, and P. J. Feltovich. Ten challenges for", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 195, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 127, + 195, + 505, + 209 + ], + "score": 1.0, + "content": "making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 207, + 197, + 218 + ], + "spans": [ + { + "bbox": [ + 127, + 207, + 197, + 218 + ], + "score": 1.0, + "content": "(6):91–95, 2004.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 219, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 236 + ], + "score": 1.0, + "content": "[38] P. Knott, M. Carroll, S. Devlin, K. Ciosek, K. Hofmann, A. Dragan, and R. Shah. Evaluating", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 231, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 506, + 245 + ], + "score": 1.0, + "content": "the robustness of collaborative agents. In International Conference on Autonomous Agents and", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 244, + 276, + 255 + ], + "spans": [ + { + "bbox": [ + 127, + 244, + 276, + 255 + ], + "score": 1.0, + "content": "Multiagent Systems (AAMAS), 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "score": 1.0, + "content": "[39] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 269, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 127, + 269, + 506, + 282 + ], + "score": 1.0, + "content": "T. Graepel. A unified game-theoretic approach to multiagent reinforcement learning. In Neural", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 280, + 317, + 294 + ], + "spans": [ + { + "bbox": [ + 127, + 280, + 317, + 294 + ], + "score": 1.0, + "content": "Information Processing Systems (NIPS), 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 296, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 107, + 296, + 505, + 307 + ], + "score": 1.0, + "content": "[40] A. Lerer and A. Peysakhovich. Learning existing social conventions via observationally", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 126, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "augmented self-play. In AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 316, + 185, + 330 + ], + "spans": [ + { + "bbox": [ + 126, + 316, + 185, + 330 + ], + "score": 1.0, + "content": "(AIES), 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 330, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 506, + 347 + ], + "score": 1.0, + "content": "[41] E. Lockhart, N. Burch, N. Bard, S. Borgeaud, T. Eccles, L. Smaira, and R. Smith. Human-agent", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 342, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 125, + 342, + 416, + 357 + ], + "score": 1.0, + "content": "cooperation in bridge bidding. arXiv preprint arXiv:2011.14124, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 357, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 357, + 506, + 372 + ], + "score": 1.0, + "content": "[42] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch. Multi-agent actor-critic", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 368, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 126, + 368, + 505, + 383 + ], + "score": 1.0, + "content": "for mixed cooperative-competitive environments. In Neural Information Processing Systems", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 380, + 186, + 394 + ], + "spans": [ + { + "bbox": [ + 126, + 380, + 186, + 394 + ], + "score": 1.0, + "content": "(NIPS), 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "[43] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. Learning to learn to communicate. In", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 406, + 352, + 419 + ], + "spans": [ + { + "bbox": [ + 126, + 406, + 352, + 419 + ], + "score": 1.0, + "content": "ICML Adaptive & Multitask Learning Workshop, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 421, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 507, + 435 + ], + "score": 1.0, + "content": "[44] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. On the interaction between super-", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 127, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "vision and self-play in emergent communication. In International Conference on Learning", + "type": "text", + "cross_page": true + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 443, + 254, + 457 + ], + "spans": [ + { + "bbox": [ + 127, + 443, + 254, + 457 + ], + "score": 1.0, + "content": "Representations (ICLR), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "[45] A. Lupu, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 468, + 434, + 482 + ], + "spans": [ + { + "bbox": [ + 126, + 468, + 434, + 482 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "[46] A. Mahajan, T. Rashid, M. Samvelyan, and S. Whiteson. Maven: Multi-agent variational", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 495, + 421, + 508 + ], + "spans": [ + { + "bbox": [ + 126, + 495, + 421, + 508 + ], + "score": 1.0, + "content": "exploration. In Neural Information Processing Systems (NeurIPS), 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 509, + 507, + 525 + ], + "spans": [ + { + "bbox": [ + 104, + 509, + 507, + 525 + ], + "score": 1.0, + "content": "[47] K. R. McKee, I. Gemp, B. McWilliams, E. A. Duèñez-Guzmán, E. Hughes, and J. Z. Leibo.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 522, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 127, + 522, + 507, + 536 + ], + "score": 1.0, + "content": "Social diversity and social preferences in mixed-motive reinforcement learning. In International", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 532, + 435, + 545 + ], + "spans": [ + { + "bbox": [ + 126, + 532, + 435, + 545 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 546, + 507, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 507, + 562 + ], + "score": 1.0, + "content": "[48] K. R. McKee, X. Bai, and S. Fiske. Understanding human impressions of artificial intelligence.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 557, + 196, + 572 + ], + "spans": [ + { + "bbox": [ + 126, + 557, + 196, + 572 + ], + "score": 1.0, + "content": "PsyArXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "[49] K. R. McKee, E. Hughes, T. O. Zhu, M. J. Chadwick, R. Koster, A. G. Castaneda, C. Beattie,", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 584, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 127, + 584, + 505, + 598 + ], + "score": 1.0, + "content": "T. Graepel, M. Botvinick, and J. Z. Leibo. Deep reinforcement learning models the emergent", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 128, + 596, + 425, + 609 + ], + "spans": [ + { + "bbox": [ + 128, + 596, + 425, + 609 + ], + "score": 1.0, + "content": "dynamics of human cooperation. arXiv preprint arXiv:2103.04982, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "[50] K. R. McKee, J. Z. Leibo, C. Beattie, and R. Everett. Quantifying environment and population", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 621, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 482, + 635 + ], + "score": 1.0, + "content": "diversity in multi-agent reinforcement learning. arXiv preprint arXiv:2102.08370, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "score": 1.0, + "content": "[51] F. Melo and A. Sardinha. Ad hoc teamwork by learning teammates’ task. International", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 648, + 441, + 662 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 441, + 662 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 46, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 662, + 507, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 507, + 677 + ], + "score": 1.0, + "content": "[52] B. Mutlu, A. Terrell, and C.-M. Huang. Coordination mechanisms in human-robot collaboration.", + "type": "text", + "cross_page": true + } + ], + "index": 47, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 673, + 343, + 687 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 343, + 687 + ], + "score": 1.0, + "content": "In HRI Collaborative Manipulation Workshop, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 48, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 687, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 704 + ], + "score": 1.0, + "content": "[53] S. Nikolaidis and J. Shah. Human-robot cross-training: Computational formulation, modeling", + "type": "text", + "cross_page": true + } + ], + "index": 49, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 699, + 507, + 715 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 507, + 715 + ], + "score": 1.0, + "content": "and evaluation of a human team training strategy. In ACM/IEEE International Conference on", + "type": "text", + "cross_page": true + } + ], + "index": 50 + }, + { + "bbox": [ + 127, + 711, + 288, + 724 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 288, + 724 + ], + "score": 1.0, + "content": "Human-Robot Interaction (HRI), 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 51, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[54] OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino,", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "score": 1.0, + "content": "M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng,", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "Q. Yuan, W. Zaremba, and L. Zhang. Solving Rubik’s Cube with a robot hand. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 105, + 232, + 118 + ], + "spans": [ + { + "bbox": [ + 126, + 105, + 232, + 118 + ], + "score": 1.0, + "content": "arXiv:1910.07113, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 119, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 505, + 134 + ], + "score": 1.0, + "content": "[55] E. Peer, L. Brandimarte, S. Samat, and A. Acquisti. Beyond the Turk: Alternative platforms for", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 132, + 507, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 507, + 146 + ], + "score": 1.0, + "content": "crowdsourcing behavioral research. Journal of Experimental Social Psychology, 70:153–163,", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 142, + 155, + 157 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 155, + 157 + ], + "score": 1.0, + "content": "2017.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 158, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 107, + 158, + 505, + 170 + ], + "score": 1.0, + "content": "[56] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 168, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 126, + 168, + 506, + 183 + ], + "score": 1.0, + "content": "control with dynamics randomization. In IEEE International Conference on Robotics and", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 180, + 236, + 194 + ], + "spans": [ + { + "bbox": [ + 125, + 180, + 236, + 194 + ], + "score": 1.0, + "content": "Automation (ICRA), 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 195, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 506, + 210 + ], + "score": 1.0, + "content": "[57] P. M. Pilarski, A. Butcher, M. Johanson, M. M. Botvinick, A. Bolt, and A. S. Parker. Learned", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 207, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 126, + 207, + 507, + 221 + ], + "score": 1.0, + "content": "human-agent decision-making, communication and joint action in a virtual reality environment.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 216, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 506, + 231 + ], + "score": 1.0, + "content": "In Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM),", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 227, + 155, + 243 + ], + "spans": [ + { + "bbox": [ + 126, + 227, + 155, + 243 + ], + "score": 1.0, + "content": "2019.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "[58] D. A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 256, + 289, + 268 + ], + "spans": [ + { + "bbox": [ + 127, + 256, + 289, + 268 + ], + "score": 1.0, + "content": "Neural Computation, 3(1):88–97, 1991.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 268, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 285 + ], + "score": 1.0, + "content": "[59] S. Russell. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin,", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 279, + 155, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 279, + 155, + 295 + ], + "score": 1.0, + "content": "2019.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "[60] D. Sadigh, N. Landolfi, S. S. Sastry, S. A. Seshia, and A. D. Dragan. Planning for cars that", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 127, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "coordinate with people: leveraging effects on human actions for planning and active information", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 319, + 459, + 332 + ], + "spans": [ + { + "bbox": [ + 126, + 319, + 459, + 332 + ], + "score": 1.0, + "content": "gathering over human internal state. Autonomous Robots, 42(7):1405–1426, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 334, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 506, + 348 + ], + "score": 1.0, + "content": "[61] N. Schurr, J. Marecki, M. Tambe, and P. Scerri. Towards flexible coordination of human-agent", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 345, + 348, + 358 + ], + "spans": [ + { + "bbox": [ + 126, + 345, + 348, + 358 + ], + "score": 1.0, + "content": "teams. Multiagent and Grid Systems, 1(1):3–16, 2005.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 107, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "[62] A. Shih, A. Sawhney, J. Kondic, S. Ermon, and D. Sadigh. On the critical role of conventions in", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "adaptive human-AI collaboration. In International Conference on Learning Representations", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 382, + 187, + 397 + ], + "spans": [ + { + "bbox": [ + 126, + 382, + 187, + 397 + ], + "score": 1.0, + "content": "(ICLR), 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 398, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 506, + 412 + ], + "score": 1.0, + "content": "[63] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 409, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 126, + 409, + 505, + 423 + ], + "score": 1.0, + "content": "M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and", + "type": "text", + "cross_page": true + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 418, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 125, + 418, + 507, + 435 + ], + "score": 1.0, + "content": "D. Hassabis. Mastering the game of go without human knowledge. Nature, 550(7676):354–359,", + "type": "text", + "cross_page": true + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 428, + 156, + 446 + ], + "spans": [ + { + "bbox": [ + 126, + 428, + 156, + 446 + ], + "score": 1.0, + "content": "2017.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "[64] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre,", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 127, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 466, + 507, + 484 + ], + "spans": [ + { + "bbox": [ + 125, + 466, + 507, + 484 + ], + "score": 1.0, + "content": "learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):", + "type": "text", + "cross_page": true + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 480, + 202, + 492 + ], + "spans": [ + { + "bbox": [ + 127, + 480, + 202, + 492 + ], + "score": 1.0, + "content": "1140–1144, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "[65] H. F. Song, A. Abdolmaleki, J. T. Springenberg, A. Clark, H. Soyer, J. W. Rae, S. Noury,", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 505, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 126, + 505, + 507, + 519 + ], + "score": 1.0, + "content": "A. Ahuja, S. Liu, D. Tirumala, N. Heess, D. Belov, M. Riedmiller, and M. M. Botvinick. V-", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 125, + 515, + 507, + 532 + ], + "spans": [ + { + "bbox": [ + 125, + 515, + 507, + 532 + ], + "score": 1.0, + "content": "MPO: On-policy maximum a posteriori policy optimization for discrete and continuous control.", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 527, + 418, + 542 + ], + "spans": [ + { + "bbox": [ + 126, + 527, + 418, + 542 + ], + "score": 1.0, + "content": "In International Conference on Learning Representations (ICLR), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 543, + 507, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 507, + 557 + ], + "score": 1.0, + "content": "[66] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Collabo-", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 125, + 554, + 470, + 568 + ], + "spans": [ + { + "bbox": [ + 125, + 554, + 470, + 568 + ], + "score": 1.0, + "content": "ration without pre-coordination. In AAAI Conference on Artificial Intelligence, 2010.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "[67] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 128, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 128, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "transferring deep neural networks from simulation to the real world. In IEEE/RSJ International", + "type": "text", + "cross_page": true + } + ], + "index": 41 + }, + { + "bbox": [ + 128, + 592, + 370, + 604 + ], + "spans": [ + { + "bbox": [ + 128, + 592, + 370, + 604 + ], + "score": 1.0, + "content": "Conference on Intelligent Robots and Systems (IROS), 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "[68] P. Tylkin, G. Radanovic, and D. C. Parkes. Learning robust helpful behaviors in two-player", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 127, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "cooperative Atari environments. In International Conference on Autonomous Agents and", + "type": "text", + "cross_page": true + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 629, + 277, + 643 + ], + "spans": [ + { + "bbox": [ + 126, + 629, + 277, + 643 + ], + "score": 1.0, + "content": "Multiagent Systems (AAMAS), 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 45, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 644, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 507, + 658 + ], + "score": 1.0, + "content": "[69] O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi,", + "type": "text", + "cross_page": true + } + ], + "index": 46, + "is_list_start_line": true + }, + { + "bbox": [ + 127, + 655, + 507, + 669 + ], + "spans": [ + { + "bbox": [ + 127, + 655, + 507, + 669 + ], + "score": 1.0, + "content": "R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre,", + "type": "text", + "cross_page": true + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 667, + 507, + 681 + ], + "spans": [ + { + "bbox": [ + 127, + 667, + 507, + 681 + ], + "score": 1.0, + "content": "T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V. Dalibard,", + "type": "text", + "cross_page": true + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 676, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 125, + 676, + 507, + 692 + ], + "score": 1.0, + "content": "D. Budden, Y. Sulsky, J. Molloy, T. L. Paine, C. Gulcehre, Z. Wang, T. Pfaff, Y. Wu, R. Ring,", + "type": "text", + "cross_page": true + } + ], + "index": 49 + }, + { + "bbox": [ + 127, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 127, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "D. Yogatama, D. Wünsch, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, K. Kavukcuoglu,", + "type": "text", + "cross_page": true + } + ], + "index": 50 + }, + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "D. Hassabis, C. Apps, and D. Silver. Grandmaster level in StarCraft II using multi-agent", + "type": "text", + "cross_page": true + } + ], + "index": 51 + }, + { + "bbox": [ + 127, + 711, + 366, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 366, + 723 + ], + "score": 1.0, + "content": "reinforcement learning. Nature, 575(7782):350–354, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 52, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "[70] R. E. Wang, S. A. Wu, J. A. Evans, J. B. Tenenbaum, D. C. Parkes, and M. Kleiman-Weiner. Too", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "many cooks: Coordinating multi-agent collaboration through inverse planning. In International", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 93, + 437, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 437, + 108 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 109, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 505, + 122 + ], + "score": 1.0, + "content": "[71] F. Wu, S. Zilberstein, and X. Chen. Online planning for ad hoc autonomous agent teams. In", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 120, + 414, + 132 + ], + "spans": [ + { + "bbox": [ + 126, + 120, + 414, + 132 + ], + "score": 1.0, + "content": "International Joint Conference on Artificial Intelligence (IJCAI), 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 135, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 505, + 148 + ], + "score": 1.0, + "content": "[72] S. Zheng, A. Trott, S. Srinivasa, N. Naik, M. Gruesbeck, D. C. Parkes, and R. Socher. The AI", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "economist: Improving equality and productivity with AI-driven tax policies. arXiv preprint", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 158, + 231, + 169 + ], + "spans": [ + { + "bbox": [ + 127, + 158, + 231, + 169 + ], + "score": 1.0, + "content": "arXiv:2004.13332, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + } + ], + "index": 32, + "bbox_fs": [ + 105, + 343, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 40, + 507, + 727 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 505, + 85 + ], + "score": 1.0, + "content": "[14] R. Choudhury, G. Swamy, D. Hadfield-Menell, and A. D. Dragan. On the utility of model", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 82, + 507, + 98 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 507, + 98 + ], + "score": 1.0, + "content": "learning in HRI. In ACM/IEEE International Conference on Human-Robot Interaction (HRI),", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 126, + 92, + 155, + 109 + ], + "spans": [ + { + "bbox": [ + 126, + 92, + 155, + 109 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 109, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 505, + 123 + ], + "score": 1.0, + "content": "[15] P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 120, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 126, + 120, + 506, + 135 + ], + "score": 1.0, + "content": "learning from human preferences. In Proceedings of the 31st International Conference on", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 131, + 392, + 145 + ], + "spans": [ + { + "bbox": [ + 126, + 131, + 392, + 145 + ], + "score": 1.0, + "content": "Neural Information Processing Systems, pages 4302–4310, 2017.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "[16] A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 126, + 157, + 477, + 172 + ], + "spans": [ + { + "bbox": [ + 126, + 157, + 477, + 172 + ], + "score": 1.0, + "content": "T. Graepel. Open problems in cooperative AI. arXiv preprint arXiv:2012.08630, 2020.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 172, + 507, + 187 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 507, + 187 + ], + "score": 1.0, + "content": "[17] L. Espeholt, H. Soyer, R. Munos, K. Simonyan, V. Mnih, T. Ward, Y. Doron, V. Firoiu,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 182, + 507, + 198 + ], + "spans": [ + { + "bbox": [ + 125, + 182, + 507, + 198 + ], + "score": 1.0, + "content": "T. Harley, I. Dunning, S. Legg, and K. Kavukcuoglu. IMPALA: Scalable distributed deep-RL", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 126, + 194, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 506, + 209 + ], + "score": 1.0, + "content": "with importance weighted actor-learner architectures. In International Conference on Machine", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 206, + 226, + 218 + ], + "spans": [ + { + "bbox": [ + 127, + 206, + 226, + 218 + ], + "score": 1.0, + "content": "Learning (ICML), 2018.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 219, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 234 + ], + "score": 1.0, + "content": "[18] P. Eyal, R. David, G. Andrew, E. Zak, and D. Ekaterina. Data quality of platforms and panels", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 231, + 447, + 245 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 447, + 245 + ], + "score": 1.0, + "content": "for online behavioral research. Behavior Research Methods, pages 1–20, 2021.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "[19] B. Eysenbach, A. Gupta, J. Ibarz, and S. Levine. Diversity is all you need: Learning skills", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 126, + 258, + 507, + 272 + ], + "spans": [ + { + "bbox": [ + 126, + 258, + 507, + 272 + ], + "score": 1.0, + "content": "without a reward function. In International Conference on Learning Representations (ICLR),", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 267, + 155, + 283 + ], + "spans": [ + { + "bbox": [ + 127, + 267, + 155, + 283 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 282, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 298 + ], + "score": 1.0, + "content": "[20] E. Fehr and K. M. Schmidt. A theory of fairness, competition, and cooperation. The Quarterly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 126, + 294, + 314, + 309 + ], + "spans": [ + { + "bbox": [ + 126, + 294, + 314, + 309 + ], + "score": 1.0, + "content": "Journal of Economics, 114(3):817–868, 1999.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 309, + 507, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 507, + 324 + ], + "score": 1.0, + "content": "[21] J. Foerster, F. Song, E. Hughes, N. Burch, I. Dunning, S. Whiteson, M. Botvinick, and M. Bowl-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 126, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "ing. Bayesian action decoder for deep multi-agent reinforcement learning. In International", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 126, + 331, + 325, + 345 + ], + "spans": [ + { + "bbox": [ + 126, + 331, + 325, + 345 + ], + "score": 1.0, + "content": "Conference on Machine Learning (ICML), 2019.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 347, + 507, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 507, + 360 + ], + "score": 1.0, + "content": "[22] M. Fontaine, Y.-C. Hsu, Y. Zhang, B. Tjanaka, and S. Nikolaidis. On the importance of", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 128, + 358, + 496, + 370 + ], + "spans": [ + { + "bbox": [ + 128, + 358, + 496, + 370 + ], + "score": 1.0, + "content": "environments in human-robot coordination. In Robotics: Science and Systems (RSS), 2021.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 374, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 506, + 385 + ], + "score": 1.0, + "content": "[23] I. Gabriel. Artificial intelligence, values, and alignment. Minds and Machines, 30(3):411–437,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 127, + 384, + 154, + 397 + ], + "spans": [ + { + "bbox": [ + 127, + 384, + 154, + 397 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 398, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 506, + 413 + ], + "score": 1.0, + "content": "[24] M. Garnelo, W. M. Czarnecki, S. Liu, D. Tirumala, J. Oh, G. Gidel, H. van Hasselt, and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 410, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 126, + 410, + 505, + 424 + ], + "score": 1.0, + "content": "D. Balduzzi. Pick your battles: Interaction graphs as population-level objectives for strategic", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 126, + 419, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 126, + 419, + 507, + 435 + ], + "score": 1.0, + "content": "diversity. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS),", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 431, + 155, + 446 + ], + "spans": [ + { + "bbox": [ + 126, + 431, + 155, + 446 + ], + "score": 1.0, + "content": "2021.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 447, + 291, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 291, + 460 + ], + "score": 1.0, + "content": "[25] Ghost Town Games. Overcooked, 2016.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "[26] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In IEEE", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 473, + 419, + 487 + ], + "spans": [ + { + "bbox": [ + 126, + 473, + 419, + 487 + ], + "score": 1.0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), 2016.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 488, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 506, + 502 + ], + "score": 1.0, + "content": "[27] J. Heinrich and D. Silver. Deep reinforcement learning from self-play in imperfect-information", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 499, + 415, + 512 + ], + "spans": [ + { + "bbox": [ + 126, + 499, + 415, + 512 + ], + "score": 1.0, + "content": "games. In NIPS Deep Reinforcement Learning (DRL) Workshop, 2016.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "[28] J. Heinrich, M. Lanctot, and D. Silver. Fictitious self-play in extensive-form games. In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 126, + 525, + 380, + 538 + ], + "spans": [ + { + "bbox": [ + 126, + 525, + 380, + 538 + ], + "score": 1.0, + "content": "International Conference on Machine Learning (ICML), 2015.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 541, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 506, + 554 + ], + "score": 1.0, + "content": "[29] S. Hochreiter and J. Schmidhuber. Long short-term memory. Neural Computation, 9(8):", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 127, + 552, + 202, + 564 + ], + "spans": [ + { + "bbox": [ + 127, + 552, + 202, + 564 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 566, + 507, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 507, + 580 + ], + "score": 1.0, + "content": "[30] M. Hoffman, B. Shahriari, J. Aslanides, G. Barth-Maron, F. Behbahani, T. Norman, A. Abdol-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 127, + 577, + 507, + 591 + ], + "spans": [ + { + "bbox": [ + 127, + 577, + 507, + 591 + ], + "score": 1.0, + "content": "maleki, A. Cassirer, F. Yang, K. Baumli, S. Henderson, A. Novikov, S. G. Colmenarejo, S. Cabi,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 588, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 127, + 588, + 506, + 603 + ], + "score": 1.0, + "content": "C. Gulcehre, T. L. Paine, A. Cowie, Z. Wang, B. Piot, and N. de Freitas. Acme: A research", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 126, + 599, + 492, + 613 + ], + "spans": [ + { + "bbox": [ + 126, + 599, + 492, + 613 + ], + "score": 1.0, + "content": "framework for distributed reinforcement learning. arXiv preprint arXiv:2006.00979, 2020.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "[31] H. Hu, A. Lerer, A. Peysakhovich, and J. Foerster. “Other-Play” for zero-shot coordination. In", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 126, + 625, + 381, + 639 + ], + "spans": [ + { + "bbox": [ + 126, + 625, + 381, + 639 + ], + "score": 1.0, + "content": "International Conference on Machine Learning (ICML), 2020.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "[32] E. Hughes, J. Z. Leibo, M. G. Phillips, K. Tuyls, E. A. Duéñez-Guzmán, A. G. Castañeda,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 126, + 651, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 126, + 651, + 506, + 666 + ], + "score": 1.0, + "content": "I. Dunning, T. Zhu, K. R. McKee, R. Koster, H. Roff, and T. Graepel. Inequity aversion", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 663, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 127, + 663, + 506, + 677 + ], + "score": 1.0, + "content": "improves cooperation in intertemporal social dilemmas. In Neural Information Processing", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 674, + 232, + 686 + ], + "spans": [ + { + "bbox": [ + 127, + 674, + 232, + 686 + ], + "score": 1.0, + "content": "Systems (NeurIPS), 2018.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 688, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 506, + 703 + ], + "score": 1.0, + "content": "[33] B. Ibarz, J. Leike, T. Pohlen, G. Irving, S. Legg, and D. Amodei. Reward learning from human", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "preferences and demonstrations in Atari. In Neural Information Processing Systems (NeurIPS),", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 126, + 709, + 155, + 725 + ], + "spans": [ + { + "bbox": [ + 126, + 709, + 155, + 725 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 25.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 40, + 507, + 727 + ], + "lines": [], + "index": 25.5, + "bbox_fs": [ + 105, + 73, + 507, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 59, + 507, + 725 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[34] M. Jaderberg, W. M. Czarnecki, I. Dunning, L. Marris, G. Lever, A. G. Castaneda, C. Beattie,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "spans": [ + { + "bbox": [ + 126, + 82, + 507, + 97 + ], + "score": 1.0, + "content": "N. C. Rabinowitz, A. S. Morcos, A. Ruderman, et al. Human-level performance in 3D mul-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 95, + 507, + 109 + ], + "spans": [ + { + "bbox": [ + 127, + 95, + 507, + 109 + ], + "score": 1.0, + "content": "tiplayer games with population-based reinforcement learning. Science, 364(6443):859–865,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 104, + 154, + 119 + ], + "spans": [ + { + "bbox": [ + 126, + 104, + 154, + 119 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 120, + 507, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 507, + 134 + ], + "score": 1.0, + "content": "[35] S. Javdani, S. S. Srinivasa, and J. A. Bagnell. Shared autonomy via hindsight optimization.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 131, + 277, + 145 + ], + "spans": [ + { + "bbox": [ + 127, + 131, + 277, + 145 + ], + "score": 1.0, + "content": "Robotics Science and Systems, 2015.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 507, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 507, + 159 + ], + "score": 1.0, + "content": "[36] M. Kleiman-Weiner, M. K. Ho, J. L. Austerweil, M. L. Littman, and J. B. Tenenbaum. Co-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 157, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 127, + 157, + 505, + 171 + ], + "score": 1.0, + "content": "ordinate to cooperate or compete: abstract goals and joint intentions in social interaction. In", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 168, + 371, + 182 + ], + "spans": [ + { + "bbox": [ + 127, + 168, + 371, + 182 + ], + "score": 1.0, + "content": "Conference of the Cognitive Science Society (CogSci), 2016.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 183, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 506, + 197 + ], + "score": 1.0, + "content": "[37] G. Klien, D. D. Woods, J. M. Bradshaw, R. R. Hoffman, and P. J. Feltovich. Ten challenges for", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 195, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 127, + 195, + 505, + 209 + ], + "score": 1.0, + "content": "making automation a “team player” in joint human-agent activity. IEEE Intelligent Systems, 19", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 127, + 207, + 197, + 218 + ], + "spans": [ + { + "bbox": [ + 127, + 207, + 197, + 218 + ], + "score": 1.0, + "content": "(6):91–95, 2004.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 219, + 506, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 506, + 236 + ], + "score": 1.0, + "content": "[38] P. Knott, M. Carroll, S. Devlin, K. Ciosek, K. Hofmann, A. Dragan, and R. Shah. Evaluating", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 231, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 126, + 231, + 506, + 245 + ], + "score": 1.0, + "content": "the robustness of collaborative agents. In International Conference on Autonomous Agents and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 127, + 244, + 276, + 255 + ], + "spans": [ + { + "bbox": [ + 127, + 244, + 276, + 255 + ], + "score": 1.0, + "content": "Multiagent Systems (AAMAS), 2021.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 506, + 272 + ], + "score": 1.0, + "content": "[39] M. Lanctot, V. Zambaldi, A. Gruslys, A. Lazaridou, K. Tuyls, J. Pérolat, D. Silver, and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 127, + 269, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 127, + 269, + 506, + 282 + ], + "score": 1.0, + "content": "T. Graepel. A unified game-theoretic approach to multiagent reinforcement learning. In Neural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 280, + 317, + 294 + ], + "spans": [ + { + "bbox": [ + 127, + 280, + 317, + 294 + ], + "score": 1.0, + "content": "Information Processing Systems (NIPS), 2017.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 296, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 107, + 296, + 505, + 307 + ], + "score": 1.0, + "content": "[40] A. Lerer and A. Peysakhovich. Learning existing social conventions via observationally", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 126, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 126, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "augmented self-play. In AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 316, + 185, + 330 + ], + "spans": [ + { + "bbox": [ + 126, + 316, + 185, + 330 + ], + "score": 1.0, + "content": "(AIES), 2019.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 330, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 330, + 506, + 347 + ], + "score": 1.0, + "content": "[41] E. Lockhart, N. Burch, N. Bard, S. Borgeaud, T. Eccles, L. Smaira, and R. Smith. Human-agent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 125, + 342, + 416, + 357 + ], + "spans": [ + { + "bbox": [ + 125, + 342, + 416, + 357 + ], + "score": 1.0, + "content": "cooperation in bridge bidding. arXiv preprint arXiv:2011.14124, 2020.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 357, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 104, + 357, + 506, + 372 + ], + "score": 1.0, + "content": "[42] R. Lowe, Y. Wu, A. Tamar, J. Harb, P. Abbeel, and I. Mordatch. Multi-agent actor-critic", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 126, + 368, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 126, + 368, + 505, + 383 + ], + "score": 1.0, + "content": "for mixed cooperative-competitive environments. In Neural Information Processing Systems", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 380, + 186, + 394 + ], + "spans": [ + { + "bbox": [ + 126, + 380, + 186, + 394 + ], + "score": 1.0, + "content": "(NIPS), 2017.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 408 + ], + "score": 1.0, + "content": "[43] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. Learning to learn to communicate. In", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 406, + 352, + 419 + ], + "spans": [ + { + "bbox": [ + 126, + 406, + 352, + 419 + ], + "score": 1.0, + "content": "ICML Adaptive & Multitask Learning Workshop, 2019.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 421, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 507, + 435 + ], + "score": 1.0, + "content": "[44] R. Lowe, A. Gupta, J. Foerster, D. Kiela, and J. Pineau. On the interaction between super-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 127, + 431, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 127, + 431, + 505, + 446 + ], + "score": 1.0, + "content": "vision and self-play in emergent communication. In International Conference on Learning", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 127, + 443, + 254, + 457 + ], + "spans": [ + { + "bbox": [ + 127, + 443, + 254, + 457 + ], + "score": 1.0, + "content": "Representations (ICLR), 2020.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "[45] A. Lupu, H. Hu, and J. Foerster. Trajectory diversity for zero-shot coordination. In International", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 126, + 468, + 434, + 482 + ], + "spans": [ + { + "bbox": [ + 126, + 468, + 434, + 482 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "[46] A. Mahajan, T. Rashid, M. Samvelyan, and S. Whiteson. Maven: Multi-agent variational", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 126, + 495, + 421, + 508 + ], + "spans": [ + { + "bbox": [ + 126, + 495, + 421, + 508 + ], + "score": 1.0, + "content": "exploration. In Neural Information Processing Systems (NeurIPS), 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 509, + 507, + 525 + ], + "spans": [ + { + "bbox": [ + 104, + 509, + 507, + 525 + ], + "score": 1.0, + "content": "[47] K. R. McKee, I. Gemp, B. McWilliams, E. A. Duèñez-Guzmán, E. Hughes, and J. Z. Leibo.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 127, + 522, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 127, + 522, + 507, + 536 + ], + "score": 1.0, + "content": "Social diversity and social preferences in mixed-motive reinforcement learning. In International", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 532, + 435, + 545 + ], + "spans": [ + { + "bbox": [ + 126, + 532, + 435, + 545 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 546, + 507, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 507, + 562 + ], + "score": 1.0, + "content": "[48] K. R. McKee, X. Bai, and S. Fiske. Understanding human impressions of artificial intelligence.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 126, + 557, + 196, + 572 + ], + "spans": [ + { + "bbox": [ + 126, + 557, + 196, + 572 + ], + "score": 1.0, + "content": "PsyArXiv, 2021.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 506, + 587 + ], + "score": 1.0, + "content": "[49] K. R. McKee, E. Hughes, T. O. Zhu, M. J. Chadwick, R. Koster, A. G. Castaneda, C. Beattie,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 127, + 584, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 127, + 584, + 505, + 598 + ], + "score": 1.0, + "content": "T. Graepel, M. Botvinick, and J. Z. Leibo. Deep reinforcement learning models the emergent", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 128, + 596, + 425, + 609 + ], + "spans": [ + { + "bbox": [ + 128, + 596, + 425, + 609 + ], + "score": 1.0, + "content": "dynamics of human cooperation. arXiv preprint arXiv:2103.04982, 2021.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "[50] K. R. McKee, J. Z. Leibo, C. Beattie, and R. Everett. Quantifying environment and population", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 621, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 127, + 621, + 482, + 635 + ], + "score": 1.0, + "content": "diversity in multi-agent reinforcement learning. arXiv preprint arXiv:2102.08370, 2021.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "score": 1.0, + "content": "[51] F. Melo and A. Sardinha. Ad hoc teamwork by learning teammates’ task. International", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 127, + 648, + 441, + 662 + ], + "spans": [ + { + "bbox": [ + 127, + 648, + 441, + 662 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multi-Agent Systems (AAMAS), 2015.", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 662, + 507, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 507, + 677 + ], + "score": 1.0, + "content": "[52] B. Mutlu, A. Terrell, and C.-M. Huang. Coordination mechanisms in human-robot collaboration.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 126, + 673, + 343, + 687 + ], + "spans": [ + { + "bbox": [ + 126, + 673, + 343, + 687 + ], + "score": 1.0, + "content": "In HRI Collaborative Manipulation Workshop, 2013.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 687, + 506, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 704 + ], + "score": 1.0, + "content": "[53] S. Nikolaidis and J. Shah. Human-robot cross-training: Computational formulation, modeling", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 126, + 699, + 507, + 715 + ], + "spans": [ + { + "bbox": [ + 126, + 699, + 507, + 715 + ], + "score": 1.0, + "content": "and evaluation of a human team training strategy. In ACM/IEEE International Conference on", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 127, + 711, + 288, + 724 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 288, + 724 + ], + "score": 1.0, + "content": "Human-Robot Interaction (HRI), 2013.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 25.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 59, + 507, + 725 + ], + "lines": [], + "index": 25.5, + "bbox_fs": [ + 104, + 72, + 507, + 724 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 45, + 507, + 723 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 507, + 86 + ], + "score": 1.0, + "content": "[54] OpenAI, I. Akkaya, M. Andrychowicz, M. Chociej, M. Litwin, B. McGrew, A. Petron, A. Paino,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "spans": [ + { + "bbox": [ + 125, + 82, + 507, + 99 + ], + "score": 1.0, + "content": "M. Plappert, G. Powell, R. Ribas, J. Schneider, N. Tezak, J. Tworek, P. Welinder, L. Weng,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "Q. Yuan, W. Zaremba, and L. Zhang. Solving Rubik’s Cube with a robot hand. arXiv preprint", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 105, + 232, + 118 + ], + "spans": [ + { + "bbox": [ + 126, + 105, + 232, + 118 + ], + "score": 1.0, + "content": "arXiv:1910.07113, 2019.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 119, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 505, + 134 + ], + "score": 1.0, + "content": "[55] E. Peer, L. Brandimarte, S. Samat, and A. Acquisti. Beyond the Turk: Alternative platforms for", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 132, + 507, + 146 + ], + "spans": [ + { + "bbox": [ + 127, + 132, + 507, + 146 + ], + "score": 1.0, + "content": "crowdsourcing behavioral research. Journal of Experimental Social Psychology, 70:153–163,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 142, + 155, + 157 + ], + "spans": [ + { + "bbox": [ + 127, + 142, + 155, + 157 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 158, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 107, + 158, + 505, + 170 + ], + "score": 1.0, + "content": "[56] X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel. Sim-to-real transfer of robotic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 126, + 168, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 126, + 168, + 506, + 183 + ], + "score": 1.0, + "content": "control with dynamics randomization. In IEEE International Conference on Robotics and", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 125, + 180, + 236, + 194 + ], + "spans": [ + { + "bbox": [ + 125, + 180, + 236, + 194 + ], + "score": 1.0, + "content": "Automation (ICRA), 2018.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 195, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 506, + 210 + ], + "score": 1.0, + "content": "[57] P. M. Pilarski, A. Butcher, M. Johanson, M. M. Botvinick, A. Bolt, and A. S. Parker. Learned", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 207, + 507, + 221 + ], + "spans": [ + { + "bbox": [ + 126, + 207, + 507, + 221 + ], + "score": 1.0, + "content": "human-agent decision-making, communication and joint action in a virtual reality environment.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 126, + 216, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 126, + 216, + 506, + 231 + ], + "score": 1.0, + "content": "In Multidisciplinary Conference on Reinforcement Learning and Decision Making (RLDM),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 126, + 227, + 155, + 243 + ], + "spans": [ + { + "bbox": [ + 126, + 227, + 155, + 243 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 257 + ], + "score": 1.0, + "content": "[58] D. A. Pomerleau. Efficient training of artificial neural networks for autonomous navigation.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 127, + 256, + 289, + 268 + ], + "spans": [ + { + "bbox": [ + 127, + 256, + 289, + 268 + ], + "score": 1.0, + "content": "Neural Computation, 3(1):88–97, 1991.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 268, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 285 + ], + "score": 1.0, + "content": "[59] S. Russell. Human Compatible: Artificial Intelligence and the Problem of Control. Penguin,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 127, + 279, + 155, + 295 + ], + "spans": [ + { + "bbox": [ + 127, + 279, + 155, + 295 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "[60] D. Sadigh, N. Landolfi, S. S. Sastry, S. A. Seshia, and A. D. Dragan. Planning for cars that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 127, + 308, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 127, + 308, + 506, + 322 + ], + "score": 1.0, + "content": "coordinate with people: leveraging effects on human actions for planning and active information", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 126, + 319, + 459, + 332 + ], + "spans": [ + { + "bbox": [ + 126, + 319, + 459, + 332 + ], + "score": 1.0, + "content": "gathering over human internal state. Autonomous Robots, 42(7):1405–1426, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 334, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 334, + 506, + 348 + ], + "score": 1.0, + "content": "[61] N. Schurr, J. Marecki, M. Tambe, and P. Scerri. Towards flexible coordination of human-agent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 126, + 345, + 348, + 358 + ], + "spans": [ + { + "bbox": [ + 126, + 345, + 348, + 358 + ], + "score": 1.0, + "content": "teams. Multiagent and Grid Systems, 1(1):3–16, 2005.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 107, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 107, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "[62] A. Shih, A. Sawhney, J. Kondic, S. Ermon, and D. Sadigh. On the critical role of conventions in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 127, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "adaptive human-AI collaboration. In International Conference on Learning Representations", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 126, + 382, + 187, + 397 + ], + "spans": [ + { + "bbox": [ + 126, + 382, + 187, + 397 + ], + "score": 1.0, + "content": "(ICLR), 2021.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 398, + 506, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 506, + 412 + ], + "score": 1.0, + "content": "[63] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 126, + 409, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 126, + 409, + 505, + 423 + ], + "score": 1.0, + "content": "M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 125, + 418, + 507, + 435 + ], + "spans": [ + { + "bbox": [ + 125, + 418, + 507, + 435 + ], + "score": 1.0, + "content": "D. Hassabis. Mastering the game of go without human knowledge. Nature, 550(7676):354–359,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 126, + 428, + 156, + 446 + ], + "spans": [ + { + "bbox": [ + 126, + 428, + 156, + 446 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 506, + 460 + ], + "score": 1.0, + "content": "[64] D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 127, + 457, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 127, + 457, + 506, + 471 + ], + "score": 1.0, + "content": "D. Kumaran, T. Graepel, T. Lillicrap, K. Simonyan, and D. Hassabis. A general reinforcement", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 125, + 466, + 507, + 484 + ], + "spans": [ + { + "bbox": [ + 125, + 466, + 507, + 484 + ], + "score": 1.0, + "content": "learning algorithm that masters chess, shogi, and go through self-play. Science, 362(6419):", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 127, + 480, + 202, + 492 + ], + "spans": [ + { + "bbox": [ + 127, + 480, + 202, + 492 + ], + "score": 1.0, + "content": "1140–1144, 2018.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 507, + 509 + ], + "score": 1.0, + "content": "[65] H. F. Song, A. Abdolmaleki, J. T. Springenberg, A. Clark, H. Soyer, J. W. Rae, S. Noury,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 126, + 505, + 507, + 519 + ], + "spans": [ + { + "bbox": [ + 126, + 505, + 507, + 519 + ], + "score": 1.0, + "content": "A. Ahuja, S. Liu, D. Tirumala, N. Heess, D. Belov, M. Riedmiller, and M. M. Botvinick. V-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 125, + 515, + 507, + 532 + ], + "spans": [ + { + "bbox": [ + 125, + 515, + 507, + 532 + ], + "score": 1.0, + "content": "MPO: On-policy maximum a posteriori policy optimization for discrete and continuous control.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 126, + 527, + 418, + 542 + ], + "spans": [ + { + "bbox": [ + 126, + 527, + 418, + 542 + ], + "score": 1.0, + "content": "In International Conference on Learning Representations (ICLR), 2020.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 543, + 507, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 507, + 557 + ], + "score": 1.0, + "content": "[66] P. Stone, G. Kaminka, S. Kraus, and J. Rosenschein. Ad hoc autonomous agent teams: Collabo-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 125, + 554, + 470, + 568 + ], + "spans": [ + { + "bbox": [ + 125, + 554, + 470, + 568 + ], + "score": 1.0, + "content": "ration without pre-coordination. In AAAI Conference on Artificial Intelligence, 2010.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 583 + ], + "score": 1.0, + "content": "[67] J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, and P. Abbeel. Domain randomization for", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 128, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 128, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "transferring deep neural networks from simulation to the real world. In IEEE/RSJ International", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 128, + 592, + 370, + 604 + ], + "spans": [ + { + "bbox": [ + 128, + 592, + 370, + 604 + ], + "score": 1.0, + "content": "Conference on Intelligent Robots and Systems (IROS), 2017.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "[68] P. Tylkin, G. Radanovic, and D. C. Parkes. Learning robust helpful behaviors in two-player", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 127, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 127, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "cooperative Atari environments. In International Conference on Autonomous Agents and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 126, + 629, + 277, + 643 + ], + "spans": [ + { + "bbox": [ + 126, + 629, + 277, + 643 + ], + "score": 1.0, + "content": "Multiagent Systems (AAMAS), 2021.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 644, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 507, + 658 + ], + "score": 1.0, + "content": "[69] O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 127, + 655, + 507, + 669 + ], + "spans": [ + { + "bbox": [ + 127, + 655, + 507, + 669 + ], + "score": 1.0, + "content": "R. Powell, T. Ewalds, P. Georgiev, J. Oh, D. Horgan, M. Kroiss, I. Danihelka, A. Huang, L. Sifre,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 127, + 667, + 507, + 681 + ], + "spans": [ + { + "bbox": [ + 127, + 667, + 507, + 681 + ], + "score": 1.0, + "content": "T. Cai, J. P. Agapiou, M. Jaderberg, A. S. Vezhnevets, R. Leblond, T. Pohlen, V. Dalibard,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 125, + 676, + 507, + 692 + ], + "spans": [ + { + "bbox": [ + 125, + 676, + 507, + 692 + ], + "score": 1.0, + "content": "D. Budden, Y. Sulsky, J. Molloy, T. L. Paine, C. Gulcehre, Z. Wang, T. Pfaff, Y. Wu, R. Ring,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 127, + 688, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 127, + 688, + 506, + 702 + ], + "score": 1.0, + "content": "D. Yogatama, D. Wünsch, K. McKinney, O. Smith, T. Schaul, T. Lillicrap, K. Kavukcuoglu,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 127, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "D. Hassabis, C. Apps, and D. Silver. Grandmaster level in StarCraft II using multi-agent", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 127, + 711, + 366, + 723 + ], + "spans": [ + { + "bbox": [ + 127, + 711, + 366, + 723 + ], + "score": 1.0, + "content": "reinforcement learning. Nature, 575(7782):350–354, 2019.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 26 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 45, + 507, + 723 + ], + "lines": [], + "index": 26, + "bbox_fs": [ + 105, + 72, + 507, + 723 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 72, + 506, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "[70] R. E. Wang, S. A. Wu, J. A. Evans, J. B. Tenenbaum, D. C. Parkes, and M. Kleiman-Weiner. Too", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 126, + 83, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 126, + 83, + 505, + 96 + ], + "score": 1.0, + "content": "many cooks: Coordinating multi-agent collaboration through inverse planning. In International", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 93, + 437, + 108 + ], + "spans": [ + { + "bbox": [ + 127, + 93, + 437, + 108 + ], + "score": 1.0, + "content": "Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2020.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 109, + 505, + 122 + ], + "spans": [ + { + "bbox": [ + 106, + 109, + 505, + 122 + ], + "score": 1.0, + "content": "[71] F. Wu, S. Zilberstein, and X. Chen. Online planning for ad hoc autonomous agent teams. In", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 120, + 414, + 132 + ], + "spans": [ + { + "bbox": [ + 126, + 120, + 414, + 132 + ], + "score": 1.0, + "content": "International Joint Conference on Artificial Intelligence (IJCAI), 2011.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 135, + 505, + 148 + ], + "spans": [ + { + "bbox": [ + 106, + 135, + 505, + 148 + ], + "score": 1.0, + "content": "[72] S. Zheng, A. Trott, S. Srinivasa, N. Naik, M. Gruesbeck, D. C. Parkes, and R. Socher. The AI", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 126, + 146, + 506, + 160 + ], + "score": 1.0, + "content": "economist: Improving equality and productivity with AI-driven tax policies. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 158, + 231, + 169 + ], + "spans": [ + { + "bbox": [ + 127, + 158, + 231, + 169 + ], + "score": 1.0, + "content": "arXiv:2004.13332, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 3.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 72, + 506, + 169 + ], + "lines": [], + "index": 3.5, + "bbox_fs": [ + 106, + 72, + 506, + 169 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_model.json b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bccd37e0d2a134b898377798d9f8dedc79804b74 --- /dev/null +++ b/parse/train/1Kof-nkmQB8/1Kof-nkmQB8_model.json @@ -0,0 +1,17440 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 397, + 684, + 1305, + 684, + 1305, + 1294, + 397, + 1294 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1424, + 1405, + 1424, + 1405, + 1667, + 298, + 1667 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1681, + 1403, + 1681, + 1403, + 1926, + 298, + 1926 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 340, + 442, + 1360, + 442, + 1360, + 537, + 340, + 537 + ], + "score": 0.967 + }, + { + "category_id": 0, + "poly": [ + 300, + 1354, + 530, + 1354, + 530, + 1391, + 300, + 1391 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 337, + 270, + 1364, + 270, + 1364, + 327, + 337, + 327 + ], + "score": 0.901 + }, + { + "category_id": 0, + "poly": [ + 787, + 614, + 912, + 614, + 912, + 651, + 787, + 651 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 296, + 2033, + 1070, + 2033, + 1070, + 2062, + 296, + 2062 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 335, + 1952, + 535, + 1952, + 535, + 1979, + 335, + 1979 + ], + "score": 0.885 + }, + { + "category_id": 13, + "poly": [ + 363, + 1863, + 411, + 1863, + 411, + 1895, + 363, + 1895 + ], + "score": 0.65, + "latex": "\\pmb { \\| 3 9 \\| }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1895, + 423, + 1895, + 423, + 1926, + 374, + 1926 + ], + "score": 0.43, + "latex": "\\pmb { \\Vert 3 4 \\Vert }" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1454, + 1254, + 1454, + 1254, + 1486, + 1208, + 1486 + ], + "score": 0.26, + "latex": "\\boxed { 6 6 }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1350.0, + 536.0, + 1350.0, + 536.0, + 1397.0, + 292.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 270.0, + 1365.0, + 270.0, + 1365.0, + 332.0, + 337.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 613.0, + 919.0, + 613.0, + 919.0, + 655.0, + 783.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2028.0, + 1074.0, + 2028.0, + 1074.0, + 2066.0, + 294.0, + 2066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1946.0, + 540.0, + 1946.0, + 540.0, + 1983.0, + 331.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 685.0, + 1309.0, + 685.0, + 1309.0, + 723.0, + 394.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 716.0, + 1309.0, + 716.0, + 1309.0, + 752.0, + 394.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 747.0, + 1306.0, + 747.0, + 1306.0, + 783.0, + 393.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 778.0, + 1310.0, + 778.0, + 1310.0, + 813.0, + 392.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 808.0, + 1308.0, + 808.0, + 1308.0, + 841.0, + 394.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 839.0, + 1306.0, + 839.0, + 1306.0, + 872.0, + 393.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 869.0, + 1305.0, + 869.0, + 1305.0, + 903.0, + 393.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 897.0, + 1307.0, + 897.0, + 1307.0, + 934.0, + 393.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 931.0, + 1305.0, + 931.0, + 1305.0, + 964.0, + 395.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 960.0, + 1309.0, + 960.0, + 1309.0, + 993.0, + 394.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 988.0, + 1308.0, + 988.0, + 1308.0, + 1026.0, + 392.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1020.0, + 1306.0, + 1020.0, + 1306.0, + 1056.0, + 393.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1050.0, + 1306.0, + 1050.0, + 1306.0, + 1086.0, + 394.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1083.0, + 1306.0, + 1083.0, + 1306.0, + 1115.0, + 393.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1111.0, + 1306.0, + 1111.0, + 1306.0, + 1146.0, + 392.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1141.0, + 1307.0, + 1141.0, + 1307.0, + 1177.0, + 392.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1173.0, + 1309.0, + 1173.0, + 1309.0, + 1206.0, + 393.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1202.0, + 1305.0, + 1202.0, + 1305.0, + 1235.0, + 394.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1234.0, + 1306.0, + 1234.0, + 1306.0, + 1267.0, + 393.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1264.0, + 1259.0, + 1264.0, + 1259.0, + 1297.0, + 393.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1424.0, + 1403.0, + 1424.0, + 1403.0, + 1457.0, + 297.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1450.0, + 1207.0, + 1450.0, + 1207.0, + 1491.0, + 293.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1450.0, + 1403.0, + 1450.0, + 1403.0, + 1491.0, + 1255.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1483.0, + 1405.0, + 1483.0, + 1405.0, + 1520.0, + 295.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1513.0, + 1405.0, + 1513.0, + 1405.0, + 1551.0, + 292.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1547.0, + 1406.0, + 1547.0, + 1406.0, + 1576.0, + 300.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1405.0, + 1574.0, + 1405.0, + 1611.0, + 293.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1603.0, + 1405.0, + 1603.0, + 1405.0, + 1642.0, + 292.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1638.0, + 523.0, + 1638.0, + 523.0, + 1667.0, + 295.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1683.0, + 1405.0, + 1683.0, + 1405.0, + 1716.0, + 296.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1831.0, + 1405.0, + 1831.0, + 1405.0, + 1869.0, + 293.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1861.0, + 362.0, + 1861.0, + 362.0, + 1902.0, + 292.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 1861.0, + 1407.0, + 1861.0, + 1407.0, + 1902.0, + 412.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1899.0, + 373.0, + 1899.0, + 373.0, + 1926.0, + 294.0, + 1926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1892.0, + 1404.0, + 1892.0, + 1404.0, + 1932.0, + 424.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 442.0, + 1336.0, + 442.0, + 1336.0, + 479.0, + 377.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 474.0, + 915.0, + 474.0, + 915.0, + 508.0, + 784.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 500.0, + 1365.0, + 500.0, + 1365.0, + 544.0, + 336.0, + 544.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 998, + 1404, + 998, + 1404, + 1272, + 297, + 1272 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 740, + 1404, + 740, + 1404, + 985, + 298, + 985 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1778, + 1403, + 1778, + 1403, + 1901, + 298, + 1901 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1914, + 1403, + 1914, + 1403, + 2008, + 300, + 2008 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 327, + 195, + 1371, + 195, + 1371, + 522, + 327, + 522 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 298, + 541, + 1405, + 541, + 1405, + 638, + 298, + 638 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 361, + 1343, + 1405, + 1343, + 1405, + 1615, + 361, + 1615 + ], + "score": 0.959 + }, + { + "category_id": 0, + "poly": [ + 299, + 1720, + 657, + 1720, + 657, + 1755, + 299, + 1755 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 298, + 1286, + 997, + 1286, + 997, + 1318, + 298, + 1318 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 301, + 664, + 1399, + 664, + 1399, + 728, + 301, + 728 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 298, + 1653, + 476, + 1653, + 476, + 1692, + 298, + 1692 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2084, + 841, + 2084 + ], + "score": 0.74 + }, + { + "category_id": 13, + "poly": [ + 620, + 803, + 648, + 803, + 648, + 829, + 620, + 829 + ], + "score": 0.78, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 577, + 1088, + 627, + 1088, + 627, + 1121, + 577, + 1121 + ], + "score": 0.66, + "latex": "[ \\mathbb { 1 2 } ]" + }, + { + "category_id": 13, + "poly": [ + 660, + 1027, + 708, + 1027, + 708, + 1060, + 660, + 1060 + ], + "score": 0.63, + "latex": "\\mathbb { \\left. \\boldsymbol { \\ 2 5 } \\right. }" + }, + { + "category_id": 13, + "poly": [ + 1084, + 694, + 1133, + 694, + 1133, + 727, + 1084, + 727 + ], + "score": 0.63, + "latex": "\\dot { \\lVert 2 4 \\rVert }" + }, + { + "category_id": 13, + "poly": [ + 924, + 1579, + 974, + 1579, + 974, + 1616, + 924, + 1616 + ], + "score": 0.57, + "latex": "\\underline { { \\overline { { 5 . 2 } } } } )" + }, + { + "category_id": 13, + "poly": [ + 735, + 1807, + 872, + 1807, + 872, + 1840, + 735, + 1840 + ], + "score": 0.43, + "latex": "\\textcircled { 1 5 4 } , \\textcircled { 5 6 } , \\textcircled { 6 7 } \\textcircled { 1 }" + }, + { + "category_id": 13, + "poly": [ + 1296, + 1807, + 1345, + 1807, + 1345, + 1840, + 1296, + 1840 + ], + "score": 0.42, + "latex": "\\mathbb { \\left. \\boldsymbol { \\mathfrak { G } } \\boldsymbol { \\mathfrak { g } } \\right. }" + }, + { + "category_id": 13, + "poly": [ + 1297, + 1912, + 1346, + 1912, + 1346, + 1946, + 1297, + 1946 + ], + "score": 0.33, + "latex": "\\pmb { \\mathbb { B } } \\mathbf { \\mathbb { 1 } }" + }, + { + "category_id": 13, + "poly": [ + 654, + 1510, + 702, + 1510, + 702, + 1547, + 654, + 1547 + ], + "score": 0.32, + "latex": "5 . 1 )" + }, + { + "category_id": 13, + "poly": [ + 563, + 953, + 600, + 953, + 600, + 982, + 563, + 982 + ], + "score": 0.26, + "latex": "\\mathbf { C o }" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 199.0, + 546.0, + 199.0, + 546.0, + 222.0, + 398.0, + 222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 196.0, + 876.0, + 196.0, + 876.0, + 224.0, + 673.0, + 224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 196.0, + 1365.0, + 196.0, + 1365.0, + 224.0, + 940.0, + 224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 220.0, + 558.0, + 220.0, + 558.0, + 243.0, + 389.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 217.0, + 885.0, + 217.0, + 885.0, + 245.0, + 666.0, + 245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 217.0, + 1360.0, + 217.0, + 1360.0, + 245.0, + 947.0, + 245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 264.0, + 832.0, + 264.0, + 832.0, + 280.0, + 793.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 260.0, + 1128.0, + 260.0, + 1128.0, + 285.0, + 1080.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 290.0, + 425.0, + 290.0, + 425.0, + 323.0, + 348.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 312.0, + 407.0, + 312.0, + 407.0, + 340.0, + 368.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 323.0, + 819.0, + 323.0, + 819.0, + 338.0, + 803.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 320.0, + 1115.0, + 320.0, + 1115.0, + 339.0, + 1092.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 321.0, + 1352.0, + 321.0, + 1352.0, + 351.0, + 1250.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 329.0, + 544.0, + 329.0, + 544.0, + 349.0, + 520.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 350.0, + 1360.0, + 350.0, + 1360.0, + 377.0, + 1242.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 406.0, + 1123.0, + 406.0, + 1123.0, + 419.0, + 1103.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 420.0, + 543.0, + 420.0, + 543.0, + 435.0, + 526.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 420.0, + 1280.0, + 420.0, + 1280.0, + 430.0, + 1271.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 434.0, + 449.0, + 434.0, + 449.0, + 461.0, + 324.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 452.0, + 406.0, + 452.0, + 406.0, + 479.0, + 368.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 468.0, + 1109.0, + 468.0, + 1109.0, + 476.0, + 1099.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 540.0, + 1403.0, + 540.0, + 1403.0, + 575.0, + 295.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 570.0, + 1405.0, + 570.0, + 1405.0, + 608.0, + 296.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 603.0, + 836.0, + 603.0, + 836.0, + 638.0, + 295.0, + 638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1716.0, + 659.0, + 1716.0, + 659.0, + 1763.0, + 292.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1647.0, + 479.0, + 1647.0, + 479.0, + 1699.0, + 290.0, + 1699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2093.0, + 839.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 996.0, + 1407.0, + 996.0, + 1407.0, + 1032.0, + 295.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1026.0, + 659.0, + 1026.0, + 659.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1026.0, + 1406.0, + 1026.0, + 1406.0, + 1064.0, + 709.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1056.0, + 1405.0, + 1056.0, + 1405.0, + 1095.0, + 292.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1087.0, + 576.0, + 1087.0, + 576.0, + 1125.0, + 292.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1125.0, + 628.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1119.0, + 1405.0, + 1119.0, + 1405.0, + 1156.0, + 294.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1144.0, + 1405.0, + 1144.0, + 1405.0, + 1188.0, + 291.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1219.0, + 292.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1209.0, + 1404.0, + 1209.0, + 1404.0, + 1245.0, + 295.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1241.0, + 1120.0, + 1241.0, + 1120.0, + 1274.0, + 296.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 738.0, + 1405.0, + 738.0, + 1405.0, + 776.0, + 293.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 770.0, + 1404.0, + 770.0, + 1404.0, + 806.0, + 293.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 800.0, + 619.0, + 800.0, + 619.0, + 837.0, + 293.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 800.0, + 1408.0, + 800.0, + 1408.0, + 837.0, + 649.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 830.0, + 1404.0, + 830.0, + 1404.0, + 867.0, + 293.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 862.0, + 1404.0, + 862.0, + 1404.0, + 896.0, + 294.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 892.0, + 1405.0, + 892.0, + 1405.0, + 926.0, + 292.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 921.0, + 1404.0, + 921.0, + 1404.0, + 957.0, + 293.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 562.0, + 951.0, + 562.0, + 987.0, + 293.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 951.0, + 1355.0, + 951.0, + 1355.0, + 987.0, + 601.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1810.0, + 296.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1806.0, + 734.0, + 1806.0, + 734.0, + 1842.0, + 293.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1806.0, + 1295.0, + 1806.0, + 1295.0, + 1842.0, + 873.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1806.0, + 1407.0, + 1806.0, + 1407.0, + 1842.0, + 1346.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1838.0, + 1407.0, + 1838.0, + 1407.0, + 1873.0, + 292.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1866.0, + 603.0, + 1866.0, + 603.0, + 1905.0, + 294.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1912.0, + 1296.0, + 1912.0, + 1296.0, + 1950.0, + 295.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1347.0, + 1912.0, + 1408.0, + 1912.0, + 1408.0, + 1950.0, + 1347.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1983.0, + 293.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1977.0, + 1403.0, + 1977.0, + 1403.0, + 2011.0, + 295.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1343.0, + 1405.0, + 1343.0, + 1405.0, + 1379.0, + 361.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1374.0, + 640.0, + 1374.0, + 640.0, + 1407.0, + 393.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1412.0, + 1406.0, + 1412.0, + 1406.0, + 1448.0, + 357.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1443.0, + 1059.0, + 1443.0, + 1059.0, + 1479.0, + 393.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1480.0, + 1407.0, + 1480.0, + 1407.0, + 1516.0, + 357.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1514.0, + 653.0, + 1514.0, + 653.0, + 1546.0, + 393.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1514.0, + 711.0, + 1514.0, + 711.0, + 1546.0, + 703.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1549.0, + 1406.0, + 1549.0, + 1406.0, + 1585.0, + 357.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1581.0, + 923.0, + 1581.0, + 923.0, + 1615.0, + 393.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1581.0, + 980.0, + 1581.0, + 980.0, + 1615.0, + 975.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1283.0, + 1000.0, + 1283.0, + 1000.0, + 1322.0, + 295.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 661.0, + 1404.0, + 661.0, + 1404.0, + 701.0, + 297.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 695.0, + 1083.0, + 695.0, + 1083.0, + 728.0, + 295.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 695.0, + 1141.0, + 695.0, + 1141.0, + 728.0, + 1134.0, + 728.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1199, + 1404, + 1199, + 1404, + 1383, + 298, + 1383 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 940, + 1405, + 940, + 1405, + 1186, + 298, + 1186 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 328, + 197, + 1377, + 197, + 1377, + 610, + 328, + 610 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 296, + 627, + 1405, + 627, + 1405, + 810, + 296, + 810 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 835, + 1399, + 835, + 1399, + 929, + 300, + 929 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 362, + 1744, + 1404, + 1744, + 1404, + 2009, + 362, + 2009 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 300, + 1472, + 1405, + 1472, + 1405, + 1535, + 300, + 1535 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 361, + 1556, + 1407, + 1556, + 1407, + 1670, + 361, + 1670 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 299, + 1417, + 640, + 1417, + 640, + 1449, + 299, + 1449 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 306, + 1689, + 1315, + 1689, + 1315, + 1723, + 306, + 1723 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2085, + 841, + 2085 + ], + "score": 0.632 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.422 + }, + { + "category_id": 13, + "poly": [ + 820, + 1907, + 907, + 1907, + 907, + 1939, + 820, + 1939 + ], + "score": 0.88, + "latex": "\\operatorname { F C P } _ { + A }" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1976, + 1252, + 1976, + 1252, + 2009, + 1111, + 2009 + ], + "score": 0.87, + "latex": "( \\mathrm { F C P } _ { - T , + A } )" + }, + { + "category_id": 13, + "poly": [ + 404, + 1806, + 493, + 1806, + 493, + 1839, + 404, + 1839 + ], + "score": 0.84, + "latex": "\\mathrm { F C P } _ { - T }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1004, + 1231, + 1004, + 1231, + 1030, + 1202, + 1030 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 1372, + 719, + 1401, + 719, + 1401, + 747, + 1372, + 747 + ], + "score": 0.68, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 615, + 971, + 644, + 971, + 644, + 1007, + 615, + 1007 + ], + "score": 0.45, + "latex": "2 ," + }, + { + "category_id": 15, + "poly": [ + 414.0, + 204.0, + 677.0, + 204.0, + 677.0, + 235.0, + 414.0, + 235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 195.0, + 1002.0, + 195.0, + 1002.0, + 225.0, + 847.0, + 225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 217.0, + 1001.0, + 217.0, + 1001.0, + 248.0, + 847.0, + 248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 235.0, + 633.0, + 235.0, + 633.0, + 265.0, + 413.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 251.0, + 1095.0, + 251.0, + 1095.0, + 273.0, + 1067.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 253.0, + 1362.0, + 253.0, + 1362.0, + 270.0, + 1337.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 272.0, + 1125.0, + 272.0, + 1125.0, + 290.0, + 1098.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 274.0, + 1330.0, + 274.0, + 1330.0, + 288.0, + 1307.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 290.0, + 1157.0, + 290.0, + 1157.0, + 307.0, + 1131.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 556.0, + 456.0, + 556.0, + 456.0, + 593.0, + 354.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 555.0, + 721.0, + 555.0, + 721.0, + 594.0, + 554.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 555.0, + 1038.0, + 555.0, + 1038.0, + 594.0, + 806.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 552.0, + 1308.0, + 552.0, + 1308.0, + 594.0, + 1125.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 581.0, + 431.0, + 581.0, + 431.0, + 616.0, + 379.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 581.0, + 666.0, + 581.0, + 666.0, + 617.0, + 611.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 581.0, + 957.0, + 581.0, + 957.0, + 614.0, + 890.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 581.0, + 1250.0, + 581.0, + 1250.0, + 616.0, + 1183.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 626.0, + 1406.0, + 626.0, + 1406.0, + 662.0, + 294.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 658.0, + 1406.0, + 658.0, + 1406.0, + 693.0, + 294.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 1405.0, + 686.0, + 1405.0, + 724.0, + 293.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 717.0, + 1371.0, + 717.0, + 1371.0, + 753.0, + 294.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 749.0, + 1405.0, + 749.0, + 1405.0, + 784.0, + 295.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 779.0, + 1217.0, + 779.0, + 1217.0, + 813.0, + 293.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1415.0, + 641.0, + 1415.0, + 641.0, + 1452.0, + 295.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2058.0, + 860.0, + 2058.0, + 860.0, + 2089.0, + 839.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1198.0, + 1406.0, + 1198.0, + 1406.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1229.0, + 1406.0, + 1229.0, + 1406.0, + 1264.0, + 293.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1258.0, + 1406.0, + 1258.0, + 1406.0, + 1295.0, + 291.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1290.0, + 1406.0, + 1290.0, + 1406.0, + 1325.0, + 293.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1321.0, + 1405.0, + 1321.0, + 1405.0, + 1353.0, + 296.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1349.0, + 1268.0, + 1349.0, + 1268.0, + 1385.0, + 293.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 940.0, + 1406.0, + 940.0, + 1406.0, + 978.0, + 293.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 972.0, + 614.0, + 972.0, + 614.0, + 1007.0, + 295.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 972.0, + 1403.0, + 972.0, + 1403.0, + 1007.0, + 645.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1002.0, + 1201.0, + 1002.0, + 1201.0, + 1038.0, + 295.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 1002.0, + 1405.0, + 1002.0, + 1405.0, + 1038.0, + 1232.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1032.0, + 1405.0, + 1032.0, + 1405.0, + 1068.0, + 293.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1097.0, + 295.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1094.0, + 1407.0, + 1094.0, + 1407.0, + 1128.0, + 295.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1121.0, + 1406.0, + 1121.0, + 1406.0, + 1160.0, + 292.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1154.0, + 1408.0, + 1154.0, + 1408.0, + 1188.0, + 295.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 1405.0, + 833.0, + 1405.0, + 871.0, + 295.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 900.0, + 295.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 897.0, + 858.0, + 897.0, + 858.0, + 934.0, + 295.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1744.0, + 1404.0, + 1744.0, + 1404.0, + 1782.0, + 360.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1775.0, + 1404.0, + 1775.0, + 1404.0, + 1810.0, + 394.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1808.0, + 403.0, + 1808.0, + 403.0, + 1837.0, + 395.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1808.0, + 760.0, + 1808.0, + 760.0, + 1837.0, + 494.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1846.0, + 1406.0, + 1846.0, + 1406.0, + 1881.0, + 361.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1878.0, + 1404.0, + 1878.0, + 1404.0, + 1910.0, + 395.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1906.0, + 819.0, + 1906.0, + 819.0, + 1938.0, + 393.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1906.0, + 1348.0, + 1906.0, + 1348.0, + 1938.0, + 908.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1944.0, + 1405.0, + 1944.0, + 1405.0, + 1981.0, + 356.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1973.0, + 1110.0, + 1973.0, + 1110.0, + 2012.0, + 393.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1973.0, + 1265.0, + 1973.0, + 1265.0, + 2012.0, + 1253.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1472.0, + 1405.0, + 1472.0, + 1405.0, + 1507.0, + 296.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1502.0, + 1370.0, + 1502.0, + 1370.0, + 1538.0, + 295.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1557.0, + 1266.0, + 1557.0, + 1266.0, + 1592.0, + 361.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1593.0, + 1409.0, + 1593.0, + 1409.0, + 1634.0, + 356.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1632.0, + 1408.0, + 1632.0, + 1408.0, + 1673.0, + 356.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1687.0, + 1318.0, + 1687.0, + 1318.0, + 1725.0, + 299.0, + 1725.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 365, + 1404, + 365, + 1404, + 519, + 297, + 519 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1853, + 1404, + 1853, + 1404, + 2008, + 297, + 2008 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 330, + 760, + 1368, + 760, + 1368, + 1164, + 330, + 1164 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 260, + 1404, + 260, + 1404, + 353, + 298, + 353 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1716, + 1404, + 1716, + 1404, + 1841, + 298, + 1841 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 300, + 637, + 1402, + 637, + 1402, + 731, + 300, + 731 + ], + "score": 0.963 + }, + { + "category_id": 3, + "poly": [ + 336, + 1275, + 1367, + 1275, + 1367, + 1449, + 336, + 1449 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 299, + 531, + 1403, + 531, + 1403, + 624, + 299, + 624 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 299, + 1184, + 1398, + 1184, + 1398, + 1247, + 299, + 1247 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 299, + 1642, + 1405, + 1642, + 1405, + 1705, + 299, + 1705 + ], + "score": 0.933 + }, + { + "category_id": 0, + "poly": [ + 299, + 1586, + 638, + 1586, + 638, + 1618, + 299, + 1618 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 298, + 203, + 519, + 203, + 519, + 235, + 298, + 235 + ], + "score": 0.914 + }, + { + "category_id": 4, + "poly": [ + 289, + 1465, + 1402, + 1465, + 1402, + 1529, + 289, + 1529 + ], + "score": 0.908 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2084, + 841, + 2084 + ], + "score": 0.775 + }, + { + "category_id": 13, + "poly": [ + 1267, + 1976, + 1354, + 1976, + 1354, + 2008, + 1267, + 2008 + ], + "score": 0.9, + "latex": "\\operatorname { F C P } _ { + A }" + }, + { + "category_id": 13, + "poly": [ + 901, + 1855, + 994, + 1855, + 994, + 1883, + 901, + 1883 + ], + "score": 0.89, + "latex": "N = 3 2" + }, + { + "category_id": 13, + "poly": [ + 1141, + 1946, + 1197, + 1946, + 1197, + 1975, + 1141, + 1975 + ], + "score": 0.87, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 432, + 1746, + 481, + 1746, + 481, + 1780, + 432, + 1780 + ], + "score": 0.7, + "latex": "\\mathbb { \\left| \\bigstar \\bigstar \\right\\| }" + }, + { + "category_id": 13, + "poly": [ + 1045, + 1777, + 1094, + 1777, + 1094, + 1810, + 1045, + 1810 + ], + "score": 0.69, + "latex": "\\textcircled { 1 1 7 }" + }, + { + "category_id": 13, + "poly": [ + 950, + 1715, + 999, + 1715, + 999, + 1749, + 950, + 1749 + ], + "score": 0.45, + "latex": "\\boldsymbol { \\left[ \\left[ 6 5 \\right] \\right] }" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 768.0, + 614.0, + 768.0, + 614.0, + 800.0, + 549.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 768.0, + 780.0, + 768.0, + 780.0, + 802.0, + 644.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 754.0, + 907.0, + 754.0, + 907.0, + 812.0, + 817.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 758.0, + 1340.0, + 758.0, + 1340.0, + 784.0, + 1058.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 865.0, + 418.0, + 865.0, + 418.0, + 892.0, + 331.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 866.0, + 1346.0, + 866.0, + 1346.0, + 886.0, + 1053.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 974.0, + 401.0, + 974.0, + 401.0, + 1003.0, + 348.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 968.0, + 1342.0, + 968.0, + 1342.0, + 992.0, + 1055.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1000.0, + 412.0, + 1000.0, + 412.0, + 1025.0, + 333.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 1078.0, + 1301.0, + 1078.0, + 1301.0, + 1101.0, + 1097.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1096.0, + 415.0, + 1096.0, + 415.0, + 1128.0, + 333.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1117.0, + 419.0, + 1117.0, + 419.0, + 1151.0, + 329.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1112.0, + 1341.0, + 1112.0, + 1341.0, + 1158.0, + 1234.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1319.0, + 703.0, + 1319.0, + 703.0, + 1329.0, + 694.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1319.0, + 1236.0, + 1319.0, + 1236.0, + 1328.0, + 1225.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1333.0, + 512.0, + 1333.0, + 512.0, + 1339.0, + 506.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1339.0, + 682.0, + 1339.0, + 682.0, + 1363.0, + 656.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1407.0, + 863.0, + 1407.0, + 863.0, + 1417.0, + 852.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1427.0, + 491.0, + 1427.0, + 491.0, + 1450.0, + 378.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1426.0, + 753.0, + 1426.0, + 753.0, + 1449.0, + 584.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 1422.0, + 951.0, + 1422.0, + 951.0, + 1453.0, + 824.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1428.0, + 1141.0, + 1428.0, + 1141.0, + 1448.0, + 1033.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1428.0, + 1358.0, + 1428.0, + 1358.0, + 1448.0, + 1218.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1182.0, + 1402.0, + 1182.0, + 1402.0, + 1218.0, + 295.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1212.0, + 706.0, + 1212.0, + 706.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1583.0, + 639.0, + 1583.0, + 639.0, + 1623.0, + 295.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 199.0, + 523.0, + 199.0, + 523.0, + 239.0, + 293.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1462.0, + 1405.0, + 1462.0, + 1405.0, + 1501.0, + 295.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1493.0, + 1405.0, + 1493.0, + 1405.0, + 1533.0, + 293.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 363.0, + 1406.0, + 363.0, + 1406.0, + 401.0, + 292.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 394.0, + 1401.0, + 394.0, + 1401.0, + 427.0, + 295.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 426.0, + 1404.0, + 426.0, + 1404.0, + 459.0, + 296.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 449.0, + 1405.0, + 449.0, + 1405.0, + 499.0, + 292.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 486.0, + 1023.0, + 486.0, + 1023.0, + 523.0, + 294.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1853.0, + 900.0, + 1853.0, + 900.0, + 1891.0, + 295.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1853.0, + 1409.0, + 1853.0, + 1409.0, + 1891.0, + 995.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1881.0, + 1407.0, + 1881.0, + 1407.0, + 1920.0, + 292.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1913.0, + 1407.0, + 1913.0, + 1407.0, + 1950.0, + 294.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1944.0, + 1140.0, + 1944.0, + 1140.0, + 1981.0, + 292.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1981.0, + 1198.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1977.0, + 1266.0, + 1977.0, + 1266.0, + 2010.0, + 296.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1977.0, + 1405.0, + 1977.0, + 1405.0, + 2010.0, + 1355.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 1405.0, + 258.0, + 1405.0, + 292.0, + 294.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 287.0, + 1406.0, + 287.0, + 1406.0, + 325.0, + 293.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 896.0, + 319.0, + 896.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1716.0, + 949.0, + 1716.0, + 949.0, + 1750.0, + 292.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1716.0, + 1400.0, + 1716.0, + 1400.0, + 1750.0, + 1000.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1745.0, + 431.0, + 1745.0, + 431.0, + 1783.0, + 292.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1745.0, + 1405.0, + 1745.0, + 1405.0, + 1783.0, + 482.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1775.0, + 1044.0, + 1775.0, + 1044.0, + 1815.0, + 292.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1775.0, + 1405.0, + 1775.0, + 1405.0, + 1815.0, + 1095.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1808.0, + 962.0, + 1808.0, + 962.0, + 1843.0, + 295.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 635.0, + 1406.0, + 635.0, + 1406.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 668.0, + 1406.0, + 668.0, + 1406.0, + 702.0, + 295.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 697.0, + 1323.0, + 697.0, + 1323.0, + 733.0, + 294.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 532.0, + 1404.0, + 532.0, + 1404.0, + 565.0, + 295.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 563.0, + 1408.0, + 563.0, + 1408.0, + 597.0, + 295.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 591.0, + 1377.0, + 591.0, + 1377.0, + 627.0, + 294.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 1409.0, + 1642.0, + 1409.0, + 1677.0, + 295.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 1278.0, + 1671.0, + 1278.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 648, + 1404, + 648, + 1404, + 925, + 297, + 925 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 307, + 1404, + 307, + 1404, + 523, + 298, + 523 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1074, + 1404, + 1074, + 1404, + 1289, + 298, + 1289 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1437, + 1404, + 1437, + 1404, + 1654, + 297, + 1654 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1664, + 1405, + 1664, + 1405, + 1850, + 298, + 1850 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1301, + 1403, + 1301, + 1403, + 1424, + 298, + 1424 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 937, + 1404, + 937, + 1404, + 1060, + 298, + 1060 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 203, + 1403, + 203, + 1403, + 295, + 298, + 295 + ], + "score": 0.96 + }, + { + "category_id": 0, + "poly": [ + 298, + 1900, + 849, + 1900, + 849, + 1938, + 298, + 1938 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 299, + 574, + 540, + 574, + 540, + 611, + 299, + 611 + ], + "score": 0.918 + }, + { + "category_id": 1, + "poly": [ + 290, + 1974, + 1355, + 1974, + 1355, + 2007, + 290, + 2007 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2085, + 841, + 2085 + ], + "score": 0.671 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.204 + }, + { + "category_id": 13, + "poly": [ + 851, + 264, + 944, + 264, + 944, + 292, + 851, + 292 + ], + "score": 0.89, + "latex": "N = 3 2" + }, + { + "category_id": 13, + "poly": [ + 297, + 204, + 424, + 204, + 424, + 237, + 297, + 237 + ], + "score": 0.85, + "latex": "\\mathrm { F C P } _ { - T , + A }" + }, + { + "category_id": 13, + "poly": [ + 716, + 308, + 763, + 308, + 763, + 341, + 716, + 341 + ], + "score": 0.79, + "latex": "\\left[ \\left[ 5 8 \\right] \\right]" + }, + { + "category_id": 13, + "poly": [ + 1192, + 710, + 1227, + 710, + 1227, + 741, + 1192, + 741 + ], + "score": 0.78, + "latex": "\\bar { \\mathbb { D } }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 308, + 1158, + 308, + 1158, + 341, + 1109, + 341 + ], + "score": 0.77, + "latex": "\\pmb { \\mathbb { B } } 0 \\|" + }, + { + "category_id": 13, + "poly": [ + 402, + 800, + 451, + 800, + 451, + 832, + 402, + 832 + ], + "score": 0.76, + "latex": "\\widehat { \\left\\| 3 6 \\right\\| }" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1467, + 1341, + 1467, + 1341, + 1500, + 1293, + 1500 + ], + "score": 0.76, + "latex": "\\boxed { \\boxed { 1 6 } }" + }, + { + "category_id": 13, + "poly": [ + 1268, + 647, + 1343, + 647, + 1343, + 681, + 1268, + 681 + ], + "score": 0.73, + "latex": " { \\mathbb { B } } , { \\mathbb { G } } 6 { \\mathbb { I } }" + }, + { + "category_id": 13, + "poly": [ + 857, + 1361, + 906, + 1361, + 906, + 1394, + 857, + 1394 + ], + "score": 0.66, + "latex": "\\bar { \\lVert 1 6 rVert }" + }, + { + "category_id": 13, + "poly": [ + 848, + 1498, + 897, + 1498, + 897, + 1530, + 848, + 1530 + ], + "score": 0.64, + "latex": "\\lVert \\rVert" + }, + { + "category_id": 13, + "poly": [ + 571, + 709, + 621, + 709, + 621, + 742, + 571, + 742 + ], + "score": 0.64, + "latex": "\\pmb { \\mathbb { Z } 1 }" + }, + { + "category_id": 13, + "poly": [ + 831, + 430, + 880, + 430, + 880, + 462, + 831, + 462 + ], + "score": 0.61, + "latex": "\\bar { \\lVert 1 2 \\rVert }" + }, + { + "category_id": 13, + "poly": [ + 1220, + 800, + 1269, + 800, + 1269, + 833, + 1220, + 833 + ], + "score": 0.6, + "latex": "\\lVert \\rVert \\dot { \\boldsymbol { \\mathrm { E } } } \\rVert" + }, + { + "category_id": 13, + "poly": [ + 1175, + 1254, + 1305, + 1254, + 1305, + 1288, + 1175, + 1288 + ], + "score": 0.6, + "latex": "[ \\overbrace { 2 4 } , \\overbrace { 3 9 } , \\overbrace { 6 9 } ]" + }, + { + "category_id": 13, + "poly": [ + 955, + 1664, + 1003, + 1664, + 1003, + 1696, + 955, + 1696 + ], + "score": 0.58, + "latex": "\\pmb { \\mathbb { I } }" + }, + { + "category_id": 13, + "poly": [ + 676, + 679, + 724, + 679, + 724, + 712, + 676, + 712 + ], + "score": 0.49, + "latex": "\\bar { \\mathbb { B } } \\mathbb { I }" + }, + { + "category_id": 13, + "poly": [ + 480, + 1724, + 531, + 1724, + 531, + 1761, + 480, + 1761 + ], + "score": 0.44, + "latex": "\\boxed { 2 . 2 }" + }, + { + "category_id": 13, + "poly": [ + 810, + 709, + 860, + 709, + 860, + 742, + 810, + 742 + ], + "score": 0.35, + "latex": "\\mathbb { \\left| \\left. \\sum \\right. { ] } \\right| }" + }, + { + "category_id": 13, + "poly": [ + 901, + 800, + 991, + 800, + 991, + 833, + 901, + 833 + ], + "score": 0.34, + "latex": "\\checkmark" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1895.0, + 852.0, + 1895.0, + 852.0, + 1946.0, + 291.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 571.0, + 545.0, + 571.0, + 545.0, + 616.0, + 292.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2058.0, + 861.0, + 2058.0, + 861.0, + 2093.0, + 839.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 649.0, + 1267.0, + 649.0, + 1267.0, + 682.0, + 296.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 682.0, + 1344.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 678.0, + 675.0, + 678.0, + 675.0, + 715.0, + 294.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 678.0, + 1405.0, + 678.0, + 1405.0, + 715.0, + 725.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 711.0, + 570.0, + 711.0, + 570.0, + 743.0, + 296.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 711.0, + 809.0, + 711.0, + 809.0, + 743.0, + 622.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 711.0, + 1191.0, + 711.0, + 1191.0, + 743.0, + 861.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 711.0, + 1405.0, + 711.0, + 1405.0, + 743.0, + 1228.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 776.0, + 295.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 769.0, + 1406.0, + 769.0, + 1406.0, + 806.0, + 292.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 797.0, + 401.0, + 797.0, + 401.0, + 840.0, + 292.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 797.0, + 900.0, + 797.0, + 900.0, + 840.0, + 452.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 797.0, + 1219.0, + 797.0, + 1219.0, + 840.0, + 992.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 840.0, + 1270.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 828.0, + 1404.0, + 828.0, + 1404.0, + 866.0, + 292.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1405.0, + 860.0, + 1405.0, + 896.0, + 294.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 894.0, + 959.0, + 894.0, + 959.0, + 927.0, + 295.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 715.0, + 306.0, + 715.0, + 345.0, + 293.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 306.0, + 1108.0, + 306.0, + 1108.0, + 345.0, + 764.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 306.0, + 1404.0, + 306.0, + 1404.0, + 345.0, + 1159.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 340.0, + 1405.0, + 340.0, + 1405.0, + 374.0, + 294.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 369.0, + 1406.0, + 369.0, + 1406.0, + 403.0, + 293.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 403.0, + 1404.0, + 403.0, + 1404.0, + 434.0, + 296.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 430.0, + 830.0, + 430.0, + 830.0, + 464.0, + 293.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 430.0, + 1404.0, + 430.0, + 1404.0, + 464.0, + 881.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 461.0, + 1405.0, + 461.0, + 1405.0, + 495.0, + 293.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 493.0, + 1257.0, + 493.0, + 1257.0, + 525.0, + 293.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1111.0, + 293.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1105.0, + 1405.0, + 1105.0, + 1405.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1133.0, + 1408.0, + 1133.0, + 1408.0, + 1170.0, + 295.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1164.0, + 1405.0, + 1164.0, + 1405.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1232.0, + 297.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1226.0, + 1404.0, + 1226.0, + 1404.0, + 1260.0, + 296.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1255.0, + 1174.0, + 1255.0, + 1174.0, + 1291.0, + 293.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1255.0, + 1312.0, + 1255.0, + 1312.0, + 1291.0, + 1306.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1433.0, + 1406.0, + 1433.0, + 1406.0, + 1475.0, + 292.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1465.0, + 1292.0, + 1465.0, + 1292.0, + 1504.0, + 293.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1465.0, + 1408.0, + 1465.0, + 1408.0, + 1504.0, + 1342.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1493.0, + 847.0, + 1493.0, + 847.0, + 1536.0, + 294.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1493.0, + 1404.0, + 1493.0, + 1404.0, + 1536.0, + 898.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1563.0, + 295.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1559.0, + 1404.0, + 1559.0, + 1404.0, + 1594.0, + 295.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1589.0, + 1406.0, + 1589.0, + 1406.0, + 1622.0, + 294.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1620.0, + 920.0, + 1620.0, + 920.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1664.0, + 954.0, + 1664.0, + 954.0, + 1700.0, + 295.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1664.0, + 1408.0, + 1664.0, + 1408.0, + 1700.0, + 1004.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1693.0, + 1410.0, + 1693.0, + 1410.0, + 1733.0, + 293.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1725.0, + 479.0, + 1725.0, + 479.0, + 1761.0, + 293.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 1725.0, + 1407.0, + 1725.0, + 1407.0, + 1761.0, + 532.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1756.0, + 1406.0, + 1756.0, + 1406.0, + 1791.0, + 295.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1786.0, + 1406.0, + 1786.0, + 1406.0, + 1822.0, + 295.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1817.0, + 1227.0, + 1817.0, + 1227.0, + 1851.0, + 294.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1301.0, + 1404.0, + 1301.0, + 1404.0, + 1336.0, + 293.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1367.0, + 292.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1360.0, + 856.0, + 1360.0, + 856.0, + 1401.0, + 292.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 1360.0, + 1404.0, + 1360.0, + 1404.0, + 1401.0, + 907.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1394.0, + 1309.0, + 1394.0, + 1309.0, + 1426.0, + 294.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 933.0, + 1405.0, + 933.0, + 1405.0, + 974.0, + 293.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 965.0, + 1410.0, + 965.0, + 1410.0, + 1004.0, + 291.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 997.0, + 1405.0, + 997.0, + 1405.0, + 1032.0, + 292.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 1186.0, + 1029.0, + 1186.0, + 1061.0, + 293.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 202.0, + 296.0, + 202.0, + 296.0, + 239.0, + 293.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 239.0, + 425.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 265.0, + 850.0, + 265.0, + 850.0, + 298.0, + 294.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 265.0, + 1388.0, + 265.0, + 1388.0, + 298.0, + 945.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1359.0, + 1971.0, + 1359.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 258, + 1405, + 258, + 1405, + 412, + 298, + 412 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 955, + 1405, + 955, + 1405, + 1107, + 298, + 1107 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 295, + 1121, + 1405, + 1121, + 1405, + 1276, + 295, + 1276 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 706, + 1406, + 706, + 1406, + 830, + 298, + 830 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 282, + 1307, + 1376, + 1307, + 1376, + 1555, + 282, + 1555 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 361, + 479, + 1406, + 479, + 1406, + 685, + 361, + 685 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 295, + 1570, + 1406, + 1570, + 1406, + 1728, + 295, + 1728 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 300, + 425, + 795, + 425, + 795, + 457, + 300, + 457 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 299, + 202, + 1128, + 202, + 1128, + 236, + 299, + 236 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 298, + 863, + 450, + 863, + 450, + 895, + 298, + 895 + ], + "score": 0.904 + }, + { + "category_id": 0, + "poly": [ + 300, + 919, + 936, + 919, + 936, + 952, + 300, + 952 + ], + "score": 0.848 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 859, + 2062, + 859, + 2085, + 840, + 2085 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 296, + 1770, + 1407, + 1770, + 1407, + 2008, + 296, + 2008 + ], + "score": 0.504 + }, + { + "category_id": 0, + "poly": [ + 299, + 1773, + 1364, + 1773, + 1364, + 1805, + 299, + 1805 + ], + "score": 0.374 + }, + { + "category_id": 13, + "poly": [ + 789, + 481, + 868, + 481, + 868, + 514, + 789, + 514 + ], + "score": 0.91, + "latex": "H _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 982, + 1946, + 1071, + 1946, + 1071, + 1978, + 982, + 1978 + ], + "score": 0.9, + "latex": "\\operatorname { F C P } _ { + A }" + }, + { + "category_id": 13, + "poly": [ + 1076, + 1601, + 1184, + 1601, + 1184, + 1630, + 1076, + 1630 + ], + "score": 0.9, + "latex": "T = 5 4 0" + }, + { + "category_id": 13, + "poly": [ + 933, + 582, + 1020, + 582, + 1020, + 613, + 933, + 613 + ], + "score": 0.9, + "latex": "\\mathrm { F C P } _ { + A }" + }, + { + "category_id": 13, + "poly": [ + 577, + 582, + 671, + 582, + 671, + 609, + 577, + 609 + ], + "score": 0.89, + "latex": "N = 3 2" + }, + { + "category_id": 13, + "poly": [ + 1317, + 1016, + 1397, + 1016, + 1397, + 1048, + 1317, + 1048 + ], + "score": 0.88, + "latex": "H _ { \\mathrm { p r o x y } }" + }, + { + "category_id": 13, + "poly": [ + 585, + 1915, + 674, + 1915, + 674, + 1947, + 585, + 1947 + ], + "score": 0.83, + "latex": "\\mathrm { F C P } _ { - T }" + }, + { + "category_id": 13, + "poly": [ + 920, + 735, + 949, + 735, + 949, + 771, + 920, + 771 + ], + "score": 0.74, + "latex": "^ 4" + }, + { + "category_id": 13, + "poly": [ + 884, + 1841, + 924, + 1841, + 924, + 1870, + 884, + 1870 + ], + "score": 0.72, + "latex": "( T )" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1840, + 1156, + 1840, + 1156, + 1870, + 1115, + 1870 + ], + "score": 0.68, + "latex": "( A )" + }, + { + "category_id": 13, + "poly": [ + 773, + 1598, + 824, + 1598, + 824, + 1635, + 773, + 1635 + ], + "score": 0.62, + "latex": "4 . 1 )" + }, + { + "category_id": 13, + "poly": [ + 900, + 983, + 928, + 983, + 928, + 1020, + 900, + 1020 + ], + "score": 0.61, + "latex": "\\boxed { 5 }" + }, + { + "category_id": 13, + "poly": [ + 408, + 765, + 462, + 765, + 462, + 801, + 408, + 801 + ], + "score": 0.51, + "latex": "\\underline { { \\overline { { ( \\mathrm { C . 2 } ) } } } }" + }, + { + "category_id": 13, + "poly": [ + 385, + 288, + 412, + 288, + 412, + 324, + 385, + 324 + ], + "score": 0.35, + "latex": "\\textcircled{5}" + }, + { + "category_id": 13, + "poly": [ + 668, + 509, + 717, + 509, + 717, + 543, + 668, + 543 + ], + "score": 0.29, + "latex": "[ \\overbrace { | 1 2 | }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1044, + 1304, + 1044, + 1304, + 1077, + 1255, + 1077 + ], + "score": 0.28, + "latex": "\\mathbb { \\lVert 1 2 \\rVert }" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1333.0, + 348.0, + 1333.0, + 348.0, + 1453.0, + 278.0, + 1453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1336.0, + 733.0, + 1336.0, + 733.0, + 1454.0, + 666.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1334.0, + 1116.0, + 1334.0, + 1116.0, + 1454.0, + 1051.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1396.0, + 345.0, + 1396.0, + 345.0, + 1426.0, + 323.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1396.0, + 731.0, + 1396.0, + 731.0, + 1425.0, + 707.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1401.0, + 1111.0, + 1401.0, + 1111.0, + 1420.0, + 1096.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1455.0, + 350.0, + 1455.0, + 350.0, + 1482.0, + 324.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1455.0, + 732.0, + 1455.0, + 732.0, + 1482.0, + 708.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1458.0, + 1114.0, + 1458.0, + 1114.0, + 1479.0, + 1095.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1477.0, + 588.0, + 1477.0, + 588.0, + 1512.0, + 363.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1474.0, + 975.0, + 1474.0, + 975.0, + 1513.0, + 747.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1476.0, + 1358.0, + 1476.0, + 1358.0, + 1511.0, + 1132.0, + 1511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1515.0, + 557.0, + 1515.0, + 557.0, + 1563.0, + 373.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 1522.0, + 989.0, + 1522.0, + 989.0, + 1555.0, + 710.0, + 1555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1522.0, + 1359.0, + 1522.0, + 1359.0, + 1557.0, + 1110.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1570.0, + 1403.0, + 1570.0, + 1403.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 772.0, + 1601.0, + 772.0, + 1634.0, + 295.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 1601.0, + 1075.0, + 1601.0, + 1075.0, + 1634.0, + 825.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1601.0, + 1404.0, + 1601.0, + 1404.0, + 1634.0, + 1185.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1665.0, + 295.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1697.0, + 293.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1687.0, + 839.0, + 1687.0, + 839.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 197.0, + 1132.0, + 197.0, + 1132.0, + 244.0, + 292.0, + 244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 861.0, + 453.0, + 861.0, + 453.0, + 899.0, + 294.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 918.0, + 941.0, + 918.0, + 941.0, + 957.0, + 296.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 862.0, + 2060.0, + 862.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1768.0, + 1365.0, + 1768.0, + 1365.0, + 1812.0, + 294.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 260.0, + 1403.0, + 260.0, + 1403.0, + 293.0, + 297.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 287.0, + 384.0, + 287.0, + 384.0, + 324.0, + 297.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 287.0, + 1406.0, + 287.0, + 1406.0, + 324.0, + 413.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 319.0, + 1406.0, + 319.0, + 1406.0, + 356.0, + 295.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 348.0, + 1405.0, + 348.0, + 1405.0, + 385.0, + 293.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 379.0, + 1405.0, + 379.0, + 1405.0, + 416.0, + 293.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 955.0, + 1405.0, + 955.0, + 1405.0, + 988.0, + 296.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 899.0, + 984.0, + 899.0, + 1021.0, + 295.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 984.0, + 1403.0, + 984.0, + 1403.0, + 1021.0, + 929.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1005.0, + 1316.0, + 1005.0, + 1316.0, + 1060.0, + 288.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1005.0, + 1412.0, + 1005.0, + 1412.0, + 1060.0, + 1398.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1043.0, + 1254.0, + 1043.0, + 1254.0, + 1081.0, + 293.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1043.0, + 1406.0, + 1043.0, + 1406.0, + 1081.0, + 1305.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1077.0, + 701.0, + 1077.0, + 701.0, + 1110.0, + 296.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1121.0, + 1406.0, + 1121.0, + 1406.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1149.0, + 1406.0, + 1149.0, + 1406.0, + 1184.0, + 294.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1180.0, + 1408.0, + 1180.0, + 1408.0, + 1220.0, + 293.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1213.0, + 1406.0, + 1213.0, + 1406.0, + 1246.0, + 295.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1335.0, + 1241.0, + 1335.0, + 1278.0, + 294.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 703.0, + 1408.0, + 703.0, + 1408.0, + 741.0, + 292.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 735.0, + 919.0, + 735.0, + 919.0, + 771.0, + 292.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 735.0, + 1406.0, + 735.0, + 1406.0, + 771.0, + 950.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 767.0, + 407.0, + 767.0, + 407.0, + 801.0, + 293.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 767.0, + 466.0, + 767.0, + 466.0, + 801.0, + 463.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 766.0, + 1404.0, + 766.0, + 1404.0, + 800.0, + 469.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 796.0, + 1409.0, + 796.0, + 1409.0, + 833.0, + 292.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 477.0, + 788.0, + 477.0, + 788.0, + 516.0, + 360.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 477.0, + 1408.0, + 477.0, + 1408.0, + 516.0, + 869.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 508.0, + 667.0, + 508.0, + 667.0, + 546.0, + 392.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 508.0, + 725.0, + 508.0, + 725.0, + 546.0, + 718.0, + 546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 547.0, + 1409.0, + 547.0, + 1409.0, + 588.0, + 356.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 580.0, + 576.0, + 580.0, + 576.0, + 617.0, + 393.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 580.0, + 932.0, + 580.0, + 932.0, + 617.0, + 672.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 580.0, + 1279.0, + 580.0, + 1279.0, + 617.0, + 1021.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 577.0, + 1405.0, + 577.0, + 1405.0, + 616.0, + 1286.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 611.0, + 1175.0, + 611.0, + 1175.0, + 647.0, + 395.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 650.0, + 1289.0, + 650.0, + 1289.0, + 686.0, + 359.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 420.0, + 797.0, + 420.0, + 797.0, + 462.0, + 296.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1769.0, + 1366.0, + 1769.0, + 1366.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1810.0, + 1404.0, + 1810.0, + 1404.0, + 1843.0, + 296.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1835.0, + 883.0, + 1835.0, + 883.0, + 1880.0, + 291.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1835.0, + 1114.0, + 1835.0, + 1114.0, + 1880.0, + 925.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1835.0, + 1407.0, + 1835.0, + 1407.0, + 1880.0, + 1157.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1865.0, + 1385.0, + 1865.0, + 1385.0, + 1911.0, + 293.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 584.0, + 1910.0, + 584.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1954.0, + 675.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 981.0, + 1942.0, + 981.0, + 1984.0, + 293.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1984.0, + 1072.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2012.0, + 294.0, + 2012.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1541, + 1405, + 1541, + 1405, + 1818, + 298, + 1818 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1376, + 1406, + 1376, + 1406, + 1529, + 298, + 1529 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 465, + 196, + 1226, + 196, + 1226, + 326, + 465, + 326 + ], + "score": 0.974, + "html": "
PartnerFCPFCP-TFCP+AFCP-T,+A
Hproxy10.6± 0.54.7± 0.49.9±0.67.0±0.8
Diverse SP11.2 ± 0.16.9 ± 0.111.1 ± 0.48.6 ± 0.4
Random8.6± 0.21.0 ± 0.18.4±0.43.2 ± 0.5
" + }, + { + "category_id": 1, + "poly": [ + 297, + 755, + 1406, + 755, + 1406, + 846, + 297, + 846 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 298, + 579, + 1403, + 579, + 1403, + 643, + 298, + 643 + ], + "score": 0.95 + }, + { + "category_id": 0, + "poly": [ + 298, + 685, + 872, + 685, + 872, + 722, + 298, + 722 + ], + "score": 0.932 + }, + { + "category_id": 0, + "poly": [ + 296, + 1319, + 1188, + 1319, + 1188, + 1352, + 296, + 1352 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 298, + 1853, + 450, + 1853, + 450, + 1885, + 298, + 1885 + ], + "score": 0.9 + }, + { + "category_id": 4, + "poly": [ + 296, + 1139, + 1407, + 1139, + 1407, + 1263, + 296, + 1263 + ], + "score": 0.9 + }, + { + "category_id": 1, + "poly": [ + 293, + 1909, + 1407, + 1909, + 1407, + 2008, + 293, + 2008 + ], + "score": 0.862 + }, + { + "category_id": 3, + "poly": [ + 255, + 875, + 1449, + 875, + 1449, + 1112, + 255, + 1112 + ], + "score": 0.814 + }, + { + "category_id": 2, + "poly": [ + 842, + 2061, + 858, + 2061, + 858, + 2084, + 842, + 2084 + ], + "score": 0.7 + }, + { + "category_id": 1, + "poly": [ + 296, + 343, + 1405, + 343, + 1405, + 530, + 296, + 530 + ], + "score": 0.614 + }, + { + "category_id": 6, + "poly": [ + 296, + 343, + 1405, + 343, + 1405, + 530, + 296, + 530 + ], + "score": 0.331 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2084, + 841, + 2084 + ], + "score": 0.142 + }, + { + "category_id": 3, + "poly": [ + 271, + 968, + 577, + 968, + 577, + 1116, + 271, + 1116 + ], + "score": 0.098 + }, + { + "category_id": 1, + "poly": [ + 296, + 1139, + 1407, + 1139, + 1407, + 1263, + 296, + 1263 + ], + "score": 0.093 + }, + { + "category_id": 13, + "poly": [ + 1126, + 497, + 1375, + 497, + 1375, + 530, + 1126, + 530 + ], + "score": 0.91, + "latex": "( \\mathrm { F C P } _ { - T , + A } > \\mathrm { F C P } _ { - T , }" + }, + { + "category_id": 13, + "poly": [ + 652, + 468, + 831, + 468, + 831, + 499, + 652, + 499 + ], + "score": 0.9, + "latex": "( \\mathrm { F C P } \\approx \\mathrm { F C P } _ { + A . }" + }, + { + "category_id": 13, + "poly": [ + 681, + 436, + 850, + 436, + 850, + 467, + 681, + 467 + ], + "score": 0.89, + "latex": "( \\mathrm { F C P } > \\mathrm { F C P } _ { - T }" + }, + { + "category_id": 13, + "poly": [ + 584, + 1634, + 693, + 1634, + 693, + 1662, + 584, + 1662 + ], + "score": 0.89, + "latex": "T = 3 0 0" + }, + { + "category_id": 13, + "poly": [ + 857, + 582, + 982, + 582, + 982, + 614, + 857, + 614 + ], + "score": 0.88, + "latex": "\\mathrm { F C P } _ { - T , + A }" + }, + { + "category_id": 13, + "poly": [ + 719, + 582, + 806, + 582, + 806, + 612, + 719, + 612 + ], + "score": 0.87, + "latex": "\\mathrm { F C P } _ { - T }" + }, + { + "category_id": 13, + "poly": [ + 914, + 1407, + 1020, + 1407, + 1020, + 1436, + 914, + 1436 + ], + "score": 0.87, + "latex": "N = 1 1 4" + }, + { + "category_id": 13, + "poly": [ + 1198, + 1407, + 1272, + 1407, + 1272, + 1436, + 1198, + 1436 + ], + "score": 0.86, + "latex": "5 9 . 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1344, + 1407, + 1403, + 1407, + 1403, + 1436, + 1344, + 1436 + ], + "score": 0.85, + "latex": "1 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1032, + 1407, + 1107, + 1407, + 1107, + 1436, + 1032, + 1436 + ], + "score": 0.83, + "latex": "3 7 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 827, + 1572, + 903, + 1572, + 903, + 1609, + 827, + 1609 + ], + "score": 0.78, + "latex": "\\dot { \\mathrm { \\bf D } } . 1 . 1" + }, + { + "category_id": 13, + "poly": [ + 444, + 377, + 467, + 377, + 467, + 404, + 444, + 404 + ], + "score": 0.63, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1054, + 377, + 1093, + 377, + 1093, + 406, + 1054, + 406 + ], + "score": 0.49, + "latex": "( A )" + }, + { + "category_id": 13, + "poly": [ + 445, + 1405, + 536, + 1405, + 536, + 1438, + 445, + 1438 + ], + "score": 0.31, + "latex": "\\mathbb { 1 1 8 } , \\lvert 5 5 \\rvert" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 682.0, + 874.0, + 682.0, + 874.0, + 728.0, + 291.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1315.0, + 1192.0, + 1315.0, + 1192.0, + 1360.0, + 293.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1851.0, + 453.0, + 1851.0, + 453.0, + 1889.0, + 294.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1139.0, + 1407.0, + 1139.0, + 1407.0, + 1175.0, + 295.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 1405.0, + 1168.0, + 1405.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1200.0, + 1409.0, + 1200.0, + 1409.0, + 1236.0, + 293.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1230.0, + 1386.0, + 1230.0, + 1386.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 251.0, + 880.0, + 588.0, + 880.0, + 588.0, + 905.0, + 251.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 879.0, + 1046.0, + 879.0, + 1046.0, + 904.0, + 685.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 883.0, + 1448.0, + 883.0, + 1448.0, + 908.0, + 1178.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 263.0, + 901.0, + 568.0, + 901.0, + 568.0, + 926.0, + 263.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 903.0, + 1034.0, + 903.0, + 1034.0, + 928.0, + 696.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 901.0, + 1453.0, + 901.0, + 1453.0, + 926.0, + 1170.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 949.0, + 810.0, + 949.0, + 810.0, + 974.0, + 677.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 949.0, + 1047.0, + 949.0, + 1047.0, + 974.0, + 911.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 993.0, + 374.0, + 993.0, + 374.0, + 1019.0, + 316.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 996.0, + 445.0, + 996.0, + 445.0, + 1016.0, + 421.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 986.0, + 524.0, + 986.0, + 524.0, + 1004.0, + 496.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 993.0, + 555.0, + 993.0, + 555.0, + 1013.0, + 531.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 983.0, + 740.0, + 983.0, + 740.0, + 992.0, + 725.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 980.0, + 766.0, + 980.0, + 766.0, + 995.0, + 746.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 982.0, + 1002.0, + 982.0, + 1002.0, + 995.0, + 961.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 974.0, + 1336.0, + 974.0, + 1336.0, + 1026.0, + 1207.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1028.0, + 1335.0, + 1028.0, + 1335.0, + 1053.0, + 1224.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1053.0, + 1336.0, + 1053.0, + 1336.0, + 1079.0, + 1269.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1071.0, + 351.0, + 1071.0, + 351.0, + 1090.0, + 318.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1071.0, + 453.0, + 1071.0, + 453.0, + 1090.0, + 423.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1079.0, + 1336.0, + 1079.0, + 1336.0, + 1108.0, + 1207.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1071.0, + 1398.0, + 1071.0, + 1398.0, + 1096.0, + 1368.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 1067.0, + 575.0, + 1067.0, + 575.0, + 1095.0, + 533.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 340.0, + 1407.0, + 340.0, + 1407.0, + 379.0, + 293.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 376.0, + 443.0, + 376.0, + 443.0, + 408.0, + 298.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 376.0, + 1053.0, + 376.0, + 1053.0, + 408.0, + 468.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 376.0, + 1405.0, + 376.0, + 1405.0, + 408.0, + 1094.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 404.0, + 1405.0, + 404.0, + 1405.0, + 440.0, + 294.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 435.0, + 680.0, + 435.0, + 680.0, + 471.0, + 294.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 435.0, + 1405.0, + 435.0, + 1405.0, + 471.0, + 851.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 466.0, + 651.0, + 466.0, + 651.0, + 502.0, + 295.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 466.0, + 1405.0, + 466.0, + 1405.0, + 502.0, + 832.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 497.0, + 1125.0, + 497.0, + 1125.0, + 533.0, + 294.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 497.0, + 1393.0, + 497.0, + 1393.0, + 533.0, + 1376.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2092.0, + 840.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 993.0, + 374.0, + 993.0, + 374.0, + 1018.0, + 315.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 996.0, + 444.0, + 996.0, + 444.0, + 1016.0, + 421.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 994.0, + 555.0, + 994.0, + 555.0, + 1013.0, + 531.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1013.0, + 301.0, + 1013.0, + 301.0, + 1021.0, + 293.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1070.0, + 353.0, + 1070.0, + 353.0, + 1090.0, + 315.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1072.0, + 454.0, + 1072.0, + 454.0, + 1089.0, + 420.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 1070.0, + 573.0, + 1070.0, + 573.0, + 1093.0, + 530.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1541.0, + 1403.0, + 1541.0, + 1403.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1572.0, + 826.0, + 1572.0, + 826.0, + 1608.0, + 295.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1572.0, + 1407.0, + 1572.0, + 1407.0, + 1608.0, + 904.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1601.0, + 1405.0, + 1601.0, + 1405.0, + 1639.0, + 292.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 583.0, + 1632.0, + 583.0, + 1669.0, + 293.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1632.0, + 1403.0, + 1632.0, + 1403.0, + 1669.0, + 694.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 1407.0, + 1662.0, + 1407.0, + 1702.0, + 292.0, + 1702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1694.0, + 1406.0, + 1694.0, + 1406.0, + 1731.0, + 295.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1724.0, + 1407.0, + 1724.0, + 1407.0, + 1761.0, + 295.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1754.0, + 1405.0, + 1754.0, + 1405.0, + 1791.0, + 293.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1783.0, + 1323.0, + 1783.0, + 1323.0, + 1821.0, + 293.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1372.0, + 1406.0, + 1372.0, + 1406.0, + 1412.0, + 293.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1405.0, + 444.0, + 1405.0, + 444.0, + 1439.0, + 296.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1405.0, + 913.0, + 1405.0, + 913.0, + 1439.0, + 537.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 1405.0, + 1031.0, + 1405.0, + 1031.0, + 1439.0, + 1021.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 1405.0, + 1197.0, + 1405.0, + 1197.0, + 1439.0, + 1108.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1405.0, + 1343.0, + 1405.0, + 1343.0, + 1439.0, + 1273.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1435.0, + 1408.0, + 1435.0, + 1408.0, + 1472.0, + 295.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1464.0, + 1408.0, + 1464.0, + 1408.0, + 1504.0, + 293.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1493.0, + 1369.0, + 1493.0, + 1369.0, + 1537.0, + 292.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 754.0, + 1406.0, + 754.0, + 1406.0, + 790.0, + 294.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 786.0, + 1406.0, + 786.0, + 1406.0, + 820.0, + 294.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 816.0, + 593.0, + 816.0, + 593.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 576.0, + 718.0, + 576.0, + 718.0, + 619.0, + 293.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 576.0, + 856.0, + 576.0, + 856.0, + 619.0, + 807.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 576.0, + 1407.0, + 576.0, + 1407.0, + 619.0, + 983.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 610.0, + 1305.0, + 610.0, + 1305.0, + 648.0, + 294.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1906.0, + 1327.0, + 1906.0, + 1327.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1409.0, + 1942.0, + 1409.0, + 1985.0, + 292.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1973.0, + 1409.0, + 1973.0, + 1409.0, + 2013.0, + 292.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 340.0, + 1407.0, + 340.0, + 1407.0, + 379.0, + 293.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 376.0, + 443.0, + 376.0, + 443.0, + 408.0, + 298.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 376.0, + 1053.0, + 376.0, + 1053.0, + 408.0, + 468.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 376.0, + 1405.0, + 376.0, + 1405.0, + 408.0, + 1094.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 404.0, + 1405.0, + 404.0, + 1405.0, + 440.0, + 294.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 435.0, + 680.0, + 435.0, + 680.0, + 471.0, + 294.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 435.0, + 1405.0, + 435.0, + 1405.0, + 471.0, + 851.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 466.0, + 651.0, + 466.0, + 651.0, + 502.0, + 295.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 466.0, + 1405.0, + 466.0, + 1405.0, + 502.0, + 832.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 497.0, + 1125.0, + 497.0, + 1125.0, + 533.0, + 294.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 497.0, + 1393.0, + 497.0, + 1393.0, + 533.0, + 1376.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1139.0, + 1407.0, + 1139.0, + 1407.0, + 1175.0, + 295.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 1405.0, + 1168.0, + 1405.0, + 1205.0, + 294.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1200.0, + 1409.0, + 1200.0, + 1409.0, + 1236.0, + 293.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1230.0, + 1386.0, + 1230.0, + 1386.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 202, + 1405, + 202, + 1405, + 387, + 297, + 387 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1886, + 1406, + 1886, + 1406, + 2008, + 299, + 2008 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 455, + 1406, + 455, + 1406, + 607, + 297, + 607 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 295, + 1226, + 1404, + 1226, + 1404, + 1290, + 295, + 1290 + ], + "score": 0.949 + }, + { + "category_id": 3, + "poly": [ + 298, + 640, + 1402, + 640, + 1402, + 978, + 298, + 978 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 297, + 418, + 927, + 418, + 927, + 450, + 297, + 450 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 299, + 1171, + 741, + 1171, + 741, + 1203, + 299, + 1203 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 299, + 1849, + 1133, + 1849, + 1133, + 1881, + 299, + 1881 + ], + "score": 0.912 + }, + { + "category_id": 4, + "poly": [ + 295, + 993, + 1407, + 993, + 1407, + 1122, + 295, + 1122 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 858, + 2061, + 858, + 2085, + 840, + 2085 + ], + "score": 0.806 + }, + { + "category_id": 1, + "poly": [ + 295, + 1672, + 1407, + 1672, + 1407, + 1827, + 295, + 1827 + ], + "score": 0.743 + }, + { + "category_id": 3, + "poly": [ + 384, + 1328, + 1317, + 1328, + 1317, + 1659, + 384, + 1659 + ], + "score": 0.519 + }, + { + "category_id": 4, + "poly": [ + 295, + 1672, + 1407, + 1672, + 1407, + 1827, + 295, + 1827 + ], + "score": 0.436 + }, + { + "category_id": 13, + "poly": [ + 394, + 576, + 511, + 576, + 511, + 607, + 394, + 607 + ], + "score": 0.9, + "latex": "p = 0 . 0 0 3" + }, + { + "category_id": 13, + "poly": [ + 787, + 356, + 903, + 356, + 903, + 387, + 787, + 387 + ], + "score": 0.9, + "latex": "p < 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 490, + 514, + 598, + 514, + 598, + 545, + 490, + 545 + ], + "score": 0.9, + "latex": "p < 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 650, + 1024, + 752, + 1024, + 752, + 1053, + 650, + 1053 + ], + "score": 0.9, + "latex": "T = 3 0 0" + }, + { + "category_id": 13, + "poly": [ + 679, + 234, + 797, + 234, + 797, + 265, + 679, + 265 + ], + "score": 0.9, + "latex": "p < 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 523, + 326, + 640, + 326, + 640, + 356, + 523, + 356 + ], + "score": 0.88, + "latex": "p < 0 . 0 { \\overline { { 0 1 } } }" + }, + { + "category_id": 13, + "poly": [ + 998, + 1024, + 1052, + 1024, + 1052, + 1053, + 998, + 1053 + ], + "score": 0.87, + "latex": "9 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1147, + 1672, + 1201, + 1672, + 1201, + 1703, + 1147, + 1703 + ], + "score": 0.85, + "latex": "3 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 898, + 947, + 980, + 947, + 980, + 976, + 898, + 976 + ], + "score": 0.85, + "latex": "\\mathrm { F C P } _ { - T . }" + }, + { + "category_id": 13, + "poly": [ + 689, + 513, + 731, + 513, + 731, + 549, + 689, + 549 + ], + "score": 0.83, + "latex": "\\dot { \\bigtriangledown } \\dot { \\mathbf { c } } { \\big \\rVert }" + }, + { + "category_id": 13, + "poly": [ + 727, + 323, + 769, + 323, + 769, + 359, + 727, + 359 + ], + "score": 0.82, + "latex": "\\textcircled { 7 6 }" + }, + { + "category_id": 13, + "poly": [ + 885, + 232, + 925, + 232, + 925, + 264, + 885, + 264 + ], + "score": 0.81, + "latex": "\\mathrm { 7 a ) }" + }, + { + "category_id": 13, + "poly": [ + 608, + 293, + 636, + 293, + 636, + 327, + 608, + 327 + ], + "score": 0.63, + "latex": "\\perp )" + }, + { + "category_id": 13, + "poly": [ + 888, + 264, + 942, + 264, + 942, + 298, + 888, + 298 + ], + "score": 0.6, + "latex": "\\mathbf { D } . 3 )" + }, + { + "category_id": 13, + "poly": [ + 1042, + 323, + 1090, + 323, + 1090, + 356, + 1042, + 356 + ], + "score": 0.58, + "latex": "[ \\left[ 1 2 \\right] ]" + }, + { + "category_id": 13, + "poly": [ + 1128, + 1882, + 1170, + 1882, + 1170, + 1920, + 1128, + 1920 + ], + "score": 0.27, + "latex": "\\textcircled { 8 \\mathrm { a } }" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 653.0, + 1127.0, + 653.0, + 1127.0, + 685.0, + 1082.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 658.0, + 1219.0, + 658.0, + 1219.0, + 681.0, + 1175.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 660.0, + 1264.0, + 660.0, + 1264.0, + 681.0, + 1224.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 659.0, + 1359.0, + 659.0, + 1359.0, + 681.0, + 1270.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 688.0, + 366.0, + 688.0, + 366.0, + 839.0, + 302.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 691.0, + 817.0, + 691.0, + 817.0, + 835.0, + 757.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 701.0, + 1175.0, + 701.0, + 1175.0, + 732.0, + 1081.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 703.0, + 1266.0, + 703.0, + 1266.0, + 729.0, + 1222.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 703.0, + 1313.0, + 703.0, + 1313.0, + 729.0, + 1269.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 706.0, + 1359.0, + 706.0, + 1359.0, + 728.0, + 1317.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 749.0, + 1221.0, + 749.0, + 1221.0, + 777.0, + 1093.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 751.0, + 1313.0, + 751.0, + 1313.0, + 777.0, + 1269.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 752.0, + 1360.0, + 752.0, + 1360.0, + 774.0, + 1318.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 791.0, + 803.0, + 814.0, + 803.0, + 814.0, + 834.0, + 791.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 791.0, + 1268.0, + 791.0, + 1268.0, + 826.0, + 1067.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1315.0, + 797.0, + 1359.0, + 797.0, + 1359.0, + 823.0, + 1315.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 840.0, + 360.0, + 840.0, + 360.0, + 870.0, + 337.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 826.0, + 816.0, + 826.0, + 816.0, + 870.0, + 790.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 841.0, + 1176.0, + 841.0, + 1176.0, + 870.0, + 1092.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 844.0, + 1221.0, + 844.0, + 1221.0, + 866.0, + 1177.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 843.0, + 1315.0, + 843.0, + 1315.0, + 867.0, + 1222.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 867.0, + 430.0, + 867.0, + 430.0, + 905.0, + 383.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 869.0, + 496.0, + 869.0, + 496.0, + 902.0, + 454.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 510.0, + 866.0, + 642.0, + 866.0, + 642.0, + 905.0, + 510.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 865.0, + 954.0, + 865.0, + 954.0, + 905.0, + 812.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 882.0, + 1221.0, + 882.0, + 1221.0, + 907.0, + 1131.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 881.0, + 1355.0, + 881.0, + 1355.0, + 910.0, + 1226.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 915.0, + 696.0, + 915.0, + 696.0, + 953.0, + 292.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 916.0, + 1039.0, + 916.0, + 1039.0, + 951.0, + 715.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 918.0, + 1406.0, + 918.0, + 1406.0, + 950.0, + 1063.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 941.0, + 513.0, + 941.0, + 513.0, + 980.0, + 291.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 945.0, + 897.0, + 945.0, + 897.0, + 979.0, + 714.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 945.0, + 999.0, + 945.0, + 999.0, + 979.0, + 981.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 946.0, + 1358.0, + 946.0, + 1358.0, + 978.0, + 1063.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 416.0, + 927.0, + 416.0, + 927.0, + 456.0, + 296.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1169.0, + 744.0, + 1169.0, + 744.0, + 1209.0, + 295.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1847.0, + 1133.0, + 1847.0, + 1133.0, + 1886.0, + 295.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 992.0, + 1410.0, + 992.0, + 1410.0, + 1028.0, + 293.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1024.0, + 649.0, + 1024.0, + 649.0, + 1055.0, + 294.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1024.0, + 997.0, + 1024.0, + 997.0, + 1055.0, + 753.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1055.0, + 1053.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 1406.0, + 1054.0, + 1406.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1081.0, + 600.0, + 1081.0, + 600.0, + 1125.0, + 293.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2058.0, + 859.0, + 2058.0, + 859.0, + 2089.0, + 836.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1337.0, + 484.0, + 1337.0, + 484.0, + 1370.0, + 404.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1336.0, + 541.0, + 1336.0, + 541.0, + 1368.0, + 492.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1337.0, + 602.0, + 1337.0, + 602.0, + 1366.0, + 557.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1339.0, + 665.0, + 1339.0, + 665.0, + 1366.0, + 621.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1339.0, + 729.0, + 1339.0, + 729.0, + 1366.0, + 684.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1337.0, + 791.0, + 1337.0, + 791.0, + 1366.0, + 747.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1344.0, + 991.0, + 1344.0, + 991.0, + 1370.0, + 894.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 1340.0, + 1051.0, + 1340.0, + 1051.0, + 1371.0, + 1003.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1340.0, + 1113.0, + 1340.0, + 1113.0, + 1372.0, + 1064.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1340.0, + 1175.0, + 1340.0, + 1175.0, + 1371.0, + 1128.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1342.0, + 1237.0, + 1342.0, + 1237.0, + 1369.0, + 1193.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1342.0, + 1300.0, + 1342.0, + 1300.0, + 1370.0, + 1256.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1379.0, + 483.0, + 1379.0, + 483.0, + 1406.0, + 382.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1376.0, + 541.0, + 1376.0, + 541.0, + 1407.0, + 493.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1376.0, + 603.0, + 1376.0, + 603.0, + 1407.0, + 556.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1378.0, + 665.0, + 1378.0, + 665.0, + 1406.0, + 620.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1378.0, + 730.0, + 1378.0, + 730.0, + 1406.0, + 684.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1378.0, + 792.0, + 1378.0, + 792.0, + 1406.0, + 747.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 1387.0, + 994.0, + 1387.0, + 994.0, + 1424.0, + 946.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1388.0, + 1051.0, + 1388.0, + 1051.0, + 1419.0, + 1002.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1390.0, + 1112.0, + 1390.0, + 1112.0, + 1417.0, + 1067.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1390.0, + 1174.0, + 1390.0, + 1174.0, + 1417.0, + 1129.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1390.0, + 1238.0, + 1390.0, + 1238.0, + 1417.0, + 1194.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1390.0, + 1301.0, + 1390.0, + 1301.0, + 1418.0, + 1257.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1417.0, + 486.0, + 1417.0, + 486.0, + 1451.0, + 435.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1418.0, + 538.0, + 1418.0, + 538.0, + 1446.0, + 494.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1418.0, + 602.0, + 1418.0, + 602.0, + 1445.0, + 557.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1419.0, + 664.0, + 1419.0, + 664.0, + 1445.0, + 622.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1418.0, + 729.0, + 1418.0, + 729.0, + 1445.0, + 684.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1416.0, + 793.0, + 1416.0, + 793.0, + 1447.0, + 746.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1437.0, + 994.0, + 1437.0, + 994.0, + 1470.0, + 932.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1436.0, + 1050.0, + 1436.0, + 1050.0, + 1468.0, + 1002.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1439.0, + 1112.0, + 1439.0, + 1112.0, + 1465.0, + 1067.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1439.0, + 1175.0, + 1439.0, + 1175.0, + 1465.0, + 1129.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1439.0, + 1238.0, + 1439.0, + 1238.0, + 1467.0, + 1193.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1439.0, + 1300.0, + 1439.0, + 1300.0, + 1465.0, + 1256.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1460.0, + 482.0, + 1460.0, + 482.0, + 1486.0, + 422.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 1458.0, + 537.0, + 1458.0, + 537.0, + 1487.0, + 494.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 1459.0, + 602.0, + 1459.0, + 602.0, + 1487.0, + 558.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1459.0, + 665.0, + 1459.0, + 665.0, + 1487.0, + 621.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1459.0, + 729.0, + 1459.0, + 729.0, + 1487.0, + 685.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1458.0, + 791.0, + 1458.0, + 791.0, + 1487.0, + 748.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1500.0, + 484.0, + 1500.0, + 484.0, + 1527.0, + 421.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1497.0, + 541.0, + 1497.0, + 541.0, + 1528.0, + 493.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1497.0, + 603.0, + 1497.0, + 603.0, + 1528.0, + 556.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1499.0, + 666.0, + 1499.0, + 666.0, + 1527.0, + 621.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1499.0, + 729.0, + 1499.0, + 729.0, + 1526.0, + 684.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1498.0, + 792.0, + 1498.0, + 792.0, + 1526.0, + 747.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1489.0, + 991.0, + 1489.0, + 991.0, + 1515.0, + 930.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1487.0, + 1048.0, + 1487.0, + 1048.0, + 1514.0, + 1004.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 1485.0, + 1113.0, + 1485.0, + 1113.0, + 1516.0, + 1066.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1488.0, + 1173.0, + 1488.0, + 1173.0, + 1514.0, + 1131.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1487.0, + 1238.0, + 1487.0, + 1238.0, + 1514.0, + 1194.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1487.0, + 1300.0, + 1487.0, + 1300.0, + 1514.0, + 1256.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1538.0, + 483.0, + 1538.0, + 483.0, + 1570.0, + 401.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 1536.0, + 541.0, + 1536.0, + 541.0, + 1568.0, + 493.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1538.0, + 602.0, + 1538.0, + 602.0, + 1566.0, + 557.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1538.0, + 665.0, + 1538.0, + 665.0, + 1566.0, + 621.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 1538.0, + 729.0, + 1538.0, + 729.0, + 1566.0, + 684.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1538.0, + 792.0, + 1538.0, + 792.0, + 1566.0, + 746.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1534.0, + 991.0, + 1534.0, + 991.0, + 1567.0, + 912.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1531.0, + 1049.0, + 1531.0, + 1049.0, + 1564.0, + 1005.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1535.0, + 1111.0, + 1535.0, + 1111.0, + 1563.0, + 1067.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1537.0, + 1173.0, + 1537.0, + 1173.0, + 1561.0, + 1130.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1535.0, + 1238.0, + 1535.0, + 1238.0, + 1562.0, + 1194.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1534.0, + 1301.0, + 1534.0, + 1301.0, + 1562.0, + 1256.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1574.0, + 550.0, + 1574.0, + 550.0, + 1605.0, + 485.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 1573.0, + 597.0, + 1573.0, + 597.0, + 1603.0, + 563.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1575.0, + 659.0, + 1575.0, + 659.0, + 1602.0, + 628.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1574.0, + 729.0, + 1574.0, + 729.0, + 1602.0, + 685.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1574.0, + 790.0, + 1574.0, + 790.0, + 1603.0, + 748.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1575.0, + 1059.0, + 1575.0, + 1059.0, + 1602.0, + 994.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1072.0, + 1573.0, + 1107.0, + 1573.0, + 1107.0, + 1603.0, + 1072.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1573.0, + 1169.0, + 1573.0, + 1169.0, + 1603.0, + 1136.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1194.0, + 1574.0, + 1237.0, + 1574.0, + 1237.0, + 1603.0, + 1194.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1574.0, + 1299.0, + 1574.0, + 1299.0, + 1603.0, + 1257.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1619.0, + 799.0, + 1619.0, + 799.0, + 1662.0, + 389.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1623.0, + 1275.0, + 1623.0, + 1275.0, + 1657.0, + 932.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 1146.0, + 1671.0, + 1146.0, + 1708.0, + 294.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1671.0, + 1409.0, + 1671.0, + 1409.0, + 1708.0, + 1202.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1703.0, + 1406.0, + 1703.0, + 1406.0, + 1739.0, + 293.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1734.0, + 1406.0, + 1734.0, + 1406.0, + 1769.0, + 293.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1759.0, + 1407.0, + 1759.0, + 1407.0, + 1803.0, + 292.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1793.0, + 1355.0, + 1793.0, + 1355.0, + 1831.0, + 297.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 204.0, + 1405.0, + 204.0, + 1405.0, + 236.0, + 296.0, + 236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 678.0, + 231.0, + 678.0, + 270.0, + 295.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 231.0, + 884.0, + 231.0, + 884.0, + 270.0, + 798.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 231.0, + 1406.0, + 231.0, + 1406.0, + 270.0, + 926.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 263.0, + 887.0, + 263.0, + 887.0, + 298.0, + 296.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 263.0, + 1405.0, + 263.0, + 1405.0, + 298.0, + 943.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 607.0, + 292.0, + 607.0, + 331.0, + 294.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 292.0, + 1406.0, + 292.0, + 1406.0, + 331.0, + 637.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 324.0, + 522.0, + 324.0, + 522.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 324.0, + 726.0, + 324.0, + 726.0, + 358.0, + 641.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 324.0, + 1041.0, + 324.0, + 1041.0, + 358.0, + 770.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 324.0, + 1406.0, + 324.0, + 1406.0, + 358.0, + 1091.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 786.0, + 352.0, + 786.0, + 390.0, + 294.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 352.0, + 915.0, + 352.0, + 915.0, + 390.0, + 904.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1883.0, + 1127.0, + 1883.0, + 1127.0, + 1920.0, + 293.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1883.0, + 1407.0, + 1883.0, + 1407.0, + 1920.0, + 1171.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1411.0, + 1912.0, + 1411.0, + 1949.0, + 293.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1946.0, + 1404.0, + 1946.0, + 1404.0, + 1978.0, + 296.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1976.0, + 1408.0, + 1976.0, + 1408.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 452.0, + 1411.0, + 452.0, + 1411.0, + 493.0, + 294.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 481.0, + 1406.0, + 481.0, + 1406.0, + 521.0, + 294.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 511.0, + 489.0, + 511.0, + 489.0, + 552.0, + 294.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 511.0, + 688.0, + 511.0, + 688.0, + 552.0, + 599.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 511.0, + 1407.0, + 511.0, + 1407.0, + 552.0, + 732.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 542.0, + 1407.0, + 542.0, + 1407.0, + 582.0, + 294.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 574.0, + 393.0, + 574.0, + 393.0, + 611.0, + 295.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 574.0, + 1181.0, + 574.0, + 1181.0, + 611.0, + 512.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1404.0, + 1223.0, + 1404.0, + 1262.0, + 294.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1257.0, + 1215.0, + 1257.0, + 1215.0, + 1293.0, + 295.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 1146.0, + 1671.0, + 1146.0, + 1708.0, + 294.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1671.0, + 1409.0, + 1671.0, + 1409.0, + 1708.0, + 1202.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1703.0, + 1406.0, + 1703.0, + 1406.0, + 1739.0, + 293.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1734.0, + 1406.0, + 1734.0, + 1406.0, + 1769.0, + 293.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1759.0, + 1407.0, + 1759.0, + 1407.0, + 1803.0, + 292.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1793.0, + 1355.0, + 1793.0, + 1355.0, + 1831.0, + 297.0, + 1831.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1005, + 1405, + 1005, + 1405, + 1280, + 298, + 1280 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1292, + 1404, + 1292, + 1404, + 1568, + 298, + 1568 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 777, + 1404, + 777, + 1404, + 991, + 297, + 991 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 535, + 1404, + 535, + 1404, + 688, + 298, + 688 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1716, + 1403, + 1716, + 1403, + 1902, + 298, + 1902 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1582, + 1403, + 1582, + 1403, + 1704, + 298, + 1704 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 399, + 1404, + 399, + 1404, + 521, + 299, + 521 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 203, + 1405, + 203, + 1405, + 326, + 296, + 326 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1914, + 1402, + 1914, + 1402, + 2007, + 300, + 2007 + ], + "score": 0.97 + }, + { + "category_id": 0, + "poly": [ + 298, + 722, + 500, + 722, + 500, + 759, + 298, + 759 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 297, + 361, + 1236, + 361, + 1236, + 393, + 297, + 393 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2084, + 840, + 2084 + ], + "score": 0.796 + }, + { + "category_id": 13, + "poly": [ + 841, + 1097, + 935, + 1097, + 935, + 1125, + 841, + 1125 + ], + "score": 0.9, + "latex": "N = 3 2" + }, + { + "category_id": 13, + "poly": [ + 642, + 595, + 719, + 595, + 719, + 625, + 642, + 625 + ], + "score": 0.88, + "latex": "5 5 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 822, + 1216, + 871, + 1216, + 871, + 1249, + 822, + 1249 + ], + "score": 0.8, + "latex": "\\mathbb { \\lVert 6 9 \\rVert }" + }, + { + "category_id": 13, + "poly": [ + 914, + 1504, + 962, + 1504, + 962, + 1536, + 914, + 1536 + ], + "score": 0.76, + "latex": "\\overline { { [ 3 8 ] } }" + }, + { + "category_id": 13, + "poly": [ + 676, + 428, + 717, + 428, + 717, + 464, + 676, + 464 + ], + "score": 0.55, + "latex": "\\textcircled { 8 6 }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1610, + 1149, + 1610, + 1149, + 1643, + 1100, + 1643 + ], + "score": 0.5, + "latex": "\\textcircled { \\lvert 3 8 \\rvert }" + }, + { + "category_id": 13, + "poly": [ + 642, + 1383, + 761, + 1383, + 761, + 1416, + 642, + 1416 + ], + "score": 0.46, + "latex": "\\textcircled { 1 1 } \\textcircled { 3 3 } \\textcircled { 5 9 }" + }, + { + "category_id": 13, + "poly": [ + 932, + 1535, + 980, + 1535, + 980, + 1567, + 932, + 1567 + ], + "score": 0.39, + "latex": "\\pmb { \\mathbb { Z } } 2 \\mathbf { l }" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 718.0, + 505.0, + 718.0, + 505.0, + 765.0, + 291.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 359.0, + 1239.0, + 359.0, + 1239.0, + 399.0, + 295.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 860.0, + 2060.0, + 860.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1003.0, + 1405.0, + 1003.0, + 1405.0, + 1041.0, + 293.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1037.0, + 1403.0, + 1037.0, + 1403.0, + 1070.0, + 296.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1066.0, + 1405.0, + 1066.0, + 1405.0, + 1102.0, + 293.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1092.0, + 840.0, + 1092.0, + 840.0, + 1135.0, + 291.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1092.0, + 1410.0, + 1092.0, + 1410.0, + 1135.0, + 936.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1127.0, + 1405.0, + 1127.0, + 1405.0, + 1163.0, + 295.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1158.0, + 1402.0, + 1158.0, + 1402.0, + 1191.0, + 296.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1185.0, + 1405.0, + 1185.0, + 1405.0, + 1225.0, + 293.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1215.0, + 821.0, + 1215.0, + 821.0, + 1254.0, + 293.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 1215.0, + 1405.0, + 1215.0, + 1405.0, + 1254.0, + 872.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 781.0, + 1247.0, + 781.0, + 1282.0, + 293.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1296.0, + 1404.0, + 1296.0, + 1404.0, + 1324.0, + 298.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1323.0, + 1408.0, + 1323.0, + 1408.0, + 1360.0, + 293.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1356.0, + 1406.0, + 1356.0, + 1406.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1383.0, + 641.0, + 1383.0, + 641.0, + 1420.0, + 293.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1383.0, + 1406.0, + 1383.0, + 1406.0, + 1420.0, + 762.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1413.0, + 1406.0, + 1413.0, + 1406.0, + 1448.0, + 292.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1446.0, + 1405.0, + 1446.0, + 1405.0, + 1476.0, + 294.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1474.0, + 1406.0, + 1474.0, + 1406.0, + 1510.0, + 293.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1506.0, + 913.0, + 1506.0, + 913.0, + 1541.0, + 293.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1506.0, + 1401.0, + 1506.0, + 1401.0, + 1541.0, + 963.0, + 1541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1537.0, + 931.0, + 1537.0, + 931.0, + 1570.0, + 294.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 779.0, + 1404.0, + 779.0, + 1404.0, + 813.0, + 295.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 809.0, + 1405.0, + 809.0, + 1405.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 837.0, + 1407.0, + 837.0, + 1407.0, + 875.0, + 294.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 905.0, + 295.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 1406.0, + 899.0, + 1406.0, + 936.0, + 292.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 928.0, + 1405.0, + 928.0, + 1405.0, + 967.0, + 292.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 961.0, + 692.0, + 961.0, + 692.0, + 996.0, + 295.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 532.0, + 1405.0, + 532.0, + 1405.0, + 569.0, + 294.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 563.0, + 1404.0, + 563.0, + 1404.0, + 601.0, + 295.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 594.0, + 641.0, + 594.0, + 641.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 594.0, + 1406.0, + 594.0, + 1406.0, + 631.0, + 720.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 626.0, + 1406.0, + 626.0, + 1406.0, + 663.0, + 292.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 655.0, + 1228.0, + 655.0, + 1228.0, + 694.0, + 292.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1752.0, + 293.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1743.0, + 1407.0, + 1743.0, + 1407.0, + 1786.0, + 292.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1813.0, + 293.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1810.0, + 1405.0, + 1810.0, + 1405.0, + 1842.0, + 296.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1840.0, + 1405.0, + 1840.0, + 1405.0, + 1874.0, + 293.0, + 1874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1869.0, + 1350.0, + 1869.0, + 1350.0, + 1905.0, + 294.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1408.0, + 1581.0, + 1408.0, + 1616.0, + 294.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1611.0, + 1099.0, + 1611.0, + 1099.0, + 1646.0, + 293.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1611.0, + 1404.0, + 1611.0, + 1404.0, + 1646.0, + 1150.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1641.0, + 1407.0, + 1641.0, + 1407.0, + 1680.0, + 292.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 1181.0, + 1674.0, + 1181.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 399.0, + 1403.0, + 399.0, + 1403.0, + 431.0, + 295.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 426.0, + 675.0, + 426.0, + 675.0, + 466.0, + 295.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 426.0, + 1408.0, + 426.0, + 1408.0, + 466.0, + 718.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 461.0, + 1405.0, + 461.0, + 1405.0, + 494.0, + 295.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 489.0, + 957.0, + 489.0, + 957.0, + 524.0, + 294.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 203.0, + 1405.0, + 203.0, + 1405.0, + 239.0, + 294.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 233.0, + 1408.0, + 233.0, + 1408.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 299.0, + 290.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 295.0, + 1315.0, + 295.0, + 1315.0, + 331.0, + 294.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1915.0, + 1404.0, + 1915.0, + 1404.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1980.0, + 295.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1976.0, + 1404.0, + 1976.0, + 1404.0, + 2010.0, + 295.0, + 2010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 203, + 1403, + 203, + 1403, + 355, + 298, + 355 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 303, + 941, + 1408, + 941, + 1408, + 2009, + 303, + 2009 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 600, + 1403, + 600, + 1403, + 723, + 299, + 723 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 369, + 1404, + 369, + 1404, + 492, + 298, + 492 + ], + "score": 0.972 + }, + { + "category_id": 0, + "poly": [ + 299, + 763, + 571, + 763, + 571, + 801, + 299, + 801 + ], + "score": 0.914 + }, + { + "category_id": 0, + "poly": [ + 300, + 534, + 575, + 534, + 575, + 571, + 300, + 571 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 301, + 830, + 1276, + 830, + 1276, + 863, + 301, + 863 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 298, + 905, + 455, + 905, + 455, + 939, + 298, + 939 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.811 + }, + { + "category_id": 13, + "poly": [ + 628, + 262, + 677, + 262, + 677, + 295, + 628, + 295 + ], + "score": 0.47, + "latex": "\\lVert \\overline { { 4 8 } } \\rVert" + }, + { + "category_id": 13, + "poly": [ + 881, + 323, + 971, + 323, + 971, + 356, + 881, + 356 + ], + "score": 0.28, + "latex": "\\pm \\mathbb { Z } 0 . \\pm \\mathbb { B } 2 \\mathbb { I }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 762.0, + 575.0, + 762.0, + 575.0, + 809.0, + 294.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 529.0, + 579.0, + 529.0, + 579.0, + 579.0, + 294.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 902.0, + 460.0, + 902.0, + 460.0, + 944.0, + 295.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 203.0, + 1405.0, + 203.0, + 1405.0, + 237.0, + 293.0, + 237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 270.0, + 293.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 261.0, + 627.0, + 261.0, + 627.0, + 299.0, + 293.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 299.0, + 678.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 291.0, + 1408.0, + 291.0, + 1408.0, + 334.0, + 291.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 880.0, + 322.0, + 880.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 322.0, + 976.0, + 322.0, + 976.0, + 358.0, + 972.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 955.0, + 1408.0, + 955.0, + 1408.0, + 990.0, + 308.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 984.0, + 1408.0, + 984.0, + 1408.0, + 1022.0, + 352.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1016.0, + 427.0, + 1016.0, + 427.0, + 1050.0, + 352.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1056.0, + 1405.0, + 1056.0, + 1405.0, + 1091.0, + 308.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1085.0, + 1405.0, + 1085.0, + 1405.0, + 1123.0, + 353.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1118.0, + 899.0, + 1118.0, + 899.0, + 1150.0, + 351.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1158.0, + 1408.0, + 1158.0, + 1408.0, + 1193.0, + 309.0, + 1193.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1186.0, + 1108.0, + 1186.0, + 1108.0, + 1225.0, + 353.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1226.0, + 1407.0, + 1226.0, + 1407.0, + 1261.0, + 309.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1256.0, + 1404.0, + 1256.0, + 1404.0, + 1291.0, + 355.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1286.0, + 715.0, + 1286.0, + 715.0, + 1321.0, + 355.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1324.0, + 1405.0, + 1324.0, + 1405.0, + 1362.0, + 307.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1358.0, + 1025.0, + 1358.0, + 1025.0, + 1391.0, + 352.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1396.0, + 1405.0, + 1396.0, + 1405.0, + 1431.0, + 309.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1425.0, + 931.0, + 1425.0, + 931.0, + 1461.0, + 352.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 1466.0, + 1405.0, + 1466.0, + 1405.0, + 1501.0, + 308.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1498.0, + 730.0, + 1498.0, + 730.0, + 1529.0, + 355.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1530.0, + 1405.0, + 1530.0, + 1405.0, + 1574.0, + 311.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1566.0, + 922.0, + 1566.0, + 922.0, + 1600.0, + 353.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 1602.0, + 1407.0, + 1602.0, + 1407.0, + 1641.0, + 307.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1637.0, + 426.0, + 1637.0, + 426.0, + 1668.0, + 355.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 1407.0, + 1674.0, + 1407.0, + 1712.0, + 294.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1707.0, + 1408.0, + 1707.0, + 1408.0, + 1741.0, + 353.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1746.0, + 1407.0, + 1746.0, + 1407.0, + 1781.0, + 296.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1777.0, + 1105.0, + 1777.0, + 1105.0, + 1810.0, + 352.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1812.0, + 1405.0, + 1812.0, + 1405.0, + 1851.0, + 294.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1842.0, + 1405.0, + 1842.0, + 1405.0, + 1885.0, + 352.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1878.0, + 648.0, + 1878.0, + 648.0, + 1910.0, + 353.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1913.0, + 1404.0, + 1913.0, + 1404.0, + 1951.0, + 293.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1947.0, + 1404.0, + 1947.0, + 1404.0, + 1982.0, + 355.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1976.0, + 818.0, + 1976.0, + 818.0, + 2011.0, + 353.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 599.0, + 1405.0, + 599.0, + 1405.0, + 634.0, + 294.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 1403.0, + 630.0, + 1403.0, + 663.0, + 295.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 662.0, + 1405.0, + 662.0, + 1405.0, + 698.0, + 294.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 693.0, + 1167.0, + 693.0, + 1167.0, + 725.0, + 294.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 1404.0, + 367.0, + 1404.0, + 404.0, + 295.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 401.0, + 1404.0, + 401.0, + 1404.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 431.0, + 1405.0, + 431.0, + 1405.0, + 467.0, + 294.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 458.0, + 1148.0, + 458.0, + 1148.0, + 498.0, + 292.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 824.0, + 1284.0, + 824.0, + 1284.0, + 872.0, + 294.0, + 872.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 292, + 112, + 1410, + 112, + 1410, + 2020, + 292, + 2020 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 835, + 2061, + 864, + 2061, + 864, + 2087, + 835, + 2087 + ], + "score": 0.84 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2097.0, + 831.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 204.0, + 1405.0, + 204.0, + 1405.0, + 238.0, + 295.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 229.0, + 1409.0, + 229.0, + 1409.0, + 273.0, + 350.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 256.0, + 433.0, + 256.0, + 433.0, + 304.0, + 352.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 305.0, + 1405.0, + 305.0, + 1405.0, + 344.0, + 293.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 334.0, + 1407.0, + 334.0, + 1407.0, + 376.0, + 352.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 365.0, + 1090.0, + 365.0, + 1090.0, + 405.0, + 352.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 407.0, + 1407.0, + 407.0, + 1407.0, + 447.0, + 295.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 438.0, + 1325.0, + 438.0, + 1325.0, + 478.0, + 352.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 480.0, + 1409.0, + 480.0, + 1409.0, + 520.0, + 295.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 508.0, + 1409.0, + 508.0, + 1409.0, + 551.0, + 348.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 541.0, + 1407.0, + 541.0, + 1407.0, + 581.0, + 352.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 574.0, + 629.0, + 574.0, + 629.0, + 608.0, + 354.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 610.0, + 1405.0, + 610.0, + 1405.0, + 650.0, + 295.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 643.0, + 1244.0, + 643.0, + 1244.0, + 683.0, + 352.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 690.0, + 1405.0, + 690.0, + 1405.0, + 723.0, + 297.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 717.0, + 1409.0, + 717.0, + 1409.0, + 756.0, + 352.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 743.0, + 431.0, + 743.0, + 431.0, + 787.0, + 354.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 828.0, + 295.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 819.0, + 874.0, + 819.0, + 874.0, + 859.0, + 352.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 861.0, + 1409.0, + 861.0, + 1409.0, + 901.0, + 297.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 892.0, + 1407.0, + 892.0, + 1407.0, + 930.0, + 352.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 922.0, + 904.0, + 922.0, + 904.0, + 959.0, + 352.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 964.0, + 1409.0, + 964.0, + 1409.0, + 1001.0, + 293.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 997.0, + 1378.0, + 997.0, + 1378.0, + 1030.0, + 356.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1039.0, + 1407.0, + 1039.0, + 1407.0, + 1072.0, + 297.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1068.0, + 430.0, + 1068.0, + 430.0, + 1104.0, + 354.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 1407.0, + 1108.0, + 1407.0, + 1148.0, + 295.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1139.0, + 1405.0, + 1139.0, + 1405.0, + 1179.0, + 352.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1166.0, + 1409.0, + 1166.0, + 1409.0, + 1210.0, + 350.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1198.0, + 432.0, + 1198.0, + 432.0, + 1240.0, + 352.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1242.0, + 811.0, + 1242.0, + 811.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1321.0, + 295.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1315.0, + 1164.0, + 1315.0, + 1164.0, + 1353.0, + 352.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1357.0, + 1407.0, + 1357.0, + 1407.0, + 1396.0, + 295.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1388.0, + 1153.0, + 1388.0, + 1153.0, + 1424.0, + 352.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1428.0, + 1407.0, + 1428.0, + 1407.0, + 1468.0, + 295.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1459.0, + 1056.0, + 1459.0, + 1056.0, + 1497.0, + 350.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1505.0, + 1407.0, + 1505.0, + 1407.0, + 1539.0, + 297.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1535.0, + 563.0, + 1535.0, + 563.0, + 1568.0, + 354.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1574.0, + 1409.0, + 1574.0, + 1409.0, + 1612.0, + 293.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1604.0, + 1409.0, + 1604.0, + 1409.0, + 1643.0, + 354.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1635.0, + 1407.0, + 1635.0, + 1407.0, + 1675.0, + 354.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1664.0, + 1367.0, + 1664.0, + 1367.0, + 1704.0, + 352.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1744.0, + 297.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1737.0, + 1060.0, + 1737.0, + 1060.0, + 1775.0, + 350.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1779.0, + 1409.0, + 1779.0, + 1409.0, + 1819.0, + 295.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1811.0, + 1407.0, + 1811.0, + 1407.0, + 1850.0, + 352.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1842.0, + 1407.0, + 1842.0, + 1407.0, + 1882.0, + 354.0, + 1882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1873.0, + 646.0, + 1873.0, + 646.0, + 1907.0, + 354.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 1407.0, + 1913.0, + 1407.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1944.0, + 1409.0, + 1944.0, + 1409.0, + 1984.0, + 352.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1972.0, + 432.0, + 1972.0, + 432.0, + 2014.0, + 352.0, + 2014.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 292, + 164, + 1409, + 164, + 1409, + 2014, + 292, + 2014 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2087, + 836, + 2087 + ], + "score": 0.836 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 869.0, + 2057.0, + 869.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 201.0, + 1409.0, + 201.0, + 1409.0, + 240.0, + 296.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 230.0, + 1409.0, + 230.0, + 1409.0, + 270.0, + 350.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 264.0, + 1409.0, + 264.0, + 1409.0, + 303.0, + 354.0, + 303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 291.0, + 430.0, + 291.0, + 430.0, + 331.0, + 350.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 335.0, + 1409.0, + 335.0, + 1409.0, + 374.0, + 296.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 366.0, + 772.0, + 366.0, + 772.0, + 404.0, + 354.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 409.0, + 1409.0, + 409.0, + 1409.0, + 443.0, + 292.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 437.0, + 1405.0, + 437.0, + 1405.0, + 476.0, + 354.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 469.0, + 1033.0, + 469.0, + 1033.0, + 508.0, + 354.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 510.0, + 1407.0, + 510.0, + 1407.0, + 549.0, + 296.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 543.0, + 1403.0, + 543.0, + 1403.0, + 581.0, + 354.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 575.0, + 548.0, + 575.0, + 548.0, + 608.0, + 354.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 610.0, + 1407.0, + 610.0, + 1407.0, + 656.0, + 292.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 644.0, + 1407.0, + 644.0, + 1407.0, + 683.0, + 352.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 679.0, + 768.0, + 679.0, + 768.0, + 711.0, + 354.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 717.0, + 1407.0, + 717.0, + 1407.0, + 756.0, + 294.0, + 756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 748.0, + 1407.0, + 748.0, + 1407.0, + 786.0, + 354.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 778.0, + 881.0, + 778.0, + 881.0, + 817.0, + 354.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 823.0, + 1403.0, + 823.0, + 1403.0, + 855.0, + 298.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 851.0, + 1403.0, + 851.0, + 1403.0, + 890.0, + 352.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 880.0, + 516.0, + 880.0, + 516.0, + 918.0, + 350.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 918.0, + 1407.0, + 918.0, + 1407.0, + 965.0, + 290.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 951.0, + 1157.0, + 951.0, + 1157.0, + 993.0, + 348.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 992.0, + 1407.0, + 992.0, + 1407.0, + 1034.0, + 290.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1065.0, + 350.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1056.0, + 518.0, + 1056.0, + 518.0, + 1095.0, + 352.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1407.0, + 1097.0, + 1407.0, + 1136.0, + 294.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1128.0, + 979.0, + 1128.0, + 979.0, + 1166.0, + 352.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1170.0, + 1409.0, + 1170.0, + 1409.0, + 1209.0, + 294.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1199.0, + 1405.0, + 1199.0, + 1405.0, + 1241.0, + 354.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1231.0, + 706.0, + 1231.0, + 706.0, + 1270.0, + 354.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1272.0, + 1407.0, + 1272.0, + 1407.0, + 1310.0, + 294.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1302.0, + 1207.0, + 1302.0, + 1207.0, + 1341.0, + 352.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 1405.0, + 1345.0, + 1405.0, + 1381.0, + 292.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1377.0, + 1171.0, + 1377.0, + 1171.0, + 1412.0, + 352.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1416.0, + 1411.0, + 1416.0, + 1411.0, + 1459.0, + 290.0, + 1459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1451.0, + 1409.0, + 1451.0, + 1409.0, + 1489.0, + 354.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1479.0, + 1211.0, + 1479.0, + 1211.0, + 1516.0, + 352.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1518.0, + 1409.0, + 1518.0, + 1409.0, + 1562.0, + 292.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1548.0, + 547.0, + 1548.0, + 547.0, + 1591.0, + 351.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1593.0, + 1407.0, + 1593.0, + 1407.0, + 1631.0, + 294.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1623.0, + 1403.0, + 1623.0, + 1403.0, + 1662.0, + 354.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1656.0, + 1181.0, + 1656.0, + 1181.0, + 1694.0, + 356.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1696.0, + 1405.0, + 1696.0, + 1405.0, + 1735.0, + 296.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1727.0, + 1339.0, + 1727.0, + 1339.0, + 1765.0, + 354.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1765.0, + 1407.0, + 1765.0, + 1407.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1800.0, + 1227.0, + 1800.0, + 1227.0, + 1839.0, + 354.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1839.0, + 1411.0, + 1839.0, + 1411.0, + 1881.0, + 292.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1871.0, + 953.0, + 1871.0, + 953.0, + 1910.0, + 352.0, + 1910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1956.0, + 292.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 350.0, + 1942.0, + 1409.0, + 1942.0, + 1409.0, + 1987.0, + 350.0, + 1987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1975.0, + 800.0, + 1975.0, + 800.0, + 2013.0, + 354.0, + 2013.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.795 + }, + { + "category_id": 1, + "poly": [ + 293, + 125, + 1410, + 125, + 1410, + 2010, + 293, + 2010 + ], + "score": 0.739 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2086, + 836, + 2086 + ], + "score": 0.201 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 201.0, + 1411.0, + 201.0, + 1411.0, + 240.0, + 296.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 228.0, + 1411.0, + 228.0, + 1411.0, + 276.0, + 349.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 263.0, + 1407.0, + 263.0, + 1407.0, + 302.0, + 355.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 292.0, + 646.0, + 292.0, + 646.0, + 329.0, + 351.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 333.0, + 1405.0, + 333.0, + 1405.0, + 373.0, + 296.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 367.0, + 1409.0, + 367.0, + 1409.0, + 406.0, + 353.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 395.0, + 433.0, + 395.0, + 433.0, + 437.0, + 353.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 441.0, + 1403.0, + 441.0, + 1403.0, + 474.0, + 298.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 468.0, + 1407.0, + 468.0, + 1407.0, + 509.0, + 351.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 501.0, + 657.0, + 501.0, + 657.0, + 540.0, + 349.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 544.0, + 1407.0, + 544.0, + 1407.0, + 584.0, + 296.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 575.0, + 1409.0, + 575.0, + 1409.0, + 615.0, + 351.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 602.0, + 1407.0, + 602.0, + 1407.0, + 644.0, + 351.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 633.0, + 431.0, + 633.0, + 431.0, + 675.0, + 351.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 677.0, + 1407.0, + 677.0, + 1407.0, + 716.0, + 294.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 712.0, + 805.0, + 712.0, + 805.0, + 745.0, + 353.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 747.0, + 1407.0, + 747.0, + 1407.0, + 793.0, + 292.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 777.0, + 432.0, + 777.0, + 432.0, + 820.0, + 355.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 826.0, + 1407.0, + 826.0, + 1407.0, + 865.0, + 296.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 857.0, + 1407.0, + 857.0, + 1407.0, + 896.0, + 355.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 888.0, + 1276.0, + 888.0, + 1276.0, + 923.0, + 351.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 929.0, + 1407.0, + 929.0, + 1407.0, + 968.0, + 296.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 960.0, + 968.0, + 960.0, + 968.0, + 997.0, + 351.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1005.0, + 1405.0, + 1005.0, + 1405.0, + 1039.0, + 298.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1036.0, + 1405.0, + 1036.0, + 1405.0, + 1070.0, + 355.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1063.0, + 520.0, + 1063.0, + 520.0, + 1103.0, + 351.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1107.0, + 1407.0, + 1107.0, + 1407.0, + 1146.0, + 296.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1138.0, + 1405.0, + 1138.0, + 1405.0, + 1177.0, + 351.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1163.0, + 1409.0, + 1163.0, + 1409.0, + 1210.0, + 349.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1191.0, + 434.0, + 1191.0, + 434.0, + 1240.0, + 351.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 1407.0, + 1241.0, + 1407.0, + 1280.0, + 294.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1272.0, + 1407.0, + 1272.0, + 1407.0, + 1311.0, + 353.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1297.0, + 1411.0, + 1297.0, + 1411.0, + 1345.0, + 349.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1334.0, + 562.0, + 1334.0, + 562.0, + 1367.0, + 355.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1376.0, + 1409.0, + 1376.0, + 1409.0, + 1415.0, + 294.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1405.0, + 1409.0, + 1405.0, + 1409.0, + 1442.0, + 351.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1431.0, + 1411.0, + 1431.0, + 1411.0, + 1479.0, + 349.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1464.0, + 1162.0, + 1464.0, + 1162.0, + 1506.0, + 351.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1510.0, + 1409.0, + 1510.0, + 1409.0, + 1549.0, + 296.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1541.0, + 1306.0, + 1541.0, + 1306.0, + 1578.0, + 349.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1582.0, + 1407.0, + 1582.0, + 1407.0, + 1620.0, + 294.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1618.0, + 1405.0, + 1618.0, + 1405.0, + 1651.0, + 357.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1646.0, + 1029.0, + 1646.0, + 1029.0, + 1680.0, + 357.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1688.0, + 1407.0, + 1688.0, + 1407.0, + 1727.0, + 296.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1719.0, + 1407.0, + 1719.0, + 1407.0, + 1758.0, + 353.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1748.0, + 771.0, + 1748.0, + 771.0, + 1787.0, + 351.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1791.0, + 1409.0, + 1791.0, + 1409.0, + 1830.0, + 296.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1822.0, + 1409.0, + 1822.0, + 1409.0, + 1861.0, + 353.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1853.0, + 1409.0, + 1853.0, + 1409.0, + 1893.0, + 353.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1880.0, + 1409.0, + 1880.0, + 1409.0, + 1924.0, + 349.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1913.0, + 1407.0, + 1913.0, + 1407.0, + 1952.0, + 353.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1983.0, + 353.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1977.0, + 1018.0, + 1977.0, + 1018.0, + 2010.0, + 353.0, + 2010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 292, + 201, + 1407, + 201, + 1407, + 471, + 292, + 471 + ], + "score": 0.733 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2095.0, + 832.0, + 2095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 238.0, + 295.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 233.0, + 1405.0, + 233.0, + 1405.0, + 269.0, + 352.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 260.0, + 1214.0, + 260.0, + 1214.0, + 301.0, + 353.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 305.0, + 1405.0, + 305.0, + 1405.0, + 341.0, + 295.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 335.0, + 1151.0, + 335.0, + 1151.0, + 369.0, + 352.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 376.0, + 1405.0, + 376.0, + 1405.0, + 412.0, + 295.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 407.0, + 1408.0, + 407.0, + 1408.0, + 445.0, + 351.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 439.0, + 642.0, + 439.0, + 642.0, + 471.0, + 355.0, + 471.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/ARFshOO1Iu/ARFshOO1Iu.md b/parse/train/ARFshOO1Iu/ARFshOO1Iu.md new file mode 100644 index 0000000000000000000000000000000000000000..815b7f6cae05b9847cf48f1f66e1a49f07682dfa --- /dev/null +++ b/parse/train/ARFshOO1Iu/ARFshOO1Iu.md @@ -0,0 +1,355 @@ +# ADAPTIVE SELF-TRAINING FOR NEURAL SEQUENCE LABELING WITH FEW LABELS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Neural sequence labeling is an important technique employed for many Natural Language Processing (NLP) tasks, such as Named Entity Recognition (NER), slot tagging for dialog systems and semantic parsing. Large-scale pre-trained language models obtain very good performance on these tasks when fine-tuned on large amounts of task-specific labeled data. However, such large-scale labeled datasets are difficult to obtain for several tasks and domains due to the high cost of human annotation as well as privacy and data access constraints for sensitive user applications. This is exacerbated for sequence labeling tasks requiring such annotations at token-level. In this work, we develop techniques to address the label scarcity challenge for neural sequence labeling models. Specifically, we develop self-training and meta-learning techniques for training neural sequence taggers with few labels. While self-training serves as an effective mechanism to learn from large amounts of unlabeled data – meta-learning helps in adaptive sample re-weighting to mitigate error propagation from noisy pseudo-labels. Extensive experiments on six benchmark datasets including two for massive multilingual NER and four slot tagging datasets for task-oriented dialog systems demonstrate the effectiveness of our method. With only 10 labeled examples for each class for each task, our method obtains $1 0 \%$ improvement over state-of-the-art systems demonstrating its effectiveness for the low-resource setting. + +# 1 INTRODUCTION + +Motivation. Deep neural networks typically require large amounts of training data to achieve stateof-the-art performance. Recent advances with pre-trained language models like BERT (Devlin et al., 2019), GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019) have reduced this annotation bottleneck. In this paradigm, large neural network models are trained on massive amounts of unlabeled data in a self-supervised manner. However, the success of these large-scale models still relies on fine-tuning them on large amounts of labeled data for downstream tasks. For instance, our experiments show $\bar { 2 } 7 \%$ relative improvement on an average when fine-tuning BERT with the full training set (2.5K-705K labels) vs. fine-tuning with only 10 labels per class. This poses several challenges for many real-world tasks. Not only is acquiring large amounts of labeled data for every task expensive and time consuming, but also not feasible in many cases due to data access and privacy constraints. This issue is exacerbated for sequence labeling tasks that require annotations at token- and slot-level as opposed to instance-level classification tasks. For example, an NER task can have slots like B-PER, I-PER, O-PER marking the beginning, intermediate and out-of-span markers for person names, and similar slots for the names of location and organization. Similarly, language understanding models for dialog systems rely on effective identification of what the user intends to do (intents) and the corresponding values as arguments (slots) for use by downstream applications. Therefore, fully supervised neural sequence taggers are expensive to train for such tasks, given the requirement of thousands of annotations for hundreds of slots for the many different intents. + +Semi-supervised learning (SSL) (Chapelle et al., 2010) is one of the promising paradigms to address labeled data scarcity by making effective use of large amounts of unlabeled data in addition to task-specific labeled data. Self-training (ST, (III, 1965)) as one of the earliest SSL approaches has recently shown state-of-the-art performance for tasks like image classification (Li et al., 2019; Xie et al., 2020) performing at par with supervised systems while using very few training labels. In contrast to such instance-level classification tasks, sequence labeling tasks have dependencies between the slots demanding different design choices for slot-level loss optimization for the limited labeled data setting. For instance, prior work (Ruder & Plank, 2018) using classic self-training techniques for sequence labeling did not find much success in the low-data regime with $1 0 \%$ labeled data for the target domain. Although there has been some success with careful task-specific data selection (Petrov & McDonald, 2012) and more recently for distant supervision (Liang et al., 2020) using external resources like knowledge bases (e.g., Wikipedia). In contrast to these prior work, we develop techniques for self-training with limited labels and without any task-specific assumption or external knowledge. + +![](images/4716dc9424f650ad710d7db8d4aff2f8dd06b88c864da63dd6e410117264cc70.jpg) +Figure 1: MetaST framework. + +For self-training, a base model (teacher) is trained on some amount of labeled data and used to pseudo-annotate (task-specific) unlabeled data. The original labeled data is augmented with the pseudo-labeled data and used to train a student model. The student-teacher training is repeated until convergence. Traditionally in self-training frameworks, the teacher model pseudo-annotates unlabeled data without any sample selection. This may result in gradual drifts from self-training on noisy pseudo-labeled instances (Zhang et al., 2017). In order to deal with noisy labels and training set biases, Ren et al. (2018) propose a meta-learning technique to automatically re-weight noisy samples by their loss changes on a held-out clean labeled validation set. We adopt a similar principle in our work and leverage meta-learning to re-weight noisy pseudo-labeled examples from the teacher. While prior techniques for learning to re-weight examples have been developed for instance-level classification tasks, we extend them to operate at token-level for discrete sequence labeling tasks. To this end, we address some key challenges on how to construct an informative held-out validation set for token-level re-weighting. Prior works (Ren et al., 2018; Shu et al., 2019) for instance classification construct this validation set by random sampling. However, sequence labeling tasks involve many slots (e.g. WikiAnn has 123 slots over 41 languages) with variable difficulty and distribution in the data. In case of random sampling, the model oversamples from the most populous category and slots. This is particularly detrimental for low-resource languages in the multilingual setting. To this end, we develop an adaptive mechanism to create the validation set on the fly considering the diversity and uncertainty of the model for different slot types. Furthermore, we leverage this validation set for token-level loss estimation and re-weighting pseudo-labeled sequences from the teacher in the meta-learning setup. While prior works (Li et al., 2019; Sun et al., 2019; Bansal et al., 2020) on meta-learning for image and text classification leverage multi-task learning to improve a target classification task based on several similar tasks, in this work we focus on a single sequence labeling task – making our setup more challenging altogether. + +Our task and framework overview. We focus on sequence labeling tasks with only a few annotated samples (e.g., $K = \{ 5 , 1 0 , 2 0 , 1 0 0 \} )$ per slot type for training and large amounts of task-specific unlabeled data. Figure 1 shows an overview of our framework with the following components: (i) Self-training: Our self-training framework leverages a pre-trained language model as a teacher and co-trains a student model with iterative knowledge exchange (ii) Adaptive labeled data acquisition for validation: Our few-shot learning setup assumes a small number of labeled training samples per slot type. The labeled data from multiple slot types are not equally informative for the student model to learn from. While prior works in meta-learning randomly sample some labeled examples for held-out validation set, we develop an adaptive mechanism to create this set on the fly. To this end, we leverage loss decay as a proxy for model uncertainty to select informative labeled samples for the student model to learn from in conjunction with the re-weighting mechanism in the next step. (iii) Meta-learning for sample re-weighting: Since pseudo-labeled samples from the teacher can be noisy, we employ meta-learning to re-weight them to improve the student model performance on the held-out validation set obtained from the previous step. In contrast to prior work (Ren et al., 2018) + +on sample re-weighting operating at instance-level, we incorporate the re-weighting mechanism at token-level for sequence labeling tasks. Here the token-level weights are determined by the student model loss on the above validation set. Finally, we learn all of the above steps jointly with end-toend learning in the self-training framework. We refer to our adaptive self-training framework with meta-learning based sample re-weighting mechanism as MetaST. + +We perform extensive experiments on six benchmark datasets for several tasks including multilingual Named Entity Recognition and slot tagging for user utterances from task-oriented dialog systems to demonstrate the generalizability of our approach across diverse tasks and languages. We adopt BERT and multilingual BERT as encoder and show that its performance can be significantly improved by nearly $1 0 \%$ for low-resource settings with few training labels (e.g., 10 labeled examples per slot type) and large amounts of unlabeled data. In summary, our work makes the following contributions. (i) Develops a self-training framework for neural sequence tagging with few labeled training examples. (ii) Leverages an acquisition strategy to adaptively select a validation set from the labeled set for meta-learning of the student model. (iii) Develops a meta-learning framework for re-weighting pseudo-labeled samples at token-level to reduce drifts from noisy teacher predictions. (iv) Integrates the aforementioned components into an end-to-end learning framework and demonstrates its effectiveness for neural sequence labeling across six benchmark datasets with multiple slots, shots, domains and languages. + +# 2 BACKGROUND + +Sequence labeling and slot tagging. This is the task identifying the entity span of several slot types (e.g., names of person, organization, location, date, etc.) in a text sequence. Formally, given a sentence with $N$ tokens $\mathbf { \bar { \cal X } } ~ = ~ \{ x _ { 1 } , . . . , x _ { N } \}$ , an entity or slot value is a span of tokens $s = [ x _ { i } , . . . , x _ { j } ] ( 0 \leq i \leq j \leq N )$ associated with a type. This task assumes a pre-defined tagging policy like BIO (Tjong et al., 1999), where B marks the beginning of the slot, I marks an intermediate token in the span, and $\bigcirc$ marks out-of-span tokens. These span markers are used to extract multi-token values for each of the slot types with phrase-level evaluation for the performance. + +Self-training. Consider $f ( \cdot ; \theta _ { t e a } )$ and $f ( \cdot ; \theta _ { s t u } )$ to denote the teacher and student models respectively in the self-training framework. The role of the teacher model (e.g., a pre-trained language model) is to assign pseudo-labels to unlabeled data that is used to train a student model. The teacher and student model can exchange knowledge and the training schedules are repeated till convergence. The success of self-training with deep neural networks in recent works (He et al., 2019; Xie et al., 2020) has been attributed to a number of factors including stochastic regularization with dropouts and data regularization with unlabeled data. Formally, given $m$ -th unlabeled sentence with $N$ tokens $X _ { m } ^ { u } \doteq \{ x _ { 1 , m } ^ { u } , . . . , x _ { N , m } ^ { u } \}$ and $C$ pre-defined labels, consider the pseudo-labels $\hat { Y } _ { m } ^ { ( t ) } = [ \hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \hat { y } _ { m , N } ^ { ( t ) } ]$ generated by the teacher model at the $t$ -th iteration where, + +$$ +\hat { y } _ { m , n } ^ { ( t ) } = \underset { c \in C } { \arg \operatorname* { m a x } } f _ { n , c } ( x _ { m , n } ^ { u } ; \theta _ { t e a } ^ { ( t ) } ) . +$$ + +The pseudo-labeled data set, denoted as $( X ^ { u } , \hat { Y } ^ { ( t ) } ) = \{ ( X _ { m } ^ { u } , \hat { Y } _ { m } ^ { ( t ) } ) \} _ { m } ^ { M }$ , is used to train the student model and learn its parameters as: + +$$ +\hat { \theta } _ { s t u } ^ { ( t ) } = \underset { \theta } { \arg \operatorname* { m i n } } \frac { 1 } { M } \sum _ { m = 1 } ^ { M } l ( \hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \theta _ { s t u } ^ { ( t - 1 ) } ) ) , +$$ + +where $l ( \cdot , \cdot )$ can be modeled as the cross-entropy loss. + +# 3 ADAPTIVE SELF TRAINING + +Given a pre-trained language model (e.g., BERT (Devlin et al., 2019)) as the teacher, we first finetune it on the small labeled data to make it aware of the underlying task. The fine-tuned teacher model is now used to pseudo-label the large unlabeled data. We consider the student model as another instantiation of the pre-trained language model that is trained over the pseudo-labeled data. However, our few-shot setting with limited labeled data results in a noisy teacher. A naive transfer of teacher knowledge to the student results in the propagation of noisy labels limiting the performance of the student model. To address this challenge, we develop an adaptive self-training framework to re-weight pseudo-labeled predictions from the teacher with a meta-learning objective that optimizes the token-level loss from the student model on a held-out labeled validation set. This held-out set is adaptively constructed via labeled data acquisition which selects labeled samples with high uncertainty for efficient data exploration. + +# 3.1 ADAPTIVE LABELED DATA ACQUISITION + +In standard meta-learning setup for instance-level classification tasks, the held-out validation set is usually constructed via random sampling (Ren et al., 2018; Shu et al., 2019). Sequence labeling tasks involve many slot types with variable difficulty and distribution in the data. For instance, NER tasks over WikiAnn operate over 123 slot types from 41 languages with additional complexity from variable model performance across different languages. A random sampling leads to oversampling instances with the most populous categories and slot types in the data. Therefore, we propose a novel labeled data acquisition strategy to construct the validation set for effective data exploration. We demonstrate its benefit over classic meta-learning approaches from prior works in experiments. + +In general, data acquisition strategies for prior works in meta-learning and active learning broadly leverage random sampling (Ren et al., 2018; Shu et al., 2019), easy (Kumar et al., 2010) and hard example mining (Shrivastava et al., 2016) or uncertainty-based methods (Chang et al., 2017a). These strategies have been compared in prior works (Chang et al., 2017a; Gal et al., 2017) that show uncertainty-based methods to have better generalizability across diverse settings. There are several approaches to uncertainty estimation including error decay (Konyushkova et al., 2017; Chang et al., 2020), Monte Carlo dropouts (Gal et al., 2017) and predictive variance (Chang et al., 2017a). We follow a similar principle of error decay to find samples that the model is uncertain about and can correspondingly benefit from knowing their labels (similar to active learning settings). To this end, we leverage stochastic loss decay from the model as a proxy for the model uncertainty to generate validation set on the fly. This is used for estimating token-level weights and re-weighting pseudo labeled data in Section 3.2. + +Consider the loss of the student model with parameters $\theta _ { s t u } ^ { ( t ) }$ on the labeled data $( X _ { m } ^ { l } , Y _ { m } )$ in the $t$ -th iteration as $l ( Y _ { m } , f ( X _ { m } ^ { l } ; \theta _ { s t u } ^ { ( t ) } ) )$ . To measure the loss decay value at any iteration, we use the difference between the current and previous loss values. Considering these values may fluctuate across iterations, we adopt the moving average of the loss values for $( X _ { m } ^ { l } , Y _ { m } )$ in the latest $R$ iterations as a baseline $l _ { b } ^ { m }$ for loss decay estimation. Baseline measure $l _ { b } ^ { m }$ is calculated as follows: + +$$ +l _ { b } ^ { m } = \frac { 1 } { R } \sum _ { r = 1 } ^ { R } l ( Y _ { m } , f ( X _ { m } ^ { l } ; \theta _ { s t u } ^ { ( t - r ) } ) ) . +$$ + +Since the loss decay values are estimated on the fly, we want to balance exploration and exploitation. To this end, we add a smoothness factor $\delta$ to prevent the low loss decay samples (i.e. samples with low uncertainty) from never being selected again. Considering all of the above factors, we obtain the sampling weight of labeled data $( X _ { m } ^ { l } , Y _ { m } ^ { l } )$ as follows: + +$$ +W _ { m } \propto \mathrm { m a x } ( l _ { b } ^ { m } - l ( Y _ { m } , f ( X _ { m } ^ { l } ; \theta _ { s t u } ^ { ( t ) } ) ) , 0 ) + \delta . +$$ + +The smoothness factor $\delta$ needs to be adaptive since the training loss is dynamic. Therefore, We adopt the maximum of the loss decay value as the smoothness factor $\delta$ to encourage exploration. + +The aforementioned acquisition function is re-estimated after a fixed number of steps to adapt to model changes. With labeled data acquisition, we rely on informative uncertain samples to improve learning efficiency. The sampled mini-batches of labeled data $\{ B _ { s } ^ { l } \}$ are used as a validation set for the student model in the next step for re-weighting pseudo-labeled data from the teacher model. We demonstrate its impact via ablation study in experiments. Note that the labeled data is only used to compute the acquisition function and not used for explicit training of the student model in this step. + +# 3.2 RE-WEIGHTING PSEUDO-LABELED DATA + +To mitigate error propagation from noisy pseudo-labeled sequences from the teacher, we leverage meta-learning to adaptively re-weight them based on the student model loss on the held-out validation set obtained via labeled data acquisition from the previous section. In contrast to prior work focusing on instance-level tasks like image classification – sequence labeling operates on discrete text sequences as input and assigns labels to each token in the sequence. Since teacher predictions vary for different slot labels and types, we adapt the meta-learning framework to re-weight samples at a token-level resolution. + +Token Re-weighting. Consider the pseudo-labels $\{ \hat { Y } _ { m } ^ { ( t ) } = [ \hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \hat { y } _ { m , N } ^ { ( t ) } ] \} _ { m = 1 } ^ { M }$ from the teacher in the $t$ -th iteration with $m$ and $n$ indexing the instance and a token in the instance, respectively. In classic self-training, we update the student parameters leveraging pseudo-labels as follows: + +$$ +\hat { \theta } _ { s t u } ^ { ( t ) } = \hat { \theta } _ { s t u } ^ { ( t - 1 ) } - \alpha \nabla \big ( \frac { 1 } { M } \sum _ { m = 1 } ^ { M } l ( \hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \theta _ { s t u } ^ { ( t - 1 ) } ) ) \big ) . +$$ + +Now, to downplay noisy token-level labels, we leverage meta-learning to re-weight the pseudolabeled data. To this end, we follow a similar analysis from (Koh & Liang, 2017) and (Ren et al., 2018) to perturb the weight for each token in the mini-batch by $\epsilon$ . Weight perturbation is used to discover data points that are most important to improve the model performance on a held-out validation set (Koh & Liang, 2017) where the sample importance is given by the magnitude of the the negative gradients. We extend prior techniques to obtain token-level perturbations as: + +$$ +\hat { \theta } _ { s t u } ^ { ( t ) } ( \epsilon ) = \hat { \theta } _ { s t u } ^ { ( t - 1 ) } - \alpha \nabla \big ( \frac { 1 } { M } \frac { 1 } { N } \sum _ { m = 1 } ^ { M } \sum _ { n = 1 } ^ { N } [ \epsilon _ { m , n } \cdot l ( \hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \hat { \theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \big ) . +$$ + +The token weights are obtained by minimizing the student model loss on the held-out validation set. Here, we employ the labeled data acquisition strategy from Eq. 4 to sample informative mini-batches of labeled data ${ \dot { B } } _ { s } ^ { l }$ locally at step $t$ . To obtain a cheap estimate of the meta-weight at step $t$ , we take a single gradient descent step for the sampled labeled mini-batch $B _ { s } ^ { l }$ : + +$$ +u _ { m , n , s } = - \frac { \partial } { \partial \epsilon _ { m , n , s } } \big ( \frac { 1 } { | \mathcal { B } _ { s } ^ { l } | } \frac { 1 } { N } \sum _ { m = 1 } ^ { | \mathcal { B } _ { s } ^ { l } | } \sum _ { n = 1 } ^ { N } [ l ( y _ { m , n } , f ( x _ { m , n } ^ { l } ; \hat { \theta } _ { s t u } ^ { ( t ) } ( \epsilon ) ) ] \big ) | _ { \epsilon _ { m , n , s } = 0 } +$$ + +We set the token weights to be proportional to the negative gradients to reflect the importance of pseudo-labeled tokens in the sequence. Since sequence labeling tasks have dependencies between the slot types and tokens, it is difficult to obtain a good estimation of the weights based on a single mini-batch of examples. Therefore, we sample $S$ mini-batches of labeled data $\{ B _ { 1 } ^ { l } , . . . , B _ { S } ^ { l } \}$ with the adaptive acquisition strategy and calculate the mean of the gradients to obtain a robust gradient estimate. Note that $S$ is a constant number that is the same for each token and the proportional sign in Eq. 8. Since a negative weight indicates a pseudo-label of poor quality that would potentially degrade the model performance, we set such weights to 0 to filter them out. The impact of $S$ is investigated in the experiments (refer to Appendix A.1). The overall meta-weight of pseudo-labeled token $\left( x _ { m , n } ^ { u } , \hat { y } _ { m , n } \right)$ is obtained as: + +$$ +w _ { m , n } \propto \mathrm { m a x } ( \sum _ { s = 1 } ^ { S } u _ { m , n , s } , 0 ) +$$ + +To further ensure the stability of the loss function in each mini-batch, we normalise the weight $w _ { m , n }$ . Finally, we update the student model parameters while accounting for token-level re-weighting as: + +$$ +\hat { \theta } _ { s t u } ^ { ( t ) } = \hat { \theta } _ { s t u } ^ { ( t - 1 ) } - \alpha \nabla \big ( \frac { 1 } { M } \frac { 1 } { N } \sum _ { m = 1 } ^ { M } \sum _ { n = 1 } ^ { N } [ w _ { m , n } \cdot l ( \hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \hat { \theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \big ) . +$$ + +We demonstrate the impact of our re-weighting mechanism with an ablation study in experiments. + +# 3.3 TEACHER MODEL ITERATIVE UPDATES + +At the end of every self-training iteration, we assign the student model as a new teacher model (i.e., $\theta _ { t e a } = \theta _ { s t u } ^ { ( T ) } )$ . Since the student model uses the labeled data only as a held-out validation setl for meta-learning, we further utilize the labeled data + +$f ( \cdot , \theta _ { t e a } ^ { ( t ) } )$ with standard supervised loss minimization. We explore the effectiveness of this step with an ablation study in experiments. The overall training procedure is summarized in Algorithm 1. + +# Algorithm 1: MetaST Algorithm. + +
Input:Labeled sequences (Xl,Y); Unlabeled sequences (X");Pre-trained BERT model with randomly initialized token classification layer f(;((O)); Batches S; Number of self-training iterations T.
Initialize teacher model 0tea =0(0) while not converged do
Fine-tune teacher model on small labeled data (Xl, Y);
Initialize the student model θ(0)=0(0);
Generate hard pseudo-labels Y(t）for unlabeled samples Xu with model f(·,0tea);
fort←1toTdo
Compute labeled data acquisition function according to Eq. 4;
function;Sample S mini-batches of labeled examples {Bl,.,B's} from (Xl,Y）based on labeled data acquisition
Randomly sample abatch of pseudo-labeled examples B from(Xu,Y(t));
Compute token-level weights in B based on the loss on {Bl,..,B's} according to Eq.8;
endstu
Update the teacher:θtea
end
+ +# 4 EXPERIMENTS + +Encoder. Pre-trained language models like BERT (Devlin et al., 2019), GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019) have shown state-of-the-art performance for various natural language processing tasks. In this work we adopt one of them as a base encoder by initializing the teacher with pre-trained BERT-base model and a randomly initialized token classification layer. + +Datasets. We perform large-scale experiments with six different datasets including user utterances for task-oriented dialog systems and multilingual Named Entity Recognition tasks as summarized in Table 1. (a) Email. This consists of natural language user utterances for email-oriented user actions like sending, receiving or searching emails with attributes like date, time, topics, people, etc. (b) SNIPS is a public benchmark + +
Dataset# Slots#Train#Test#Lang
Email202.5K1kEN
SNIPS3913K0.7KEN
MIT Movie128.8K2.4KEN
MIT Restaurant86.9K1.5KEN
Wikiann (EN)320K10KEN
CoNLL03 (EN)415K3.6KEN
CoNLL031638K15K4
Wikiann123705K329K41
+ +Table 1: Dataset summary. + +dataset (Coucke et al., 2018) of user queries from multiple domains including music, media, and weather. (c) MIT Movie and Restaurant corpus (Liu et al., 2013) consist of similar user utterances for movie and restaurant domains. (d) CoNLL03 (Sang & Meulder, 2003) and Wikiann (Pan et al., 2017) are public benchmark datasets for multilingual Named Entity Recognition. CoNLL03 is a collection of news wire articles from the Reuters Corpus from 4 languages with manual annotations, whereas Wikiann comprises of extractions from Wikipedia articles from 41 languages with automatic annotation leveraging meta-data for different entity types like ORG, PER, LOC etc. For every dataset, we sample $K \in \{ 5 , 1 0 , 2 0 , 1 0 0 \}$ labeled sequences for each slot type from the Train data, and add the remaining to the unlabeled set while ignoring their labels – following standard setups for semi-supervised learning. We repeatedly sample $K$ labeled instances three times for multiple runs to report average performance with standard deviation across the runs. + +Baselines. The first baseline we consider is the fully supervised BERT model trained on all available training data which provides the ceiling performance for every task. Each of the other models are trained on $K$ training labels per slot type. We adopt several state-of-the-art semi-supervised methods as baselines: (1) CVT (Clark et al., 2018) is a semi-supervised sequence labeling method based on cross-view training; (2) SeqVAT (Chen et al., 2020) incorporates adversarial training with conditional random field layer for semi-supervised sequence labeling; (3) Mean Teacher (MT) (Tarvainen & Valpola, 2017) averages model weights to obtain an aggregated teacher; (4) VAT (Miyato et al., 2018) adopts virtual adversarial training to make the model robust to noise; (5) classic ST (III, 1965) is simple self-training method with hard pseudo-labels; (6) BOND (Liang et al., 2020) is the most recent work on self-training for sequence labeling with confidence-based sample selection and forms a strong baseline for our work. We implement our framework in Pytorch and use Tesla V100 gpus for experiments. Hyper-parameter configurations with model settings presented in Appendix. + +Neural sequence labeling performance with few training labels. Table 2 shows the performance comparison among different models with $\mathrm { K } { = } 1 0$ labeled examples per slot type. The fully supervised BERT trained on thousands of labeled examples provides the ceiling performance for the few-shot setting. We observe our method MetaST to significantly outperform all methods across all datasets including the models that also use the same BERT encoder as ours like MT, VAT, Classic ST and BOND with corresponding average performance improvements as $1 4 . 2 2 \%$ , $1 4 . 9 0 \%$ , $8 . 4 6 \%$ and $8 . 8 2 \%$ . Non BERT models like CVT and ${ \tt S e q V A T }$ are consistently worse than other baselines. + +
MethodSNIPSEmailMovieRestaurantCoNLL03 (EN)Wikiann (EN)
# Slots392012843
Full-supervision
BERT95.8094.4487.8778.9592.4084.04
Few-shot supervision (10 labels per slot)69.5054.0671.1545.61
BERT 79.01 87.85
Few-shot supervision (10 labels per slot) + unlabeled data
CVT78.2378.2462.7342.5754.3127.89
SeqVAT78.6772.6567.1051.5567.2135.16
MT79.4889.5367.6251.7568.6741.43
VAT79.0889.7170.1753.3465.0338.81
Classic ST83.2690.7071.8856.8070.9946.15
BOND83.5489.7570.9155.7869.5648.73
MetaST88.2392.1877.6763.8376.65
(0.04;↑12%)(0.47;↑4.93%)(0.10;↑11.76%)(1.62;↑18.07%)(0.73;17.73%)56.61 (0.4;↑24.12%)
+ +Table 2: F1 score comparison of models for sequence labeling on different datasets. All models (except CVT and ${ \tt S e q V A T } ,$ use the same BERT encoder. F1 score of our model for each task is followed by standard deviation and percentage improvement (↑) over BERT with few-shot supervision. + +We also observe variable performance of the models across different tasks. Specifically, the performance gap between the best few-shot model and the fully supervised model varies significantly. MetaST achieves close performance to the fully-supervised model in some datasets (e.g. SNIPS and Email) but has bigger room for improvement in others (e.g. CoNLL03 (EN) and Wikiann (EN)). This can be attributed to the following factors. (i) Labeled training examples and slots. The total number of labeled training instances for our K-shot setting is given by $K \times \# S l o t s$ . Therefore, for tasks with higher number of slots and consequently more training labels, most of the models perform better including MetaST. Task-oriented dialog systems with more slots and inherent dependency between the slot types benefit more than NER tasks. (ii) Task difficulty: User utterances from task-oriented dialog systems for some of the domains like weather, music and emails contain predictive query patterns and limited diversity. In contrast, Named Entity Recognition datasets are comparatively diverse and require more training labels to generalize well. Similar observations are also depicted in Table 3 for multilingual NER tasks with more slots and consequently more training labels from multiple languages as well as richer interactions across the slots from different languages. + +
Dataset#Lang#SlotsFull Sup.Few-shot Sup.Few-shot supervision + unlabeled data
BERTBERTMTVATClassic STBONDMetaST
CoNLL0341687.6770.7768.3467.6372.6972.7976.41 (0.47) (↑ 7.97%)
Wikiann4112387.1779.6780.2378.8280.2479.5781.61 (0.14) (↑ 2.42%)
+ +Table 3: F1 score comparison of models for sequence labeling on multilingual datasets using the same BERT-Multilingual-Base encoder. F1 score of MetaST for each task is followed by standard deviation in parentheses and percentage improvement (↑) over BERT with few-shot supervision. + +Controlling for the total amount of labeled data. In order to control for the variable amount of training labels across different datasets, we perform another experiment where we vary the number of labels for different slot types while keeping the total number of labeled instances for each dataset similar (ca. 200). Results are shown in Table 4. To better illustrate the effect of the number of training labels, we choose tasks with lower performance in Table 2 for this experiment. Comparing the results in Tables 2 and 4, we observe the performance of MetaST to improve with more training labels for all the tasks . + +Effect of varying the number of labels $K$ per slot. Table 5 shows the improvement in the performance of MetaST when increasing the number of labels for each slot type in the SNIPS dataset. + +Table 4: F1 scores of different models with 200 labeled samples for each task. The percentage improvement $( \uparrow )$ is over the BERT model with few-shot supervision. + +
DatasetBERT (Full Supervision)BERT (Few-shot Supervision)MetaST（%Improvement）
MIT Movie87.8775.8180.33 (↑5.96%)
MIT Restaurant78.9560.1267.86 (↑12.87%)
CoNLL03 (EN)92.4077.4881.61 (↑ 5.33%)
Wikiann (EN)84.0462.0471.27 (个14.88%)
Average85.8268.8675.27 (↑ 9.31%)
+ +Similar trends can be found on other datasets (results in Appendix). As we increase the amount of labeled training instances, the performance of BERT also improves, and correspondingly the margin between MetaST and these baselines decreases although MetaST still improves over all of them. In the self-training framework, given the ceiling performance for every task and the improved performance of the teacher with more training labels, there is less room for (relative) improvement of the student over the teacher model. Consider SNIPS for example. Our model obtains $12 \%$ and $2 \%$ improvement over the few-shot BERT model for the 10-shot and 100-shot setting with F1-scores as $8 8 . 2 2 \%$ and $9 5 . 3 9 \%$ , respectively. The ceiling performance for this task is $9 5 . 8 \%$ on training BERT on the entire dataset with 13K labeled examples. This demonstrates that MetaST is most impactful for low-resource settings with few training labels for a given task. + +
#SlotsFew-shot SupervisionFew-shot supervision + unlabeled data
BERTCVTSeqVATMTVATClassic STBONDMetaST(%Improvement)
570.6369.8269.3470.8571.3472.5972.8581.56 (↑15%)
1079.0178.2378.6779.4879.0883.2683.5488.22 (↑12%)
2086.8188.0485.0587.3188.1988.3288.9391.99 (↑6%)
10093.9094.6191.4694.2694.5393.9294.2295.39 (12%)
+ +Table 5: Variation in model performance on varying $K$ labels / slot on SNIPS dataset with 39 slots. The percentage improvement $( \uparrow )$ is relative to the BERT model with few-shot supervision. + +Ablation analysis. Table 6 demonstrates the impact of different MetaST components with ablation analysis. We observe that soft pseudo-labels hurt the model performance compared to hard pseudolabels, as also shown in recent work (Kumar et al., 2020). Such a performance drop may be attributed to soft labels being less informative compared to sharpened ones. Removing the iterative teacher fine-tuning step (Section 3.1) also hurts the overall performance. + +Table 6: Ablation analysis of our framework MetaST with 10 labeled examples per slot on SNIPS and CoNLL03 (EN). + +
MethodDatasets
SNIPSCoNLL03
BERT w/ Continued Pre-training +
Few-shot Supervision83.9669.84
Classic ST83.2670.99
Classic ST w/ Soft Pseudo-Labels81.1771.87
MetaST(ours) w/Hard Pseudo-Labels88.2376.65
MetaST w/ Soft Pseudo-Labels86.1675.84
MetaST w/o Iterative Teacher Fine-tune85.6472.74
MetaST w/o Labeled Data Acq.86.6375.02
Pseudo-labeled Data Re-weighting
MetaST w/o Re-weighting85.4873.02
MetaST (Easy)85.5674.53
MetaST (Difficult)86.3468.06
+ +![](images/6e003691b635563a0994ab8d5c2a1f4aff6b7c20f56ba95beaa0a166af22b000.jpg) +Figure 2: Visualization of MetaST reweighting on CoNLL03 (EN). + +Continued pre-training v.s. self-training. To contrast continued pre-training with self-training, we further pre-train BERT on in-domain unlabeled data and then fine-tune it with few labeled examples denoted as “BERT (Continued Pre-training $^ +$ Few-shot Supervision)”. The pre-training step improves the BERT performance over the baseline on SNIPS but degrades the performance on CoNLL03. This indicates that continued pre-training can improve the performance of few-shot supervised BERT on specialized tasks (e.g., SNIPS) with different data distribution than the original pre-training data (e.g., Wikipedia), but may not help for general domain ones like CoNLL03 with overlapping data from Wikipedia. In contrast to the above baseline, MetaST brings significant improvements on both datasets. This demonstrates the generality and flexibility of self-training over pre-training as also observed in contemporary work (Zoph et al., 2020) on image classification. + +Adaptive labeled data acquisition. We perform an ablation study by removing adaptive labeled data acquisition from MetaST (denoted as “MetaST w/o Labeled Data Acq.”). Removing this component leads to around $2 \%$ performance drop on an average demonstrating the impact of labeled data acquisition. Moreover, the performance drop on SNIPS (39 slots) is larger than that on CoNLL03 (4 slots). This demonstrates that adaptive acquisition is more helpful for tasks with more slot types – where diversity and data distribution necessitate a better exploration strategy in contrast to random sampling employed in prior meta-learning works. + +Re-weighting strategies. To explore the role of token-level re-weighting for pseudo-labeled sequences (discussed in Section 3.2), we replace our meta-learning component with different sample selection strategies based on the model confidence for different tokens. One sampling strategy chooses samples uniformly without any re-weighting (referred to as “MetaST w/o Re-weighting”). The sampling strategy with weights proportional to the model confidence favors easy samples (referred to as “MetaST-Easy”), whereas the converse favors difficult ones (referred to as “MetaSTDifficult”).We observe the meta-learning based re-weighting strategy to perform the best. Interestingly, MetaST-Easy outperforms MetaST-Difficult significantly on CoNLL03 (EN) but achieves slightly lower performance on SNIPS. This demonstrates that difficult samples are more helpful when the quality of pseudo-labeled data is relatively high. On the converse, the sample selection strategy focusing on difficult samples introduces noisy examples with lower pseudo-label quality. Therefore, sampling strategies may need to vary for different datasets, thereby, demonstrating the necessity of adaptive data re-weighting as in our framework MetaST. Moreover, MetaST significantly outperforms classic self-training strategies with hard and soft pseudo-labels demonstrating the effectiveness of our design. + +Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting mechanism, we illustrate token-level re-weighting of MetaST on CoNLL03 (EN) dataset with ${ \mathrm { K } } { = } 1 0$ shot at step 100 in Fig. 2. We include the re-weighting visualisation on SNIPS in Appendix A.1. We observe that the selection mechanism filters out most of the noisy pseudo-labels (colored in blue) even those with high teacher confidence as shown in Fig. 2. + +# 5 RELATED WORK + +Semi-supervised learning has been widely used for consistency training (Bachman et al., 2014; Rasmus et al., 2015; Laine & Aila, 2017; Tarvainen & Valpola, 2017; Miyato et al., 2018), latent variable models (Kingma et al., 2014) for sentence compression (Miao & Blunsom, 2016) and code generation (Yin et al., 2018). More recently, methods like UDA (Xie et al., 2019) leverage consistency training for few-shot learning of instance-classification tasks leveraging auxiliary resources like paraphrasing and back-translation (BT) (Sennrich et al., 2016). + +Sample selection. Curriculum learning (Bengio et al., 2009) techniques are based on the idea of learning easier aspects of the task first followed by the more complex ones. Prior work leveraging self-paced learning (Kumar et al., 2010) and more recently self-paced co-training (Ma et al., 2017) leverage teacher confidence to select easy samples during training. Sample selection for image classification tasks have been explored in recent works with meta-learning (Ren et al., 2018; Li et al., 2019) and active learning (Panagiota Mastoropoulou, 2019; Chang et al., 2017b). However, all of these techniques rely on only the model outputs applied to instance-level classification tasks. + +Semi-supervised sequence labeling. Miller et al. (2004); Peters et al. (2017) leverage large amounts of unlabeled data to improve token representation for sequence labeling tasks. Another line of research introduces latent variable modeling (Chen et al., 2019; Zhou & Neubig, 2017) to learn interpretable and structured latent representations. Recently, adversarial training based model SeqVAT (Chen et al., 2020) and cross-view training method CVT (Clark et al., 2018) have shown promising results for sequence labeling tasks. + +# 6 CONCLUSIONS + +In this work, we develop an adaptive self-training framework MetaST that leverages self-training and meta-learning for few-shot training of neural sequence taggers. We address the issue of error propagation from noisy pseudo-labels from the teacher in the self-training framework by adaptive sample selection and re-weighting with meta-learning. Extensive experiments on six benchmark datasets and different tasks including multilingual NER and slot tagging for task-oriented dialog systems demonstrate the effectiveness of the proposed method particularly for low-resource settings. + +# REFERENCES + +Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Zoubin Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 3365–3373, 2014. + +Trapit Bansal, Rishikesh Jha, and Andrew McCallum. Learning to few-shot learn across diverse natural language classification tasks, 2020. + +Yoshua Bengio, Jer´ ome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ˆ Andrea Pohoreckyj Danyluk, Leon Bottou, and Michael L. Littman (eds.), ´ Proceedings of the 26th Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009, volume 382 of ACM International Conference Proceeding Series, pp. 41–48. ACM, 2009. doi: 10.1145/1553374.1553380. + +Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems 30, pp. 1002–1012. Curran Associates, Inc., 2017a. + +Haw-Shiuan Chang, Erik G. Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp. 1002–1012, 2017b. + +Haw-Shiuan Chang, Shankar Vembu, Sunil Mohan, Rheeya Uppaal, and Andrew McCallum. Using error decay prediction to overcome practical issues of deep active learning for named entity recognition. Machine Learning, 109(9):1749–1778, 2020. + +Olivier Chapelle, Bernhard Schlkopf, and Alexander Zien. Semi-supervised learning. 2010. + +Luoxin Chen, Weitong Ruan, Xinyue Liu, and Jianhua Lu. Seqvat: Virtual adversarial training for semi-supervised sequence labeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 8801–8811, 2020. + +Mingda Chen, Qingming Tang, Karen Livescu, and Kevin Gimpel. Variational sequential labelers for semi-supervised learning. arXiv preprint arXiv:1906.09535, 2019. + +Kevin Clark, Minh-Thang Luong, Christopher D Manning, and Quoc V Le. Semi-supervised sequence modeling with cross-view training. arXiv preprint arXiv:1809.08370, 2018. + +Alice Coucke, Alaa Saade, Adrien Ball, Theodore Bluche, Alexandre Caulier, David Leroy, Cl ´ ement ´ Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, Mael Primet, and ¨ Joseph Dureau. Snips voice platform: an embedded spoken language understanding system for private-by-design voice interfaces. In Privacy in Machine Learning and Artificial Intelligence workshop, ICML2018, 2018. + +Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pp. 4171–4186, 2019. + +Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pp. 1183–1192. PMLR, 2017. + +Junxian He, Jiatao Gu, Jiajun Shen, and Marc’Aurelio Ranzato. Revisiting self-training for neural sequence generation, 2019. + +H. J. Scudder III. Probability of error of some adaptive pattern-recognition machines. IEEE Trans. Inf. Theory, 11(3):363–371, 1965. doi: 10.1109/TIT.1965.1053799. + +Diederik P. Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semisupervised learning with deep generative models. In Zoubin Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 3581–3589, 2014. + +Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. arXiv preprint arXiv:1703.04730, 2017. + +Ksenia Konyushkova, Raphael Sznitman, and Pascal Fua. Learning active learning from data. In Advances in Neural Information Processing Systems, pp. 4225–4235, 2017. + +Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. arXiv preprint arXiv:2002.11361, 2020. + +M. P. Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta (eds.), Advances in Neural Information Processing Systems 23, pp. 1189–1197. Curran Associates, Inc., 2010. + +Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. + +Xinzhe Li, Qianru Sun, Yaoyao Liu, Qin Zhou, Shibao Zheng, Tat-Seng Chua, and Bernt Schiele. Learning to self-train for semi-supervised few-shot classification. In Advances in Neural Information Processing Systems 32, pp. 10276–10286. Curran Associates, Inc., 2019. + +Chen Liang, Yue Yu, Haoming Jiang, Siawpeng Er, Ruijia Wang, Tuo Zhao, and Chao Zhang. Bond: Bert-assisted open-domain named entity recognition with distant supervision. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 1054–1064, 2020. + +J. Liu, Panupong Pasupat, D. Cyphers, and James R. Glass. Asgard: A portable architecture for multilingual dialogue systems. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 8386–8390, 2013. + +Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. + +Fan Ma, Deyu Meng, Qi Xie, Zina Li, and Xuanyi Dong. Self-paced co-training. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp. 2275–2284, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR. + +Yishu Miao and Phil Blunsom. Language as a latent variable: Discrete generative models for sentence compression. In Jian Su, Xavier Carreras, and Kevin Duh (eds.), Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pp. 319–328. The Association for Computational Linguistics, 2016. doi: 10.18653/v1/d16-1031. + +Scott Miller, Jethran Guinness, and Alex Zamanian. Name tagging with word clusters and discriminative training. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 337– 342, 2004. + +Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. + +Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. Crosslingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1946–1958, Vancouver, Canada, July 2017. Association for Computational Linguistics. + +Emmeleia Panagiota Mastoropoulou. Enhancing deep active learning using selective self-training for image classification. Master’s thesis, KTH, School of Electrical Engineering and Computer Science (EECS), 2019. + +Matthew E Peters, Waleed Ammar, Chandra Bhagavatula, and Russell Power. Semi-supervised sequence tagging with bidirectional language models. arXiv preprint arXiv:1705.00108, 2017. + +Slav Petrov and Ryan McDonald. Overview of the 2012 shared task on parsing the web. 2012. + +Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. + +Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semisupervised learning with ladder networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pp. 3546–3554, 2015. + +Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International Conference on Machine Learning, pp. 4334–4343, 2018. + +Sebastian Ruder and Barbara Plank. Strong baselines for neural semi-supervised learning under domain shift. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1044–1054, 2018. + +Erik F. Tjong Kim Sang and Fien De Meulder. Introduction to the conll-2003 shared task: Languageindependent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, 2003. + +Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics, 2016. doi: 10.18653/v1/p16-1009. + +Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 761–769, 2016. + +Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weightnet: Learning an explicit mapping for sample weighting. In Advances in Neural Information Processing Systems, pp. 1919–1930, 2019. + +Q. Sun, Y. Liu, T. Chua, and B. Schiele. Meta-transfer learning for few-shot learning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019. + +Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings. OpenReview.net, 2017. + +Erik F Tjong, Kim Sang, and Jorn Veenstra. Representing text chunks. In Ninth Conference of the European Chapter of the Association for Computational Linguistics, 1999. + +Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised data augmentation for consistency training, 2019. + +Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. + +Pengcheng Yin, Chunting Zhou, Junxian He, and Graham Neubig. Structvae: Tree-structured latent variable models for semi-supervised semantic parsing. In Iryna Gurevych and Yusuke Miyao (eds.), Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pp. 754–765. Association for Computational Linguistics, 2018. doi: 10.18653/v1/P18-1070. + +Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. + +Chunting Zhou and Graham Neubig. Multi-space variational encoder-decoders for semi-supervised labeled sequence transduction. arXiv preprint arXiv:1704.01691, 2017. + +Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pre-training and self-training. Advances in Neural Information Processing Systems, 33, 2020. + +# A APPENDIX + +# A.1 EXPLORATIONS ON UNLABELED DATA AND MINI-BATCH S + +Variation in model performance with unlabeled data. Table 12 shows the improvement in model performance as we inject more unlabeled data with diminishing returns after a certain point. + +Variation in model performance with mini-batch S. We set the value of $S$ in Eq. 8 to $\{ 1 , 3 , 5 \}$ respectively to explore its impact on the re-weighting mechanism. From Figure 3 we observe that the model is not super sensitive to hyper-parameter $S$ but can achieve a better estimate of the weights of the pseudo-labeled data with increasing mini-batch values. + +Table 7: Varying proportion of unlabeled data for MetaST with 10 labels per slot. + +
Ratio of Unlabeled DataDatasets
SNIPSCoNLL03
5%84.4772.92
25%87.1076.46
75%87.5076.56
+ +![](images/4acf5be292e2775588f2a99e982a0a9dfdc3fa3dd074d857771804a5f6d9f1ce.jpg) +Figure 3: Varying $S$ mini-batch labeled data for re-weighting. + +# A.2 ANALYSIS OF RE-WEIGHTING ON SNIPS AND CONLL03 + +Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting mechanism, we illustrate token re-weighting of MetaST on CoNLL03 and SNIPS datasets with ${ \mathrm { K } } { = } 1 0$ shot at step 100 in Fig. 4. Besides the observation in the experimental section, we observe that many difficult and correct pseudo-labeled samples (low teacher confidence) are selected according to Fig. 4a. + +![](images/a1ce0ac42b92d98b8512da697f6121d4c41ab38cbc3fd9329203574c7580c82c.jpg) +Figure 4: Visualization of MetaST re-weighting examples on SNIPS and CoNLL03 (EN). + +# A.3 K-SHOTS + +Effect of varying the number of few-shots K. We show the performance changes with respect to varying number of few-shots K $\{ 5 , 1 0 , 2 0 , 1 0 0 \}$ on Wikiann (en), MIT movie, MIT Restaurant, CoNLL2003 (En), Multilingual CoNLL and Multilingual Wikiann in Table 9-13. Since the number of labeled examples for some slots in Email dataset is around 10, we only show 5 and 10 shots for Email dataset in Table 8. + +Table 8: Email Dataset. + +
MethodShots
510
Full-supervision
BERT0.9444
Few-shot Supervision
BERT0.82110.8785
Few-shot Supervision + unlabeled data
CVT67.4478.24
SeqVAT64.6772.65
Mean Teacher84.1089.53
VAT83.2489.71
Classic ST86.8890.70
BOND84.9289.75
MetaST89.2192.18
+ +Table 9: Wikiann (En) Dataset. + +
MethodShots (3 Slot Types)
51020100
Full-supervision
BERT84.04
Few-shot Supervision
BERT37.0145.6154.5367.87
Few-shot Supervision +unlabeled data
CVT16.0527.8946.4266.36
SeqVAT21.1135.1642.2662.37
Mean Teacher30.9241.4350.6167.16
VAT24.7238.8150.1566.31
Classic ST32.7246.1554.4168.64
BOND34.2248.7352.4568.89
MetaST55.0456.6160.3873.20
+ +Figure 5: MIT Movie Dataset. + +
MethodShots (12 Slot Types)
51020100
Full-supervision87.87
BERT
Few-shot Supervision
BERT62.8069.5075.8182.49
Few-shot Supervision +unlabeled data
CVT57.4862.7370.2081.82
SeqVAT60.9467.1074.1582.73
Mean Teacher58.9267.6275.2482.20
VAT60.7570.1775.4182.39
Classic ST63.3971.8876.5883.06
BOND62.5070.9175.5282.65
MetaST72.5777.6780.3384.35
+ +Table 10: MIT Restaurant Dataset. + +
MethodShots (8 Slot Types)
51020100
Full-supervision
BERT78.95
Few-shot Supervision
BERT41.3954.0660.1272.24
Few-shot Supervision +unlabeled data
CVT33.7442.5751.3370.84
SeqVAT41.9451.5556.1571.39
Mean Teacher40.3751.7557.3472.40
VAT41.2953.3459.6872.65
Classic ST44.3556.8060.2873.13
BOND43.0155.7859.9673.60
MetaST53.0263.8367.8675.25
+ +Table 11: CoNLL2003 (EN) + +
MethodShots (4 Slot Types)
51020100
Full-supervision92.40
BERT
Few-shot Supervision
BERT63.8771.1573.5784.36
Few-shot Supervision +unlabeled data
CVT51.1554.3166.1181.99
SeqVAT58.0267.2174.1582.20
Mean Teacher59.0468.6772.6284.17
VAT57.0365.0372.6984.43
Classic ST64.0470.9974.6584.93
BOND62.5269.5674.1983.87
MetaST71.4976.6578.5485.77
+ +Table 12: Multilingual CoNLL03. + +
MethodShots (4 Slot Types)
51020100
Full-supervision
BERT
Few-shot Supervision
BERT64.8070.7773.8980.61
Few-shot Supervision +unlabeled data
Mean Teacher64.5568.3473.8779.21
VAT64.9767.6374.2680.70
Classic ST67.9572.6973.7981.82
BOND69.4272.7976.0280.62
MetaST73.3476.6577.0182.11
+ +Table 13: Multilingual Wikiann + +
MethodShots (3 Slot Types X 41 languages)
51020100
Full-supervision
BERT
Few-shot Supervision82.3385.70
BERT77.6879.67
Few-shot Supervision +unlabeled data
Mean Teacher77.0980.2382.1985.34
VAT74.7178.8282.6085.82
Classic ST76.7380.2482.3986.08
BOND78.8179.5782.1986.14
MetaST79.1081.6183.1485.57
+ +# A.4 IMPLEMENTATIONS AND HYPER-PARAMETER + +We do not perform any hyper-parameter tuning for different datasets. The batch size and maximum sequence length varies due to data characteristics and are as shown in Tbale 14. The hyperparameters are as shown in Table 14. + +Also, we retain parameters from original BERT implementation from https://github.com/ huggingface/transformers. + +We implement SeqVAT based on https://github.com/jiesutd/NCRFpp and implement CVT following https://github.com/tensorflow/models/tree/master/ research/cvt_text. + +Table 14: Batch size, sequence length and BERT encoder choices across datasets + +
DatasetSequence LengthBatch SizeLabeled data sample size B|Unlabeled Batch SizeBERT Encoder
SNIPS64163232BERT-base-uncased
Email64163232BERT-base-cased
Movie64163232BERT-base-uncased
Restaurant64161632BERT-base-uncased
CoNLL03 (EN)12816832BERT-base-cased
Wikiann (EN)12816832BERT-base-cased
CoNLLO3 (multilingual)128163232BERT-multilingual-base-cased
Wikiann (multilingaul)128163232BERT-multilingual-base-cased
+ +Table 15: Hyper-parameters. + +
BERT attention dropout BERT hidden dropout Latest Iteration R in labeled data acquisition BERT output hidden size h0.3 0.3 5 768
Steps for fine-tuning teacher model on labeled data2000
Steps T for self-training model on unlabeled data3000
Mini-batch S Re-initialize Student5
Pseudo-label TypeY Hard
Warmup steps20
learning rate α5e-5
Weight_decay5e-6
\ No newline at end of file diff --git a/parse/train/ARFshOO1Iu/ARFshOO1Iu_content_list.json b/parse/train/ARFshOO1Iu/ARFshOO1Iu_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cc67c6ca6efba51cf7afca32b14d7392f9df0ac2 --- /dev/null +++ b/parse/train/ARFshOO1Iu/ARFshOO1Iu_content_list.json @@ -0,0 +1,1838 @@ +[ + { + "type": "text", + "text": "ADAPTIVE SELF-TRAINING FOR NEURAL SEQUENCE LABELING WITH FEW LABELS ", + "text_level": 1, + "bbox": [ + 176, + 98, + 823, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 236, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural sequence labeling is an important technique employed for many Natural Language Processing (NLP) tasks, such as Named Entity Recognition (NER), slot tagging for dialog systems and semantic parsing. Large-scale pre-trained language models obtain very good performance on these tasks when fine-tuned on large amounts of task-specific labeled data. However, such large-scale labeled datasets are difficult to obtain for several tasks and domains due to the high cost of human annotation as well as privacy and data access constraints for sensitive user applications. This is exacerbated for sequence labeling tasks requiring such annotations at token-level. In this work, we develop techniques to address the label scarcity challenge for neural sequence labeling models. Specifically, we develop self-training and meta-learning techniques for training neural sequence taggers with few labels. While self-training serves as an effective mechanism to learn from large amounts of unlabeled data – meta-learning helps in adaptive sample re-weighting to mitigate error propagation from noisy pseudo-labels. Extensive experiments on six benchmark datasets including two for massive multilingual NER and four slot tagging datasets for task-oriented dialog systems demonstrate the effectiveness of our method. With only 10 labeled examples for each class for each task, our method obtains $1 0 \\%$ improvement over state-of-the-art systems demonstrating its effectiveness for the low-resource setting. ", + "bbox": [ + 233, + 263, + 764, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 553, + 336, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivation. Deep neural networks typically require large amounts of training data to achieve stateof-the-art performance. Recent advances with pre-trained language models like BERT (Devlin et al., 2019), GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019) have reduced this annotation bottleneck. In this paradigm, large neural network models are trained on massive amounts of unlabeled data in a self-supervised manner. However, the success of these large-scale models still relies on fine-tuning them on large amounts of labeled data for downstream tasks. For instance, our experiments show $\\bar { 2 } 7 \\%$ relative improvement on an average when fine-tuning BERT with the full training set (2.5K-705K labels) vs. fine-tuning with only 10 labels per class. This poses several challenges for many real-world tasks. Not only is acquiring large amounts of labeled data for every task expensive and time consuming, but also not feasible in many cases due to data access and privacy constraints. This issue is exacerbated for sequence labeling tasks that require annotations at token- and slot-level as opposed to instance-level classification tasks. For example, an NER task can have slots like B-PER, I-PER, O-PER marking the beginning, intermediate and out-of-span markers for person names, and similar slots for the names of location and organization. Similarly, language understanding models for dialog systems rely on effective identification of what the user intends to do (intents) and the corresponding values as arguments (slots) for use by downstream applications. Therefore, fully supervised neural sequence taggers are expensive to train for such tasks, given the requirement of thousands of annotations for hundreds of slots for the many different intents. ", + "bbox": [ + 174, + 584, + 823, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semi-supervised learning (SSL) (Chapelle et al., 2010) is one of the promising paradigms to address labeled data scarcity by making effective use of large amounts of unlabeled data in addition to task-specific labeled data. Self-training (ST, (III, 1965)) as one of the earliest SSL approaches has recently shown state-of-the-art performance for tasks like image classification (Li et al., 2019; Xie et al., 2020) performing at par with supervised systems while using very few training labels. In contrast to such instance-level classification tasks, sequence labeling tasks have dependencies between the slots demanding different design choices for slot-level loss optimization for the limited labeled data setting. For instance, prior work (Ruder & Plank, 2018) using classic self-training techniques for sequence labeling did not find much success in the low-data regime with $1 0 \\%$ labeled data for the target domain. Although there has been some success with careful task-specific data selection (Petrov & McDonald, 2012) and more recently for distant supervision (Liang et al., 2020) using external resources like knowledge bases (e.g., Wikipedia). In contrast to these prior work, we develop techniques for self-training with limited labels and without any task-specific assumption or external knowledge. ", + "bbox": [ + 174, + 840, + 823, + 922 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4716dc9424f650ad710d7db8d4aff2f8dd06b88c864da63dd6e410117264cc70.jpg", + "image_caption": [ + "Figure 1: MetaST framework. " + ], + "image_footnote": [], + "bbox": [ + 307, + 87, + 678, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 271, + 825, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For self-training, a base model (teacher) is trained on some amount of labeled data and used to pseudo-annotate (task-specific) unlabeled data. The original labeled data is augmented with the pseudo-labeled data and used to train a student model. The student-teacher training is repeated until convergence. Traditionally in self-training frameworks, the teacher model pseudo-annotates unlabeled data without any sample selection. This may result in gradual drifts from self-training on noisy pseudo-labeled instances (Zhang et al., 2017). In order to deal with noisy labels and training set biases, Ren et al. (2018) propose a meta-learning technique to automatically re-weight noisy samples by their loss changes on a held-out clean labeled validation set. We adopt a similar principle in our work and leverage meta-learning to re-weight noisy pseudo-labeled examples from the teacher. While prior techniques for learning to re-weight examples have been developed for instance-level classification tasks, we extend them to operate at token-level for discrete sequence labeling tasks. To this end, we address some key challenges on how to construct an informative held-out validation set for token-level re-weighting. Prior works (Ren et al., 2018; Shu et al., 2019) for instance classification construct this validation set by random sampling. However, sequence labeling tasks involve many slots (e.g. WikiAnn has 123 slots over 41 languages) with variable difficulty and distribution in the data. In case of random sampling, the model oversamples from the most populous category and slots. This is particularly detrimental for low-resource languages in the multilingual setting. To this end, we develop an adaptive mechanism to create the validation set on the fly considering the diversity and uncertainty of the model for different slot types. Furthermore, we leverage this validation set for token-level loss estimation and re-weighting pseudo-labeled sequences from the teacher in the meta-learning setup. While prior works (Li et al., 2019; Sun et al., 2019; Bansal et al., 2020) on meta-learning for image and text classification leverage multi-task learning to improve a target classification task based on several similar tasks, in this work we focus on a single sequence labeling task – making our setup more challenging altogether. ", + "bbox": [ + 173, + 390, + 825, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our task and framework overview. We focus on sequence labeling tasks with only a few annotated samples (e.g., $K = \\{ 5 , 1 0 , 2 0 , 1 0 0 \\} )$ per slot type for training and large amounts of task-specific unlabeled data. Figure 1 shows an overview of our framework with the following components: (i) Self-training: Our self-training framework leverages a pre-trained language model as a teacher and co-trains a student model with iterative knowledge exchange (ii) Adaptive labeled data acquisition for validation: Our few-shot learning setup assumes a small number of labeled training samples per slot type. The labeled data from multiple slot types are not equally informative for the student model to learn from. While prior works in meta-learning randomly sample some labeled examples for held-out validation set, we develop an adaptive mechanism to create this set on the fly. To this end, we leverage loss decay as a proxy for model uncertainty to select informative labeled samples for the student model to learn from in conjunction with the re-weighting mechanism in the next step. (iii) Meta-learning for sample re-weighting: Since pseudo-labeled samples from the teacher can be noisy, we employ meta-learning to re-weight them to improve the student model performance on the held-out validation set obtained from the previous step. In contrast to prior work (Ren et al., 2018) ", + "bbox": [ + 173, + 729, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on sample re-weighting operating at instance-level, we incorporate the re-weighting mechanism at token-level for sequence labeling tasks. Here the token-level weights are determined by the student model loss on the above validation set. Finally, we learn all of the above steps jointly with end-toend learning in the self-training framework. We refer to our adaptive self-training framework with meta-learning based sample re-weighting mechanism as MetaST. ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We perform extensive experiments on six benchmark datasets for several tasks including multilingual Named Entity Recognition and slot tagging for user utterances from task-oriented dialog systems to demonstrate the generalizability of our approach across diverse tasks and languages. We adopt BERT and multilingual BERT as encoder and show that its performance can be significantly improved by nearly $1 0 \\%$ for low-resource settings with few training labels (e.g., 10 labeled examples per slot type) and large amounts of unlabeled data. In summary, our work makes the following contributions. (i) Develops a self-training framework for neural sequence tagging with few labeled training examples. (ii) Leverages an acquisition strategy to adaptively select a validation set from the labeled set for meta-learning of the student model. (iii) Develops a meta-learning framework for re-weighting pseudo-labeled samples at token-level to reduce drifts from noisy teacher predictions. (iv) Integrates the aforementioned components into an end-to-end learning framework and demonstrates its effectiveness for neural sequence labeling across six benchmark datasets with multiple slots, shots, domains and languages. ", + "bbox": [ + 173, + 180, + 825, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 174, + 372, + 326, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sequence labeling and slot tagging. This is the task identifying the entity span of several slot types (e.g., names of person, organization, location, date, etc.) in a text sequence. Formally, given a sentence with $N$ tokens $\\mathbf { \\bar { \\cal X } } ~ = ~ \\{ x _ { 1 } , . . . , x _ { N } \\}$ , an entity or slot value is a span of tokens $s = [ x _ { i } , . . . , x _ { j } ] ( 0 \\leq i \\leq j \\leq N )$ associated with a type. This task assumes a pre-defined tagging policy like BIO (Tjong et al., 1999), where B marks the beginning of the slot, I marks an intermediate token in the span, and $\\bigcirc$ marks out-of-span tokens. These span markers are used to extract multi-token values for each of the slot types with phrase-level evaluation for the performance. ", + "bbox": [ + 173, + 398, + 825, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-training. Consider $f ( \\cdot ; \\theta _ { t e a } )$ and $f ( \\cdot ; \\theta _ { s t u } )$ to denote the teacher and student models respectively in the self-training framework. The role of the teacher model (e.g., a pre-trained language model) is to assign pseudo-labels to unlabeled data that is used to train a student model. The teacher and student model can exchange knowledge and the training schedules are repeated till convergence. The success of self-training with deep neural networks in recent works (He et al., 2019; Xie et al., 2020) has been attributed to a number of factors including stochastic regularization with dropouts and data regularization with unlabeled data. Formally, given $m$ -th unlabeled sentence with $N$ tokens $X _ { m } ^ { u } \\doteq \\{ x _ { 1 , m } ^ { u } , . . . , x _ { N , m } ^ { u } \\}$ and $C$ pre-defined labels, consider the pseudo-labels $\\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ]$ generated by the teacher model at the $t$ -th iteration where, ", + "bbox": [ + 173, + 503, + 825, + 637 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/0e5a23d228b8b84c3f9f8d50a1a2a8b9f8936c649b392433f95ac84053ec7c07.jpg", + "text": "$$\n\\hat { y } _ { m , n } ^ { ( t ) } = \\underset { c \\in C } { \\arg \\operatorname* { m a x } } f _ { n , c } ( x _ { m , n } ^ { u } ; \\theta _ { t e a } ^ { ( t ) } ) .\n$$", + "text_format": "latex", + "bbox": [ + 383, + 645, + 612, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pseudo-labeled data set, denoted as $( X ^ { u } , \\hat { Y } ^ { ( t ) } ) = \\{ ( X _ { m } ^ { u } , \\hat { Y } _ { m } ^ { ( t ) } ) \\} _ { m } ^ { M }$ , is used to train the student model and learn its parameters as: ", + "bbox": [ + 174, + 684, + 823, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/529a8dc877aeeae9554ba104386433ff96b02e57164c11e0a3697f446980baa2.jpg", + "text": "$$\n\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\underset { \\theta } { \\arg \\operatorname* { m i n } } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) ,\n$$", + "text_format": "latex", + "bbox": [ + 339, + 722, + 656, + 766 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $l ( \\cdot , \\cdot )$ can be modeled as the cross-entropy loss. ", + "bbox": [ + 174, + 772, + 529, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 ADAPTIVE SELF TRAINING ", + "text_level": 1, + "bbox": [ + 176, + 803, + 433, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pre-trained language model (e.g., BERT (Devlin et al., 2019)) as the teacher, we first finetune it on the small labeled data to make it aware of the underlying task. The fine-tuned teacher model is now used to pseudo-label the large unlabeled data. We consider the student model as another instantiation of the pre-trained language model that is trained over the pseudo-labeled data. However, our few-shot setting with limited labeled data results in a noisy teacher. A naive transfer of teacher knowledge to the student results in the propagation of noisy labels limiting the performance of the student model. To address this challenge, we develop an adaptive self-training framework to re-weight pseudo-labeled predictions from the teacher with a meta-learning objective that optimizes the token-level loss from the student model on a held-out labeled validation set. This held-out set is adaptively constructed via labeled data acquisition which selects labeled samples with high uncertainty for efficient data exploration. ", + "bbox": [ + 173, + 825, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 ADAPTIVE LABELED DATA ACQUISITION ", + "text_level": 1, + "bbox": [ + 176, + 167, + 500, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In standard meta-learning setup for instance-level classification tasks, the held-out validation set is usually constructed via random sampling (Ren et al., 2018; Shu et al., 2019). Sequence labeling tasks involve many slot types with variable difficulty and distribution in the data. For instance, NER tasks over WikiAnn operate over 123 slot types from 41 languages with additional complexity from variable model performance across different languages. A random sampling leads to oversampling instances with the most populous categories and slot types in the data. Therefore, we propose a novel labeled data acquisition strategy to construct the validation set for effective data exploration. We demonstrate its benefit over classic meta-learning approaches from prior works in experiments. ", + "bbox": [ + 174, + 184, + 825, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, data acquisition strategies for prior works in meta-learning and active learning broadly leverage random sampling (Ren et al., 2018; Shu et al., 2019), easy (Kumar et al., 2010) and hard example mining (Shrivastava et al., 2016) or uncertainty-based methods (Chang et al., 2017a). These strategies have been compared in prior works (Chang et al., 2017a; Gal et al., 2017) that show uncertainty-based methods to have better generalizability across diverse settings. There are several approaches to uncertainty estimation including error decay (Konyushkova et al., 2017; Chang et al., 2020), Monte Carlo dropouts (Gal et al., 2017) and predictive variance (Chang et al., 2017a). We follow a similar principle of error decay to find samples that the model is uncertain about and can correspondingly benefit from knowing their labels (similar to active learning settings). To this end, we leverage stochastic loss decay from the model as a proxy for the model uncertainty to generate validation set on the fly. This is used for estimating token-level weights and re-weighting pseudo labeled data in Section 3.2. ", + "bbox": [ + 173, + 303, + 825, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider the loss of the student model with parameters $\\theta _ { s t u } ^ { ( t ) }$ on the labeled data $( X _ { m } ^ { l } , Y _ { m } )$ in the $t$ -th iteration as $l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) )$ . To measure the loss decay value at any iteration, we use the difference between the current and previous loss values. Considering these values may fluctuate across iterations, we adopt the moving average of the loss values for $( X _ { m } ^ { l } , Y _ { m } )$ in the latest $R$ iterations as a baseline $l _ { b } ^ { m }$ for loss decay estimation. Baseline measure $l _ { b } ^ { m }$ is calculated as follows: ", + "bbox": [ + 173, + 473, + 825, + 549 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/081689b844fe1c90c78194dc94909054d58f72d968de98028221f68df20e6fd6.jpg", + "text": "$$\nl _ { b } ^ { m } = \\frac { 1 } { R } \\sum _ { r = 1 } ^ { R } l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t - r ) } ) ) .\n$$", + "text_format": "latex", + "bbox": [ + 380, + 554, + 617, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the loss decay values are estimated on the fly, we want to balance exploration and exploitation. To this end, we add a smoothness factor $\\delta$ to prevent the low loss decay samples (i.e. samples with low uncertainty) from never being selected again. Considering all of the above factors, we obtain the sampling weight of labeled data $( X _ { m } ^ { l } , Y _ { m } ^ { l } )$ as follows: ", + "bbox": [ + 174, + 598, + 825, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/68c406599266593bd337a243fd21d5fe019742eef2d3d18fac7ca85f0ad01bfc.jpg", + "text": "$$\nW _ { m } \\propto \\mathrm { m a x } ( l _ { b } ^ { m } - l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) ) , 0 ) + \\delta .\n$$", + "text_format": "latex", + "bbox": [ + 343, + 659, + 653, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The smoothness factor $\\delta$ needs to be adaptive since the training loss is dynamic. Therefore, We adopt the maximum of the loss decay value as the smoothness factor $\\delta$ to encourage exploration. ", + "bbox": [ + 174, + 683, + 828, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The aforementioned acquisition function is re-estimated after a fixed number of steps to adapt to model changes. With labeled data acquisition, we rely on informative uncertain samples to improve learning efficiency. The sampled mini-batches of labeled data $\\{ B _ { s } ^ { l } \\}$ are used as a validation set for the student model in the next step for re-weighting pseudo-labeled data from the teacher model. We demonstrate its impact via ablation study in experiments. Note that the labeled data is only used to compute the acquisition function and not used for explicit training of the student model in this step. ", + "bbox": [ + 174, + 718, + 825, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 RE-WEIGHTING PSEUDO-LABELED DATA", + "text_level": 1, + "bbox": [ + 174, + 809, + 501, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate error propagation from noisy pseudo-labeled sequences from the teacher, we leverage meta-learning to adaptively re-weight them based on the student model loss on the held-out validation set obtained via labeled data acquisition from the previous section. In contrast to prior work focusing on instance-level tasks like image classification – sequence labeling operates on discrete text sequences as input and assigns labels to each token in the sequence. Since teacher predictions vary for different slot labels and types, we adapt the meta-learning framework to re-weight samples at a token-level resolution. ", + "bbox": [ + 174, + 827, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token Re-weighting. Consider the pseudo-labels $\\{ \\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ] \\} _ { m = 1 } ^ { M }$ from the teacher in the $t$ -th iteration with $m$ and $n$ indexing the instance and a token in the instance, respectively. In classic self-training, we update the student parameters leveraging pseudo-labels as follows: ", + "bbox": [ + 173, + 101, + 826, + 148 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a93edb77cbec55bb3f10a1c0c2f2e3059edbfed5a03d2bd61bb287549a51d107.jpg", + "text": "$$\n\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 318, + 161, + 678, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now, to downplay noisy token-level labels, we leverage meta-learning to re-weight the pseudolabeled data. To this end, we follow a similar analysis from (Koh & Liang, 2017) and (Ren et al., 2018) to perturb the weight for each token in the mini-batch by $\\epsilon$ . Weight perturbation is used to discover data points that are most important to improve the model performance on a held-out validation set (Koh & Liang, 2017) where the sample importance is given by the magnitude of the the negative gradients. We extend prior techniques to obtain token-level perturbations as: ", + "bbox": [ + 173, + 217, + 826, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/37247b0bb74268012fe5868440cb8e374fa97cafb08a7177e5a490724cedc343.jpg", + "text": "$$\n\\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ \\epsilon _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 256, + 314, + 741, + 358 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The token weights are obtained by minimizing the student model loss on the held-out validation set. Here, we employ the labeled data acquisition strategy from Eq. 4 to sample informative mini-batches of labeled data ${ \\dot { B } } _ { s } ^ { l }$ locally at step $t$ . To obtain a cheap estimate of the meta-weight at step $t$ , we take a single gradient descent step for the sampled labeled mini-batch $B _ { s } ^ { l }$ : ", + "bbox": [ + 174, + 369, + 825, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a54b73da96f95cdd452f67d2ea837739062736d61fa7d9339910bd9d96896b5e.jpg", + "text": "$$\nu _ { m , n , s } = - \\frac { \\partial } { \\partial \\epsilon _ { m , n , s } } \\big ( \\frac { 1 } { | \\mathcal { B } _ { s } ^ { l } | } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { | \\mathcal { B } _ { s } ^ { l } | } \\sum _ { n = 1 } ^ { N } [ l ( y _ { m , n } , f ( x _ { m , n } ^ { l } ; \\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) ) ] \\big ) | _ { \\epsilon _ { m , n , s } = 0 }\n$$", + "text_format": "latex", + "bbox": [ + 254, + 441, + 743, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set the token weights to be proportional to the negative gradients to reflect the importance of pseudo-labeled tokens in the sequence. Since sequence labeling tasks have dependencies between the slot types and tokens, it is difficult to obtain a good estimation of the weights based on a single mini-batch of examples. Therefore, we sample $S$ mini-batches of labeled data $\\{ B _ { 1 } ^ { l } , . . . , B _ { S } ^ { l } \\}$ with the adaptive acquisition strategy and calculate the mean of the gradients to obtain a robust gradient estimate. Note that $S$ is a constant number that is the same for each token and the proportional sign in Eq. 8. Since a negative weight indicates a pseudo-label of poor quality that would potentially degrade the model performance, we set such weights to 0 to filter them out. The impact of $S$ is investigated in the experiments (refer to Appendix A.1). The overall meta-weight of pseudo-labeled token $\\left( x _ { m , n } ^ { u } , \\hat { y } _ { m , n } \\right)$ is obtained as: ", + "bbox": [ + 173, + 507, + 825, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/158474a5f605855bf2ce03c2add7d17b0784b5e88eeaa7a49b97d059a0c1afa4.jpg", + "text": "$$\nw _ { m , n } \\propto \\mathrm { m a x } ( \\sum _ { s = 1 } ^ { S } u _ { m , n , s } , 0 )\n$$", + "text_format": "latex", + "bbox": [ + 403, + 662, + 594, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further ensure the stability of the loss function in each mini-batch, we normalise the weight $w _ { m , n }$ . Finally, we update the student model parameters while accounting for token-level re-weighting as: ", + "bbox": [ + 171, + 718, + 825, + 747 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/7eeb7ee5259c8235f204598ace160923989343c9104ae62454eb42a1a21e4cc3.jpg", + "text": "$$\n\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ w _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) .\n$$", + "text_format": "latex", + "bbox": [ + 263, + 760, + 735, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We demonstrate the impact of our re-weighting mechanism with an ablation study in experiments. ", + "bbox": [ + 173, + 823, + 813, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 TEACHER MODEL ITERATIVE UPDATES ", + "text_level": 1, + "bbox": [ + 174, + 853, + 486, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At the end of every self-training iteration, we assign the student model as a new teacher model (i.e., $\\theta _ { t e a } = \\theta _ { s t u } ^ { ( T ) } )$ . Since the student model uses the labeled data only as a held-out validation setl for meta-learning, we further utilize the labeled data ", + "bbox": [ + 174, + 877, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$f ( \\cdot , \\theta _ { t e a } ^ { ( t ) } )$ with standard supervised loss minimization. We explore the effectiveness of this step with an ablation study in experiments. The overall training procedure is summarized in Algorithm 1. ", + "bbox": [ + 171, + 103, + 825, + 133 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1: MetaST Algorithm. ", + "text_level": 1, + "bbox": [ + 176, + 145, + 398, + 160 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d4822d52a3210ec1b53952b0ae2190d99dbdf6149e0f2ef6f0fb9f27a125e7d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input:Labeled sequences (Xl,Y); Unlabeled sequences (X");Pre-trained BERT model with randomly initialized token classification layer f(;((O)); Batches S; Number of self-training iterations T.
Initialize teacher model 0tea =0(0) while not converged do
Fine-tune teacher model on small labeled data (Xl, Y);
Initialize the student model θ(0)=0(0);
Generate hard pseudo-labels Y(t）for unlabeled samples Xu with model f(·,0tea);
fort←1toTdo
Compute labeled data acquisition function according to Eq. 4;
function;Sample S mini-batches of labeled examples {Bl,.,B's} from (Xl,Y）based on labeled data acquisition
Randomly sample abatch of pseudo-labeled examples B from(Xu,Y(t));
Compute token-level weights in B based on the loss on {Bl,..,B's} according to Eq.8;
endstu
Update the teacher:θtea
end
", + "bbox": [ + 173, + 164, + 746, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 372, + 328, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Encoder. Pre-trained language models like BERT (Devlin et al., 2019), GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019) have shown state-of-the-art performance for various natural language processing tasks. In this work we adopt one of them as a base encoder by initializing the teacher with pre-trained BERT-base model and a randomly initialized token classification layer. ", + "bbox": [ + 173, + 396, + 825, + 452 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We perform large-scale experiments with six different datasets including user utterances for task-oriented dialog systems and multilingual Named Entity Recognition tasks as summarized in Table 1. (a) Email. This consists of natural language user utterances for email-oriented user actions like sending, receiving or searching emails with attributes like date, time, topics, people, etc. (b) SNIPS is a public benchmark ", + "bbox": [ + 174, + 459, + 473, + 597 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8275e215d6daaa4c5857892a43aa7a0317755556a49b5761a434b233c2aeffbf.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1: Dataset summary. " + ], + "table_body": "
Dataset# Slots#Train#Test#Lang
Email202.5K1kEN
SNIPS3913K0.7KEN
MIT Movie128.8K2.4KEN
MIT Restaurant86.9K1.5KEN
Wikiann (EN)320K10KEN
CoNLL03 (EN)415K3.6KEN
CoNLL031638K15K4
Wikiann123705K329K41
", + "bbox": [ + 486, + 465, + 812, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset (Coucke et al., 2018) of user queries from multiple domains including music, media, and weather. (c) MIT Movie and Restaurant corpus (Liu et al., 2013) consist of similar user utterances for movie and restaurant domains. (d) CoNLL03 (Sang & Meulder, 2003) and Wikiann (Pan et al., 2017) are public benchmark datasets for multilingual Named Entity Recognition. CoNLL03 is a collection of news wire articles from the Reuters Corpus from 4 languages with manual annotations, whereas Wikiann comprises of extractions from Wikipedia articles from 41 languages with automatic annotation leveraging meta-data for different entity types like ORG, PER, LOC etc. For every dataset, we sample $K \\in \\{ 5 , 1 0 , 2 0 , 1 0 0 \\}$ labeled sequences for each slot type from the Train data, and add the remaining to the unlabeled set while ignoring their labels – following standard setups for semi-supervised learning. We repeatedly sample $K$ labeled instances three times for multiple runs to report average performance with standard deviation across the runs. ", + "bbox": [ + 174, + 598, + 825, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. The first baseline we consider is the fully supervised BERT model trained on all available training data which provides the ceiling performance for every task. Each of the other models are trained on $K$ training labels per slot type. We adopt several state-of-the-art semi-supervised methods as baselines: (1) CVT (Clark et al., 2018) is a semi-supervised sequence labeling method based on cross-view training; (2) SeqVAT (Chen et al., 2020) incorporates adversarial training with conditional random field layer for semi-supervised sequence labeling; (3) Mean Teacher (MT) (Tarvainen & Valpola, 2017) averages model weights to obtain an aggregated teacher; (4) VAT (Miyato et al., 2018) adopts virtual adversarial training to make the model robust to noise; (5) classic ST (III, 1965) is simple self-training method with hard pseudo-labels; (6) BOND (Liang et al., 2020) is the most recent work on self-training for sequence labeling with confidence-based sample selection and forms a strong baseline for our work. We implement our framework in Pytorch and use Tesla V100 gpus for experiments. Hyper-parameter configurations with model settings presented in Appendix. ", + "bbox": [ + 173, + 757, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Neural sequence labeling performance with few training labels. Table 2 shows the performance comparison among different models with $\\mathrm { K } { = } 1 0$ labeled examples per slot type. The fully supervised BERT trained on thousands of labeled examples provides the ceiling performance for the few-shot setting. We observe our method MetaST to significantly outperform all methods across all datasets including the models that also use the same BERT encoder as ours like MT, VAT, Classic ST and BOND with corresponding average performance improvements as $1 4 . 2 2 \\%$ , $1 4 . 9 0 \\%$ , $8 . 4 6 \\%$ and $8 . 8 2 \\%$ . Non BERT models like CVT and ${ \\tt S e q V A T }$ are consistently worse than other baselines. ", + "bbox": [ + 173, + 103, + 825, + 202 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/807ffdbed937b94ff4fc00672dd73d1c64c56dace02c37ba83e830ba3369cdbb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSNIPSEmailMovieRestaurantCoNLL03 (EN)Wikiann (EN)
# Slots392012843
Full-supervision
BERT95.8094.4487.8778.9592.4084.04
Few-shot supervision (10 labels per slot)69.5054.0671.1545.61
BERT 79.01 87.85
Few-shot supervision (10 labels per slot) + unlabeled data
CVT78.2378.2462.7342.5754.3127.89
SeqVAT78.6772.6567.1051.5567.2135.16
MT79.4889.5367.6251.7568.6741.43
VAT79.0889.7170.1753.3465.0338.81
Classic ST83.2690.7071.8856.8070.9946.15
BOND83.5489.7570.9155.7869.5648.73
MetaST88.2392.1877.6763.8376.65
(0.04;↑12%)(0.47;↑4.93%)(0.10;↑11.76%)(1.62;↑18.07%)(0.73;17.73%)56.61 (0.4;↑24.12%)
", + "bbox": [ + 173, + 210, + 821, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2: F1 score comparison of models for sequence labeling on different datasets. All models (except CVT and ${ \\tt S e q V A T } ,$ use the same BERT encoder. F1 score of our model for each task is followed by standard deviation and percentage improvement (↑) over BERT with few-shot supervision. ", + "bbox": [ + 173, + 417, + 825, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also observe variable performance of the models across different tasks. Specifically, the performance gap between the best few-shot model and the fully supervised model varies significantly. MetaST achieves close performance to the fully-supervised model in some datasets (e.g. SNIPS and Email) but has bigger room for improvement in others (e.g. CoNLL03 (EN) and Wikiann (EN)). This can be attributed to the following factors. (i) Labeled training examples and slots. The total number of labeled training instances for our K-shot setting is given by $K \\times \\# S l o t s$ . Therefore, for tasks with higher number of slots and consequently more training labels, most of the models perform better including MetaST. Task-oriented dialog systems with more slots and inherent dependency between the slot types benefit more than NER tasks. (ii) Task difficulty: User utterances from task-oriented dialog systems for some of the domains like weather, music and emails contain predictive query patterns and limited diversity. In contrast, Named Entity Recognition datasets are comparatively diverse and require more training labels to generalize well. Similar observations are also depicted in Table 3 for multilingual NER tasks with more slots and consequently more training labels from multiple languages as well as richer interactions across the slots from different languages. ", + "bbox": [ + 173, + 468, + 825, + 662 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0163b49f589eb8049dc62232c52dfb0b36f2fa1d5e795dc8ccb45a6b066ed870.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Lang#SlotsFull Sup.Few-shot Sup.Few-shot supervision + unlabeled data
BERTBERTMTVATClassic STBONDMetaST
CoNLL0341687.6770.7768.3467.6372.6972.7976.41 (0.47) (↑ 7.97%)
Wikiann4112387.1779.6780.2378.8280.2479.5781.61 (0.14) (↑ 2.42%)
", + "bbox": [ + 176, + 669, + 821, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3: F1 score comparison of models for sequence labeling on multilingual datasets using the same BERT-Multilingual-Base encoder. F1 score of MetaST for each task is followed by standard deviation in parentheses and percentage improvement (↑) over BERT with few-shot supervision. ", + "bbox": [ + 176, + 741, + 821, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Controlling for the total amount of labeled data. In order to control for the variable amount of training labels across different datasets, we perform another experiment where we vary the number of labels for different slot types while keeping the total number of labeled instances for each dataset similar (ca. 200). Results are shown in Table 4. To better illustrate the effect of the number of training labels, we choose tasks with lower performance in Table 2 for this experiment. Comparing the results in Tables 2 and 4, we observe the performance of MetaST to improve with more training labels for all the tasks . ", + "bbox": [ + 174, + 790, + 825, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of varying the number of labels $K$ per slot. Table 5 shows the improvement in the performance of MetaST when increasing the number of labels for each slot type in the SNIPS dataset. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/341320487be14546e4f7fd3e1dfa7a103dea98dca4596616889fdc0c97e73bd2.jpg", + "table_caption": [ + "Table 4: F1 scores of different models with 200 labeled samples for each task. The percentage improvement $( \\uparrow )$ is over the BERT model with few-shot supervision. " + ], + "table_footnote": [], + "table_body": "
DatasetBERT (Full Supervision)BERT (Few-shot Supervision)MetaST（%Improvement）
MIT Movie87.8775.8180.33 (↑5.96%)
MIT Restaurant78.9560.1267.86 (↑12.87%)
CoNLL03 (EN)92.4077.4881.61 (↑ 5.33%)
Wikiann (EN)84.0462.0471.27 (个14.88%)
Average85.8268.8675.27 (↑ 9.31%)
", + "bbox": [ + 223, + 92, + 772, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Similar trends can be found on other datasets (results in Appendix). As we increase the amount of labeled training instances, the performance of BERT also improves, and correspondingly the margin between MetaST and these baselines decreases although MetaST still improves over all of them. In the self-training framework, given the ceiling performance for every task and the improved performance of the teacher with more training labels, there is less room for (relative) improvement of the student over the teacher model. Consider SNIPS for example. Our model obtains $12 \\%$ and $2 \\%$ improvement over the few-shot BERT model for the 10-shot and 100-shot setting with F1-scores as $8 8 . 2 2 \\%$ and $9 5 . 3 9 \\%$ , respectively. The ceiling performance for this task is $9 5 . 8 \\%$ on training BERT on the entire dataset with 13K labeled examples. This demonstrates that MetaST is most impactful for low-resource settings with few training labels for a given task. ", + "bbox": [ + 174, + 218, + 825, + 358 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d87a88b88eea368e0ed209527bab0ae1bc88c0732dee080f36daf2334a029db1.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5: Variation in model performance on varying $K$ labels / slot on SNIPS dataset with 39 slots. The percentage improvement $( \\uparrow )$ is relative to the BERT model with few-shot supervision. " + ], + "table_body": "
#SlotsFew-shot SupervisionFew-shot supervision + unlabeled data
BERTCVTSeqVATMTVATClassic STBONDMetaST(%Improvement)
570.6369.8269.3470.8571.3472.5972.8581.56 (↑15%)
1079.0178.2378.6779.4879.0883.2683.5488.22 (↑12%)
2086.8188.0485.0587.3188.1988.3288.9391.99 (↑6%)
10093.9094.6191.4694.2694.5393.9294.2295.39 (12%)
", + "bbox": [ + 209, + 366, + 792, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation analysis. Table 6 demonstrates the impact of different MetaST components with ablation analysis. We observe that soft pseudo-labels hurt the model performance compared to hard pseudolabels, as also shown in recent work (Kumar et al., 2020). Such a performance drop may be attributed to soft labels being less informative compared to sharpened ones. Removing the iterative teacher fine-tuning step (Section 3.1) also hurts the overall performance. ", + "bbox": [ + 174, + 483, + 823, + 554 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/97f50f81144a3b1bc21a414d82218b82ee1a37c17c4b6780a77ef271131c1485.jpg", + "table_caption": [ + "Table 6: Ablation analysis of our framework MetaST with 10 labeled examples per slot on SNIPS and CoNLL03 (EN). " + ], + "table_footnote": [], + "table_body": "
MethodDatasets
SNIPSCoNLL03
BERT w/ Continued Pre-training +
Few-shot Supervision83.9669.84
Classic ST83.2670.99
Classic ST w/ Soft Pseudo-Labels81.1771.87
MetaST(ours) w/Hard Pseudo-Labels88.2376.65
MetaST w/ Soft Pseudo-Labels86.1675.84
MetaST w/o Iterative Teacher Fine-tune85.6472.74
MetaST w/o Labeled Data Acq.86.6375.02
Pseudo-labeled Data Re-weighting
MetaST w/o Re-weighting85.4873.02
MetaST (Easy)85.5674.53
MetaST (Difficult)86.3468.06
", + "bbox": [ + 207, + 558, + 535, + 742 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e003691b635563a0994ab8d5c2a1f4aff6b7c20f56ba95beaa0a166af22b000.jpg", + "image_caption": [ + "Figure 2: Visualization of MetaST reweighting on CoNLL03 (EN). " + ], + "image_footnote": [], + "bbox": [ + 570, + 583, + 825, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Continued pre-training v.s. self-training. To contrast continued pre-training with self-training, we further pre-train BERT on in-domain unlabeled data and then fine-tune it with few labeled examples denoted as “BERT (Continued Pre-training $^ +$ Few-shot Supervision)”. The pre-training step improves the BERT performance over the baseline on SNIPS but degrades the performance on CoNLL03. This indicates that continued pre-training can improve the performance of few-shot supervised BERT on specialized tasks (e.g., SNIPS) with different data distribution than the original pre-training data (e.g., Wikipedia), but may not help for general domain ones like CoNLL03 with overlapping data from Wikipedia. In contrast to the above baseline, MetaST brings significant improvements on both datasets. This demonstrates the generality and flexibility of self-training over pre-training as also observed in contemporary work (Zoph et al., 2020) on image classification. ", + "bbox": [ + 173, + 786, + 825, + 925 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive labeled data acquisition. We perform an ablation study by removing adaptive labeled data acquisition from MetaST (denoted as “MetaST w/o Labeled Data Acq.”). Removing this component leads to around $2 \\%$ performance drop on an average demonstrating the impact of labeled data acquisition. Moreover, the performance drop on SNIPS (39 slots) is larger than that on CoNLL03 (4 slots). This demonstrates that adaptive acquisition is more helpful for tasks with more slot types – where diversity and data distribution necessitate a better exploration strategy in contrast to random sampling employed in prior meta-learning works. ", + "bbox": [ + 174, + 103, + 825, + 202 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Re-weighting strategies. To explore the role of token-level re-weighting for pseudo-labeled sequences (discussed in Section 3.2), we replace our meta-learning component with different sample selection strategies based on the model confidence for different tokens. One sampling strategy chooses samples uniformly without any re-weighting (referred to as “MetaST w/o Re-weighting”). The sampling strategy with weights proportional to the model confidence favors easy samples (referred to as “MetaST-Easy”), whereas the converse favors difficult ones (referred to as “MetaSTDifficult”).We observe the meta-learning based re-weighting strategy to perform the best. Interestingly, MetaST-Easy outperforms MetaST-Difficult significantly on CoNLL03 (EN) but achieves slightly lower performance on SNIPS. This demonstrates that difficult samples are more helpful when the quality of pseudo-labeled data is relatively high. On the converse, the sample selection strategy focusing on difficult samples introduces noisy examples with lower pseudo-label quality. Therefore, sampling strategies may need to vary for different datasets, thereby, demonstrating the necessity of adaptive data re-weighting as in our framework MetaST. Moreover, MetaST significantly outperforms classic self-training strategies with hard and soft pseudo-labels demonstrating the effectiveness of our design. ", + "bbox": [ + 174, + 204, + 825, + 411 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting mechanism, we illustrate token-level re-weighting of MetaST on CoNLL03 (EN) dataset with ${ \\mathrm { K } } { = } 1 0$ shot at step 100 in Fig. 2. We include the re-weighting visualisation on SNIPS in Appendix A.1. We observe that the selection mechanism filters out most of the noisy pseudo-labels (colored in blue) even those with high teacher confidence as shown in Fig. 2. ", + "bbox": [ + 174, + 419, + 825, + 488 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 496, + 346, + 512 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Semi-supervised learning has been widely used for consistency training (Bachman et al., 2014; Rasmus et al., 2015; Laine & Aila, 2017; Tarvainen & Valpola, 2017; Miyato et al., 2018), latent variable models (Kingma et al., 2014) for sentence compression (Miao & Blunsom, 2016) and code generation (Yin et al., 2018). More recently, methods like UDA (Xie et al., 2019) leverage consistency training for few-shot learning of instance-classification tasks leveraging auxiliary resources like paraphrasing and back-translation (BT) (Sennrich et al., 2016). ", + "bbox": [ + 174, + 522, + 825, + 606 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sample selection. Curriculum learning (Bengio et al., 2009) techniques are based on the idea of learning easier aspects of the task first followed by the more complex ones. Prior work leveraging self-paced learning (Kumar et al., 2010) and more recently self-paced co-training (Ma et al., 2017) leverage teacher confidence to select easy samples during training. Sample selection for image classification tasks have been explored in recent works with meta-learning (Ren et al., 2018; Li et al., 2019) and active learning (Panagiota Mastoropoulou, 2019; Chang et al., 2017b). However, all of these techniques rely on only the model outputs applied to instance-level classification tasks. ", + "bbox": [ + 174, + 613, + 825, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Semi-supervised sequence labeling. Miller et al. (2004); Peters et al. (2017) leverage large amounts of unlabeled data to improve token representation for sequence labeling tasks. Another line of research introduces latent variable modeling (Chen et al., 2019; Zhou & Neubig, 2017) to learn interpretable and structured latent representations. Recently, adversarial training based model SeqVAT (Chen et al., 2020) and cross-view training method CVT (Clark et al., 2018) have shown promising results for sequence labeling tasks. ", + "bbox": [ + 174, + 717, + 825, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 814, + 328, + 830 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In this work, we develop an adaptive self-training framework MetaST that leverages self-training and meta-learning for few-shot training of neural sequence taggers. We address the issue of error propagation from noisy pseudo-labels from the teacher in the self-training framework by adaptive sample selection and re-weighting with meta-learning. Extensive experiments on six benchmark datasets and different tasks including multilingual NER and slot tagging for task-oriented dialog systems demonstrate the effectiveness of the proposed method particularly for low-resource settings. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Zoubin Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 3365–3373, 2014. ", + "bbox": [ + 174, + 126, + 825, + 194 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Trapit Bansal, Rishikesh Jha, and Andrew McCallum. Learning to few-shot learn across diverse natural language classification tasks, 2020. ", + "bbox": [ + 171, + 204, + 823, + 232 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yoshua Bengio, Jer´ ome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ˆ Andrea Pohoreckyj Danyluk, Leon Bottou, and Michael L. Littman (eds.), ´ Proceedings of the 26th Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada, June 14-18, 2009, volume 382 of ACM International Conference Proceeding Series, pp. 41–48. ACM, 2009. doi: 10.1145/1553374.1553380. ", + "bbox": [ + 174, + 241, + 825, + 310 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural Information Processing Systems 30, pp. 1002–1012. Curran Associates, Inc., 2017a. ", + "bbox": [ + 173, + 319, + 825, + 376 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Haw-Shiuan Chang, Erik G. Learned-Miller, and Andrew McCallum. Active bias: Training more accurate neural networks by emphasizing high variance samples. In Isabelle Guyon, Ulrike von Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp. 1002–1012, 2017b. ", + "bbox": [ + 173, + 383, + 825, + 467 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Haw-Shiuan Chang, Shankar Vembu, Sunil Mohan, Rheeya Uppaal, and Andrew McCallum. Using error decay prediction to overcome practical issues of deep active learning for named entity recognition. Machine Learning, 109(9):1749–1778, 2020. ", + "bbox": [ + 173, + 476, + 823, + 518 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Olivier Chapelle, Bernhard Schlkopf, and Alexander Zien. Semi-supervised learning. 2010. ", + "bbox": [ + 171, + 526, + 774, + 541 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Luoxin Chen, Weitong Ruan, Xinyue Liu, and Jianhua Lu. Seqvat: Virtual adversarial training for semi-supervised sequence labeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 8801–8811, 2020. ", + "bbox": [ + 173, + 549, + 826, + 592 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mingda Chen, Qingming Tang, Karen Livescu, and Kevin Gimpel. Variational sequential labelers for semi-supervised learning. arXiv preprint arXiv:1906.09535, 2019. ", + "bbox": [ + 173, + 599, + 823, + 628 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kevin Clark, Minh-Thang Luong, Christopher D Manning, and Quoc V Le. Semi-supervised sequence modeling with cross-view training. arXiv preprint arXiv:1809.08370, 2018. ", + "bbox": [ + 173, + 637, + 823, + 666 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alice Coucke, Alaa Saade, Adrien Ball, Theodore Bluche, Alexandre Caulier, David Leroy, Cl ´ ement ´ Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, Mael Primet, and ¨ Joseph Dureau. Snips voice platform: an embedded spoken language understanding system for private-by-design voice interfaces. In Privacy in Machine Learning and Artificial Intelligence workshop, ICML2018, 2018. ", + "bbox": [ + 173, + 674, + 825, + 744 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), pp. 4171–4186, 2019. ", + "bbox": [ + 174, + 752, + 825, + 823 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of Proceedings of Machine Learning Research, pp. 1183–1192. PMLR, 2017. ", + "bbox": [ + 174, + 830, + 825, + 887 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junxian He, Jiatao Gu, Jiajun Shen, and Marc’Aurelio Ranzato. Revisiting self-training for neural sequence generation, 2019. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "H. J. Scudder III. Probability of error of some adaptive pattern-recognition machines. IEEE Trans. Inf. Theory, 11(3):363–371, 1965. doi: 10.1109/TIT.1965.1053799. ", + "bbox": [ + 169, + 103, + 821, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Diederik P. Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semisupervised learning with deep generative models. In Zoubin Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.), Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 3581–3589, 2014. ", + "bbox": [ + 173, + 142, + 825, + 213 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. arXiv preprint arXiv:1703.04730, 2017. ", + "bbox": [ + 173, + 222, + 823, + 251 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ksenia Konyushkova, Raphael Sznitman, and Pascal Fua. Learning active learning from data. In Advances in Neural Information Processing Systems, pp. 4225–4235, 2017. ", + "bbox": [ + 173, + 261, + 823, + 291 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. arXiv preprint arXiv:2002.11361, 2020. ", + "bbox": [ + 173, + 299, + 823, + 329 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "M. P. Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta (eds.), Advances in Neural Information Processing Systems 23, pp. 1189–1197. Curran Associates, Inc., 2010. ", + "bbox": [ + 174, + 338, + 823, + 382 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. ", + "bbox": [ + 174, + 391, + 823, + 434 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xinzhe Li, Qianru Sun, Yaoyao Liu, Qin Zhou, Shibao Zheng, Tat-Seng Chua, and Bernt Schiele. Learning to self-train for semi-supervised few-shot classification. In Advances in Neural Information Processing Systems 32, pp. 10276–10286. Curran Associates, Inc., 2019. ", + "bbox": [ + 173, + 443, + 825, + 487 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chen Liang, Yue Yu, Haoming Jiang, Siawpeng Er, Ruijia Wang, Tuo Zhao, and Chao Zhang. Bond: Bert-assisted open-domain named entity recognition with distant supervision. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp. 1054–1064, 2020. ", + "bbox": [ + 173, + 496, + 825, + 553 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "J. Liu, Panupong Pasupat, D. Cyphers, and James R. Glass. Asgard: A portable architecture for multilingual dialogue systems. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 8386–8390, 2013. ", + "bbox": [ + 173, + 563, + 825, + 606 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining approach. CoRR, abs/1907.11692, 2019. ", + "bbox": [ + 174, + 614, + 823, + 659 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fan Ma, Deyu Meng, Qi Xie, Zina Li, and Xuanyi Dong. Self-paced co-training. In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pp. 2275–2284, International Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR. ", + "bbox": [ + 173, + 667, + 825, + 724 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yishu Miao and Phil Blunsom. Language as a latent variable: Discrete generative models for sentence compression. In Jian Su, Xavier Carreras, and Kevin Duh (eds.), Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016, pp. 319–328. The Association for Computational Linguistics, 2016. doi: 10.18653/v1/d16-1031. ", + "bbox": [ + 174, + 734, + 825, + 805 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Scott Miller, Jethran Guinness, and Alex Zamanian. Name tagging with word clusters and discriminative training. In Proceedings of the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 337– 342, 2004. ", + "bbox": [ + 174, + 815, + 825, + 871 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979–1993, 2018. ", + "bbox": [ + 176, + 882, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. Crosslingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1946–1958, Vancouver, Canada, July 2017. Association for Computational Linguistics. ", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Emmeleia Panagiota Mastoropoulou. Enhancing deep active learning using selective self-training for image classification. Master’s thesis, KTH, School of Electrical Engineering and Computer Science (EECS), 2019. ", + "bbox": [ + 173, + 167, + 825, + 210 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Matthew E Peters, Waleed Ammar, Chandra Bhagavatula, and Russell Power. Semi-supervised sequence tagging with bidirectional language models. arXiv preprint arXiv:1705.00108, 2017. ", + "bbox": [ + 173, + 219, + 823, + 248 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Slav Petrov and Ryan McDonald. Overview of the 2012 shared task on parsing the web. 2012. ", + "bbox": [ + 173, + 257, + 794, + 272 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. ", + "bbox": [ + 171, + 280, + 823, + 309 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semisupervised learning with ladder networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett (eds.), Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pp. 3546–3554, 2015. ", + "bbox": [ + 174, + 318, + 825, + 388 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International Conference on Machine Learning, pp. 4334–4343, 2018. ", + "bbox": [ + 173, + 396, + 823, + 426 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Sebastian Ruder and Barbara Plank. Strong baselines for neural semi-supervised learning under domain shift. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1044–1054, 2018. ", + "bbox": [ + 174, + 434, + 825, + 477 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Erik F. Tjong Kim Sang and Fien De Meulder. Introduction to the conll-2003 shared task: Languageindependent named entity recognition. In Proceedings of the Seventh Conference on Natural Language Learning at HLT-NAACL 2003, 2003. ", + "bbox": [ + 174, + 486, + 825, + 529 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers. The Association for Computer Linguistics, 2016. doi: 10.18653/v1/p16-1009. ", + "bbox": [ + 174, + 536, + 825, + 593 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 761–769, 2016. ", + "bbox": [ + 174, + 601, + 823, + 645 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weightnet: Learning an explicit mapping for sample weighting. In Advances in Neural Information Processing Systems, pp. 1919–1930, 2019. ", + "bbox": [ + 173, + 652, + 823, + 695 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Q. Sun, Y. Liu, T. Chua, and B. Schiele. Meta-transfer learning for few-shot learning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019. ", + "bbox": [ + 169, + 704, + 823, + 733 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track Proceedings. OpenReview.net, 2017. ", + "bbox": [ + 174, + 741, + 825, + 797 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Erik F Tjong, Kim Sang, and Jorn Veenstra. Representing text chunks. In Ninth Conference of the European Chapter of the Association for Computational Linguistics, 1999. ", + "bbox": [ + 173, + 806, + 821, + 835 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised data augmentation for consistency training, 2019. ", + "bbox": [ + 173, + 843, + 821, + 873 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. ", + "bbox": [ + 176, + 881, + 825, + 924 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Pengcheng Yin, Chunting Zhou, Junxian He, and Graham Neubig. Structvae: Tree-structured latent variable models for semi-supervised semantic parsing. In Iryna Gurevych and Yusuke Miyao (eds.), Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pp. 754–765. Association for Computational Linguistics, 2018. doi: 10.18653/v1/P18-1070. ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning requires rethinking generalization. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net, 2017. ", + "bbox": [ + 174, + 181, + 825, + 238 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Chunting Zhou and Graham Neubig. Multi-space variational encoder-decoders for semi-supervised labeled sequence transduction. arXiv preprint arXiv:1704.01691, 2017. ", + "bbox": [ + 173, + 247, + 825, + 276 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pre-training and self-training. Advances in Neural Information Processing Systems, 33, 2020. ", + "bbox": [ + 174, + 285, + 825, + 328 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 297, + 117 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.1 EXPLORATIONS ON UNLABELED DATA AND MINI-BATCH S ", + "text_level": 1, + "bbox": [ + 174, + 133, + 619, + 148 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Variation in model performance with unlabeled data. Table 12 shows the improvement in model performance as we inject more unlabeled data with diminishing returns after a certain point. ", + "bbox": [ + 173, + 159, + 825, + 188 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Variation in model performance with mini-batch S. We set the value of $S$ in Eq. 8 to $\\{ 1 , 3 , 5 \\}$ respectively to explore its impact on the re-weighting mechanism. From Figure 3 we observe that the model is not super sensitive to hyper-parameter $S$ but can achieve a better estimate of the weights of the pseudo-labeled data with increasing mini-batch values. ", + "bbox": [ + 174, + 194, + 825, + 251 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/230f8633595e21b91bcb56d6a4b5450e066fa0bfd2a45bfdbb9e8e404f33f8c9.jpg", + "table_caption": [ + "Table 7: Varying proportion of unlabeled data for MetaST with 10 labels per slot. " + ], + "table_footnote": [], + "table_body": "
Ratio of Unlabeled DataDatasets
SNIPSCoNLL03
5%84.4772.92
25%87.1076.46
75%87.5076.56
", + "bbox": [ + 196, + 289, + 473, + 363 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/4acf5be292e2775588f2a99e982a0a9dfdc3fa3dd074d857771804a5f6d9f1ce.jpg", + "image_caption": [ + "Figure 3: Varying $S$ mini-batch labeled data for re-weighting. " + ], + "image_footnote": [], + "bbox": [ + 549, + 268, + 745, + 388 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.2 ANALYSIS OF RE-WEIGHTING ON SNIPS AND CONLL03 ", + "text_level": 1, + "bbox": [ + 174, + 452, + 611, + 467 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting mechanism, we illustrate token re-weighting of MetaST on CoNLL03 and SNIPS datasets with ${ \\mathrm { K } } { = } 1 0$ shot at step 100 in Fig. 4. Besides the observation in the experimental section, we observe that many difficult and correct pseudo-labeled samples (low teacher confidence) are selected according to Fig. 4a. ", + "bbox": [ + 173, + 478, + 825, + 547 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/a1ce0ac42b92d98b8512da697f6121d4c41ab38cbc3fd9329203574c7580c82c.jpg", + "image_caption": [ + "Figure 4: Visualization of MetaST re-weighting examples on SNIPS and CoNLL03 (EN). " + ], + "image_footnote": [], + "bbox": [ + 189, + 575, + 763, + 728 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A.3 K-SHOTS ", + "text_level": 1, + "bbox": [ + 176, + 781, + 282, + 795 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Effect of varying the number of few-shots K. We show the performance changes with respect to varying number of few-shots K $\\{ 5 , 1 0 , 2 0 , 1 0 0 \\}$ on Wikiann (en), MIT movie, MIT Restaurant, CoNLL2003 (En), Multilingual CoNLL and Multilingual Wikiann in Table 9-13. Since the number of labeled examples for some slots in Email dataset is around 10, we only show 5 and 10 shots for Email dataset in Table 8. ", + "bbox": [ + 174, + 806, + 825, + 877 + ], + "page_idx": 13 + }, + { + "type": "table", + "img_path": "images/a0a70b768721fcc3b76448f14f149d4e4414c737bad75fd505c92a646c5b05ab.jpg", + "table_caption": [ + "Table 8: Email Dataset. " + ], + "table_footnote": [], + "table_body": "
MethodShots
510
Full-supervision
BERT0.9444
Few-shot Supervision
BERT0.82110.8785
Few-shot Supervision + unlabeled data
CVT67.4478.24
SeqVAT64.6772.65
Mean Teacher84.1089.53
VAT83.2489.71
Classic ST86.8890.70
BOND84.9289.75
MetaST89.2192.18
", + "bbox": [ + 370, + 130, + 627, + 333 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/95fcfbdbbf909669be074fbba1a4c6a7b11a739db36ef810a02b75c6dcd33573.jpg", + "table_caption": [ + "Table 9: Wikiann (En) Dataset. " + ], + "table_footnote": [], + "table_body": "
MethodShots (3 Slot Types)
51020100
Full-supervision
BERT84.04
Few-shot Supervision
BERT37.0145.6154.5367.87
Few-shot Supervision +unlabeled data
CVT16.0527.8946.4266.36
SeqVAT21.1135.1642.2662.37
Mean Teacher30.9241.4350.6167.16
VAT24.7238.8150.1566.31
Classic ST32.7246.1554.4168.64
BOND34.2248.7352.4568.89
MetaST55.0456.6160.3873.20
", + "bbox": [ + 202, + 347, + 482, + 507 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/2e756955727d5af6ceff79d49d5a8c00dbdb2ebe481f26438fc9a3cd5983d3fb.jpg", + "table_caption": [ + "Figure 5: MIT Movie Dataset. " + ], + "table_footnote": [], + "table_body": "
MethodShots (12 Slot Types)
51020100
Full-supervision87.87
BERT
Few-shot Supervision
BERT62.8069.5075.8182.49
Few-shot Supervision +unlabeled data
CVT57.4862.7370.2081.82
SeqVAT60.9467.1074.1582.73
Mean Teacher58.9267.6275.2482.20
VAT60.7570.1775.4182.39
Classic ST63.3971.8876.5883.06
BOND62.5070.9175.5282.65
MetaST72.5777.6780.3384.35
", + "bbox": [ + 517, + 347, + 795, + 507 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/f476962ad01d07f77ba58f1fdd395dfa4759197269c394736f1bb1c3e19b84bf.jpg", + "table_caption": [ + "Table 10: MIT Restaurant Dataset. " + ], + "table_footnote": [], + "table_body": "
MethodShots (8 Slot Types)
51020100
Full-supervision
BERT78.95
Few-shot Supervision
BERT41.3954.0660.1272.24
Few-shot Supervision +unlabeled data
CVT33.7442.5751.3370.84
SeqVAT41.9451.5556.1571.39
Mean Teacher40.3751.7557.3472.40
VAT41.2953.3459.6872.65
Classic ST44.3556.8060.2873.13
BOND43.0155.7859.9673.60
MetaST53.0263.8367.8675.25
", + "bbox": [ + 202, + 549, + 482, + 708 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/a5c5c774b1339fd4c329a18b099f6042e9d44fe22e679ea606230aedb95f867e.jpg", + "table_caption": [ + "Table 11: CoNLL2003 (EN) " + ], + "table_footnote": [], + "table_body": "
MethodShots (4 Slot Types)
51020100
Full-supervision92.40
BERT
Few-shot Supervision
BERT63.8771.1573.5784.36
Few-shot Supervision +unlabeled data
CVT51.1554.3166.1181.99
SeqVAT58.0267.2174.1582.20
Mean Teacher59.0468.6772.6284.17
VAT57.0365.0372.6984.43
Classic ST64.0470.9974.6584.93
BOND62.5269.5674.1983.87
MetaST71.4976.6578.5485.77
", + "bbox": [ + 517, + 549, + 794, + 708 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/25284224f76b660bd08133cc6918cc8de0525a602474f8bf588087c496594e31.jpg", + "table_caption": [ + "Table 12: Multilingual CoNLL03. " + ], + "table_footnote": [], + "table_body": "
MethodShots (4 Slot Types)
51020100
Full-supervision
BERT
Few-shot Supervision
BERT64.8070.7773.8980.61
Few-shot Supervision +unlabeled data
Mean Teacher64.5568.3473.8779.21
VAT64.9767.6374.2680.70
Classic ST67.9572.6973.7981.82
BOND69.4272.7976.0280.62
MetaST73.3476.6577.0182.11
", + "bbox": [ + 202, + 750, + 482, + 891 + ], + "page_idx": 14 + }, + { + "type": "table", + "img_path": "images/bdb9bde3a099c1cd2efc59fc41a63bf1059ccedbcf2cff1774df674b185df69c.jpg", + "table_caption": [ + "Table 13: Multilingual Wikiann " + ], + "table_footnote": [], + "table_body": "
MethodShots (3 Slot Types X 41 languages)
51020100
Full-supervision
BERT
Few-shot Supervision82.3385.70
BERT77.6879.67
Few-shot Supervision +unlabeled data
Mean Teacher77.0980.2382.1985.34
VAT74.7178.8282.6085.82
Classic ST76.7380.2482.3986.08
BOND78.8179.5782.1986.14
MetaST79.1081.6183.1485.57
", + "bbox": [ + 517, + 750, + 794, + 891 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "A.4 IMPLEMENTATIONS AND HYPER-PARAMETER ", + "text_level": 1, + "bbox": [ + 176, + 103, + 529, + 117 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We do not perform any hyper-parameter tuning for different datasets. The batch size and maximum sequence length varies due to data characteristics and are as shown in Tbale 14. The hyperparameters are as shown in Table 14. ", + "bbox": [ + 176, + 128, + 821, + 171 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Also, we retain parameters from original BERT implementation from https://github.com/ huggingface/transformers. ", + "bbox": [ + 173, + 178, + 820, + 207 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We implement SeqVAT based on https://github.com/jiesutd/NCRFpp and implement CVT following https://github.com/tensorflow/models/tree/master/ research/cvt_text. ", + "bbox": [ + 176, + 213, + 821, + 256 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/7d8a5ff09a19797def76c476a6247747396bbda92f80f3b10f8a54a82da45d85.jpg", + "table_caption": [ + "Table 14: Batch size, sequence length and BERT encoder choices across datasets " + ], + "table_footnote": [], + "table_body": "
DatasetSequence LengthBatch SizeLabeled data sample size B|Unlabeled Batch SizeBERT Encoder
SNIPS64163232BERT-base-uncased
Email64163232BERT-base-cased
Movie64163232BERT-base-uncased
Restaurant64161632BERT-base-uncased
CoNLL03 (EN)12816832BERT-base-cased
Wikiann (EN)12816832BERT-base-cased
CoNLLO3 (multilingual)128163232BERT-multilingual-base-cased
Wikiann (multilingaul)128163232BERT-multilingual-base-cased
", + "bbox": [ + 176, + 267, + 820, + 359 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/61aff8079c20ff5289ce3220cb1b1edc76f284a5836c724cdde2f73eb1a1b91e.jpg", + "table_caption": [ + "Table 15: Hyper-parameters. " + ], + "table_footnote": [], + "table_body": "
BERT attention dropout BERT hidden dropout Latest Iteration R in labeled data acquisition BERT output hidden size h0.3 0.3 5 768
Steps for fine-tuning teacher model on labeled data2000
Steps T for self-training model on unlabeled data3000
Mini-batch S Re-initialize Student5
Pseudo-label TypeY Hard
Warmup steps20
learning rate α5e-5
Weight_decay5e-6
", + "bbox": [ + 178, + 406, + 821, + 575 + ], + "page_idx": 15 + } +] \ No newline at end of file diff --git a/parse/train/ARFshOO1Iu/ARFshOO1Iu_middle.json b/parse/train/ARFshOO1Iu/ARFshOO1Iu_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..8c8139759f5e2d2465cf9d7fe5bb6022a90f2b00 --- /dev/null +++ b/parse/train/ARFshOO1Iu/ARFshOO1Iu_middle.json @@ -0,0 +1,40132 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 99 + ], + "score": 1.0, + "content": "ADAPTIVE SELF-TRAINING FOR NEURAL SEQUENCE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 332, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 332, + 118 + ], + "score": 1.0, + "content": "LABELING WITH FEW LABELS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 209, + 468, + 419 + ], + "lines": [ + { + "bbox": [ + 142, + 210, + 469, + 222 + ], + "spans": [ + { + "bbox": [ + 142, + 210, + 469, + 222 + ], + "score": 1.0, + "content": "Neural sequence labeling is an important technique employed for many Natural", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 221, + 470, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 221, + 470, + 234 + ], + "score": 1.0, + "content": "Language Processing (NLP) tasks, such as Named Entity Recognition (NER), slot", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 232, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 470, + 246 + ], + "score": 1.0, + "content": "tagging for dialog systems and semantic parsing. Large-scale pre-trained language", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 242, + 470, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 470, + 256 + ], + "score": 1.0, + "content": "models obtain very good performance on these tasks when fine-tuned on large", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 470, + 266 + ], + "score": 1.0, + "content": "amounts of task-specific labeled data. However, such large-scale labeled datasets", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 266, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 142, + 266, + 469, + 276 + ], + "score": 1.0, + "content": "are difficult to obtain for several tasks and domains due to the high cost of hu-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 277, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 142, + 277, + 469, + 288 + ], + "score": 1.0, + "content": "man annotation as well as privacy and data access constraints for sensitive user", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 287, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 287, + 469, + 299 + ], + "score": 1.0, + "content": "applications. This is exacerbated for sequence labeling tasks requiring such an-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 298, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 469, + 309 + ], + "score": 1.0, + "content": "notations at token-level. In this work, we develop techniques to address the label", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "score": 1.0, + "content": "scarcity challenge for neural sequence labeling models. Specifically, we develop", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "score": 1.0, + "content": "self-training and meta-learning techniques for training neural sequence taggers", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "score": 1.0, + "content": "with few labels. While self-training serves as an effective mechanism to learn", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 342, + 469, + 354 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 469, + 354 + ], + "score": 1.0, + "content": "from large amounts of unlabeled data – meta-learning helps in adaptive sample", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 353, + 470, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 470, + 365 + ], + "score": 1.0, + "content": "re-weighting to mitigate error propagation from noisy pseudo-labels. Extensive", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 364, + 470, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 470, + 375 + ], + "score": 1.0, + "content": "experiments on six benchmark datasets including two for massive multilingual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 374, + 470, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 374, + 470, + 387 + ], + "score": 1.0, + "content": "NER and four slot tagging datasets for task-oriented dialog systems demonstrate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 385, + 469, + 397 + ], + "spans": [ + { + "bbox": [ + 142, + 385, + 469, + 397 + ], + "score": 1.0, + "content": "the effectiveness of our method. With only 10 labeled examples for each class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 396, + 469, + 409 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 277, + 409 + ], + "score": 1.0, + "content": "for each task, our method obtains", + "type": "text" + }, + { + "bbox": [ + 278, + 396, + 297, + 407 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 396, + 469, + 409 + ], + "score": 1.0, + "content": "improvement over state-of-the-art systems", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 407, + 382, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 382, + 421 + ], + "score": 1.0, + "content": "demonstrating its effectiveness for the low-resource setting.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 108, + 438, + 206, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 208, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 208, + 453 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 504, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Motivation. Deep neural networks typically require large amounts of training data to achieve state-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 473, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 486 + ], + "score": 1.0, + "content": "of-the-art performance. Recent advances with pre-trained language models like BERT (Devlin et al.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "2019), GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019) have reduced this annotation", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "bottleneck. In this paradigm, large neural network models are trained on massive amounts of un-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "labeled data in a self-supervised manner. However, the success of these large-scale models still", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "relies on fine-tuning them on large amounts of labeled data for downstream tasks. For instance, our", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 182, + 541 + ], + "score": 1.0, + "content": "experiments show", + "type": "text" + }, + { + "bbox": [ + 182, + 528, + 202, + 539 + ], + "score": 0.87, + "content": "\\bar { 2 } 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "relative improvement on an average when fine-tuning BERT with the full", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "training set (2.5K-705K labels) vs. fine-tuning with only 10 labels per class. This poses several", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "score": 1.0, + "content": "challenges for many real-world tasks. Not only is acquiring large amounts of labeled data for ev-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "score": 1.0, + "content": "ery task expensive and time consuming, but also not feasible in many cases due to data access and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "privacy constraints. This issue is exacerbated for sequence labeling tasks that require annotations at", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "token- and slot-level as opposed to instance-level classification tasks. For example, an NER task can", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "have slots like B-PER, I-PER, O-PER marking the beginning, intermediate and out-of-span markers", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "for person names, and similar slots for the names of location and organization. Similarly, language", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "understanding models for dialog systems rely on effective identification of what the user intends to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "do (intents) and the corresponding values as arguments (slots) for use by downstream applications.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "Therefore, fully supervised neural sequence taggers are expensive to train for such tasks, given the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 476, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 476, + 662 + ], + "score": 1.0, + "content": "requirement of thousands of annotations for hundreds of slots for the many different intents.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Semi-supervised learning (SSL) (Chapelle et al., 2010) is one of the promising paradigms to address", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "labeled data scarcity by making effective use of large amounts of unlabeled data in addition to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "task-specific labeled data. Self-training (ST, (III, 1965)) as one of the earliest SSL approaches", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 710 + ], + "score": 1.0, + "content": "has recently shown state-of-the-art performance for tasks like image classification (Li et al., 2019;", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Xie et al., 2020) performing at par with supervised systems while using very few training labels.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "In contrast to such instance-level classification tasks, sequence labeling tasks have dependencies", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 505, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 505, + 99 + ], + "score": 1.0, + "content": "ADAPTIVE SELF-TRAINING FOR NEURAL SEQUENCE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 332, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 332, + 118 + ], + "score": 1.0, + "content": "LABELING WITH FEW LABELS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 187, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 209, + 468, + 419 + ], + "lines": [ + { + "bbox": [ + 142, + 210, + 469, + 222 + ], + "spans": [ + { + "bbox": [ + 142, + 210, + 469, + 222 + ], + "score": 1.0, + "content": "Neural sequence labeling is an important technique employed for many Natural", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 221, + 470, + 234 + ], + "spans": [ + { + "bbox": [ + 141, + 221, + 470, + 234 + ], + "score": 1.0, + "content": "Language Processing (NLP) tasks, such as Named Entity Recognition (NER), slot", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 232, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 232, + 470, + 246 + ], + "score": 1.0, + "content": "tagging for dialog systems and semantic parsing. Large-scale pre-trained language", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 242, + 470, + 256 + ], + "spans": [ + { + "bbox": [ + 141, + 242, + 470, + 256 + ], + "score": 1.0, + "content": "models obtain very good performance on these tasks when fine-tuned on large", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 470, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 470, + 266 + ], + "score": 1.0, + "content": "amounts of task-specific labeled data. However, such large-scale labeled datasets", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 266, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 142, + 266, + 469, + 276 + ], + "score": 1.0, + "content": "are difficult to obtain for several tasks and domains due to the high cost of hu-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 277, + 469, + 288 + ], + "spans": [ + { + "bbox": [ + 142, + 277, + 469, + 288 + ], + "score": 1.0, + "content": "man annotation as well as privacy and data access constraints for sensitive user", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 287, + 469, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 287, + 469, + 299 + ], + "score": 1.0, + "content": "applications. This is exacerbated for sequence labeling tasks requiring such an-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 298, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 298, + 469, + 309 + ], + "score": 1.0, + "content": "notations at token-level. In this work, we develop techniques to address the label", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "spans": [ + { + "bbox": [ + 141, + 309, + 469, + 322 + ], + "score": 1.0, + "content": "scarcity challenge for neural sequence labeling models. Specifically, we develop", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "spans": [ + { + "bbox": [ + 141, + 320, + 469, + 333 + ], + "score": 1.0, + "content": "self-training and meta-learning techniques for training neural sequence taggers", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "spans": [ + { + "bbox": [ + 141, + 330, + 469, + 343 + ], + "score": 1.0, + "content": "with few labels. While self-training serves as an effective mechanism to learn", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 342, + 469, + 354 + ], + "spans": [ + { + "bbox": [ + 141, + 342, + 469, + 354 + ], + "score": 1.0, + "content": "from large amounts of unlabeled data – meta-learning helps in adaptive sample", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 353, + 470, + 365 + ], + "spans": [ + { + "bbox": [ + 141, + 353, + 470, + 365 + ], + "score": 1.0, + "content": "re-weighting to mitigate error propagation from noisy pseudo-labels. Extensive", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 364, + 470, + 375 + ], + "spans": [ + { + "bbox": [ + 141, + 364, + 470, + 375 + ], + "score": 1.0, + "content": "experiments on six benchmark datasets including two for massive multilingual", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 374, + 470, + 387 + ], + "spans": [ + { + "bbox": [ + 141, + 374, + 470, + 387 + ], + "score": 1.0, + "content": "NER and four slot tagging datasets for task-oriented dialog systems demonstrate", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 142, + 385, + 469, + 397 + ], + "spans": [ + { + "bbox": [ + 142, + 385, + 469, + 397 + ], + "score": 1.0, + "content": "the effectiveness of our method. With only 10 labeled examples for each class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 396, + 469, + 409 + ], + "spans": [ + { + "bbox": [ + 141, + 396, + 277, + 409 + ], + "score": 1.0, + "content": "for each task, our method obtains", + "type": "text" + }, + { + "bbox": [ + 278, + 396, + 297, + 407 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 396, + 469, + 409 + ], + "score": 1.0, + "content": "improvement over state-of-the-art systems", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 407, + 382, + 421 + ], + "spans": [ + { + "bbox": [ + 141, + 407, + 382, + 421 + ], + "score": 1.0, + "content": "demonstrating its effectiveness for the low-resource setting.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 14, + "bbox_fs": [ + 141, + 210, + 470, + 421 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 438, + 206, + 450 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 208, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 208, + 453 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 504, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 475 + ], + "score": 1.0, + "content": "Motivation. Deep neural networks typically require large amounts of training data to achieve state-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 473, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 506, + 486 + ], + "score": 1.0, + "content": "of-the-art performance. Recent advances with pre-trained language models like BERT (Devlin et al.,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "2019), GPT-2 (Radford et al., 2019) and RoBERTa (Liu et al., 2019) have reduced this annotation", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 508 + ], + "score": 1.0, + "content": "bottleneck. In this paradigm, large neural network models are trained on massive amounts of un-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 519 + ], + "score": 1.0, + "content": "labeled data in a self-supervised manner. However, the success of these large-scale models still", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 517, + 506, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 506, + 530 + ], + "score": 1.0, + "content": "relies on fine-tuning them on large amounts of labeled data for downstream tasks. For instance, our", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 182, + 541 + ], + "score": 1.0, + "content": "experiments show", + "type": "text" + }, + { + "bbox": [ + 182, + 528, + 202, + 539 + ], + "score": 0.87, + "content": "\\bar { 2 } 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "relative improvement on an average when fine-tuning BERT with the full", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 505, + 552 + ], + "score": 1.0, + "content": "training set (2.5K-705K labels) vs. fine-tuning with only 10 labels per class. This poses several", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 563 + ], + "score": 1.0, + "content": "challenges for many real-world tasks. Not only is acquiring large amounts of labeled data for ev-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 506, + 573 + ], + "score": 1.0, + "content": "ery task expensive and time consuming, but also not feasible in many cases due to data access and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "privacy constraints. This issue is exacerbated for sequence labeling tasks that require annotations at", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "token- and slot-level as opposed to instance-level classification tasks. For example, an NER task can", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "have slots like B-PER, I-PER, O-PER marking the beginning, intermediate and out-of-span markers", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 506, + 619 + ], + "score": 1.0, + "content": "for person names, and similar slots for the names of location and organization. Similarly, language", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "understanding models for dialog systems rely on effective identification of what the user intends to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "do (intents) and the corresponding values as arguments (slots) for use by downstream applications.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "Therefore, fully supervised neural sequence taggers are expensive to train for such tasks, given the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 649, + 476, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 476, + 662 + ], + "score": 1.0, + "content": "requirement of thousands of annotations for hundreds of slots for the many different intents.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 462, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Semi-supervised learning (SSL) (Chapelle et al., 2010) is one of the promising paradigms to address", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 689 + ], + "score": 1.0, + "content": "labeled data scarcity by making effective use of large amounts of unlabeled data in addition to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "task-specific labeled data. Self-training (ST, (III, 1965)) as one of the earliest SSL approaches", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 699, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 710 + ], + "score": 1.0, + "content": "has recently shown state-of-the-art performance for tasks like image classification (Li et al., 2019;", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Xie et al., 2020) performing at par with supervised systems while using very few training labels.", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 720, + 506, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 506, + 732 + ], + "score": 1.0, + "content": "In contrast to such instance-level classification tasks, sequence labeling tasks have dependencies", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 216, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 226 + ], + "score": 1.0, + "content": "between the slots demanding different design choices for slot-level loss optimization for the limited", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "labeled data setting. For instance, prior work (Ruder & Plank, 2018) using classic self-training", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 453, + 249 + ], + "score": 1.0, + "content": "techniques for sequence labeling did not find much success in the low-data regime with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 453, + 237, + 473, + 248 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 473, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "labeled", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 249, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 260 + ], + "score": 1.0, + "content": "data for the target domain. Although there has been some success with careful task-specific data", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 259, + 504, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 504, + 270 + ], + "score": 1.0, + "content": "selection (Petrov & McDonald, 2012) and more recently for distant supervision (Liang et al., 2020)", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 271, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 504, + 282 + ], + "score": 1.0, + "content": "using external resources like knowledge bases (e.g., Wikipedia). In contrast to these prior work, we", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "develop techniques for self-training with limited labels and without any task-specific assumption or", + "type": "text", + "cross_page": true + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 189, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 189, + 304 + ], + "score": 1.0, + "content": "external knowledge.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 665, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 69, + 415, + 185 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 188, + 69, + 415, + 185 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 69, + 415, + 185 + ], + "spans": [ + { + "bbox": [ + 188, + 69, + 415, + 185 + ], + "score": 0.97, + "type": "image", + "image_path": "4716dc9424f650ad710d7db8d4aff2f8dd06b88c864da63dd6e410117264cc70.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 69, + 415, + 83.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 188, + 83.5, + 415, + 98.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 188, + 98.0, + 415, + 112.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 188, + 112.5, + 415, + 127.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 188, + 127.0, + 415, + 141.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 188, + 141.5, + 415, + 156.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 156.0, + 415, + 170.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 170.5, + 415, + 185.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 245, + 190, + 366, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 244, + 188, + 366, + 202 + ], + "spans": [ + { + "bbox": [ + 244, + 188, + 366, + 202 + ], + "score": 1.0, + "content": "Figure 1: MetaST framework.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 226 + ], + "score": 1.0, + "content": "between the slots demanding different design choices for slot-level loss optimization for the limited", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 239 + ], + "score": 1.0, + "content": "labeled data setting. For instance, prior work (Ruder & Plank, 2018) using classic self-training", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 453, + 249 + ], + "score": 1.0, + "content": "techniques for sequence labeling did not find much success in the low-data regime with", + "type": "text" + }, + { + "bbox": [ + 453, + 237, + 473, + 248 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "labeled", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 249, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 505, + 260 + ], + "score": 1.0, + "content": "data for the target domain. Although there has been some success with careful task-specific data", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 259, + 504, + 270 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 504, + 270 + ], + "score": 1.0, + "content": "selection (Petrov & McDonald, 2012) and more recently for distant supervision (Liang et al., 2020)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 271, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 504, + 282 + ], + "score": 1.0, + "content": "using external resources like knowledge bases (e.g., Wikipedia). In contrast to these prior work, we", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 280, + 505, + 293 + ], + "score": 1.0, + "content": "develop techniques for self-training with limited labels and without any task-specific assumption or", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 291, + 189, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 189, + 304 + ], + "score": 1.0, + "content": "external knowledge.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 309, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 320 + ], + "score": 1.0, + "content": "For self-training, a base model (teacher) is trained on some amount of labeled data and used to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "pseudo-annotate (task-specific) unlabeled data. The original labeled data is augmented with the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "pseudo-labeled data and used to train a student model. The student-teacher training is repeated until", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "convergence. Traditionally in self-training frameworks, the teacher model pseudo-annotates unla-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "beled data without any sample selection. This may result in gradual drifts from self-training on noisy", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "pseudo-labeled instances (Zhang et al., 2017). In order to deal with noisy labels and training set bi-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "ases, Ren et al. (2018) propose a meta-learning technique to automatically re-weight noisy samples", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "by their loss changes on a held-out clean labeled validation set. We adopt a similar principle in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "our work and leverage meta-learning to re-weight noisy pseudo-labeled examples from the teacher.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "While prior techniques for learning to re-weight examples have been developed for instance-level", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "score": 1.0, + "content": "classification tasks, we extend them to operate at token-level for discrete sequence labeling tasks.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "To this end, we address some key challenges on how to construct an informative held-out validation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "set for token-level re-weighting. Prior works (Ren et al., 2018; Shu et al., 2019) for instance classi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "score": 1.0, + "content": "fication construct this validation set by random sampling. However, sequence labeling tasks involve", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "many slots (e.g. WikiAnn has 123 slots over 41 languages) with variable difficulty and distribution", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "score": 1.0, + "content": "in the data. In case of random sampling, the model oversamples from the most populous category", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "score": 1.0, + "content": "and slots. This is particularly detrimental for low-resource languages in the multilingual setting.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 493, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 509 + ], + "score": 1.0, + "content": "To this end, we develop an adaptive mechanism to create the validation set on the fly considering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "the diversity and uncertainty of the model for different slot types. Furthermore, we leverage this", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "score": 1.0, + "content": "validation set for token-level loss estimation and re-weighting pseudo-labeled sequences from the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "teacher in the meta-learning setup. While prior works (Li et al., 2019; Sun et al., 2019; Bansal et al.,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "2020) on meta-learning for image and text classification leverage multi-task learning to improve a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "target classification task based on several similar tasks, in this work we focus on a single sequence", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 562, + 355, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 355, + 573 + ], + "score": 1.0, + "content": "labeling task – making our setup more challenging altogether.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "Our task and framework overview. We focus on sequence labeling tasks with only a few annotated", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 165, + 602 + ], + "score": 1.0, + "content": "samples (e.g.,", + "type": "text" + }, + { + "bbox": [ + 165, + 589, + 257, + 601 + ], + "score": 0.91, + "content": "K = \\{ 5 , 1 0 , 2 0 , 1 0 0 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "per slot type for training and large amounts of task-specific", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "unlabeled data. Figure 1 shows an overview of our framework with the following components: (i)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "Self-training: Our self-training framework leverages a pre-trained language model as a teacher and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "score": 1.0, + "content": "co-trains a student model with iterative knowledge exchange (ii) Adaptive labeled data acquisition", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "for validation: Our few-shot learning setup assumes a small number of labeled training samples", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "per slot type. The labeled data from multiple slot types are not equally informative for the student", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "model to learn from. While prior works in meta-learning randomly sample some labeled examples", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "for held-out validation set, we develop an adaptive mechanism to create this set on the fly. To this", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "end, we leverage loss decay as a proxy for model uncertainty to select informative labeled samples", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "score": 1.0, + "content": "for the student model to learn from in conjunction with the re-weighting mechanism in the next step.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "(iii) Meta-learning for sample re-weighting: Since pseudo-labeled samples from the teacher can be", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 104, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 104, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "noisy, we employ meta-learning to re-weight them to improve the student model performance on the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "held-out validation set obtained from the previous step. In contrast to prior work (Ren et al., 2018)", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 47.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 188, + 69, + 415, + 185 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 188, + 69, + 415, + 185 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 188, + 69, + 415, + 185 + ], + "spans": [ + { + "bbox": [ + 188, + 69, + 415, + 185 + ], + "score": 0.97, + "type": "image", + "image_path": "4716dc9424f650ad710d7db8d4aff2f8dd06b88c864da63dd6e410117264cc70.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 188, + 69, + 415, + 83.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 188, + 83.5, + 415, + 98.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 188, + 98.0, + 415, + 112.5 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 188, + 112.5, + 415, + 127.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 188, + 127.0, + 415, + 141.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 188, + 141.5, + 415, + 156.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 188, + 156.0, + 415, + 170.5 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 188, + 170.5, + 415, + 185.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 245, + 190, + 366, + 201 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 244, + 188, + 366, + 202 + ], + "spans": [ + { + "bbox": [ + 244, + 188, + 366, + 202 + ], + "score": 1.0, + "content": "Figure 1: MetaST framework.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + }, + { + "type": "text", + "bbox": [ + 107, + 215, + 505, + 303 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 105, + 216, + 505, + 304 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 309, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 309, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 320 + ], + "score": 1.0, + "content": "For self-training, a base model (teacher) is trained on some amount of labeled data and used to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 332 + ], + "score": 1.0, + "content": "pseudo-annotate (task-specific) unlabeled data. The original labeled data is augmented with the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "pseudo-labeled data and used to train a student model. The student-teacher training is repeated until", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "convergence. Traditionally in self-training frameworks, the teacher model pseudo-annotates unla-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "beled data without any sample selection. This may result in gradual drifts from self-training on noisy", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 376 + ], + "score": 1.0, + "content": "pseudo-labeled instances (Zhang et al., 2017). In order to deal with noisy labels and training set bi-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "ases, Ren et al. (2018) propose a meta-learning technique to automatically re-weight noisy samples", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "by their loss changes on a held-out clean labeled validation set. We adopt a similar principle in", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "our work and leverage meta-learning to re-weight noisy pseudo-labeled examples from the teacher.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "While prior techniques for learning to re-weight examples have been developed for instance-level", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "score": 1.0, + "content": "classification tasks, we extend them to operate at token-level for discrete sequence labeling tasks.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "To this end, we address some key challenges on how to construct an informative held-out validation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "set for token-level re-weighting. Prior works (Ren et al., 2018; Shu et al., 2019) for instance classi-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 464 + ], + "score": 1.0, + "content": "fication construct this validation set by random sampling. However, sequence labeling tasks involve", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "many slots (e.g. WikiAnn has 123 slots over 41 languages) with variable difficulty and distribution", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 487 + ], + "score": 1.0, + "content": "in the data. In case of random sampling, the model oversamples from the most populous category", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 505, + 498 + ], + "score": 1.0, + "content": "and slots. This is particularly detrimental for low-resource languages in the multilingual setting.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 493, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 509 + ], + "score": 1.0, + "content": "To this end, we develop an adaptive mechanism to create the validation set on the fly considering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "the diversity and uncertainty of the model for different slot types. Furthermore, we leverage this", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "score": 1.0, + "content": "validation set for token-level loss estimation and re-weighting pseudo-labeled sequences from the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "teacher in the meta-learning setup. While prior works (Li et al., 2019; Sun et al., 2019; Bansal et al.,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "2020) on meta-learning for image and text classification leverage multi-task learning to improve a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "target classification task based on several similar tasks, in this work we focus on a single sequence", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 562, + 355, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 355, + 573 + ], + "score": 1.0, + "content": "labeling task – making our setup more challenging altogether.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 309, + 506, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "Our task and framework overview. We focus on sequence labeling tasks with only a few annotated", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 588, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 165, + 602 + ], + "score": 1.0, + "content": "samples (e.g.,", + "type": "text" + }, + { + "bbox": [ + 165, + 589, + 257, + 601 + ], + "score": 0.91, + "content": "K = \\{ 5 , 1 0 , 2 0 , 1 0 0 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 588, + 506, + 602 + ], + "score": 1.0, + "content": "per slot type for training and large amounts of task-specific", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "unlabeled data. Figure 1 shows an overview of our framework with the following components: (i)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 506, + 623 + ], + "score": 1.0, + "content": "Self-training: Our self-training framework leverages a pre-trained language model as a teacher and", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 623, + 505, + 634 + ], + "score": 1.0, + "content": "co-trains a student model with iterative knowledge exchange (ii) Adaptive labeled data acquisition", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "for validation: Our few-shot learning setup assumes a small number of labeled training samples", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 643, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 643, + 506, + 656 + ], + "score": 1.0, + "content": "per slot type. The labeled data from multiple slot types are not equally informative for the student", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "model to learn from. While prior works in meta-learning randomly sample some labeled examples", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "for held-out validation set, we develop an adaptive mechanism to create this set on the fly. To this", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "end, we leverage loss decay as a proxy for model uncertainty to select informative labeled samples", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 702 + ], + "score": 1.0, + "content": "for the student model to learn from in conjunction with the re-weighting mechanism in the next step.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "(iii) Meta-learning for sample re-weighting: Since pseudo-labeled samples from the teacher can be", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 104, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 104, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "noisy, we employ meta-learning to re-weight them to improve the student model performance on the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "held-out validation set obtained from the previous step. In contrast to prior work (Ren et al., 2018)", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 47.5, + "bbox_fs": [ + 104, + 577, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "on sample re-weighting operating at instance-level, we incorporate the re-weighting mechanism at", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "token-level for sequence labeling tasks. Here the token-level weights are determined by the student", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "model loss on the above validation set. Finally, we learn all of the above steps jointly with end-to-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "end learning in the self-training framework. We refer to our adaptive self-training framework with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 372, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 372, + 139 + ], + "score": 1.0, + "content": "meta-learning based sample re-weighting mechanism as MetaST.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 504, + 155 + ], + "score": 1.0, + "content": "We perform extensive experiments on six benchmark datasets for several tasks including multi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "lingual Named Entity Recognition and slot tagging for user utterances from task-oriented dialog", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "systems to demonstrate the generalizability of our approach across diverse tasks and languages. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "adopt BERT and multilingual BERT as encoder and show that its performance can be significantly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 187, + 201 + ], + "score": 1.0, + "content": "improved by nearly", + "type": "text" + }, + { + "bbox": [ + 188, + 187, + 208, + 198 + ], + "score": 0.89, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "for low-resource settings with few training labels (e.g., 10 labeled exam-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "score": 1.0, + "content": "ples per slot type) and large amounts of unlabeled data. In summary, our work makes the following", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "contributions. (i) Develops a self-training framework for neural sequence tagging with few labeled", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "training examples. (ii) Leverages an acquisition strategy to adaptively select a validation set from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "the labeled set for meta-learning of the student model. (iii) Develops a meta-learning framework for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 241, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 505, + 256 + ], + "score": 1.0, + "content": "re-weighting pseudo-labeled samples at token-level to reduce drifts from noisy teacher predictions.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "(iv) Integrates the aforementioned components into an end-to-end learning framework and demon-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "strates its effectiveness for neural sequence labeling across six benchmark datasets with multiple", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 253, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 253, + 288 + ], + "score": 1.0, + "content": "slots, shots, domains and languages.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 295, + 200, + 308 + ], + "lines": [ + { + "bbox": [ + 104, + 294, + 201, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 201, + 311 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "Sequence labeling and slot tagging. This is the task identifying the entity span of several slot", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "types (e.g., names of person, organization, location, date, etc.) in a text sequence. Formally,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 200, + 352 + ], + "score": 1.0, + "content": "given a sentence with", + "type": "text" + }, + { + "bbox": [ + 200, + 340, + 211, + 349 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 339, + 243, + 352 + ], + "score": 1.0, + "content": "tokens", + "type": "text" + }, + { + "bbox": [ + 243, + 339, + 322, + 351 + ], + "score": 0.93, + "content": "\\mathbf { \\bar { \\cal X } } ~ = ~ \\{ x _ { 1 } , . . . , x _ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 339, + 506, + 352 + ], + "score": 1.0, + "content": ", an entity or slot value is a span of tokens", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 349, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 244, + 363 + ], + "score": 0.83, + "content": "s = [ x _ { i } , . . . , x _ { j } ] ( 0 \\leq i \\leq j \\leq N )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 349, + 506, + 365 + ], + "score": 1.0, + "content": "associated with a type. This task assumes a pre-defined tagging", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "policy like BIO (Tjong et al., 1999), where B marks the beginning of the slot, I marks an interme-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 222, + 385 + ], + "score": 1.0, + "content": "diate token in the span, and", + "type": "text" + }, + { + "bbox": [ + 223, + 374, + 231, + 383 + ], + "score": 0.42, + "content": "\\bigcirc", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "marks out-of-span tokens. These span markers are used to extract", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 383, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 482, + 396 + ], + "score": 1.0, + "content": "multi-token values for each of the slot types with phrase-level evaluation for the performance.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 207, + 414 + ], + "score": 1.0, + "content": "Self-training. Consider", + "type": "text" + }, + { + "bbox": [ + 207, + 400, + 245, + 412 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\theta _ { t e a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 398, + 264, + 414 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 264, + 400, + 303, + 412 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\theta _ { s t u } )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 398, + 506, + 414 + ], + "score": 1.0, + "content": "to denote the teacher and student models respec-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 410, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 424 + ], + "score": 1.0, + "content": "tively in the self-training framework. The role of the teacher model (e.g., a pre-trained language", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 423, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 433 + ], + "score": 1.0, + "content": "model) is to assign pseudo-labels to unlabeled data that is used to train a student model. The", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "teacher and student model can exchange knowledge and the training schedules are repeated till", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "convergence. The success of self-training with deep neural networks in recent works (He et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "2019; Xie et al., 2020) has been attributed to a number of factors including stochastic regularization", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 420, + 478 + ], + "score": 1.0, + "content": "with dropouts and data regularization with unlabeled data. Formally, given", + "type": "text" + }, + { + "bbox": [ + 420, + 467, + 430, + 476 + ], + "score": 0.77, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "-th unlabeled sen-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 103, + 473, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 103, + 473, + 150, + 494 + ], + "score": 1.0, + "content": "tence with", + "type": "text" + }, + { + "bbox": [ + 151, + 477, + 161, + 487 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 473, + 191, + 494 + ], + "score": 1.0, + "content": "tokens", + "type": "text" + }, + { + "bbox": [ + 191, + 477, + 291, + 490 + ], + "score": 0.92, + "content": "X _ { m } ^ { u } \\doteq \\{ x _ { 1 , m } ^ { u } , . . . , x _ { N , m } ^ { u } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 473, + 309, + 494 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 309, + 477, + 318, + 487 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 473, + 507, + 494 + ], + "score": 1.0, + "content": "pre-defined labels, consider the pseudo-labels", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 488, + 438, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 202, + 506 + ], + "score": 0.92, + "content": "\\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 488, + 354, + 509 + ], + "score": 1.0, + "content": "generated by the teacher model at the", + "type": "text" + }, + { + "bbox": [ + 355, + 493, + 360, + 502 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 488, + 438, + 509 + ], + "score": 1.0, + "content": "-th iteration where,", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 511, + 375, + 534 + ], + "lines": [ + { + "bbox": [ + 235, + 511, + 375, + 534 + ], + "spans": [ + { + "bbox": [ + 235, + 511, + 375, + 534 + ], + "score": 0.92, + "content": "\\hat { y } _ { m , n } ^ { ( t ) } = \\underset { c \\in C } { \\arg \\operatorname* { m a x } } f _ { n , c } ( x _ { m , n } ^ { u } ; \\theta _ { t e a } ^ { ( t ) } ) .", + "type": "interline_equation", + "image_path": "0e5a23d228b8b84c3f9f8d50a1a2a8b9f8936c649b392433f95ac84053ec7c07.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 235, + 511, + 375, + 534 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 102, + 535, + 508, + 562 + ], + "spans": [ + { + "bbox": [ + 102, + 535, + 269, + 562 + ], + "score": 1.0, + "content": "The pseudo-labeled data set, denoted as", + "type": "text" + }, + { + "bbox": [ + 269, + 541, + 393, + 556 + ], + "score": 0.92, + "content": "( X ^ { u } , \\hat { Y } ^ { ( t ) } ) = \\{ ( X _ { m } ^ { u } , \\hat { Y } _ { m } ^ { ( t ) } ) \\} _ { m } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 535, + 508, + 562 + ], + "score": 1.0, + "content": ", is used to train the student", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 554, + 245, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 245, + 568 + ], + "score": 1.0, + "content": "model and learn its parameters as:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 572, + 402, + 607 + ], + "lines": [ + { + "bbox": [ + 208, + 572, + 402, + 607 + ], + "spans": [ + { + "bbox": [ + 208, + 572, + 402, + 607 + ], + "score": 0.95, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\underset { \\theta } { \\arg \\operatorname* { m i n } } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) ,", + "type": "interline_equation", + "image_path": "529a8dc877aeeae9554ba104386433ff96b02e57164c11e0a3697f446980baa2.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 572, + 402, + 589.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 208, + 589.5, + 402, + 607.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 324, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 324, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 133, + 626 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 612, + 155, + 624 + ], + "score": 0.91, + "content": "l ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 611, + 324, + 626 + ], + "score": 1.0, + "content": "can be modeled as the cross-entropy loss.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "title", + "bbox": [ + 108, + 636, + 265, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 266, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 266, + 651 + ], + "score": 1.0, + "content": "3 ADAPTIVE SELF TRAINING", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Given a pre-trained language model (e.g., BERT (Devlin et al., 2019)) as the teacher, we first fine-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "tune it on the small labeled data to make it aware of the underlying task. The fine-tuned teacher", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "model is now used to pseudo-label the large unlabeled data. We consider the student model as", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "another instantiation of the pre-trained language model that is trained over the pseudo-labeled data.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "However, our few-shot setting with limited labeled data results in a noisy teacher. A naive transfer of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "teacher knowledge to the student results in the propagation of noisy labels limiting the performance", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "of the student model. To address this challenge, we develop an adaptive self-training framework to", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "on sample re-weighting operating at instance-level, we incorporate the re-weighting mechanism at", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "token-level for sequence labeling tasks. Here the token-level weights are determined by the student", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "model loss on the above validation set. Finally, we learn all of the above steps jointly with end-to-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "end learning in the self-training framework. We refer to our adaptive self-training framework with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 372, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 372, + 139 + ], + "score": 1.0, + "content": "meta-learning based sample re-weighting mechanism as MetaST.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 83, + 505, + 139 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 143, + 505, + 286 + ], + "lines": [ + { + "bbox": [ + 106, + 144, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 504, + 155 + ], + "score": 1.0, + "content": "We perform extensive experiments on six benchmark datasets for several tasks including multi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 166 + ], + "score": 1.0, + "content": "lingual Named Entity Recognition and slot tagging for user utterances from task-oriented dialog", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 165, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 506, + 178 + ], + "score": 1.0, + "content": "systems to demonstrate the generalizability of our approach across diverse tasks and languages. We", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "adopt BERT and multilingual BERT as encoder and show that its performance can be significantly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 505, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 187, + 201 + ], + "score": 1.0, + "content": "improved by nearly", + "type": "text" + }, + { + "bbox": [ + 188, + 187, + 208, + 198 + ], + "score": 0.89, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 187, + 505, + 201 + ], + "score": 1.0, + "content": "for low-resource settings with few training labels (e.g., 10 labeled exam-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "spans": [ + { + "bbox": [ + 104, + 196, + 505, + 212 + ], + "score": 1.0, + "content": "ples per slot type) and large amounts of unlabeled data. In summary, our work makes the following", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "contributions. (i) Develops a self-training framework for neural sequence tagging with few labeled", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 505, + 232 + ], + "score": 1.0, + "content": "training examples. (ii) Leverages an acquisition strategy to adaptively select a validation set from", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 505, + 244 + ], + "score": 1.0, + "content": "the labeled set for meta-learning of the student model. (iii) Develops a meta-learning framework for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 241, + 505, + 256 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 505, + 256 + ], + "score": 1.0, + "content": "re-weighting pseudo-labeled samples at token-level to reduce drifts from noisy teacher predictions.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 265 + ], + "score": 1.0, + "content": "(iv) Integrates the aforementioned components into an end-to-end learning framework and demon-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 505, + 277 + ], + "score": 1.0, + "content": "strates its effectiveness for neural sequence labeling across six benchmark datasets with multiple", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 274, + 253, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 253, + 288 + ], + "score": 1.0, + "content": "slots, shots, domains and languages.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 144, + 506, + 288 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 295, + 200, + 308 + ], + "lines": [ + { + "bbox": [ + 104, + 294, + 201, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 294, + 201, + 311 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 316, + 505, + 395 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "Sequence labeling and slot tagging. This is the task identifying the entity span of several slot", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 505, + 341 + ], + "score": 1.0, + "content": "types (e.g., names of person, organization, location, date, etc.) in a text sequence. Formally,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 200, + 352 + ], + "score": 1.0, + "content": "given a sentence with", + "type": "text" + }, + { + "bbox": [ + 200, + 340, + 211, + 349 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 339, + 243, + 352 + ], + "score": 1.0, + "content": "tokens", + "type": "text" + }, + { + "bbox": [ + 243, + 339, + 322, + 351 + ], + "score": 0.93, + "content": "\\mathbf { \\bar { \\cal X } } ~ = ~ \\{ x _ { 1 } , . . . , x _ { N } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 339, + 506, + 352 + ], + "score": 1.0, + "content": ", an entity or slot value is a span of tokens", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 349, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 244, + 363 + ], + "score": 0.83, + "content": "s = [ x _ { i } , . . . , x _ { j } ] ( 0 \\leq i \\leq j \\leq N )", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 349, + 506, + 365 + ], + "score": 1.0, + "content": "associated with a type. This task assumes a pre-defined tagging", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 505, + 374 + ], + "score": 1.0, + "content": "policy like BIO (Tjong et al., 1999), where B marks the beginning of the slot, I marks an interme-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 372, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 222, + 385 + ], + "score": 1.0, + "content": "diate token in the span, and", + "type": "text" + }, + { + "bbox": [ + 223, + 374, + 231, + 383 + ], + "score": 0.42, + "content": "\\bigcirc", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 372, + 505, + 385 + ], + "score": 1.0, + "content": "marks out-of-span tokens. These span markers are used to extract", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 383, + 482, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 482, + 396 + ], + "score": 1.0, + "content": "multi-token values for each of the slot types with phrase-level evaluation for the performance.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 317, + 506, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 505, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 398, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 207, + 414 + ], + "score": 1.0, + "content": "Self-training. Consider", + "type": "text" + }, + { + "bbox": [ + 207, + 400, + 245, + 412 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\theta _ { t e a } )", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 398, + 264, + 414 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 264, + 400, + 303, + 412 + ], + "score": 0.92, + "content": "f ( \\cdot ; \\theta _ { s t u } )", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 398, + 506, + 414 + ], + "score": 1.0, + "content": "to denote the teacher and student models respec-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 410, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 424 + ], + "score": 1.0, + "content": "tively in the self-training framework. The role of the teacher model (e.g., a pre-trained language", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 423, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 433 + ], + "score": 1.0, + "content": "model) is to assign pseudo-labels to unlabeled data that is used to train a student model. The", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 433, + 505, + 444 + ], + "score": 1.0, + "content": "teacher and student model can exchange knowledge and the training schedules are repeated till", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 505, + 456 + ], + "score": 1.0, + "content": "convergence. The success of self-training with deep neural networks in recent works (He et al.,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "2019; Xie et al., 2020) has been attributed to a number of factors including stochastic regularization", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 420, + 478 + ], + "score": 1.0, + "content": "with dropouts and data regularization with unlabeled data. Formally, given", + "type": "text" + }, + { + "bbox": [ + 420, + 467, + 430, + 476 + ], + "score": 0.77, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "-th unlabeled sen-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 103, + 473, + 507, + 494 + ], + "spans": [ + { + "bbox": [ + 103, + 473, + 150, + 494 + ], + "score": 1.0, + "content": "tence with", + "type": "text" + }, + { + "bbox": [ + 151, + 477, + 161, + 487 + ], + "score": 0.82, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 473, + 191, + 494 + ], + "score": 1.0, + "content": "tokens", + "type": "text" + }, + { + "bbox": [ + 191, + 477, + 291, + 490 + ], + "score": 0.92, + "content": "X _ { m } ^ { u } \\doteq \\{ x _ { 1 , m } ^ { u } , . . . , x _ { N , m } ^ { u } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 473, + 309, + 494 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 309, + 477, + 318, + 487 + ], + "score": 0.81, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 473, + 507, + 494 + ], + "score": 1.0, + "content": "pre-defined labels, consider the pseudo-labels", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 488, + 438, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 202, + 506 + ], + "score": 0.92, + "content": "\\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 488, + 354, + 509 + ], + "score": 1.0, + "content": "generated by the teacher model at the", + "type": "text" + }, + { + "bbox": [ + 355, + 493, + 360, + 502 + ], + "score": 0.77, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 488, + 438, + 509 + ], + "score": 1.0, + "content": "-th iteration where,", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30, + "bbox_fs": [ + 103, + 398, + 507, + 509 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 511, + 375, + 534 + ], + "lines": [ + { + "bbox": [ + 235, + 511, + 375, + 534 + ], + "spans": [ + { + "bbox": [ + 235, + 511, + 375, + 534 + ], + "score": 0.92, + "content": "\\hat { y } _ { m , n } ^ { ( t ) } = \\underset { c \\in C } { \\arg \\operatorname* { m a x } } f _ { n , c } ( x _ { m , n } ^ { u } ; \\theta _ { t e a } ^ { ( t ) } ) .", + "type": "interline_equation", + "image_path": "0e5a23d228b8b84c3f9f8d50a1a2a8b9f8936c649b392433f95ac84053ec7c07.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 235, + 511, + 375, + 534 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 542, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 102, + 535, + 508, + 562 + ], + "spans": [ + { + "bbox": [ + 102, + 535, + 269, + 562 + ], + "score": 1.0, + "content": "The pseudo-labeled data set, denoted as", + "type": "text" + }, + { + "bbox": [ + 269, + 541, + 393, + 556 + ], + "score": 0.92, + "content": "( X ^ { u } , \\hat { Y } ^ { ( t ) } ) = \\{ ( X _ { m } ^ { u } , \\hat { Y } _ { m } ^ { ( t ) } ) \\} _ { m } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 535, + 508, + 562 + ], + "score": 1.0, + "content": ", is used to train the student", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 554, + 245, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 245, + 568 + ], + "score": 1.0, + "content": "model and learn its parameters as:", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 102, + 535, + 508, + 568 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 572, + 402, + 607 + ], + "lines": [ + { + "bbox": [ + 208, + 572, + 402, + 607 + ], + "spans": [ + { + "bbox": [ + 208, + 572, + 402, + 607 + ], + "score": 0.95, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\underset { \\theta } { \\arg \\operatorname* { m i n } } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) ,", + "type": "interline_equation", + "image_path": "529a8dc877aeeae9554ba104386433ff96b02e57164c11e0a3697f446980baa2.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 572, + 402, + 589.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 208, + 589.5, + 402, + 607.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 324, + 624 + ], + "lines": [ + { + "bbox": [ + 105, + 611, + 324, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 611, + 133, + 626 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 612, + 155, + 624 + ], + "score": 0.91, + "content": "l ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 611, + 324, + 626 + ], + "score": 1.0, + "content": "can be modeled as the cross-entropy loss.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 611, + 324, + 626 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 636, + 265, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 266, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 266, + 651 + ], + "score": 1.0, + "content": "3 ADAPTIVE SELF TRAINING", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "Given a pre-trained language model (e.g., BERT (Devlin et al., 2019)) as the teacher, we first fine-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "tune it on the small labeled data to make it aware of the underlying task. The fine-tuned teacher", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "model is now used to pseudo-label the large unlabeled data. We consider the student model as", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "another instantiation of the pre-trained language model that is trained over the pseudo-labeled data.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 506, + 711 + ], + "score": 1.0, + "content": "However, our few-shot setting with limited labeled data results in a noisy teacher. A naive transfer of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "teacher knowledge to the student results in the propagation of noisy labels limiting the performance", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "of the student model. To address this challenge, we develop an adaptive self-training framework to", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "re-weight pseudo-labeled predictions from the teacher with a meta-learning objective that optimizes", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "the token-level loss from the student model on a held-out labeled validation set. This held-out", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "set is adaptively constructed via labeled data acquisition which selects labeled samples with high", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 272, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 272, + 127 + ], + "score": 1.0, + "content": "uncertainty for efficient data exploration.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 655, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 95 + ], + "score": 1.0, + "content": "re-weight pseudo-labeled predictions from the teacher with a meta-learning objective that optimizes", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "the token-level loss from the student model on a held-out labeled validation set. This held-out", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "set is adaptively constructed via labeled data acquisition which selects labeled samples with high", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 116, + 272, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 272, + 127 + ], + "score": 1.0, + "content": "uncertainty for efficient data exploration.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 108, + 133, + 306, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 307, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 307, + 145 + ], + "score": 1.0, + "content": "3.1 ADAPTIVE LABELED DATA ACQUISITION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 146, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "score": 1.0, + "content": "In standard meta-learning setup for instance-level classification tasks, the held-out validation set is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 505, + 169 + ], + "score": 1.0, + "content": "usually constructed via random sampling (Ren et al., 2018; Shu et al., 2019). Sequence labeling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 168, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 179 + ], + "score": 1.0, + "content": "tasks involve many slot types with variable difficulty and distribution in the data. For instance, NER", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 180, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 190 + ], + "score": 1.0, + "content": "tasks over WikiAnn operate over 123 slot types from 41 languages with additional complexity from", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "variable model performance across different languages. A random sampling leads to oversampling", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 200, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 506, + 214 + ], + "score": 1.0, + "content": "instances with the most populous categories and slot types in the data. Therefore, we propose a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "novel labeled data acquisition strategy to construct the validation set for effective data exploration.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 222, + 503, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 503, + 235 + ], + "score": 1.0, + "content": "We demonstrate its benefit over classic meta-learning approaches from prior works in experiments.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 504, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 504, + 252 + ], + "score": 1.0, + "content": "In general, data acquisition strategies for prior works in meta-learning and active learning broadly", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "leverage random sampling (Ren et al., 2018; Shu et al., 2019), easy (Kumar et al., 2010) and hard", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 261, + 504, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 504, + 273 + ], + "score": 1.0, + "content": "example mining (Shrivastava et al., 2016) or uncertainty-based methods (Chang et al., 2017a). These", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 273, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 504, + 284 + ], + "score": 1.0, + "content": "strategies have been compared in prior works (Chang et al., 2017a; Gal et al., 2017) that show", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "uncertainty-based methods to have better generalizability across diverse settings. There are several", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 504, + 306 + ], + "score": 1.0, + "content": "approaches to uncertainty estimation including error decay (Konyushkova et al., 2017; Chang et al.,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 306, + 504, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 504, + 316 + ], + "score": 1.0, + "content": "2020), Monte Carlo dropouts (Gal et al., 2017) and predictive variance (Chang et al., 2017a). We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 316, + 504, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 504, + 328 + ], + "score": 1.0, + "content": "follow a similar principle of error decay to find samples that the model is uncertain about and can", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 504, + 339 + ], + "score": 1.0, + "content": "correspondingly benefit from knowing their labels (similar to active learning settings). To this end,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 339, + 504, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 504, + 350 + ], + "score": 1.0, + "content": "we leverage stochastic loss decay from the model as a proxy for the model uncertainty to generate", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "validation set on the fly. This is used for estimating token-level weights and re-weighting pseudo", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 359, + 217, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 217, + 372 + ], + "score": 1.0, + "content": "labeled data in Section 3.2.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 102, + 370, + 509, + 395 + ], + "spans": [ + { + "bbox": [ + 102, + 370, + 334, + 395 + ], + "score": 1.0, + "content": "Consider the loss of the student model with parameters", + "type": "text" + }, + { + "bbox": [ + 335, + 374, + 352, + 389 + ], + "score": 0.93, + "content": "\\theta _ { s t u } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 370, + 435, + 395 + ], + "score": 1.0, + "content": "on the labeled data", + "type": "text" + }, + { + "bbox": [ + 435, + 376, + 477, + 390 + ], + "score": 0.92, + "content": "( X _ { m } ^ { l } , Y _ { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 370, + 509, + 395 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 385, + 509, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 111, + 401 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 385, + 172, + 408 + ], + "score": 1.0, + "content": "-th iteration as", + "type": "text" + }, + { + "bbox": [ + 172, + 389, + 252, + 403 + ], + "score": 0.93, + "content": "l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 385, + 509, + 408 + ], + "score": 1.0, + "content": ". To measure the loss decay value at any iteration, we use the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "difference between the current and previous loss values. Considering these values may fluctuate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 412, + 504, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 398, + 426 + ], + "score": 1.0, + "content": "across iterations, we adopt the moving average of the loss values for", + "type": "text" + }, + { + "bbox": [ + 398, + 412, + 439, + 425 + ], + "score": 0.92, + "content": "( X _ { m } ^ { l } , Y _ { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 412, + 495, + 426 + ], + "score": 1.0, + "content": "in the latest", + "type": "text" + }, + { + "bbox": [ + 495, + 413, + 504, + 423 + ], + "score": 0.79, + "content": "R", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 423, + 502, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 199, + 437 + ], + "score": 1.0, + "content": "iterations as a baseline", + "type": "text" + }, + { + "bbox": [ + 199, + 424, + 211, + 436 + ], + "score": 0.9, + "content": "l _ { b } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 423, + 389, + 437 + ], + "score": 1.0, + "content": "for loss decay estimation. Baseline measure", + "type": "text" + }, + { + "bbox": [ + 390, + 425, + 401, + 436 + ], + "score": 0.9, + "content": "l _ { b } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 423, + 502, + 437 + ], + "score": 1.0, + "content": "is calculated as follows:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 439, + 378, + 473 + ], + "lines": [ + { + "bbox": [ + 233, + 439, + 378, + 473 + ], + "spans": [ + { + "bbox": [ + 233, + 439, + 378, + 473 + ], + "score": 0.96, + "content": "l _ { b } ^ { m } = \\frac { 1 } { R } \\sum _ { r = 1 } ^ { R } l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t - r ) } ) ) .", + "type": "interline_equation", + "image_path": "081689b844fe1c90c78194dc94909054d58f72d968de98028221f68df20e6fd6.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 439, + 378, + 456.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 233, + 456.0, + 378, + 473.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "Since the loss decay values are estimated on the fly, we want to balance exploration and exploitation.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 270, + 499 + ], + "score": 1.0, + "content": "To this end, we add a smoothness factor", + "type": "text" + }, + { + "bbox": [ + 270, + 487, + 276, + 496 + ], + "score": 0.82, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 486, + 505, + 499 + ], + "score": 1.0, + "content": "to prevent the low loss decay samples (i.e. samples with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "low uncertainty) from never being selected again. Considering all of the above factors, we obtain", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 342, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 252, + 522 + ], + "score": 1.0, + "content": "the sampling weight of labeled data", + "type": "text" + }, + { + "bbox": [ + 252, + 507, + 293, + 520 + ], + "score": 0.92, + "content": "( X _ { m } ^ { l } , Y _ { m } ^ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 507, + 342, + 522 + ], + "score": 1.0, + "content": "as follows:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 522, + 400, + 539 + ], + "lines": [ + { + "bbox": [ + 210, + 522, + 400, + 539 + ], + "spans": [ + { + "bbox": [ + 210, + 522, + 400, + 539 + ], + "score": 0.92, + "content": "W _ { m } \\propto \\mathrm { m a x } ( l _ { b } ^ { m } - l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) ) , 0 ) + \\delta .", + "type": "interline_equation", + "image_path": "68c406599266593bd337a243fd21d5fe019742eef2d3d18fac7ca85f0ad01bfc.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 210, + 522, + 400, + 539 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 507, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 202, + 554 + ], + "score": 1.0, + "content": "The smoothness factor", + "type": "text" + }, + { + "bbox": [ + 202, + 542, + 208, + 551 + ], + "score": 0.8, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "needs to be adaptive since the training loss is dynamic. Therefore, We", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 552, + 492, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 381, + 565 + ], + "score": 1.0, + "content": "adopt the maximum of the loss decay value as the smoothness factor", + "type": "text" + }, + { + "bbox": [ + 382, + 553, + 388, + 562 + ], + "score": 0.81, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 552, + 492, + 565 + ], + "score": 1.0, + "content": "to encourage exploration.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "The aforementioned acquisition function is re-estimated after a fixed number of steps to adapt to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "model changes. With labeled data acquisition, we rely on informative uncertain samples to improve", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 357, + 604 + ], + "score": 1.0, + "content": "learning efficiency. The sampled mini-batches of labeled data", + "type": "text" + }, + { + "bbox": [ + 357, + 591, + 379, + 603 + ], + "score": 0.92, + "content": "\\{ B _ { s } ^ { l } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "are used as a validation set for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "the student model in the next step for re-weighting pseudo-labeled data from the teacher model. We", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "demonstrate its impact via ablation study in experiments. Note that the labeled data is only used to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 624, + 504, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 504, + 637 + ], + "score": 1.0, + "content": "compute the acquisition function and not used for explicit training of the student model in this step.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "title", + "bbox": [ + 107, + 641, + 307, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 640, + 307, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 307, + 654 + ], + "score": 1.0, + "content": "3.2 RE-WEIGHTING PSEUDO-LABELED DATA", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "To mitigate error propagation from noisy pseudo-labeled sequences from the teacher, we leverage", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "meta-learning to adaptively re-weight them based on the student model loss on the held-out valida-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "tion set obtained via labeled data acquisition from the previous section. In contrast to prior work", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "focusing on instance-level tasks like image classification – sequence labeling operates on discrete", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "text sequences as input and assigns labels to each token in the sequence. Since teacher predictions", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "vary for different slot labels and types, we adapt the meta-learning framework to re-weight samples", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 215, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 215, + 732 + ], + "score": 1.0, + "content": "at a token-level resolution.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 127 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 133, + 306, + 144 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 307, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 307, + 145 + ], + "score": 1.0, + "content": "3.1 ADAPTIVE LABELED DATA ACQUISITION", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 146, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 506, + 158 + ], + "score": 1.0, + "content": "In standard meta-learning setup for instance-level classification tasks, the held-out validation set is", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 157, + 505, + 169 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 505, + 169 + ], + "score": 1.0, + "content": "usually constructed via random sampling (Ren et al., 2018; Shu et al., 2019). Sequence labeling", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 168, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 179 + ], + "score": 1.0, + "content": "tasks involve many slot types with variable difficulty and distribution in the data. For instance, NER", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 180, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 190 + ], + "score": 1.0, + "content": "tasks over WikiAnn operate over 123 slot types from 41 languages with additional complexity from", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 189, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 106, + 189, + 505, + 202 + ], + "score": 1.0, + "content": "variable model performance across different languages. A random sampling leads to oversampling", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 200, + 506, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 506, + 214 + ], + "score": 1.0, + "content": "instances with the most populous categories and slot types in the data. Therefore, we propose a", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 224 + ], + "score": 1.0, + "content": "novel labeled data acquisition strategy to construct the validation set for effective data exploration.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 222, + 503, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 503, + 235 + ], + "score": 1.0, + "content": "We demonstrate its benefit over classic meta-learning approaches from prior works in experiments.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 146, + 506, + 235 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 240, + 505, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 240, + 504, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 504, + 252 + ], + "score": 1.0, + "content": "In general, data acquisition strategies for prior works in meta-learning and active learning broadly", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "leverage random sampling (Ren et al., 2018; Shu et al., 2019), easy (Kumar et al., 2010) and hard", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 261, + 504, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 504, + 273 + ], + "score": 1.0, + "content": "example mining (Shrivastava et al., 2016) or uncertainty-based methods (Chang et al., 2017a). These", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 273, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 504, + 284 + ], + "score": 1.0, + "content": "strategies have been compared in prior works (Chang et al., 2017a; Gal et al., 2017) that show", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 283, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 505, + 295 + ], + "score": 1.0, + "content": "uncertainty-based methods to have better generalizability across diverse settings. There are several", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 295, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 504, + 306 + ], + "score": 1.0, + "content": "approaches to uncertainty estimation including error decay (Konyushkova et al., 2017; Chang et al.,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 306, + 504, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 504, + 316 + ], + "score": 1.0, + "content": "2020), Monte Carlo dropouts (Gal et al., 2017) and predictive variance (Chang et al., 2017a). We", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 316, + 504, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 504, + 328 + ], + "score": 1.0, + "content": "follow a similar principle of error decay to find samples that the model is uncertain about and can", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 327, + 504, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 327, + 504, + 339 + ], + "score": 1.0, + "content": "correspondingly benefit from knowing their labels (similar to active learning settings). To this end,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 339, + 504, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 504, + 350 + ], + "score": 1.0, + "content": "we leverage stochastic loss decay from the model as a proxy for the model uncertainty to generate", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "validation set on the fly. This is used for estimating token-level weights and re-weighting pseudo", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 359, + 217, + 372 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 217, + 372 + ], + "score": 1.0, + "content": "labeled data in Section 3.2.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 240, + 505, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 102, + 370, + 509, + 395 + ], + "spans": [ + { + "bbox": [ + 102, + 370, + 334, + 395 + ], + "score": 1.0, + "content": "Consider the loss of the student model with parameters", + "type": "text" + }, + { + "bbox": [ + 335, + 374, + 352, + 389 + ], + "score": 0.93, + "content": "\\theta _ { s t u } ^ { ( t ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 370, + 435, + 395 + ], + "score": 1.0, + "content": "on the labeled data", + "type": "text" + }, + { + "bbox": [ + 435, + 376, + 477, + 390 + ], + "score": 0.92, + "content": "( X _ { m } ^ { l } , Y _ { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 370, + 509, + 395 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 385, + 509, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 111, + 401 + ], + "score": 0.72, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 385, + 172, + 408 + ], + "score": 1.0, + "content": "-th iteration as", + "type": "text" + }, + { + "bbox": [ + 172, + 389, + 252, + 403 + ], + "score": 0.93, + "content": "l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 385, + 509, + 408 + ], + "score": 1.0, + "content": ". To measure the loss decay value at any iteration, we use the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 414 + ], + "score": 1.0, + "content": "difference between the current and previous loss values. Considering these values may fluctuate", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 412, + 504, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 398, + 426 + ], + "score": 1.0, + "content": "across iterations, we adopt the moving average of the loss values for", + "type": "text" + }, + { + "bbox": [ + 398, + 412, + 439, + 425 + ], + "score": 0.92, + "content": "( X _ { m } ^ { l } , Y _ { m } )", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 412, + 495, + 426 + ], + "score": 1.0, + "content": "in the latest", + "type": "text" + }, + { + "bbox": [ + 495, + 413, + 504, + 423 + ], + "score": 0.79, + "content": "R", + "type": "inline_equation" + } + ], + "index": 28 + }, + { + "bbox": [ + 104, + 423, + 502, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 199, + 437 + ], + "score": 1.0, + "content": "iterations as a baseline", + "type": "text" + }, + { + "bbox": [ + 199, + 424, + 211, + 436 + ], + "score": 0.9, + "content": "l _ { b } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 423, + 389, + 437 + ], + "score": 1.0, + "content": "for loss decay estimation. Baseline measure", + "type": "text" + }, + { + "bbox": [ + 390, + 425, + 401, + 436 + ], + "score": 0.9, + "content": "l _ { b } ^ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 423, + 502, + 437 + ], + "score": 1.0, + "content": "is calculated as follows:", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 102, + 370, + 509, + 437 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 439, + 378, + 473 + ], + "lines": [ + { + "bbox": [ + 233, + 439, + 378, + 473 + ], + "spans": [ + { + "bbox": [ + 233, + 439, + 378, + 473 + ], + "score": 0.96, + "content": "l _ { b } ^ { m } = \\frac { 1 } { R } \\sum _ { r = 1 } ^ { R } l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t - r ) } ) ) .", + "type": "interline_equation", + "image_path": "081689b844fe1c90c78194dc94909054d58f72d968de98028221f68df20e6fd6.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 233, + 439, + 378, + 456.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 233, + 456.0, + 378, + 473.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 520 + ], + "lines": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "Since the loss decay values are estimated on the fly, we want to balance exploration and exploitation.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 486, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 270, + 499 + ], + "score": 1.0, + "content": "To this end, we add a smoothness factor", + "type": "text" + }, + { + "bbox": [ + 270, + 487, + 276, + 496 + ], + "score": 0.82, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 486, + 505, + 499 + ], + "score": 1.0, + "content": "to prevent the low loss decay samples (i.e. samples with", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "low uncertainty) from never being selected again. Considering all of the above factors, we obtain", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 507, + 342, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 252, + 522 + ], + "score": 1.0, + "content": "the sampling weight of labeled data", + "type": "text" + }, + { + "bbox": [ + 252, + 507, + 293, + 520 + ], + "score": 0.92, + "content": "( X _ { m } ^ { l } , Y _ { m } ^ { l } )", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 507, + 342, + 522 + ], + "score": 1.0, + "content": "as follows:", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 475, + 505, + 522 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 522, + 400, + 539 + ], + "lines": [ + { + "bbox": [ + 210, + 522, + 400, + 539 + ], + "spans": [ + { + "bbox": [ + 210, + 522, + 400, + 539 + ], + "score": 0.92, + "content": "W _ { m } \\propto \\mathrm { m a x } ( l _ { b } ^ { m } - l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) ) , 0 ) + \\delta .", + "type": "interline_equation", + "image_path": "68c406599266593bd337a243fd21d5fe019742eef2d3d18fac7ca85f0ad01bfc.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 210, + 522, + 400, + 539 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 507, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 202, + 554 + ], + "score": 1.0, + "content": "The smoothness factor", + "type": "text" + }, + { + "bbox": [ + 202, + 542, + 208, + 551 + ], + "score": 0.8, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "needs to be adaptive since the training loss is dynamic. Therefore, We", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 552, + 492, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 381, + 565 + ], + "score": 1.0, + "content": "adopt the maximum of the loss decay value as the smoothness factor", + "type": "text" + }, + { + "bbox": [ + 382, + 553, + 388, + 562 + ], + "score": 0.81, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 552, + 492, + 565 + ], + "score": 1.0, + "content": "to encourage exploration.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37.5, + "bbox_fs": [ + 106, + 541, + 505, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 636 + ], + "lines": [ + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 106, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "The aforementioned acquisition function is re-estimated after a fixed number of steps to adapt to", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 506, + 593 + ], + "score": 1.0, + "content": "model changes. With labeled data acquisition, we rely on informative uncertain samples to improve", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 591, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 357, + 604 + ], + "score": 1.0, + "content": "learning efficiency. The sampled mini-batches of labeled data", + "type": "text" + }, + { + "bbox": [ + 357, + 591, + 379, + 603 + ], + "score": 0.92, + "content": "\\{ B _ { s } ^ { l } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 591, + 506, + 604 + ], + "score": 1.0, + "content": "are used as a validation set for", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "the student model in the next step for re-weighting pseudo-labeled data from the teacher model. We", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 506, + 626 + ], + "score": 1.0, + "content": "demonstrate its impact via ablation study in experiments. Note that the labeled data is only used to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 624, + 504, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 504, + 637 + ], + "score": 1.0, + "content": "compute the acquisition function and not used for explicit training of the student model in this step.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 569, + 506, + 637 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 641, + 307, + 653 + ], + "lines": [ + { + "bbox": [ + 105, + 640, + 307, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 307, + 654 + ], + "score": 1.0, + "content": "3.2 RE-WEIGHTING PSEUDO-LABELED DATA", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "To mitigate error propagation from noisy pseudo-labeled sequences from the teacher, we leverage", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "meta-learning to adaptively re-weight them based on the student model loss on the held-out valida-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "tion set obtained via labeled data acquisition from the previous section. In contrast to prior work", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "focusing on instance-level tasks like image classification – sequence labeling operates on discrete", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "text sequences as input and assigns labels to each token in the sequence. Since teacher predictions", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "vary for different slot labels and types, we adapt the meta-learning framework to re-weight samples", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 720, + 215, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 215, + 732 + ], + "score": 1.0, + "content": "at a token-level resolution.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 49, + "bbox_fs": [ + 105, + 654, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 80, + 506, + 118 + ], + "lines": [ + { + "bbox": [ + 101, + 78, + 507, + 101 + ], + "spans": [ + { + "bbox": [ + 101, + 78, + 310, + 101 + ], + "score": 1.0, + "content": "Token Re-weighting. Consider the pseudo-labels", + "type": "text" + }, + { + "bbox": [ + 311, + 80, + 434, + 96 + ], + "score": 0.94, + "content": "\\{ \\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ] \\} _ { m = 1 } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 79, + 507, + 98 + ], + "score": 1.0, + "content": "from the teacher", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 131, + 108 + ], + "score": 1.0, + "content": "in the", + "type": "text" + }, + { + "bbox": [ + 132, + 96, + 137, + 105 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 94, + 205, + 108 + ], + "score": 1.0, + "content": "-th iteration with", + "type": "text" + }, + { + "bbox": [ + 205, + 97, + 215, + 105 + ], + "score": 0.75, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 94, + 233, + 108 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 234, + 97, + 241, + 105 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "indexing the instance and a token in the instance, respectively. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 106, + 471, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 471, + 118 + ], + "score": 1.0, + "content": "classic self-training, we update the student parameters leveraging pseudo-labels as follows:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 128, + 415, + 163 + ], + "lines": [ + { + "bbox": [ + 195, + 128, + 415, + 163 + ], + "spans": [ + { + "bbox": [ + 195, + 128, + 415, + 163 + ], + "score": 0.94, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) \\big ) .", + "type": "interline_equation", + "image_path": "a93edb77cbec55bb3f10a1c0c2f2e3059edbfed5a03d2bd61bb287549a51d107.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 195, + 128, + 415, + 145.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 195, + 145.5, + 415, + 163.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 506, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Now, to downplay noisy token-level labels, we leverage meta-learning to re-weight the pseudo-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 182, + 503, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 503, + 196 + ], + "score": 1.0, + "content": "labeled data. To this end, we follow a similar analysis from (Koh & Liang, 2017) and (Ren et al.,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 376, + 207 + ], + "score": 1.0, + "content": "2018) to perturb the weight for each token in the mini-batch by", + "type": "text" + }, + { + "bbox": [ + 376, + 198, + 381, + 204 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 194, + 505, + 207 + ], + "score": 1.0, + "content": ". Weight perturbation is used", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "to discover data points that are most important to improve the model performance on a held-out", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "score": 1.0, + "content": "validation set (Koh & Liang, 2017) where the sample importance is given by the magnitude of the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 226, + 464, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 464, + 240 + ], + "score": 1.0, + "content": "the negative gradients. We extend prior techniques to obtain token-level perturbations as:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 157, + 249, + 454, + 284 + ], + "lines": [ + { + "bbox": [ + 157, + 249, + 454, + 284 + ], + "spans": [ + { + "bbox": [ + 157, + 249, + 454, + 284 + ], + "score": 0.94, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ \\epsilon _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) .", + "type": "interline_equation", + "image_path": "37247b0bb74268012fe5868440cb8e374fa97cafb08a7177e5a490724cedc343.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 157, + 249, + 454, + 260.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 157, + 260.6666666666667, + 454, + 272.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 157, + 272.33333333333337, + 454, + 284.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 504, + 306 + ], + "score": 1.0, + "content": "The token weights are obtained by minimizing the student model loss on the held-out validation set.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 305, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 317 + ], + "score": 1.0, + "content": "Here, we employ the labeled data acquisition strategy from Eq. 4 to sample informative mini-batches", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 168, + 329 + ], + "score": 1.0, + "content": "of labeled data", + "type": "text" + }, + { + "bbox": [ + 168, + 316, + 180, + 328 + ], + "score": 0.9, + "content": "{ \\dot { B } } _ { s } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 315, + 239, + 329 + ], + "score": 1.0, + "content": "locally at step", + "type": "text" + }, + { + "bbox": [ + 239, + 317, + 244, + 326 + ], + "score": 0.66, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 315, + 463, + 329 + ], + "score": 1.0, + "content": ". To obtain a cheap estimate of the meta-weight at step", + "type": "text" + }, + { + "bbox": [ + 464, + 318, + 468, + 326 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 315, + 505, + 329 + ], + "score": 1.0, + "content": ", we take", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 326, + 387, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 367, + 340 + ], + "score": 1.0, + "content": "a single gradient descent step for the sampled labeled mini-batch", + "type": "text" + }, + { + "bbox": [ + 367, + 326, + 379, + 339 + ], + "score": 0.89, + "content": "B _ { s } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 326, + 387, + 340 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 350, + 455, + 387 + ], + "lines": [ + { + "bbox": [ + 156, + 350, + 455, + 387 + ], + "spans": [ + { + "bbox": [ + 156, + 350, + 455, + 387 + ], + "score": 0.93, + "content": "u _ { m , n , s } = - \\frac { \\partial } { \\partial \\epsilon _ { m , n , s } } \\big ( \\frac { 1 } { | \\mathcal { B } _ { s } ^ { l } | } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { | \\mathcal { B } _ { s } ^ { l } | } \\sum _ { n = 1 } ^ { N } [ l ( y _ { m , n } , f ( x _ { m , n } ^ { l } ; \\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) ) ] \\big ) | _ { \\epsilon _ { m , n , s } = 0 }", + "type": "interline_equation", + "image_path": "a54b73da96f95cdd452f67d2ea837739062736d61fa7d9339910bd9d96896b5e.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 156, + 350, + 455, + 362.3333333333333 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 156, + 362.3333333333333, + 455, + 374.66666666666663 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 156, + 374.66666666666663, + 455, + 386.99999999999994 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "We set the token weights to be proportional to the negative gradients to reflect the importance of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "pseudo-labeled tokens in the sequence. Since sequence labeling tasks have dependencies between", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "the slot types and tokens, it is difficult to obtain a good estimation of the weights based on a single", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 302, + 447 + ], + "score": 1.0, + "content": "mini-batch of examples. Therefore, we sample", + "type": "text" + }, + { + "bbox": [ + 303, + 437, + 310, + 445 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 434, + 431, + 447 + ], + "score": 1.0, + "content": "mini-batches of labeled data", + "type": "text" + }, + { + "bbox": [ + 432, + 436, + 483, + 448 + ], + "score": 0.94, + "content": "\\{ B _ { 1 } ^ { l } , . . . , B _ { S } ^ { l } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 445, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 459 + ], + "score": 1.0, + "content": "the adaptive acquisition strategy and calculate the mean of the gradients to obtain a robust gradient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 457, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 185, + 469 + ], + "score": 1.0, + "content": "estimate. Note that", + "type": "text" + }, + { + "bbox": [ + 185, + 459, + 192, + 467 + ], + "score": 0.84, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 457, + 504, + 469 + ], + "score": 1.0, + "content": "is a constant number that is the same for each token and the proportional sign", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 468, + 504, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 504, + 480 + ], + "score": 1.0, + "content": "in Eq. 8. Since a negative weight indicates a pseudo-label of poor quality that would potentially", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 107, + 479, + 486, + 491 + ], + "score": 1.0, + "content": "degrade the model performance, we set such weights to 0 to filter them out. The impact of", + "type": "text" + }, + { + "bbox": [ + 486, + 482, + 493, + 489 + ], + "score": 0.89, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 107, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "investigated in the experiments (refer to Appendix A.1). The overall meta-weight of pseudo-labeled", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 500, + 247, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 131, + 515 + ], + "score": 1.0, + "content": "token", + "type": "text" + }, + { + "bbox": [ + 132, + 502, + 184, + 515 + ], + "score": 0.91, + "content": "\\left( x _ { m , n } ^ { u } , \\hat { y } _ { m , n } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 500, + 247, + 515 + ], + "score": 1.0, + "content": "is obtained as:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 525, + 364, + 559 + ], + "lines": [ + { + "bbox": [ + 247, + 525, + 364, + 559 + ], + "spans": [ + { + "bbox": [ + 247, + 525, + 364, + 559 + ], + "score": 0.94, + "content": "w _ { m , n } \\propto \\mathrm { m a x } ( \\sum _ { s = 1 } ^ { S } u _ { m , n , s } , 0 )", + "type": "interline_equation", + "image_path": "158474a5f605855bf2ce03c2add7d17b0784b5e88eeaa7a49b97d059a0c1afa4.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 525, + 364, + 542.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 247, + 542.0, + 364, + 559.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 569, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 478, + 585 + ], + "score": 1.0, + "content": "To further ensure the stability of the loss function in each mini-batch, we normalise the weight", + "type": "text" + }, + { + "bbox": [ + 478, + 571, + 501, + 582 + ], + "score": 0.87, + "content": "w _ { m , n }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 566, + 505, + 585 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 580, + 500, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 500, + 594 + ], + "score": 1.0, + "content": "Finally, we update the student model parameters while accounting for token-level re-weighting as:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 602, + 450, + 637 + ], + "lines": [ + { + "bbox": [ + 161, + 602, + 450, + 637 + ], + "spans": [ + { + "bbox": [ + 161, + 602, + 450, + 637 + ], + "score": 0.94, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ w _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) .", + "type": "interline_equation", + "image_path": "7eeb7ee5259c8235f204598ace160923989343c9104ae62454eb42a1a21e4cc3.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 161, + 602, + 450, + 613.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 161, + 613.6666666666666, + 450, + 625.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 161, + 625.3333333333333, + 450, + 636.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 498, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 651, + 499, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 499, + 666 + ], + "score": 1.0, + "content": "We demonstrate the impact of our re-weighting mechanism with an ablation study in experiments.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 676, + 298, + 688 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 299, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 299, + 689 + ], + "score": 1.0, + "content": "3.3 TEACHER MODEL ITERATIVE UPDATES", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 695, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 695, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 106, + 695, + 505, + 708 + ], + "score": 1.0, + "content": "At the end of every self-training iteration, we assign the student model as a new teacher model", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 101, + 702, + 510, + 729 + ], + "spans": [ + { + "bbox": [ + 101, + 702, + 127, + 729 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 127, + 707, + 179, + 722 + ], + "score": 0.92, + "content": "\\theta _ { t e a } = \\theta _ { s t u } ^ { ( T ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 702, + 510, + 729 + ], + "score": 1.0, + "content": ". Since the student model uses the labeled data only as a held-out validation setl", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 325, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 325, + 732 + ], + "score": 1.0, + "content": "for meta-learning, we further utilize the labeled data", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 26, + 307, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 307, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 307, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 80, + 506, + 118 + ], + "lines": [ + { + "bbox": [ + 101, + 78, + 507, + 101 + ], + "spans": [ + { + "bbox": [ + 101, + 78, + 310, + 101 + ], + "score": 1.0, + "content": "Token Re-weighting. Consider the pseudo-labels", + "type": "text" + }, + { + "bbox": [ + 311, + 80, + 434, + 96 + ], + "score": 0.94, + "content": "\\{ \\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ] \\} _ { m = 1 } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 79, + 507, + 98 + ], + "score": 1.0, + "content": "from the teacher", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 131, + 108 + ], + "score": 1.0, + "content": "in the", + "type": "text" + }, + { + "bbox": [ + 132, + 96, + 137, + 105 + ], + "score": 0.76, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 94, + 205, + 108 + ], + "score": 1.0, + "content": "-th iteration with", + "type": "text" + }, + { + "bbox": [ + 205, + 97, + 215, + 105 + ], + "score": 0.75, + "content": "m", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 94, + 233, + 108 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 234, + 97, + 241, + 105 + ], + "score": 0.77, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 94, + 506, + 108 + ], + "score": 1.0, + "content": "indexing the instance and a token in the instance, respectively. In", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 106, + 471, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 471, + 118 + ], + "score": 1.0, + "content": "classic self-training, we update the student parameters leveraging pseudo-labels as follows:", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 101, + 78, + 507, + 118 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 128, + 415, + 163 + ], + "lines": [ + { + "bbox": [ + 195, + 128, + 415, + 163 + ], + "spans": [ + { + "bbox": [ + 195, + 128, + 415, + 163 + ], + "score": 0.94, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) \\big ) .", + "type": "interline_equation", + "image_path": "a93edb77cbec55bb3f10a1c0c2f2e3059edbfed5a03d2bd61bb287549a51d107.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 195, + 128, + 415, + 145.5 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 195, + 145.5, + 415, + 163.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 506, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Now, to downplay noisy token-level labels, we leverage meta-learning to re-weight the pseudo-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 182, + 503, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 503, + 196 + ], + "score": 1.0, + "content": "labeled data. To this end, we follow a similar analysis from (Koh & Liang, 2017) and (Ren et al.,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 376, + 207 + ], + "score": 1.0, + "content": "2018) to perturb the weight for each token in the mini-batch by", + "type": "text" + }, + { + "bbox": [ + 376, + 198, + 381, + 204 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 194, + 505, + 207 + ], + "score": 1.0, + "content": ". Weight perturbation is used", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "to discover data points that are most important to improve the model performance on a held-out", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 506, + 229 + ], + "score": 1.0, + "content": "validation set (Koh & Liang, 2017) where the sample importance is given by the magnitude of the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 226, + 464, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 464, + 240 + ], + "score": 1.0, + "content": "the negative gradients. We extend prior techniques to obtain token-level perturbations as:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 172, + 506, + 240 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 157, + 249, + 454, + 284 + ], + "lines": [ + { + "bbox": [ + 157, + 249, + 454, + 284 + ], + "spans": [ + { + "bbox": [ + 157, + 249, + 454, + 284 + ], + "score": 0.94, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ \\epsilon _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) .", + "type": "interline_equation", + "image_path": "37247b0bb74268012fe5868440cb8e374fa97cafb08a7177e5a490724cedc343.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 157, + 249, + 454, + 260.6666666666667 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 157, + 260.6666666666667, + 454, + 272.33333333333337 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 157, + 272.33333333333337, + 454, + 284.00000000000006 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 293, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 294, + 504, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 504, + 306 + ], + "score": 1.0, + "content": "The token weights are obtained by minimizing the student model loss on the held-out validation set.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 305, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 317 + ], + "score": 1.0, + "content": "Here, we employ the labeled data acquisition strategy from Eq. 4 to sample informative mini-batches", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 168, + 329 + ], + "score": 1.0, + "content": "of labeled data", + "type": "text" + }, + { + "bbox": [ + 168, + 316, + 180, + 328 + ], + "score": 0.9, + "content": "{ \\dot { B } } _ { s } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 315, + 239, + 329 + ], + "score": 1.0, + "content": "locally at step", + "type": "text" + }, + { + "bbox": [ + 239, + 317, + 244, + 326 + ], + "score": 0.66, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 315, + 463, + 329 + ], + "score": 1.0, + "content": ". To obtain a cheap estimate of the meta-weight at step", + "type": "text" + }, + { + "bbox": [ + 464, + 318, + 468, + 326 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 315, + 505, + 329 + ], + "score": 1.0, + "content": ", we take", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 326, + 387, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 367, + 340 + ], + "score": 1.0, + "content": "a single gradient descent step for the sampled labeled mini-batch", + "type": "text" + }, + { + "bbox": [ + 367, + 326, + 379, + 339 + ], + "score": 0.89, + "content": "B _ { s } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 326, + 387, + 340 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 294, + 505, + 340 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 156, + 350, + 455, + 387 + ], + "lines": [ + { + "bbox": [ + 156, + 350, + 455, + 387 + ], + "spans": [ + { + "bbox": [ + 156, + 350, + 455, + 387 + ], + "score": 0.93, + "content": "u _ { m , n , s } = - \\frac { \\partial } { \\partial \\epsilon _ { m , n , s } } \\big ( \\frac { 1 } { | \\mathcal { B } _ { s } ^ { l } | } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { | \\mathcal { B } _ { s } ^ { l } | } \\sum _ { n = 1 } ^ { N } [ l ( y _ { m , n } , f ( x _ { m , n } ^ { l } ; \\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) ) ] \\big ) | _ { \\epsilon _ { m , n , s } = 0 }", + "type": "interline_equation", + "image_path": "a54b73da96f95cdd452f67d2ea837739062736d61fa7d9339910bd9d96896b5e.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 156, + 350, + 455, + 362.3333333333333 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 156, + 362.3333333333333, + 455, + 374.66666666666663 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 156, + 374.66666666666663, + 455, + 386.99999999999994 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 505, + 514 + ], + "lines": [ + { + "bbox": [ + 106, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "We set the token weights to be proportional to the negative gradients to reflect the importance of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "pseudo-labeled tokens in the sequence. Since sequence labeling tasks have dependencies between", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "the slot types and tokens, it is difficult to obtain a good estimation of the weights based on a single", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 434, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 302, + 447 + ], + "score": 1.0, + "content": "mini-batch of examples. Therefore, we sample", + "type": "text" + }, + { + "bbox": [ + 303, + 437, + 310, + 445 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 434, + 431, + 447 + ], + "score": 1.0, + "content": "mini-batches of labeled data", + "type": "text" + }, + { + "bbox": [ + 432, + 436, + 483, + 448 + ], + "score": 0.94, + "content": "\\{ B _ { 1 } ^ { l } , . . . , B _ { S } ^ { l } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 445, + 506, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 445, + 506, + 459 + ], + "score": 1.0, + "content": "the adaptive acquisition strategy and calculate the mean of the gradients to obtain a robust gradient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 457, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 185, + 469 + ], + "score": 1.0, + "content": "estimate. Note that", + "type": "text" + }, + { + "bbox": [ + 185, + 459, + 192, + 467 + ], + "score": 0.84, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 457, + 504, + 469 + ], + "score": 1.0, + "content": "is a constant number that is the same for each token and the proportional sign", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 468, + 504, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 468, + 504, + 480 + ], + "score": 1.0, + "content": "in Eq. 8. Since a negative weight indicates a pseudo-label of poor quality that would potentially", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 107, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 107, + 479, + 486, + 491 + ], + "score": 1.0, + "content": "degrade the model performance, we set such weights to 0 to filter them out. The impact of", + "type": "text" + }, + { + "bbox": [ + 486, + 482, + 493, + 489 + ], + "score": 0.89, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 107, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "investigated in the experiments (refer to Appendix A.1). The overall meta-weight of pseudo-labeled", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 500, + 247, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 131, + 515 + ], + "score": 1.0, + "content": "token", + "type": "text" + }, + { + "bbox": [ + 132, + 502, + 184, + 515 + ], + "score": 0.91, + "content": "\\left( x _ { m , n } ^ { u } , \\hat { y } _ { m , n } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 500, + 247, + 515 + ], + "score": 1.0, + "content": "is obtained as:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 401, + 506, + 515 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 247, + 525, + 364, + 559 + ], + "lines": [ + { + "bbox": [ + 247, + 525, + 364, + 559 + ], + "spans": [ + { + "bbox": [ + 247, + 525, + 364, + 559 + ], + "score": 0.94, + "content": "w _ { m , n } \\propto \\mathrm { m a x } ( \\sum _ { s = 1 } ^ { S } u _ { m , n , s } , 0 )", + "type": "interline_equation", + "image_path": "158474a5f605855bf2ce03c2add7d17b0784b5e88eeaa7a49b97d059a0c1afa4.jpg" + } + ] + } + ], + "index": 31.5, + "virtual_lines": [ + { + "bbox": [ + 247, + 525, + 364, + 542.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 247, + 542.0, + 364, + 559.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 569, + 505, + 592 + ], + "lines": [ + { + "bbox": [ + 105, + 566, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 478, + 585 + ], + "score": 1.0, + "content": "To further ensure the stability of the loss function in each mini-batch, we normalise the weight", + "type": "text" + }, + { + "bbox": [ + 478, + 571, + 501, + 582 + ], + "score": 0.87, + "content": "w _ { m , n }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 566, + 505, + 585 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 580, + 500, + 594 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 500, + 594 + ], + "score": 1.0, + "content": "Finally, we update the student model parameters while accounting for token-level re-weighting as:", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 566, + 505, + 594 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 161, + 602, + 450, + 637 + ], + "lines": [ + { + "bbox": [ + 161, + 602, + 450, + 637 + ], + "spans": [ + { + "bbox": [ + 161, + 602, + 450, + 637 + ], + "score": 0.94, + "content": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ w _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) .", + "type": "interline_equation", + "image_path": "7eeb7ee5259c8235f204598ace160923989343c9104ae62454eb42a1a21e4cc3.jpg" + } + ] + } + ], + "index": 36, + "virtual_lines": [ + { + "bbox": [ + 161, + 602, + 450, + 613.6666666666666 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 161, + 613.6666666666666, + 450, + 625.3333333333333 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 161, + 625.3333333333333, + 450, + 636.9999999999999 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 652, + 498, + 664 + ], + "lines": [ + { + "bbox": [ + 106, + 651, + 499, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 499, + 666 + ], + "score": 1.0, + "content": "We demonstrate the impact of our re-weighting mechanism with an ablation study in experiments.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38, + "bbox_fs": [ + 106, + 651, + 499, + 666 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 676, + 298, + 688 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 299, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 299, + 689 + ], + "score": 1.0, + "content": "3.3 TEACHER MODEL ITERATIVE UPDATES", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 695, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 695, + 505, + 708 + ], + "spans": [ + { + "bbox": [ + 106, + 695, + 505, + 708 + ], + "score": 1.0, + "content": "At the end of every self-training iteration, we assign the student model as a new teacher model", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 101, + 702, + 510, + 729 + ], + "spans": [ + { + "bbox": [ + 101, + 702, + 127, + 729 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 127, + 707, + 179, + 722 + ], + "score": 0.92, + "content": "\\theta _ { t e a } = \\theta _ { s t u } ^ { ( T ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 702, + 510, + 729 + ], + "score": 1.0, + "content": ". Since the student model uses the labeled data only as a held-out validation setl", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 325, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 325, + 732 + ], + "score": 1.0, + "content": "for meta-learning, we further utilize the labeled data", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 101, + 695, + 510, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 107, + 78, + 507, + 100 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 145, + 95 + ], + "score": 0.92, + "content": "f ( \\cdot , \\theta _ { t e a } ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 78, + 507, + 100 + ], + "score": 1.0, + "content": "with standard supervised loss minimization. We explore the effectiveness of this step with", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 489, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 489, + 106 + ], + "score": 1.0, + "content": "an ablation study in experiments. The overall training procedure is summarized in Algorithm 1.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 115, + 244, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 244, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 244, + 129 + ], + "score": 1.0, + "content": "Algorithm 1: MetaST Algorithm.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 106, + 130, + 457, + 286 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 130, + 457, + 286 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 130, + 457, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 457, + 286 + ], + "score": 0.323, + "html": "
Input:Labeled sequences (Xl,Y); Unlabeled sequences (X");Pre-trained BERT model with randomly initialized token classification layer f(;((O)); Batches S; Number of self-training iterations T.
Initialize teacher model 0tea =0(0) while not converged do
Fine-tune teacher model on small labeled data (Xl, Y);
Initialize the student model θ(0)=0(0);
Generate hard pseudo-labels Y(t）for unlabeled samples Xu with model f(·,0tea);
fort←1toTdo
Compute labeled data acquisition function according to Eq. 4;
function;Sample S mini-batches of labeled examples {Bl,.,B's} from (Xl,Y）based on labeled data acquisition
Randomly sample abatch of pseudo-labeled examples B from(Xu,Y(t));
Compute token-level weights in B based on the loss on {Bl,..,B's} according to Eq.8;
endstu
Update the teacher:θtea
end
", + "type": "table", + "image_path": "2d4822d52a3210ec1b53952b0ae2190d99dbdf6149e0f2ef6f0fb9f27a125e7d.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 130, + 457, + 182.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 182.0, + 457, + 234.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 234.0, + 457, + 286.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 295, + 201, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 201, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 201, + 309 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "Encoder. Pre-trained language models like BERT (Devlin et al., 2019), GPT-2 (Radford et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "2019) and RoBERTa (Liu et al., 2019) have shown state-of-the-art performance for various natural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "language processing tasks. In this work we adopt one of them as a base encoder by initializing the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 347, + 489, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 489, + 360 + ], + "score": 1.0, + "content": "teacher with pre-trained BERT-base model and a randomly initialized token classification layer.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 290, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 290, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 290, + 376 + ], + "score": 1.0, + "content": "Datasets. We perform large-scale experi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 374, + 291, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 291, + 387 + ], + "score": 1.0, + "content": "ments with six different datasets including", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 385, + 291, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 291, + 398 + ], + "score": 1.0, + "content": "user utterances for task-oriented dialog sys-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 395, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 291, + 410 + ], + "score": 1.0, + "content": "tems and multilingual Named Entity Recog-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 407, + 291, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 291, + 419 + ], + "score": 1.0, + "content": "nition tasks as summarized in Table 1. (a)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 418, + 291, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 291, + 431 + ], + "score": 1.0, + "content": "Email. This consists of natural language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 429, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 291, + 441 + ], + "score": 1.0, + "content": "user utterances for email-oriented user actions", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 439, + 291, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 291, + 453 + ], + "score": 1.0, + "content": "like sending, receiving or searching emails", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 451, + 291, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 291, + 464 + ], + "score": 1.0, + "content": "with attributes like date, time, topics, peo-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 462, + 291, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 291, + 474 + ], + "score": 1.0, + "content": "ple, etc. (b) SNIPS is a public benchmark", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5 + }, + { + "type": "table", + "bbox": [ + 298, + 369, + 497, + 461 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 298, + 369, + 497, + 461 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 298, + 369, + 497, + 461 + ], + "spans": [ + { + "bbox": [ + 298, + 369, + 497, + 461 + ], + "score": 0.974, + "html": "
Dataset# Slots#Train#Test#Lang
Email202.5K1kEN
SNIPS3913K0.7KEN
MIT Movie128.8K2.4KEN
MIT Restaurant86.9K1.5KEN
Wikiann (EN)320K10KEN
CoNLL03 (EN)415K3.6KEN
CoNLL031638K15K4
Wikiann123705K329K41
", + "type": "table", + "image_path": "8275e215d6daaa4c5857892a43aa7a0317755556a49b5761a434b233c2aeffbf.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 298, + 369, + 497, + 382.14285714285717 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 298, + 382.14285714285717, + 497, + 395.28571428571433 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 298, + 395.28571428571433, + 497, + 408.4285714285715 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 298, + 408.4285714285715, + 497, + 421.57142857142867 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 298, + 421.57142857142867, + 497, + 434.71428571428584 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 298, + 434.71428571428584, + 497, + 447.857142857143 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 298, + 447.857142857143, + 497, + 461.00000000000017 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 348, + 462, + 455, + 473 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 347, + 461, + 456, + 475 + ], + "spans": [ + { + "bbox": [ + 347, + 461, + 456, + 475 + ], + "score": 1.0, + "content": "Table 1: Dataset summary.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + } + ], + "index": 26.0 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "score": 1.0, + "content": "dataset (Coucke et al., 2018) of user queries from multiple domains including music, media, and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "weather. (c) MIT Movie and Restaurant corpus (Liu et al., 2013) consist of similar user utterances", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "for movie and restaurant domains. (d) CoNLL03 (Sang & Meulder, 2003) and Wikiann (Pan et al.,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "2017) are public benchmark datasets for multilingual Named Entity Recognition. CoNLL03 is a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "collection of news wire articles from the Reuters Corpus from 4 languages with manual annotations,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "whereas Wikiann comprises of extractions from Wikipedia articles from 41 languages with auto-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "matic annotation leveraging meta-data for different entity types like ORG, PER, LOC etc. For every", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 185, + 562 + ], + "score": 1.0, + "content": "dataset, we sample", + "type": "text" + }, + { + "bbox": [ + 186, + 551, + 272, + 562 + ], + "score": 0.91, + "content": "K \\in \\{ 5 , 1 0 , 2 0 , 1 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "labeled sequences for each slot type from the Train data,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "and add the remaining to the unlabeled set while ignoring their labels – following standard setups", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 322, + 585 + ], + "score": 1.0, + "content": "for semi-supervised learning. We repeatedly sample", + "type": "text" + }, + { + "bbox": [ + 322, + 573, + 333, + 582 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "labeled instances three times for multiple", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 583, + 409, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 409, + 595 + ], + "score": 1.0, + "content": "runs to report average performance with standard deviation across the runs.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "Baselines. The first baseline we consider is the fully supervised BERT model trained on all avail-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "able training data which provides the ceiling performance for every task. Each of the other models", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 167, + 634 + ], + "score": 1.0, + "content": "are trained on", + "type": "text" + }, + { + "bbox": [ + 167, + 623, + 178, + 632 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "training labels per slot type. We adopt several state-of-the-art semi-supervised", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "methods as baselines: (1) CVT (Clark et al., 2018) is a semi-supervised sequence labeling method", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "based on cross-view training; (2) SeqVAT (Chen et al., 2020) incorporates adversarial training with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "conditional random field layer for semi-supervised sequence labeling; (3) Mean Teacher (MT) (Tar-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "vainen & Valpola, 2017) averages model weights to obtain an aggregated teacher; (4) VAT (Miyato", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 677, + 504, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 688 + ], + "score": 1.0, + "content": "et al., 2018) adopts virtual adversarial training to make the model robust to noise; (5) classic ST (III,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "1965) is simple self-training method with hard pseudo-labels; (6) BOND (Liang et al., 2020) is the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "most recent work on self-training for sequence labeling with confidence-based sample selection and", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "forms a strong baseline for our work. We implement our framework in Pytorch and use Tesla V100", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 502, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 502, + 733 + ], + "score": 1.0, + "content": "gpus for experiments. Hyper-parameter configurations with model settings presented in Appendix.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 106 + ], + "lines": [ + { + "bbox": [ + 107, + 78, + 507, + 100 + ], + "spans": [ + { + "bbox": [ + 107, + 81, + 145, + 95 + ], + "score": 0.92, + "content": "f ( \\cdot , \\theta _ { t e a } ^ { ( t ) } )", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 78, + 507, + 100 + ], + "score": 1.0, + "content": "with standard supervised loss minimization. We explore the effectiveness of this step with", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 489, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 489, + 106 + ], + "score": 1.0, + "content": "an ablation study in experiments. The overall training procedure is summarized in Algorithm 1.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 78, + 507, + 106 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 115, + 244, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 114, + 244, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 244, + 129 + ], + "score": 1.0, + "content": "Algorithm 1: MetaST Algorithm.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 106, + 130, + 457, + 286 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 130, + 457, + 286 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 130, + 457, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 130, + 457, + 286 + ], + "score": 0.323, + "html": "
Input:Labeled sequences (Xl,Y); Unlabeled sequences (X");Pre-trained BERT model with randomly initialized token classification layer f(;((O)); Batches S; Number of self-training iterations T.
Initialize teacher model 0tea =0(0) while not converged do
Fine-tune teacher model on small labeled data (Xl, Y);
Initialize the student model θ(0)=0(0);
Generate hard pseudo-labels Y(t）for unlabeled samples Xu with model f(·,0tea);
fort←1toTdo
Compute labeled data acquisition function according to Eq. 4;
function;Sample S mini-batches of labeled examples {Bl,.,B's} from (Xl,Y）based on labeled data acquisition
Randomly sample abatch of pseudo-labeled examples B from(Xu,Y(t));
Compute token-level weights in B based on the loss on {Bl,..,B's} according to Eq.8;
endstu
Update the teacher:θtea
end
", + "type": "table", + "image_path": "2d4822d52a3210ec1b53952b0ae2190d99dbdf6149e0f2ef6f0fb9f27a125e7d.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 106, + 130, + 457, + 182.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 182.0, + 457, + 234.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 234.0, + 457, + 286.0 + ], + "spans": [], + "index": 5 + } + ] + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 295, + 201, + 307 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 201, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 201, + 309 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 505, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 505, + 326 + ], + "score": 1.0, + "content": "Encoder. Pre-trained language models like BERT (Devlin et al., 2019), GPT-2 (Radford et al.,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "2019) and RoBERTa (Liu et al., 2019) have shown state-of-the-art performance for various natural", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 336, + 505, + 348 + ], + "score": 1.0, + "content": "language processing tasks. In this work we adopt one of them as a base encoder by initializing the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 347, + 489, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 489, + 360 + ], + "score": 1.0, + "content": "teacher with pre-trained BERT-base model and a randomly initialized token classification layer.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 314, + 505, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 364, + 290, + 473 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 290, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 290, + 376 + ], + "score": 1.0, + "content": "Datasets. We perform large-scale experi-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 374, + 291, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 291, + 387 + ], + "score": 1.0, + "content": "ments with six different datasets including", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 385, + 291, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 291, + 398 + ], + "score": 1.0, + "content": "user utterances for task-oriented dialog sys-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 395, + 291, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 291, + 410 + ], + "score": 1.0, + "content": "tems and multilingual Named Entity Recog-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 407, + 291, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 291, + 419 + ], + "score": 1.0, + "content": "nition tasks as summarized in Table 1. (a)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 418, + 291, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 291, + 431 + ], + "score": 1.0, + "content": "Email. This consists of natural language", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 429, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 291, + 441 + ], + "score": 1.0, + "content": "user utterances for email-oriented user actions", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 439, + 291, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 291, + 453 + ], + "score": 1.0, + "content": "like sending, receiving or searching emails", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 451, + 291, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 291, + 464 + ], + "score": 1.0, + "content": "with attributes like date, time, topics, peo-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 462, + 291, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 291, + 474 + ], + "score": 1.0, + "content": "ple, etc. (b) SNIPS is a public benchmark", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 363, + 291, + 474 + ] + }, + { + "type": "table", + "bbox": [ + 298, + 369, + 497, + 461 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 298, + 369, + 497, + 461 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 298, + 369, + 497, + 461 + ], + "spans": [ + { + "bbox": [ + 298, + 369, + 497, + 461 + ], + "score": 0.974, + "html": "
Dataset# Slots#Train#Test#Lang
Email202.5K1kEN
SNIPS3913K0.7KEN
MIT Movie128.8K2.4KEN
MIT Restaurant86.9K1.5KEN
Wikiann (EN)320K10KEN
CoNLL03 (EN)415K3.6KEN
CoNLL031638K15K4
Wikiann123705K329K41
", + "type": "table", + "image_path": "8275e215d6daaa4c5857892a43aa7a0317755556a49b5761a434b233c2aeffbf.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 298, + 369, + 497, + 382.14285714285717 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 298, + 382.14285714285717, + 497, + 395.28571428571433 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 298, + 395.28571428571433, + 497, + 408.4285714285715 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 298, + 408.4285714285715, + 497, + 421.57142857142867 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 298, + 421.57142857142867, + 497, + 434.71428571428584 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 298, + 434.71428571428584, + 497, + 447.857142857143 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 298, + 447.857142857143, + 497, + 461.00000000000017 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 348, + 462, + 455, + 473 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 347, + 461, + 456, + 475 + ], + "spans": [ + { + "bbox": [ + 347, + 461, + 456, + 475 + ], + "score": 1.0, + "content": "Table 1: Dataset summary.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + } + ], + "index": 26.0 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 594 + ], + "lines": [ + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 473, + 506, + 485 + ], + "score": 1.0, + "content": "dataset (Coucke et al., 2018) of user queries from multiple domains including music, media, and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "weather. (c) MIT Movie and Restaurant corpus (Liu et al., 2013) consist of similar user utterances", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "for movie and restaurant domains. (d) CoNLL03 (Sang & Meulder, 2003) and Wikiann (Pan et al.,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "2017) are public benchmark datasets for multilingual Named Entity Recognition. CoNLL03 is a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 506, + 531 + ], + "score": 1.0, + "content": "collection of news wire articles from the Reuters Corpus from 4 languages with manual annotations,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "whereas Wikiann comprises of extractions from Wikipedia articles from 41 languages with auto-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "matic annotation leveraging meta-data for different entity types like ORG, PER, LOC etc. For every", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 185, + 562 + ], + "score": 1.0, + "content": "dataset, we sample", + "type": "text" + }, + { + "bbox": [ + 186, + 551, + 272, + 562 + ], + "score": 0.91, + "content": "K \\in \\{ 5 , 1 0 , 2 0 , 1 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "labeled sequences for each slot type from the Train data,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "and add the remaining to the unlabeled set while ignoring their labels – following standard setups", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 322, + 585 + ], + "score": 1.0, + "content": "for semi-supervised learning. We repeatedly sample", + "type": "text" + }, + { + "bbox": [ + 322, + 573, + 333, + 582 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "labeled instances three times for multiple", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 583, + 409, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 409, + 595 + ], + "score": 1.0, + "content": "runs to report average performance with standard deviation across the runs.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 473, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 600, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 613 + ], + "score": 1.0, + "content": "Baselines. The first baseline we consider is the fully supervised BERT model trained on all avail-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 505, + 622 + ], + "score": 1.0, + "content": "able training data which provides the ceiling performance for every task. Each of the other models", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 167, + 634 + ], + "score": 1.0, + "content": "are trained on", + "type": "text" + }, + { + "bbox": [ + 167, + 623, + 178, + 632 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "training labels per slot type. We adopt several state-of-the-art semi-supervised", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "methods as baselines: (1) CVT (Clark et al., 2018) is a semi-supervised sequence labeling method", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 656 + ], + "score": 1.0, + "content": "based on cross-view training; (2) SeqVAT (Chen et al., 2020) incorporates adversarial training with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "conditional random field layer for semi-supervised sequence labeling; (3) Mean Teacher (MT) (Tar-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "vainen & Valpola, 2017) averages model weights to obtain an aggregated teacher; (4) VAT (Miyato", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 677, + 504, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 688 + ], + "score": 1.0, + "content": "et al., 2018) adopts virtual adversarial training to make the model robust to noise; (5) classic ST (III,", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 699 + ], + "score": 1.0, + "content": "1965) is simple self-training method with hard pseudo-labels; (6) BOND (Liang et al., 2020) is the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "most recent work on self-training for sequence labeling with confidence-based sample selection and", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "forms a strong baseline for our work. We implement our framework in Pytorch and use Tesla V100", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 720, + 502, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 502, + 733 + ], + "score": 1.0, + "content": "gpus for experiments. Hyper-parameter configurations with model settings presented in Appendix.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 599, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Neural sequence labeling performance with few training labels. Table 2 shows the performance", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 271, + 106 + ], + "score": 1.0, + "content": "comparison among different models with", + "type": "text" + }, + { + "bbox": [ + 271, + 94, + 296, + 104 + ], + "score": 0.85, + "content": "\\mathrm { K } { = } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "labeled examples per slot type. The fully supervised", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "BERT trained on thousands of labeled examples provides the ceiling performance for the few-shot", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "setting. We observe our method MetaST to significantly outperform all methods across all datasets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "including the models that also use the same BERT encoder as ours like MT, VAT, Classic ST and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 382, + 150 + ], + "score": 1.0, + "content": "BOND with corresponding average performance improvements as", + "type": "text" + }, + { + "bbox": [ + 383, + 137, + 415, + 148 + ], + "score": 0.88, + "content": "1 4 . 2 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 137, + 420, + 150 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 420, + 137, + 453, + 148 + ], + "score": 0.88, + "content": "1 4 . 9 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 137, + 457, + 150 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 458, + 137, + 486, + 148 + ], + "score": 0.87, + "content": "8 . 4 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 147, + 484, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 134, + 159 + ], + "score": 0.86, + "content": "8 . 8 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 147, + 273, + 161 + ], + "score": 1.0, + "content": ". Non BERT models like CVT and", + "type": "text" + }, + { + "bbox": [ + 273, + 149, + 308, + 160 + ], + "score": 0.62, + "content": "{ \\tt S e q V A T }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 147, + 484, + 161 + ], + "score": 1.0, + "content": "are consistently worse than other baselines.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 106, + 167, + 503, + 326 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 167, + 503, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 167, + 503, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 503, + 326 + ], + "score": 0.984, + "html": "
MethodSNIPSEmailMovieRestaurantCoNLL03 (EN)Wikiann (EN)
# Slots392012843
Full-supervision
BERT95.8094.4487.8778.9592.4084.04
Few-shot supervision (10 labels per slot)69.5054.0671.1545.61
BERT 79.01 87.85
Few-shot supervision (10 labels per slot) + unlabeled data
CVT78.2378.2462.7342.5754.3127.89
SeqVAT78.6772.6567.1051.5567.2135.16
MT79.4889.5367.6251.7568.6741.43
VAT79.0889.7170.1753.3465.0338.81
Classic ST83.2690.7071.8856.8070.9946.15
BOND83.5489.7570.9155.7869.5648.73
MetaST88.2392.1877.6763.8376.65
(0.04;↑12%)(0.47;↑4.93%)(0.10;↑11.76%)(1.62;↑18.07%)(0.73;17.73%)56.61 (0.4;↑24.12%)
", + "type": "table", + "image_path": "807ffdbed937b94ff4fc00672dd73d1c64c56dace02c37ba83e830ba3369cdbb.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 106, + 167, + 503, + 220.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 220.0, + 503, + 273.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 273.0, + 503, + 326.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "Table 2: F1 score comparison of models for sequence labeling on different datasets. All models (ex-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 343, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 164, + 355 + ], + "score": 1.0, + "content": "cept CVT and", + "type": "text" + }, + { + "bbox": [ + 164, + 343, + 199, + 354 + ], + "score": 0.32, + "content": "{ \\tt S e q V A T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 343, + 506, + 355 + ], + "score": 1.0, + "content": "use the same BERT encoder. F1 score of our model for each task is followed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 354, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 481, + 366 + ], + "score": 1.0, + "content": "by standard deviation and percentage improvement (↑) over BERT with few-shot supervision.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 371, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "We also observe variable performance of the models across different tasks. Specifically, the per-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "formance gap between the best few-shot model and the fully supervised model varies significantly.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "score": 1.0, + "content": "MetaST achieves close performance to the fully-supervised model in some datasets (e.g. SNIPS and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "Email) but has bigger room for improvement in others (e.g. CoNLL03 (EN) and Wikiann (EN)). This", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "can be attributed to the following factors. (i) Labeled training examples and slots. The total number", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 352, + 439 + ], + "score": 1.0, + "content": "of labeled training instances for our K-shot setting is given by", + "type": "text" + }, + { + "bbox": [ + 352, + 426, + 401, + 437 + ], + "score": 0.89, + "content": "K \\times \\# S l o t s", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 426, + 505, + 439 + ], + "score": 1.0, + "content": ". Therefore, for tasks with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "higher number of slots and consequently more training labels, most of the models perform better", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "including MetaST. Task-oriented dialog systems with more slots and inherent dependency between", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "the slot types benefit more than NER tasks. (ii) Task difficulty: User utterances from task-oriented", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "score": 1.0, + "content": "dialog systems for some of the domains like weather, music and emails contain predictive query", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "score": 1.0, + "content": "patterns and limited diversity. In contrast, Named Entity Recognition datasets are comparatively", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "diverse and require more training labels to generalize well. Similar observations are also depicted", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "in Table 3 for multilingual NER tasks with more slots and consequently more training labels from", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 513, + 467, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 467, + 528 + ], + "score": 1.0, + "content": "multiple languages as well as richer interactions across the slots from different languages.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19.5 + }, + { + "type": "table", + "bbox": [ + 108, + 530, + 503, + 579 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 530, + 503, + 579 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 530, + 503, + 579 + ], + "spans": [ + { + "bbox": [ + 108, + 530, + 503, + 579 + ], + "score": 0.971, + "html": "
Dataset#Lang#SlotsFull Sup.Few-shot Sup.Few-shot supervision + unlabeled data
BERTBERTMTVATClassic STBONDMetaST
CoNLL0341687.6770.7768.3467.6372.6972.7976.41 (0.47) (↑ 7.97%)
Wikiann4112387.1779.6780.2378.8280.2479.5781.61 (0.14) (↑ 2.42%)
", + "type": "table", + "image_path": "0163b49f589eb8049dc62232c52dfb0b36f2fa1d5e795dc8ccb45a6b066ed870.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 108, + 530, + 503, + 546.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 108, + 546.3333333333334, + 503, + 562.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 108, + 562.6666666666667, + 503, + 579.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 587, + 503, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "Table 3: F1 score comparison of models for sequence labeling on multilingual datasets using the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "same BERT-Multilingual-Base encoder. F1 score of MetaST for each task is followed by standard", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 610, + 492, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 492, + 622 + ], + "score": 1.0, + "content": "deviation in parentheses and percentage improvement (↑) over BERT with few-shot supervision.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 107, + 627, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 107, + 627, + 505, + 638 + ], + "score": 1.0, + "content": "Controlling for the total amount of labeled data. In order to control for the variable amount of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "training labels across different datasets, we perform another experiment where we vary the number", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "of labels for different slot types while keeping the total number of labeled instances for each dataset", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "similar (ca. 200). Results are shown in Table 4. To better illustrate the effect of the number of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 669, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 685 + ], + "score": 1.0, + "content": "training labels, we choose tasks with lower performance in Table 2 for this experiment. Comparing", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 680, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 695 + ], + "score": 1.0, + "content": "the results in Tables 2 and 4, we observe the performance of MetaST to improve with more training", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 693, + 200, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 200, + 704 + ], + "score": 1.0, + "content": "labels for all the tasks .", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 275, + 723 + ], + "score": 1.0, + "content": "Effect of varying the number of labels", + "type": "text" + }, + { + "bbox": [ + 275, + 710, + 286, + 720 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 707, + 505, + 723 + ], + "score": 1.0, + "content": "per slot. Table 5 shows the improvement in the per-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 719, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 505, + 733 + ], + "score": 1.0, + "content": "formance of MetaST when increasing the number of labels for each slot type in the SNIPS dataset.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Neural sequence labeling performance with few training labels. Table 2 shows the performance", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 271, + 106 + ], + "score": 1.0, + "content": "comparison among different models with", + "type": "text" + }, + { + "bbox": [ + 271, + 94, + 296, + 104 + ], + "score": 0.85, + "content": "\\mathrm { K } { = } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "labeled examples per slot type. The fully supervised", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "BERT trained on thousands of labeled examples provides the ceiling performance for the few-shot", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "setting. We observe our method MetaST to significantly outperform all methods across all datasets", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 138 + ], + "score": 1.0, + "content": "including the models that also use the same BERT encoder as ours like MT, VAT, Classic ST and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 382, + 150 + ], + "score": 1.0, + "content": "BOND with corresponding average performance improvements as", + "type": "text" + }, + { + "bbox": [ + 383, + 137, + 415, + 148 + ], + "score": 0.88, + "content": "1 4 . 2 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 137, + 420, + 150 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 420, + 137, + 453, + 148 + ], + "score": 0.88, + "content": "1 4 . 9 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 453, + 137, + 457, + 150 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 458, + 137, + 486, + 148 + ], + "score": 0.87, + "content": "8 . 4 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 147, + 484, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 134, + 159 + ], + "score": 0.86, + "content": "8 . 8 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 147, + 273, + 161 + ], + "score": 1.0, + "content": ". Non BERT models like CVT and", + "type": "text" + }, + { + "bbox": [ + 273, + 149, + 308, + 160 + ], + "score": 0.62, + "content": "{ \\tt S e q V A T }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 147, + 484, + 161 + ], + "score": 1.0, + "content": "are consistently worse than other baselines.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 82, + 506, + 161 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 167, + 503, + 326 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 167, + 503, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 167, + 503, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 167, + 503, + 326 + ], + "score": 0.984, + "html": "
MethodSNIPSEmailMovieRestaurantCoNLL03 (EN)Wikiann (EN)
# Slots392012843
Full-supervision
BERT95.8094.4487.8778.9592.4084.04
Few-shot supervision (10 labels per slot)69.5054.0671.1545.61
BERT 79.01 87.85
Few-shot supervision (10 labels per slot) + unlabeled data
CVT78.2378.2462.7342.5754.3127.89
SeqVAT78.6772.6567.1051.5567.2135.16
MT79.4889.5367.6251.7568.6741.43
VAT79.0889.7170.1753.3465.0338.81
Classic ST83.2690.7071.8856.8070.9946.15
BOND83.5489.7570.9155.7869.5648.73
MetaST88.2392.1877.6763.8376.65
(0.04;↑12%)(0.47;↑4.93%)(0.10;↑11.76%)(1.62;↑18.07%)(0.73;17.73%)56.61 (0.4;↑24.12%)
", + "type": "table", + "image_path": "807ffdbed937b94ff4fc00672dd73d1c64c56dace02c37ba83e830ba3369cdbb.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 106, + 167, + 503, + 220.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 220.0, + 503, + 273.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 273.0, + 503, + 326.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "Table 2: F1 score comparison of models for sequence labeling on different datasets. All models (ex-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 343, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 164, + 355 + ], + "score": 1.0, + "content": "cept CVT and", + "type": "text" + }, + { + "bbox": [ + 164, + 343, + 199, + 354 + ], + "score": 0.32, + "content": "{ \\tt S e q V A T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 343, + 506, + 355 + ], + "score": 1.0, + "content": "use the same BERT encoder. F1 score of our model for each task is followed", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 354, + 481, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 481, + 366 + ], + "score": 1.0, + "content": "by standard deviation and percentage improvement (↑) over BERT with few-shot supervision.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 330, + 506, + 366 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 371, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 384 + ], + "score": 1.0, + "content": "We also observe variable performance of the models across different tasks. Specifically, the per-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "formance gap between the best few-shot model and the fully supervised model varies significantly.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 506, + 406 + ], + "score": 1.0, + "content": "MetaST achieves close performance to the fully-supervised model in some datasets (e.g. SNIPS and", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 505, + 417 + ], + "score": 1.0, + "content": "Email) but has bigger room for improvement in others (e.g. CoNLL03 (EN) and Wikiann (EN)). This", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 428 + ], + "score": 1.0, + "content": "can be attributed to the following factors. (i) Labeled training examples and slots. The total number", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 352, + 439 + ], + "score": 1.0, + "content": "of labeled training instances for our K-shot setting is given by", + "type": "text" + }, + { + "bbox": [ + 352, + 426, + 401, + 437 + ], + "score": 0.89, + "content": "K \\times \\# S l o t s", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 426, + 505, + 439 + ], + "score": 1.0, + "content": ". Therefore, for tasks with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 506, + 450 + ], + "score": 1.0, + "content": "higher number of slots and consequently more training labels, most of the models perform better", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 448, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 506, + 461 + ], + "score": 1.0, + "content": "including MetaST. Task-oriented dialog systems with more slots and inherent dependency between", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "score": 1.0, + "content": "the slot types benefit more than NER tasks. (ii) Task difficulty: User utterances from task-oriented", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 484 + ], + "score": 1.0, + "content": "dialog systems for some of the domains like weather, music and emails contain predictive query", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "score": 1.0, + "content": "patterns and limited diversity. In contrast, Named Entity Recognition datasets are comparatively", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 504 + ], + "score": 1.0, + "content": "diverse and require more training labels to generalize well. Similar observations are also depicted", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "in Table 3 for multilingual NER tasks with more slots and consequently more training labels from", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 513, + 467, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 467, + 528 + ], + "score": 1.0, + "content": "multiple languages as well as richer interactions across the slots from different languages.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 371, + 506, + 528 + ] + }, + { + "type": "table", + "bbox": [ + 108, + 530, + 503, + 579 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 530, + 503, + 579 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 530, + 503, + 579 + ], + "spans": [ + { + "bbox": [ + 108, + 530, + 503, + 579 + ], + "score": 0.971, + "html": "
Dataset#Lang#SlotsFull Sup.Few-shot Sup.Few-shot supervision + unlabeled data
BERTBERTMTVATClassic STBONDMetaST
CoNLL0341687.6770.7768.3467.6372.6972.7976.41 (0.47) (↑ 7.97%)
Wikiann4112387.1779.6780.2378.8280.2479.5781.61 (0.14) (↑ 2.42%)
", + "type": "table", + "image_path": "0163b49f589eb8049dc62232c52dfb0b36f2fa1d5e795dc8ccb45a6b066ed870.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 108, + 530, + 503, + 546.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 108, + 546.3333333333334, + 503, + 562.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 108, + 562.6666666666667, + 503, + 579.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 587, + 503, + 621 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "Table 3: F1 score comparison of models for sequence labeling on multilingual datasets using the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 610 + ], + "score": 1.0, + "content": "same BERT-Multilingual-Base encoder. F1 score of MetaST for each task is followed by standard", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 610, + 492, + 622 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 492, + 622 + ], + "score": 1.0, + "content": "deviation in parentheses and percentage improvement (↑) over BERT with few-shot supervision.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 587, + 505, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 626, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 107, + 627, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 107, + 627, + 505, + 638 + ], + "score": 1.0, + "content": "Controlling for the total amount of labeled data. In order to control for the variable amount of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "training labels across different datasets, we perform another experiment where we vary the number", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 506, + 661 + ], + "score": 1.0, + "content": "of labels for different slot types while keeping the total number of labeled instances for each dataset", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "similar (ca. 200). Results are shown in Table 4. To better illustrate the effect of the number of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 669, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 669, + 505, + 685 + ], + "score": 1.0, + "content": "training labels, we choose tasks with lower performance in Table 2 for this experiment. Comparing", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 680, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 505, + 695 + ], + "score": 1.0, + "content": "the results in Tables 2 and 4, we observe the performance of MetaST to improve with more training", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 693, + 200, + 704 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 200, + 704 + ], + "score": 1.0, + "content": "labels for all the tasks .", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 627, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 707, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 707, + 275, + 723 + ], + "score": 1.0, + "content": "Effect of varying the number of labels", + "type": "text" + }, + { + "bbox": [ + 275, + 710, + 286, + 720 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 707, + 505, + 723 + ], + "score": 1.0, + "content": "per slot. Table 5 shows the improvement in the per-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 719, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 505, + 733 + ], + "score": 1.0, + "content": "formance of MetaST when increasing the number of labels for each slot type in the SNIPS dataset.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 707, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 73, + 473, + 137 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 137, + 73, + 473, + 137 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 73, + 473, + 137 + ], + "spans": [ + { + "bbox": [ + 137, + 73, + 473, + 137 + ], + "score": 0.98, + "html": "
DatasetBERT (Full Supervision)BERT (Few-shot Supervision)MetaST（%Improvement）
MIT Movie87.8775.8180.33 (↑5.96%)
MIT Restaurant78.9560.1267.86 (↑12.87%)
CoNLL03 (EN)92.4077.4881.61 (↑ 5.33%)
Wikiann (EN)84.0462.0471.27 (个14.88%)
Average85.8268.8675.27 (↑ 9.31%)
", + "type": "table", + "image_path": "341320487be14546e4f7fd3e1dfa7a103dea98dca4596616889fdc0c97e73bd2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 137, + 73, + 473, + 94.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 137, + 94.33333333333333, + 473, + 115.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 137, + 115.66666666666666, + 473, + 137.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 145, + 503, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 159 + ], + "score": 1.0, + "content": "Table 4: F1 scores of different models with 200 labeled samples for each task. The percentage", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 155, + 383, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 162, + 168 + ], + "score": 1.0, + "content": "improvement", + "type": "text" + }, + { + "bbox": [ + 162, + 157, + 174, + 167 + ], + "score": 0.29, + "content": "( \\uparrow )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 155, + 383, + 168 + ], + "score": 1.0, + "content": "is over the BERT model with few-shot supervision.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 173, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "Similar trends can be found on other datasets (results in Appendix). As we increase the amount of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "labeled training instances, the performance of BERT also improves, and correspondingly the margin", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "between MetaST and these baselines decreases although MetaST still improves over all of them.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "score": 1.0, + "content": "In the self-training framework, given the ceiling performance for every task and the improved per-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "formance of the teacher with more training labels, there is less room for (relative) improvement of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 228, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 451, + 240 + ], + "score": 1.0, + "content": "the student over the teacher model. Consider SNIPS for example. Our model obtains", + "type": "text" + }, + { + "bbox": [ + 451, + 228, + 471, + 239 + ], + "score": 0.86, + "content": "12 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 228, + 489, + 240 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 489, + 228, + 504, + 239 + ], + "score": 0.84, + "content": "2 \\%", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "score": 1.0, + "content": "improvement over the few-shot BERT model for the 10-shot and 100-shot setting with F1-scores as", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 138, + 261 + ], + "score": 0.88, + "content": "8 8 . 2 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 250, + 156, + 262 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 250, + 188, + 261 + ], + "score": 0.88, + "content": "9 5 . 3 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 250, + 403, + 262 + ], + "score": 1.0, + "content": ", respectively. The ceiling performance for this task is", + "type": "text" + }, + { + "bbox": [ + 403, + 250, + 431, + 261 + ], + "score": 0.87, + "content": "9 5 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "on training BERT", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 261, + 504, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 504, + 272 + ], + "score": 1.0, + "content": "on the entire dataset with 13K labeled examples. This demonstrates that MetaST is most impactful", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 370, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 370, + 284 + ], + "score": 1.0, + "content": "for low-resource settings with few training labels for a given task.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 128, + 290, + 485, + 352 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 128, + 290, + 485, + 352 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 128, + 290, + 485, + 352 + ], + "spans": [ + { + "bbox": [ + 128, + 290, + 485, + 352 + ], + "score": 0.98, + "html": "
#SlotsFew-shot SupervisionFew-shot supervision + unlabeled data
BERTCVTSeqVATMTVATClassic STBONDMetaST(%Improvement)
570.6369.8269.3470.8571.3472.5972.8581.56 (↑15%)
1079.0178.2378.6779.4879.0883.2683.5488.22 (↑12%)
2086.8188.0485.0587.3188.1988.3288.9391.99 (↑6%)
10093.9094.6191.4694.2694.5393.9294.2295.39 (12%)
", + "type": "table", + "image_path": "d87a88b88eea368e0ed209527bab0ae1bc88c0732dee080f36daf2334a029db1.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 128, + 290, + 485, + 310.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 128, + 310.6666666666667, + 485, + 331.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 128, + 331.33333333333337, + 485, + 352.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 357, + 504, + 379 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 316, + 369 + ], + "score": 1.0, + "content": "Table 5: Variation in model performance on varying", + "type": "text" + }, + { + "bbox": [ + 317, + 357, + 327, + 367 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "labels / slot on SNIPS dataset with 39 slots.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 367, + 468, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 225, + 380 + ], + "score": 1.0, + "content": "The percentage improvement", + "type": "text" + }, + { + "bbox": [ + 226, + 368, + 238, + 379 + ], + "score": 0.46, + "content": "( \\uparrow )", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 367, + 468, + 380 + ], + "score": 1.0, + "content": "is relative to the BERT model with few-shot supervision.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + } + ], + "index": 17.25 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Ablation analysis. Table 6 demonstrates the impact of different MetaST components with ablation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "analysis. We observe that soft pseudo-labels hurt the model performance compared to hard pseudo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "labels, as also shown in recent work (Kumar et al., 2020). Such a performance drop may be attributed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "to soft labels being less informative compared to sharpened ones. Removing the iterative teacher", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 427, + 366, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 366, + 441 + ], + "score": 1.0, + "content": "fine-tuning step (Section 3.1) also hurts the overall performance.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 127, + 442, + 328, + 588 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 127, + 442, + 328, + 588 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 127, + 442, + 328, + 588 + ], + "spans": [ + { + "bbox": [ + 127, + 442, + 328, + 588 + ], + "score": 0.973, + "html": "
MethodDatasets
SNIPSCoNLL03
BERT w/ Continued Pre-training +
Few-shot Supervision83.9669.84
Classic ST83.2670.99
Classic ST w/ Soft Pseudo-Labels81.1771.87
MetaST(ours) w/Hard Pseudo-Labels88.2376.65
MetaST w/ Soft Pseudo-Labels86.1675.84
MetaST w/o Iterative Teacher Fine-tune85.6472.74
MetaST w/o Labeled Data Acq.86.6375.02
Pseudo-labeled Data Re-weighting
MetaST w/o Re-weighting85.4873.02
MetaST (Easy)85.5674.53
MetaST (Difficult)86.3468.06
", + "type": "table", + "image_path": "97f50f81144a3b1bc21a414d82218b82ee1a37c17c4b6780a77ef271131c1485.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 127, + 442, + 328, + 454.1666666666667 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 127, + 454.1666666666667, + 328, + 466.33333333333337 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 127, + 466.33333333333337, + 328, + 478.50000000000006 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 127, + 478.50000000000006, + 328, + 490.66666666666674 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 127, + 490.66666666666674, + 328, + 502.8333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 127, + 502.8333333333334, + 328, + 515.0000000000001 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 127, + 515.0000000000001, + 328, + 527.1666666666667 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 127, + 527.1666666666667, + 328, + 539.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 127, + 539.3333333333334, + 328, + 551.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 127, + 551.5, + 328, + 563.6666666666666 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 127, + 563.6666666666666, + 328, + 575.8333333333333 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 127, + 575.8333333333333, + 328, + 587.9999999999999 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 108, + 596, + 347, + 619 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 596, + 348, + 608 + ], + "spans": [ + { + "bbox": [ + 108, + 596, + 348, + 608 + ], + "score": 1.0, + "content": "Table 6: Ablation analysis of our framework MetaST with", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 108, + 606, + 348, + 619 + ], + "spans": [ + { + "bbox": [ + 108, + 606, + 348, + 619 + ], + "score": 1.0, + "content": "10 labeled examples per slot on SNIPS and CoNLL03 (EN).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "index": 39.0 + }, + { + "type": "image", + "bbox": [ + 349, + 462, + 505, + 568 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 349, + 462, + 505, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 349, + 462, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 349, + 462, + 505, + 568 + ], + "score": 0.965, + "type": "image", + "image_path": "6e003691b635563a0994ab8d5c2a1f4aff6b7c20f56ba95beaa0a166af22b000.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 349, + 462, + 505, + 475.25 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 349, + 475.25, + 505, + 488.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 349, + 488.5, + 505, + 501.75 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 349, + 501.75, + 505, + 515.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 349, + 515.0, + 505, + 528.25 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 349, + 528.25, + 505, + 541.5 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 349, + 541.5, + 505, + 554.75 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 349, + 554.75, + 505, + 568.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 349, + 577, + 500, + 600 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 347, + 577, + 501, + 589 + ], + "spans": [ + { + "bbox": [ + 347, + 577, + 501, + 589 + ], + "score": 1.0, + "content": "Figure 2: Visualization of MetaST re-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 348, + 588, + 471, + 600 + ], + "spans": [ + { + "bbox": [ + 348, + 588, + 471, + 600 + ], + "score": 1.0, + "content": "weighting on CoNLL03 (EN).", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "index": 43.0 + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Continued pre-training v.s. self-training. To contrast continued pre-training with self-training,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 634, + 504, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 504, + 646 + ], + "score": 1.0, + "content": "we further pre-train BERT on in-domain unlabeled data and then fine-tune it with few labeled ex-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 319, + 658 + ], + "score": 1.0, + "content": "amples denoted as “BERT (Continued Pre-training", + "type": "text" + }, + { + "bbox": [ + 319, + 646, + 328, + 655 + ], + "score": 0.72, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "Few-shot Supervision)”. The pre-training", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "step improves the BERT performance over the baseline on SNIPS but degrades the performance", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "on CoNLL03. This indicates that continued pre-training can improve the performance of few-shot", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "supervised BERT on specialized tasks (e.g., SNIPS) with different data distribution than the original", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "pre-training data (e.g., Wikipedia), but may not help for general domain ones like CoNLL03 with", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "overlapping data from Wikipedia. In contrast to the above baseline, MetaST brings significant im-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "provements on both datasets. This demonstrates the generality and flexibility of self-training over", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 722, + 488, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 722, + 488, + 734 + ], + "score": 1.0, + "content": "pre-training as also observed in contemporary work (Zoph et al., 2020) on image classification.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 53.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 137, + 73, + 473, + 137 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 137, + 73, + 473, + 137 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 137, + 73, + 473, + 137 + ], + "spans": [ + { + "bbox": [ + 137, + 73, + 473, + 137 + ], + "score": 0.98, + "html": "
DatasetBERT (Full Supervision)BERT (Few-shot Supervision)MetaST（%Improvement）
MIT Movie87.8775.8180.33 (↑5.96%)
MIT Restaurant78.9560.1267.86 (↑12.87%)
CoNLL03 (EN)92.4077.4881.61 (↑ 5.33%)
Wikiann (EN)84.0462.0471.27 (个14.88%)
Average85.8268.8675.27 (↑ 9.31%)
", + "type": "table", + "image_path": "341320487be14546e4f7fd3e1dfa7a103dea98dca4596616889fdc0c97e73bd2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 137, + 73, + 473, + 94.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 137, + 94.33333333333333, + 473, + 115.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 137, + 115.66666666666666, + 473, + 137.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 145, + 503, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 142, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 505, + 159 + ], + "score": 1.0, + "content": "Table 4: F1 scores of different models with 200 labeled samples for each task. The percentage", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 155, + 383, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 155, + 162, + 168 + ], + "score": 1.0, + "content": "improvement", + "type": "text" + }, + { + "bbox": [ + 162, + 157, + 174, + 167 + ], + "score": 0.29, + "content": "( \\uparrow )", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 155, + 383, + 168 + ], + "score": 1.0, + "content": "is over the BERT model with few-shot supervision.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 173, + 505, + 284 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 505, + 185 + ], + "score": 1.0, + "content": "Similar trends can be found on other datasets (results in Appendix). As we increase the amount of", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "labeled training instances, the performance of BERT also improves, and correspondingly the margin", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 208 + ], + "score": 1.0, + "content": "between MetaST and these baselines decreases although MetaST still improves over all of them.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 505, + 219 + ], + "score": 1.0, + "content": "In the self-training framework, given the ceiling performance for every task and the improved per-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 506, + 230 + ], + "score": 1.0, + "content": "formance of the teacher with more training labels, there is less room for (relative) improvement of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 228, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 228, + 451, + 240 + ], + "score": 1.0, + "content": "the student over the teacher model. Consider SNIPS for example. Our model obtains", + "type": "text" + }, + { + "bbox": [ + 451, + 228, + 471, + 239 + ], + "score": 0.86, + "content": "12 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 228, + 489, + 240 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 489, + 228, + 504, + 239 + ], + "score": 0.84, + "content": "2 \\%", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 251 + ], + "score": 1.0, + "content": "improvement over the few-shot BERT model for the 10-shot and 100-shot setting with F1-scores as", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 250, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 138, + 261 + ], + "score": 0.88, + "content": "8 8 . 2 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 250, + 156, + 262 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 250, + 188, + 261 + ], + "score": 0.88, + "content": "9 5 . 3 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 250, + 403, + 262 + ], + "score": 1.0, + "content": ", respectively. The ceiling performance for this task is", + "type": "text" + }, + { + "bbox": [ + 403, + 250, + 431, + 261 + ], + "score": 0.87, + "content": "9 5 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 250, + 505, + 262 + ], + "score": 1.0, + "content": "on training BERT", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 261, + 504, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 504, + 272 + ], + "score": 1.0, + "content": "on the entire dataset with 13K labeled examples. This demonstrates that MetaST is most impactful", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 272, + 370, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 370, + 284 + ], + "score": 1.0, + "content": "for low-resource settings with few training labels for a given task.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 173, + 506, + 284 + ] + }, + { + "type": "table", + "bbox": [ + 128, + 290, + 485, + 352 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 128, + 290, + 485, + 352 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 128, + 290, + 485, + 352 + ], + "spans": [ + { + "bbox": [ + 128, + 290, + 485, + 352 + ], + "score": 0.98, + "html": "
#SlotsFew-shot SupervisionFew-shot supervision + unlabeled data
BERTCVTSeqVATMTVATClassic STBONDMetaST(%Improvement)
570.6369.8269.3470.8571.3472.5972.8581.56 (↑15%)
1079.0178.2378.6779.4879.0883.2683.5488.22 (↑12%)
2086.8188.0485.0587.3188.1988.3288.9391.99 (↑6%)
10093.9094.6191.4694.2694.5393.9294.2295.39 (12%)
", + "type": "table", + "image_path": "d87a88b88eea368e0ed209527bab0ae1bc88c0732dee080f36daf2334a029db1.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 128, + 290, + 485, + 310.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 128, + 310.6666666666667, + 485, + 331.33333333333337 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 128, + 331.33333333333337, + 485, + 352.00000000000006 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 107, + 357, + 504, + 379 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 106, + 356, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 356, + 316, + 369 + ], + "score": 1.0, + "content": "Table 5: Variation in model performance on varying", + "type": "text" + }, + { + "bbox": [ + 317, + 357, + 327, + 367 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 356, + 505, + 369 + ], + "score": 1.0, + "content": "labels / slot on SNIPS dataset with 39 slots.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 367, + 468, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 225, + 380 + ], + "score": 1.0, + "content": "The percentage improvement", + "type": "text" + }, + { + "bbox": [ + 226, + 368, + 238, + 379 + ], + "score": 0.46, + "content": "( \\uparrow )", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 367, + 468, + 380 + ], + "score": 1.0, + "content": "is relative to the BERT model with few-shot supervision.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + } + ], + "index": 17.25 + }, + { + "type": "text", + "bbox": [ + 107, + 383, + 504, + 439 + ], + "lines": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 505, + 396 + ], + "score": 1.0, + "content": "Ablation analysis. Table 6 demonstrates the impact of different MetaST components with ablation", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "analysis. We observe that soft pseudo-labels hurt the model performance compared to hard pseudo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 418 + ], + "score": 1.0, + "content": "labels, as also shown in recent work (Kumar et al., 2020). Such a performance drop may be attributed", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 505, + 429 + ], + "score": 1.0, + "content": "to soft labels being less informative compared to sharpened ones. Removing the iterative teacher", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 427, + 366, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 366, + 441 + ], + "score": 1.0, + "content": "fine-tuning step (Section 3.1) also hurts the overall performance.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 384, + 505, + 441 + ] + }, + { + "type": "table", + "bbox": [ + 127, + 442, + 328, + 588 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 127, + 442, + 328, + 588 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 127, + 442, + 328, + 588 + ], + "spans": [ + { + "bbox": [ + 127, + 442, + 328, + 588 + ], + "score": 0.973, + "html": "
MethodDatasets
SNIPSCoNLL03
BERT w/ Continued Pre-training +
Few-shot Supervision83.9669.84
Classic ST83.2670.99
Classic ST w/ Soft Pseudo-Labels81.1771.87
MetaST(ours) w/Hard Pseudo-Labels88.2376.65
MetaST w/ Soft Pseudo-Labels86.1675.84
MetaST w/o Iterative Teacher Fine-tune85.6472.74
MetaST w/o Labeled Data Acq.86.6375.02
Pseudo-labeled Data Re-weighting
MetaST w/o Re-weighting85.4873.02
MetaST (Easy)85.5674.53
MetaST (Difficult)86.3468.06
", + "type": "table", + "image_path": "97f50f81144a3b1bc21a414d82218b82ee1a37c17c4b6780a77ef271131c1485.jpg" + } + ] + } + ], + "index": 30.5, + "virtual_lines": [ + { + "bbox": [ + 127, + 442, + 328, + 454.1666666666667 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 127, + 454.1666666666667, + 328, + 466.33333333333337 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 127, + 466.33333333333337, + 328, + 478.50000000000006 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 127, + 478.50000000000006, + 328, + 490.66666666666674 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 127, + 490.66666666666674, + 328, + 502.8333333333334 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 127, + 502.8333333333334, + 328, + 515.0000000000001 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 127, + 515.0000000000001, + 328, + 527.1666666666667 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 127, + 527.1666666666667, + 328, + 539.3333333333334 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 127, + 539.3333333333334, + 328, + 551.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 127, + 551.5, + 328, + 563.6666666666666 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 127, + 563.6666666666666, + 328, + 575.8333333333333 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 127, + 575.8333333333333, + 328, + 587.9999999999999 + ], + "spans": [], + "index": 36 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 108, + 596, + 347, + 619 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 596, + 348, + 608 + ], + "spans": [ + { + "bbox": [ + 108, + 596, + 348, + 608 + ], + "score": 1.0, + "content": "Table 6: Ablation analysis of our framework MetaST with", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 108, + 606, + 348, + 619 + ], + "spans": [ + { + "bbox": [ + 108, + 606, + 348, + 619 + ], + "score": 1.0, + "content": "10 labeled examples per slot on SNIPS and CoNLL03 (EN).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + } + ], + "index": 39.0 + }, + { + "type": "image", + "bbox": [ + 349, + 462, + 505, + 568 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 349, + 462, + 505, + 568 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 349, + 462, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 349, + 462, + 505, + 568 + ], + "score": 0.965, + "type": "image", + "image_path": "6e003691b635563a0994ab8d5c2a1f4aff6b7c20f56ba95beaa0a166af22b000.jpg" + } + ] + } + ], + "index": 40.5, + "virtual_lines": [ + { + "bbox": [ + 349, + 462, + 505, + 475.25 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 349, + 475.25, + 505, + 488.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 349, + 488.5, + 505, + 501.75 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 349, + 501.75, + 505, + 515.0 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 349, + 515.0, + 505, + 528.25 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 349, + 528.25, + 505, + 541.5 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 349, + 541.5, + 505, + 554.75 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 349, + 554.75, + 505, + 568.0 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 349, + 577, + 500, + 600 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 347, + 577, + 501, + 589 + ], + "spans": [ + { + "bbox": [ + 347, + 577, + 501, + 589 + ], + "score": 1.0, + "content": "Figure 2: Visualization of MetaST re-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 348, + 588, + 471, + 600 + ], + "spans": [ + { + "bbox": [ + 348, + 588, + 471, + 600 + ], + "score": 1.0, + "content": "weighting on CoNLL03 (EN).", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "index": 43.0 + }, + { + "type": "text", + "bbox": [ + 106, + 623, + 505, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Continued pre-training v.s. self-training. To contrast continued pre-training with self-training,", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 106, + 634, + 504, + 646 + ], + "spans": [ + { + "bbox": [ + 106, + 634, + 504, + 646 + ], + "score": 1.0, + "content": "we further pre-train BERT on in-domain unlabeled data and then fine-tune it with few labeled ex-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 319, + 658 + ], + "score": 1.0, + "content": "amples denoted as “BERT (Continued Pre-training", + "type": "text" + }, + { + "bbox": [ + 319, + 646, + 328, + 655 + ], + "score": 0.72, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "Few-shot Supervision)”. The pre-training", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "step improves the BERT performance over the baseline on SNIPS but degrades the performance", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 680 + ], + "score": 1.0, + "content": "on CoNLL03. This indicates that continued pre-training can improve the performance of few-shot", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 690 + ], + "score": 1.0, + "content": "supervised BERT on specialized tasks (e.g., SNIPS) with different data distribution than the original", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 701 + ], + "score": 1.0, + "content": "pre-training data (e.g., Wikipedia), but may not help for general domain ones like CoNLL03 with", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "overlapping data from Wikipedia. In contrast to the above baseline, MetaST brings significant im-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "provements on both datasets. This demonstrates the generality and flexibility of self-training over", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 722, + 488, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 722, + 488, + 734 + ], + "score": 1.0, + "content": "pre-training as also observed in contemporary work (Zoph et al., 2020) on image classification.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 53.5, + "bbox_fs": [ + 105, + 622, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Adaptive labeled data acquisition. We perform an ablation study by removing adaptive labeled", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "data acquisition from MetaST (denoted as “MetaST w/o Labeled Data Acq.”). Removing this com-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 198, + 116 + ], + "score": 1.0, + "content": "ponent leads to around", + "type": "text" + }, + { + "bbox": [ + 198, + 105, + 212, + 115 + ], + "score": 0.85, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "performance drop on an average demonstrating the impact of labeled data", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 126 + ], + "score": 1.0, + "content": "acquisition. Moreover, the performance drop on SNIPS (39 slots) is larger than that on CoNLL03 (4", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "score": 1.0, + "content": "slots). This demonstrates that adaptive acquisition is more helpful for tasks with more slot types –", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "where diversity and data distribution necessitate a better exploration strategy in contrast to random", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 306, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 306, + 160 + ], + "score": 1.0, + "content": "sampling employed in prior meta-learning works.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 504, + 173 + ], + "score": 1.0, + "content": "Re-weighting strategies. To explore the role of token-level re-weighting for pseudo-labeled se-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "quences (discussed in Section 3.2), we replace our meta-learning component with different sam-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "score": 1.0, + "content": "ple selection strategies based on the model confidence for different tokens. One sampling strategy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "chooses samples uniformly without any re-weighting (referred to as “MetaST w/o Re-weighting”).", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "The sampling strategy with weights proportional to the model confidence favors easy samples (re-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "ferred to as “MetaST-Easy”), whereas the converse favors difficult ones (referred to as “MetaST-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "Difficult”).We observe the meta-learning based re-weighting strategy to perform the best. Inter-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "estingly, MetaST-Easy outperforms MetaST-Difficult significantly on CoNLL03 (EN) but achieves", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "slightly lower performance on SNIPS. This demonstrates that difficult samples are more helpful", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "when the quality of pseudo-labeled data is relatively high. On the converse, the sample selection", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "score": 1.0, + "content": "strategy focusing on difficult samples introduces noisy examples with lower pseudo-label quality.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "Therefore, sampling strategies may need to vary for different datasets, thereby, demonstrating the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 294, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 504, + 305 + ], + "score": 1.0, + "content": "necessity of adaptive data re-weighting as in our framework MetaST. Moreover, MetaST signifi-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "cantly outperforms classic self-training strategies with hard and soft pseudo-labels demonstrating", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 314, + 233, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 233, + 328 + ], + "score": 1.0, + "content": "the effectiveness of our design.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 346 + ], + "score": 1.0, + "content": "Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 342, + 504, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 479, + 355 + ], + "score": 1.0, + "content": "mechanism, we illustrate token-level re-weighting of MetaST on CoNLL03 (EN) dataset with", + "type": "text" + }, + { + "bbox": [ + 480, + 343, + 504, + 353 + ], + "score": 0.79, + "content": "{ \\mathrm { K } } { = } 1 0", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "shot at step 100 in Fig. 2. We include the re-weighting visualisation on SNIPS in Appendix A.1. We", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "observe that the selection mechanism filters out most of the noisy pseudo-labels (colored in blue)", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 376, + 346, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 346, + 389 + ], + "score": 1.0, + "content": "even those with high teacher confidence as shown in Fig. 2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 212, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 213, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 213, + 408 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "Semi-supervised learning has been widely used for consistency training (Bachman et al., 2014;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "Rasmus et al., 2015; Laine & Aila, 2017; Tarvainen & Valpola, 2017; Miyato et al., 2018), latent", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "score": 1.0, + "content": "variable models (Kingma et al., 2014) for sentence compression (Miao & Blunsom, 2016) and code", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "generation (Yin et al., 2018). More recently, methods like UDA (Xie et al., 2019) leverage consis-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "tency training for few-shot learning of instance-classification tasks leveraging auxiliary resources", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 469, + 377, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 377, + 482 + ], + "score": 1.0, + "content": "like paraphrasing and back-translation (BT) (Sennrich et al., 2016).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "score": 1.0, + "content": "Sample selection. Curriculum learning (Bengio et al., 2009) techniques are based on the idea of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "score": 1.0, + "content": "learning easier aspects of the task first followed by the more complex ones. Prior work leveraging", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "self-paced learning (Kumar et al., 2010) and more recently self-paced co-training (Ma et al., 2017)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "leverage teacher confidence to select easy samples during training. Sample selection for image", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "classification tasks have been explored in recent works with meta-learning (Ren et al., 2018; Li", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "et al., 2019) and active learning (Panagiota Mastoropoulou, 2019; Chang et al., 2017b). However,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 552, + 502, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 502, + 564 + ], + "score": 1.0, + "content": "all of these techniques rely on only the model outputs applied to instance-level classification tasks.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 583 + ], + "score": 1.0, + "content": "Semi-supervised sequence labeling. Miller et al. (2004); Peters et al. (2017) leverage large", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "amounts of unlabeled data to improve token representation for sequence labeling tasks. Another", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 589, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 604 + ], + "score": 1.0, + "content": "line of research introduces latent variable modeling (Chen et al., 2019; Zhou & Neubig, 2017) to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "learn interpretable and structured latent representations. Recently, adversarial training based model", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "SeqVAT (Chen et al., 2020) and cross-view training method CVT (Clark et al., 2018) have shown", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 624, + 290, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 290, + 636 + ], + "score": 1.0, + "content": "promising results for sequence labeling tasks.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 201, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 203, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 203, + 661 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "In this work, we develop an adaptive self-training framework MetaST that leverages self-training and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "meta-learning for few-shot training of neural sequence taggers. We address the issue of error propa-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "gation from noisy pseudo-labels from the teacher in the self-training framework by adaptive sample", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "selection and re-weighting with meta-learning. Extensive experiments on six benchmark datasets", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "and different tasks including multilingual NER and slot tagging for task-oriented dialog systems", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 719, + 475, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 475, + 734 + ], + "score": 1.0, + "content": "demonstrate the effectiveness of the proposed method particularly for low-resource settings.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 306, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Adaptive labeled data acquisition. We perform an ablation study by removing adaptive labeled", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 105 + ], + "score": 1.0, + "content": "data acquisition from MetaST (denoted as “MetaST w/o Labeled Data Acq.”). Removing this com-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 198, + 116 + ], + "score": 1.0, + "content": "ponent leads to around", + "type": "text" + }, + { + "bbox": [ + 198, + 105, + 212, + 115 + ], + "score": 0.85, + "content": "2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 104, + 505, + 116 + ], + "score": 1.0, + "content": "performance drop on an average demonstrating the impact of labeled data", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 126 + ], + "score": 1.0, + "content": "acquisition. Moreover, the performance drop on SNIPS (39 slots) is larger than that on CoNLL03 (4", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 505, + 139 + ], + "score": 1.0, + "content": "slots). This demonstrates that adaptive acquisition is more helpful for tasks with more slot types –", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 138, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 107, + 138, + 505, + 149 + ], + "score": 1.0, + "content": "where diversity and data distribution necessitate a better exploration strategy in contrast to random", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 306, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 306, + 160 + ], + "score": 1.0, + "content": "sampling employed in prior meta-learning works.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 81, + 505, + 160 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 162, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 162, + 504, + 173 + ], + "score": 1.0, + "content": "Re-weighting strategies. To explore the role of token-level re-weighting for pseudo-labeled se-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "quences (discussed in Section 3.2), we replace our meta-learning component with different sam-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 197 + ], + "score": 1.0, + "content": "ple selection strategies based on the model confidence for different tokens. One sampling strategy", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 505, + 207 + ], + "score": 1.0, + "content": "chooses samples uniformly without any re-weighting (referred to as “MetaST w/o Re-weighting”).", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 106, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "The sampling strategy with weights proportional to the model confidence favors easy samples (re-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "ferred to as “MetaST-Easy”), whereas the converse favors difficult ones (referred to as “MetaST-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "Difficult”).We observe the meta-learning based re-weighting strategy to perform the best. Inter-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 505, + 250 + ], + "score": 1.0, + "content": "estingly, MetaST-Easy outperforms MetaST-Difficult significantly on CoNLL03 (EN) but achieves", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 505, + 262 + ], + "score": 1.0, + "content": "slightly lower performance on SNIPS. This demonstrates that difficult samples are more helpful", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 505, + 273 + ], + "score": 1.0, + "content": "when the quality of pseudo-labeled data is relatively high. On the converse, the sample selection", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 504, + 284 + ], + "score": 1.0, + "content": "strategy focusing on difficult samples introduces noisy examples with lower pseudo-label quality.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "Therefore, sampling strategies may need to vary for different datasets, thereby, demonstrating the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 294, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 504, + 305 + ], + "score": 1.0, + "content": "necessity of adaptive data re-weighting as in our framework MetaST. Moreover, MetaST signifi-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 317 + ], + "score": 1.0, + "content": "cantly outperforms classic self-training strategies with hard and soft pseudo-labels demonstrating", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 314, + 233, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 233, + 328 + ], + "score": 1.0, + "content": "the effectiveness of our design.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 162, + 506, + 328 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 332, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 506, + 346 + ], + "score": 1.0, + "content": "Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 342, + 504, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 479, + 355 + ], + "score": 1.0, + "content": "mechanism, we illustrate token-level re-weighting of MetaST on CoNLL03 (EN) dataset with", + "type": "text" + }, + { + "bbox": [ + 480, + 343, + 504, + 353 + ], + "score": 0.79, + "content": "{ \\mathrm { K } } { = } 1 0", + "type": "inline_equation" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "shot at step 100 in Fig. 2. We include the re-weighting visualisation on SNIPS in Appendix A.1. We", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 505, + 377 + ], + "score": 1.0, + "content": "observe that the selection mechanism filters out most of the noisy pseudo-labels (colored in blue)", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 376, + 346, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 346, + 389 + ], + "score": 1.0, + "content": "even those with high teacher confidence as shown in Fig. 2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 330, + 506, + 389 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 212, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 213, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 213, + 408 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 414, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "score": 1.0, + "content": "Semi-supervised learning has been widely used for consistency training (Bachman et al., 2014;", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "Rasmus et al., 2015; Laine & Aila, 2017; Tarvainen & Valpola, 2017; Miyato et al., 2018), latent", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 506, + 450 + ], + "score": 1.0, + "content": "variable models (Kingma et al., 2014) for sentence compression (Miao & Blunsom, 2016) and code", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "generation (Yin et al., 2018). More recently, methods like UDA (Xie et al., 2019) leverage consis-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 459, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 505, + 471 + ], + "score": 1.0, + "content": "tency training for few-shot learning of instance-classification tasks leveraging auxiliary resources", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 469, + 377, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 377, + 482 + ], + "score": 1.0, + "content": "like paraphrasing and back-translation (BT) (Sennrich et al., 2016).", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 414, + 506, + 482 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 506, + 498 + ], + "score": 1.0, + "content": "Sample selection. Curriculum learning (Bengio et al., 2009) techniques are based on the idea of", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 511 + ], + "score": 1.0, + "content": "learning easier aspects of the task first followed by the more complex ones. Prior work leveraging", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "self-paced learning (Kumar et al., 2010) and more recently self-paced co-training (Ma et al., 2017)", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 533 + ], + "score": 1.0, + "content": "leverage teacher confidence to select easy samples during training. Sample selection for image", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 542 + ], + "score": 1.0, + "content": "classification tasks have been explored in recent works with meta-learning (Ren et al., 2018; Li", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "et al., 2019) and active learning (Panagiota Mastoropoulou, 2019; Chang et al., 2017b). However,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 552, + 502, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 502, + 564 + ], + "score": 1.0, + "content": "all of these techniques rely on only the model outputs applied to instance-level classification tasks.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 485, + 506, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 567, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 583 + ], + "score": 1.0, + "content": "Semi-supervised sequence labeling. Miller et al. (2004); Peters et al. (2017) leverage large", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "amounts of unlabeled data to improve token representation for sequence labeling tasks. Another", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 589, + 506, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 604 + ], + "score": 1.0, + "content": "line of research introduces latent variable modeling (Chen et al., 2019; Zhou & Neubig, 2017) to", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "learn interpretable and structured latent representations. Recently, adversarial training based model", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "SeqVAT (Chen et al., 2020) and cross-view training method CVT (Clark et al., 2018) have shown", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 624, + 290, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 290, + 636 + ], + "score": 1.0, + "content": "promising results for sequence labeling tasks.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 567, + 506, + 636 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 201, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 642, + 203, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 203, + 661 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 677 + ], + "score": 1.0, + "content": "In this work, we develop an adaptive self-training framework MetaST that leverages self-training and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "meta-learning for few-shot training of neural sequence taggers. We address the issue of error propa-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "gation from noisy pseudo-labels from the teacher in the self-training framework by adaptive sample", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "selection and re-weighting with meta-learning. Extensive experiments on six benchmark datasets", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "and different tasks including multilingual NER and slot tagging for task-oriented dialog systems", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 719, + 475, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 475, + 734 + ], + "score": 1.0, + "content": "demonstrate the effectiveness of the proposed method particularly for low-resource settings.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5, + "bbox_fs": [ + 105, + 665, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Zoubin", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 505, + 124 + ], + "score": 1.0, + "content": "Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 505, + 135 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems 27: Annual Conference on Neural Informa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "tion Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 3365–3373,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 142, + 142, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 142, + 156 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 175 + ], + "score": 1.0, + "content": "Trapit Bansal, Rishikesh Jha, and Andrew McCallum. Learning to few-shot learn across diverse", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 289, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 289, + 184 + ], + "score": 1.0, + "content": "natural language classification tasks, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "Yoshua Bengio, Jer´ ome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ˆ", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 202, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 505, + 214 + ], + "score": 1.0, + "content": "Andrea Pohoreckyj Danyluk, Leon Bottou, and Michael L. Littman (eds.), ´ Proceedings of the 26th", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 214, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 504, + 226 + ], + "score": 1.0, + "content": "Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 223, + 504, + 237 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 504, + 237 + ], + "score": 1.0, + "content": "June 14-18, 2009, volume 382 of ACM International Conference Proceeding Series, pp. 41–48.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 235, + 300, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 300, + 246 + ], + "score": 1.0, + "content": "ACM, 2009. doi: 10.1145/1553374.1553380.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 263, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 263, + 505, + 278 + ], + "score": 1.0, + "content": "accurate neural networks by emphasizing high variance samples. In I. Guyon, U. V. Luxburg,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 287, + 453, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 287, + 453, + 299 + ], + "score": 1.0, + "content": "Information Processing Systems 30, pp. 1002–1012. Curran Associates, Inc., 2017a.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 304, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Haw-Shiuan Chang, Erik G. Learned-Miller, and Andrew McCallum. Active bias: Training more", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 316, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 505, + 328 + ], + "score": 1.0, + "content": "accurate neural networks by emphasizing high variance samples. In Isabelle Guyon, Ulrike von", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 336, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 505, + 350 + ], + "score": 1.0, + "content": "Garnett (eds.), Advances in Neural Information Processing Systems 30: Annual Conference on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 347, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 506, + 362 + ], + "score": 1.0, + "content": "Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 359, + 196, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 359, + 196, + 370 + ], + "score": 1.0, + "content": "1002–1012, 2017b.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 390 + ], + "score": 1.0, + "content": "Haw-Shiuan Chang, Shankar Vembu, Sunil Mohan, Rheeya Uppaal, and Andrew McCallum. Us-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 389, + 504, + 401 + ], + "spans": [ + { + "bbox": [ + 116, + 389, + 504, + 401 + ], + "score": 1.0, + "content": "ing error decay prediction to overcome practical issues of deep active learning for named entity", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 400, + 351, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 400, + 351, + 411 + ], + "score": 1.0, + "content": "recognition. Machine Learning, 109(9):1749–1778, 2020.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 417, + 474, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 475, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 475, + 431 + ], + "score": 1.0, + "content": "Olivier Chapelle, Bernhard Schlkopf, and Alexander Zien. Semi-supervised learning. 2010.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 506, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "Luoxin Chen, Weitong Ruan, Xinyue Liu, and Jianhua Lu. Seqvat: Virtual adversarial training for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "semi-supervised sequence labeling. In Proceedings of the 58th Annual Meeting of the Association", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 457, + 331, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 457, + 331, + 470 + ], + "score": 1.0, + "content": "for Computational Linguistics, pp. 8801–8811, 2020.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 504, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 489 + ], + "score": 1.0, + "content": "Mingda Chen, Qingming Tang, Karen Livescu, and Kevin Gimpel. Variational sequential labelers", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 487, + 399, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 399, + 499 + ], + "score": 1.0, + "content": "for semi-supervised learning. arXiv preprint arXiv:1906.09535, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 504, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "score": 1.0, + "content": "Kevin Clark, Minh-Thang Luong, Christopher D Manning, and Quoc V Le. Semi-supervised se-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 516, + 453, + 529 + ], + "spans": [ + { + "bbox": [ + 115, + 516, + 453, + 529 + ], + "score": 1.0, + "content": "quence modeling with cross-view training. arXiv preprint arXiv:1809.08370, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "Alice Coucke, Alaa Saade, Adrien Ball, Theodore Bluche, Alexandre Caulier, David Leroy, Cl ´ ement ´", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 116, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, Mael Primet, and ¨", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "Joseph Dureau. Snips voice platform: an embedded spoken language understanding system for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "private-by-design voice interfaces. In Privacy in Machine Learning and Artificial Intelligence", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 577, + 234, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 234, + 591 + ], + "score": 1.0, + "content": "workshop, ICML2018, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 610 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 617, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 506, + 632 + ], + "score": 1.0, + "content": "the North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 640, + 266, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 640, + 266, + 653 + ], + "score": 1.0, + "content": "Short Papers), pp. 4171–4186, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 114, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 691, + 419, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 419, + 704 + ], + "score": 1.0, + "content": "Proceedings of Machine Learning Research, pp. 1183–1192. PMLR, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 721 + ], + "score": 1.0, + "content": "Junxian He, Jiatao Gu, Jiajun Shen, and Marc’Aurelio Ranzato. Revisiting self-training for neural", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 721, + 228, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 228, + 732 + ], + "score": 1.0, + "content": "sequence generation, 2019.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 505, + 154 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Philip Bachman, Ouais Alsharif, and Doina Precup. Learning with pseudo-ensembles. In Zoubin", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 505, + 124 + ], + "score": 1.0, + "content": "Ghahramani, Max Welling, Corinna Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.),", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 505, + 135 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems 27: Annual Conference on Neural Informa-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "tion Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, pp. 3365–3373,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 142, + 142, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 142, + 156 + ], + "score": 1.0, + "content": "2014.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 100, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 504, + 184 + ], + "lines": [ + { + "bbox": [ + 105, + 160, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 175 + ], + "score": 1.0, + "content": "Trapit Bansal, Rishikesh Jha, and Andrew McCallum. Learning to few-shot learn across diverse", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 289, + 184 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 289, + 184 + ], + "score": 1.0, + "content": "natural language classification tasks, 2020.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 160, + 506, + 184 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 191, + 505, + 246 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "Yoshua Bengio, Jer´ ome Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ˆ", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 202, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 116, + 202, + 505, + 214 + ], + "score": 1.0, + "content": "Andrea Pohoreckyj Danyluk, Leon Bottou, and Michael L. Littman (eds.), ´ Proceedings of the 26th", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 214, + 504, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 504, + 226 + ], + "score": 1.0, + "content": "Annual International Conference on Machine Learning, ICML 2009, Montreal, Quebec, Canada,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 223, + 504, + 237 + ], + "spans": [ + { + "bbox": [ + 115, + 223, + 504, + 237 + ], + "score": 1.0, + "content": "June 14-18, 2009, volume 382 of ACM International Conference Proceeding Series, pp. 41–48.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 235, + 300, + 246 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 300, + 246 + ], + "score": 1.0, + "content": "ACM, 2009. doi: 10.1145/1553374.1553380.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 106, + 190, + 505, + 246 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 505, + 266 + ], + "score": 1.0, + "content": "Haw-Shiuan Chang, Erik Learned-Miller, and Andrew McCallum. Active bias: Training more", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 263, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 263, + 505, + 278 + ], + "score": 1.0, + "content": "accurate neural networks by emphasizing high variance samples. In I. Guyon, U. V. Luxburg,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 275, + 506, + 288 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 506, + 288 + ], + "score": 1.0, + "content": "S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (eds.), Advances in Neural", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 287, + 453, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 287, + 453, + 299 + ], + "score": 1.0, + "content": "Information Processing Systems 30, pp. 1002–1012. Curran Associates, Inc., 2017a.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 252, + 506, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 304, + 505, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Haw-Shiuan Chang, Erik G. Learned-Miller, and Andrew McCallum. Active bias: Training more", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 316, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 115, + 316, + 505, + 328 + ], + "score": 1.0, + "content": "accurate neural networks by emphasizing high variance samples. In Isabelle Guyon, Ulrike von", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 326, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 115, + 326, + 505, + 339 + ], + "score": 1.0, + "content": "Luxburg, Samy Bengio, Hanna M. Wallach, Rob Fergus, S. V. N. Vishwanathan, and Roman", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 336, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 336, + 505, + 350 + ], + "score": 1.0, + "content": "Garnett (eds.), Advances in Neural Information Processing Systems 30: Annual Conference on", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 347, + 506, + 362 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 506, + 362 + ], + "score": 1.0, + "content": "Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 116, + 359, + 196, + 370 + ], + "spans": [ + { + "bbox": [ + 116, + 359, + 196, + 370 + ], + "score": 1.0, + "content": "1002–1012, 2017b.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 303, + 506, + 370 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 377, + 504, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 390 + ], + "score": 1.0, + "content": "Haw-Shiuan Chang, Shankar Vembu, Sunil Mohan, Rheeya Uppaal, and Andrew McCallum. Us-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 389, + 504, + 401 + ], + "spans": [ + { + "bbox": [ + 116, + 389, + 504, + 401 + ], + "score": 1.0, + "content": "ing error decay prediction to overcome practical issues of deep active learning for named entity", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 400, + 351, + 411 + ], + "spans": [ + { + "bbox": [ + 115, + 400, + 351, + 411 + ], + "score": 1.0, + "content": "recognition. Machine Learning, 109(9):1749–1778, 2020.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 377, + 506, + 411 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 417, + 474, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 475, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 475, + 431 + ], + "score": 1.0, + "content": "Olivier Chapelle, Bernhard Schlkopf, and Alexander Zien. Semi-supervised learning. 2010.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 106, + 417, + 475, + 431 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 435, + 506, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "Luoxin Chen, Weitong Ruan, Xinyue Liu, and Jianhua Lu. Seqvat: Virtual adversarial training for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "semi-supervised sequence labeling. In Proceedings of the 58th Annual Meeting of the Association", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 457, + 331, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 457, + 331, + 470 + ], + "score": 1.0, + "content": "for Computational Linguistics, pp. 8801–8811, 2020.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 435, + 505, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 504, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 489 + ], + "score": 1.0, + "content": "Mingda Chen, Qingming Tang, Karen Livescu, and Kevin Gimpel. Variational sequential labelers", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 487, + 399, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 487, + 399, + 499 + ], + "score": 1.0, + "content": "for semi-supervised learning. arXiv preprint arXiv:1906.09535, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 475, + 505, + 499 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 505, + 504, + 528 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 504, + 518 + ], + "score": 1.0, + "content": "Kevin Clark, Minh-Thang Luong, Christopher D Manning, and Quoc V Le. Semi-supervised se-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 516, + 453, + 529 + ], + "spans": [ + { + "bbox": [ + 115, + 516, + 453, + 529 + ], + "score": 1.0, + "content": "quence modeling with cross-view training. arXiv preprint arXiv:1809.08370, 2018.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 504, + 504, + 529 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 534, + 505, + 590 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "Alice Coucke, Alaa Saade, Adrien Ball, Theodore Bluche, Alexandre Caulier, David Leroy, Cl ´ ement ´", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 116, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Doumouro, Thibault Gisselbrecht, Francesco Caltagirone, Thibaut Lavril, Mael Primet, and ¨", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 555, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 505, + 569 + ], + "score": 1.0, + "content": "Joseph Dureau. Snips voice platform: an embedded spoken language understanding system for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 115, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "private-by-design voice interfaces. In Privacy in Machine Learning and Artificial Intelligence", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 577, + 234, + 591 + ], + "spans": [ + { + "bbox": [ + 115, + 577, + 234, + 591 + ], + "score": 1.0, + "content": "workshop, ICML2018, 2018.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 534, + 506, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 505, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 505, + 610 + ], + "score": 1.0, + "content": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 607, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 607, + 506, + 621 + ], + "score": 1.0, + "content": "bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 617, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 506, + 632 + ], + "score": 1.0, + "content": "the North American Chapter of the Association for Computational Linguistics: Human Language", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "Technologies, NAACL-HLT 2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 640, + 266, + 653 + ], + "spans": [ + { + "bbox": [ + 115, + 640, + 266, + 653 + ], + "score": 1.0, + "content": "Short Papers), pp. 4171–4186, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 595, + 506, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 505, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 671 + ], + "score": 1.0, + "content": "Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 114, + 669, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 114, + 669, + 506, + 682 + ], + "score": 1.0, + "content": "In Doina Precup and Yee Whye Teh (eds.), Proceedings of the 34th International Conference", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 681, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 681, + 506, + 693 + ], + "score": 1.0, + "content": "on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017, volume 70 of", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 115, + 691, + 419, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 419, + 704 + ], + "score": 1.0, + "content": "Proceedings of Machine Learning Research, pp. 1183–1192. PMLR, 2017.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 657, + 506, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 721 + ], + "score": 1.0, + "content": "Junxian He, Jiatao Gu, Jiajun Shen, and Marc’Aurelio Ranzato. Revisiting self-training for neural", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 115, + 721, + 228, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 228, + 732 + ], + "score": 1.0, + "content": "sequence generation, 2019.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 48.5, + "bbox_fs": [ + 106, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "H. J. Scudder III. Probability of error of some adaptive pattern-recognition machines. IEEE Trans.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 389, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 389, + 105 + ], + "score": 1.0, + "content": "Inf. Theory, 11(3):363–371, 1965. doi: 10.1109/TIT.1965.1053799.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 113, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 113, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 504, + 125 + ], + "score": 1.0, + "content": "Diederik P. Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 506, + 137 + ], + "score": 1.0, + "content": "supervised learning with deep generative models. In Zoubin Ghahramani, Max Welling, Corinna", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 506, + 148 + ], + "score": 1.0, + "content": "Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.), Advances in Neural Information", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 116, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 409, + 170 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 409, + 170 + ], + "score": 1.0, + "content": "December 8-13 2014, Montreal, Quebec, Canada, pp. 3581–3589, 2014.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 504, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. arXiv", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 188, + 254, + 200 + ], + "spans": [ + { + "bbox": [ + 115, + 188, + 254, + 200 + ], + "score": 1.0, + "content": "preprint arXiv:1703.04730, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 207, + 504, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "Ksenia Konyushkova, Raphael Sznitman, and Pascal Fua. Learning active learning from data. In", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 218, + 420, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 218, + 420, + 231 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 4225–4235, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 326, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 326, + 261 + ], + "score": 1.0, + "content": "adaptation. arXiv preprint arXiv:2002.11361, 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "M. P. Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "In J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta (eds.), Advances", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 290, + 490, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 290, + 490, + 304 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems 23, pp. 1189–1197. Curran Associates, Inc., 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 504, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 504, + 324 + ], + "score": 1.0, + "content": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In 5th Interna-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 116, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "tional Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 332, + 340, + 345 + ], + "spans": [ + { + "bbox": [ + 117, + 332, + 340, + 345 + ], + "score": 1.0, + "content": "Conference Track Proceedings. OpenReview.net, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "score": 1.0, + "content": "Xinzhe Li, Qianru Sun, Yaoyao Liu, Qin Zhou, Shibao Zheng, Tat-Seng Chua, and Bernt Schiele.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 116, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Learning to self-train for semi-supervised few-shot classification. In Advances in Neural Infor-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 374, + 440, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 440, + 386 + ], + "score": 1.0, + "content": "mation Processing Systems 32, pp. 10276–10286. Curran Associates, Inc., 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "Chen Liang, Yue Yu, Haoming Jiang, Siawpeng Er, Ruijia Wang, Tuo Zhao, and Chao Zhang. Bond:", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 114, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "Bert-assisted open-domain named entity recognition with distant supervision. In Proceedings of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 414, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 505, + 430 + ], + "score": 1.0, + "content": "the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 427, + 191, + 438 + ], + "spans": [ + { + "bbox": [ + 117, + 427, + 191, + 438 + ], + "score": 1.0, + "content": "1054–1064, 2020.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "J. Liu, Panupong Pasupat, D. Cyphers, and James R. Glass. Asgard: A portable architecture for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "multilingual dialogue systems. 2013 IEEE International Conference on Acoustics, Speech and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 468, + 283, + 480 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 283, + 480 + ], + "score": 1.0, + "content": "Signal Processing, pp. 8386–8390, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 114, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 511, + 281, + 522 + ], + "spans": [ + { + "bbox": [ + 116, + 511, + 281, + 522 + ], + "score": 1.0, + "content": "approach. CoRR, abs/1907.11692, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 529, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 543 + ], + "score": 1.0, + "content": "Fan Ma, Deyu Meng, Qi Xie, Zina Li, and Xuanyi Dong. Self-paced co-training. In Doina Precup", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "ing, volume 70 of Proceedings of Machine Learning Research, pp. 2275–2284, International", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 563, + 377, + 575 + ], + "spans": [ + { + "bbox": [ + 116, + 563, + 377, + 575 + ], + "score": 1.0, + "content": "Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 107, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 107, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "Yishu Miao and Phil Blunsom. Language as a latent variable: Discrete generative models for sen-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 505, + 605 + ], + "score": 1.0, + "content": "tence compression. In Jian Su, Xavier Carreras, and Kevin Duh (eds.), Proceedings of the 2016", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 603, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 116, + 603, + 505, + 618 + ], + "score": 1.0, + "content": "Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "USA, November 1-4, 2016, pp. 319–328. The Association for Computational Linguistics, 2016.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 626, + 233, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 233, + 638 + ], + "score": 1.0, + "content": "doi: 10.18653/v1/d16-1031.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "Scott Miller, Jethran Guinness, and Alex Zamanian. Name tagging with word clusters and discrim-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 657, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 657, + 506, + 670 + ], + "score": 1.0, + "content": "inative training. In Proceedings of the Human Language Technology Conference of the North", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 114, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 337–", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 678, + 162, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 162, + 691 + ], + "score": 1.0, + "content": "342, 2004.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "regularization method for supervised and semi-supervised learning. IEEE transactions on pattern", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 356, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 356, + 732 + ], + "score": 1.0, + "content": "analysis and machine intelligence, 41(8):1979–1993, 2018.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 503, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "H. J. Scudder III. Probability of error of some adaptive pattern-recognition machines. IEEE Trans.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 389, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 389, + 105 + ], + "score": 1.0, + "content": "Inf. Theory, 11(3):363–371, 1965. doi: 10.1109/TIT.1965.1053799.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 504, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 113, + 505, + 169 + ], + "lines": [ + { + "bbox": [ + 106, + 113, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 504, + 125 + ], + "score": 1.0, + "content": "Diederik P. Kingma, Shakir Mohamed, Danilo Jimenez Rezende, and Max Welling. Semi-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 506, + 137 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 506, + 137 + ], + "score": 1.0, + "content": "supervised learning with deep generative models. In Zoubin Ghahramani, Max Welling, Corinna", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 506, + 148 + ], + "score": 1.0, + "content": "Cortes, Neil D. Lawrence, and Kilian Q. Weinberger (eds.), Advances in Neural Information", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 116, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014,", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 156, + 409, + 170 + ], + "spans": [ + { + "bbox": [ + 115, + 156, + 409, + 170 + ], + "score": 1.0, + "content": "December 8-13 2014, Montreal, Quebec, Canada, pp. 3581–3589, 2014.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 113, + 506, + 170 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 176, + 504, + 199 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 189 + ], + "score": 1.0, + "content": "Pang Wei Koh and Percy Liang. Understanding black-box predictions via influence functions. arXiv", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 188, + 254, + 200 + ], + "spans": [ + { + "bbox": [ + 115, + 188, + 254, + 200 + ], + "score": 1.0, + "content": "preprint arXiv:1703.04730, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 176, + 505, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 207, + 504, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 505, + 220 + ], + "score": 1.0, + "content": "Ksenia Konyushkova, Raphael Sznitman, and Pascal Fua. Learning active learning from data. In", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 218, + 420, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 218, + 420, + 231 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems, pp. 4225–4235, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 207, + 505, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 504, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 326, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 326, + 261 + ], + "score": 1.0, + "content": "adaptation. arXiv preprint arXiv:2002.11361, 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 237, + 505, + 261 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 504, + 303 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "M. P. Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 506, + 292 + ], + "score": 1.0, + "content": "In J. D. Lafferty, C. K. I. Williams, J. Shawe-Taylor, R. S. Zemel, and A. Culotta (eds.), Advances", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 290, + 490, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 290, + 490, + 304 + ], + "score": 1.0, + "content": "in Neural Information Processing Systems 23, pp. 1189–1197. Curran Associates, Inc., 2010.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 268, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 504, + 344 + ], + "lines": [ + { + "bbox": [ + 105, + 310, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 504, + 324 + ], + "score": 1.0, + "content": "Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In 5th Interna-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 322, + 505, + 334 + ], + "spans": [ + { + "bbox": [ + 116, + 322, + 505, + 334 + ], + "score": 1.0, + "content": "tional Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 117, + 332, + 340, + 345 + ], + "spans": [ + { + "bbox": [ + 117, + 332, + 340, + 345 + ], + "score": 1.0, + "content": "Conference Track Proceedings. OpenReview.net, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 310, + 505, + 345 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 365 + ], + "score": 1.0, + "content": "Xinzhe Li, Qianru Sun, Yaoyao Liu, Qin Zhou, Shibao Zheng, Tat-Seng Chua, and Bernt Schiele.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 116, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Learning to self-train for semi-supervised few-shot classification. In Advances in Neural Infor-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 374, + 440, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 440, + 386 + ], + "score": 1.0, + "content": "mation Processing Systems 32, pp. 10276–10286. Curran Associates, Inc., 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 351, + 505, + 386 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 393, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "Chen Liang, Yue Yu, Haoming Jiang, Siawpeng Er, Ruijia Wang, Tuo Zhao, and Chao Zhang. Bond:", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 404, + 506, + 418 + ], + "spans": [ + { + "bbox": [ + 114, + 404, + 506, + 418 + ], + "score": 1.0, + "content": "Bert-assisted open-domain named entity recognition with distant supervision. In Proceedings of", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 414, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 505, + 430 + ], + "score": 1.0, + "content": "the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 117, + 427, + 191, + 438 + ], + "spans": [ + { + "bbox": [ + 117, + 427, + 191, + 438 + ], + "score": 1.0, + "content": "1054–1064, 2020.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 394, + 506, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "J. Liu, Panupong Pasupat, D. Cyphers, and James R. Glass. Asgard: A portable architecture for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "multilingual dialogue systems. 2013 IEEE International Conference on Acoustics, Speech and", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 468, + 283, + 480 + ], + "spans": [ + { + "bbox": [ + 116, + 468, + 283, + 480 + ], + "score": 1.0, + "content": "Signal Processing, pp. 8386–8390, 2013.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 446, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 487, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 505, + 501 + ], + "score": 1.0, + "content": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 114, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized BERT pretraining", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 511, + 281, + 522 + ], + "spans": [ + { + "bbox": [ + 116, + 511, + 281, + 522 + ], + "score": 1.0, + "content": "approach. CoRR, abs/1907.11692, 2019.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 487, + 505, + 522 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 574 + ], + "lines": [ + { + "bbox": [ + 105, + 529, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 543 + ], + "score": 1.0, + "content": "Fan Ma, Deyu Meng, Qi Xie, Zina Li, and Xuanyi Dong. Self-paced co-training. In Doina Precup", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "and Yee Whye Teh (eds.), Proceedings of the 34th International Conference on Machine Learn-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 551, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 505, + 564 + ], + "score": 1.0, + "content": "ing, volume 70 of Proceedings of Machine Learning Research, pp. 2275–2284, International", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 563, + 377, + 575 + ], + "spans": [ + { + "bbox": [ + 116, + 563, + 377, + 575 + ], + "score": 1.0, + "content": "Convention Centre, Sydney, Australia, 06–11 Aug 2017. PMLR.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 529, + 505, + 575 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 107, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 107, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "Yishu Miao and Phil Blunsom. Language as a latent variable: Discrete generative models for sen-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 115, + 594, + 505, + 605 + ], + "score": 1.0, + "content": "tence compression. In Jian Su, Xavier Carreras, and Kevin Duh (eds.), Proceedings of the 2016", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 603, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 116, + 603, + 505, + 618 + ], + "score": 1.0, + "content": "Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 116, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "USA, November 1-4, 2016, pp. 319–328. The Association for Computational Linguistics, 2016.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 626, + 233, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 626, + 233, + 638 + ], + "score": 1.0, + "content": "doi: 10.18653/v1/d16-1031.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 107, + 583, + 505, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 505, + 690 + ], + "lines": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 505, + 658 + ], + "score": 1.0, + "content": "Scott Miller, Jethran Guinness, and Alex Zamanian. Name tagging with word clusters and discrim-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 657, + 506, + 670 + ], + "spans": [ + { + "bbox": [ + 115, + 657, + 506, + 670 + ], + "score": 1.0, + "content": "inative training. In Proceedings of the Human Language Technology Conference of the North", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 114, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 114, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "American Chapter of the Association for Computational Linguistics: HLT-NAACL 2004, pp. 337–", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 678, + 162, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 162, + 691 + ], + "score": 1.0, + "content": "342, 2004.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 645, + 506, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "regularization method for supervised and semi-supervised learning. IEEE transactions on pattern", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 116, + 720, + 356, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 720, + 356, + 732 + ], + "score": 1.0, + "content": "analysis and machine intelligence, 41(8):1979–1993, 2018.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. Cross-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1946–1958, Van-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 115, + 401, + 128 + ], + "spans": [ + { + "bbox": [ + 116, + 115, + 401, + 128 + ], + "score": 1.0, + "content": "couver, Canada, July 2017. Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 133, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Emmeleia Panagiota Mastoropoulou. Enhancing deep active learning using selective self-training", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "score": 1.0, + "content": "for image classification. Master’s thesis, KTH, School of Electrical Engineering and Computer", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 156, + 211, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 211, + 168 + ], + "score": 1.0, + "content": "Science (EECS), 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 504, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "Matthew E Peters, Waleed Ammar, Chandra Bhagavatula, and Russell Power. Semi-supervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 496, + 198 + ], + "score": 1.0, + "content": "sequence tagging with bidirectional language models. arXiv preprint arXiv:1705.00108, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 486, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 486, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 486, + 218 + ], + "score": 1.0, + "content": "Slav Petrov and Ryan McDonald. Overview of the 2012 shared task on parsing the web. 2012.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 105, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 237 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 320, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 320, + 245 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 252, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "score": 1.0, + "content": "Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 262, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 506, + 277 + ], + "score": 1.0, + "content": "supervised learning with ladder networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "Masashi Sugiyama, and Roman Garnett (eds.), Advances in Neural Information Processing Sys-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "tems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 296, + 345, + 309 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 345, + 309 + ], + "score": 1.0, + "content": "2015, Montreal, Quebec, Canada, pp. 3546–3554, 2015.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 504, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 326, + 498, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 498, + 339 + ], + "score": 1.0, + "content": "robust deep learning. In International Conference on Machine Learning, pp. 4334–4343, 2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Sebastian Ruder and Barbara Plank. Strong baselines for neural semi-supervised learning under", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 116, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "domain shift. In Proceedings of the 56th Annual Meeting of the Association for Computational", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 366, + 358, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 358, + 379 + ], + "score": 1.0, + "content": "Linguistics (Volume 1: Long Papers), pp. 1044–1054, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. Introduction to the conll-2003 shared task: Language-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "independent named entity recognition. In Proceedings of the Seventh Conference on Natural", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 407, + 311, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 311, + 419 + ], + "score": 1.0, + "content": "Language Learning at HLT-NAACL 2003, 2003.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 116, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Com-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "putational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 459, + 431, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 459, + 431, + 470 + ], + "score": 1.0, + "content": "The Association for Computer Linguistics, 2016. doi: 10.18653/v1/p16-1009.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 116, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "with online hard example mining. In Proceedings of the IEEE conference on computer vision and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 500, + 277, + 511 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 277, + 511 + ], + "score": 1.0, + "content": "pattern recognition, pp. 761–769, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 517, + 504, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 504, + 531 + ], + "score": 1.0, + "content": "Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "net: Learning an explicit mapping for sample weighting. In Advances in Neural Information", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 540, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 289, + 552 + ], + "score": 1.0, + "content": "Processing Systems, pp. 1919–1930, 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 104, + 558, + 504, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "score": 1.0, + "content": "Q. Sun, Y. Liu, T. Chua, and B. Schiele. Meta-transfer learning for few-shot learning. In 2019", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged con-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "sistency targets improve semi-supervised deep learning results. In 5th International Conference", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 116, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 621, + 266, + 633 + ], + "spans": [ + { + "bbox": [ + 116, + 621, + 266, + 633 + ], + "score": 1.0, + "content": "Proceedings. OpenReview.net, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 639, + 503, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 653 + ], + "score": 1.0, + "content": "Erik F Tjong, Kim Sang, and Jorn Veenstra. Representing text chunks. In Ninth Conference of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 650, + 417, + 663 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 417, + 663 + ], + "score": 1.0, + "content": "European Chapter of the Association for Computational Linguistics, 1999.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 503, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised data", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 297, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 297, + 692 + ], + "score": 1.0, + "content": "augmentation for consistency training, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 117, + 720, + 325, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 720, + 325, + 732 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition (CVPR), June 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Xiaoman Pan, Boliang Zhang, Jonathan May, Joel Nothman, Kevin Knight, and Heng Ji. Cross-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 506, + 107 + ], + "score": 1.0, + "content": "lingual name tagging and linking for 282 languages. In Proceedings of the 55th Annual Meeting", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1946–1958, Van-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 115, + 401, + 128 + ], + "spans": [ + { + "bbox": [ + 116, + 115, + 401, + 128 + ], + "score": 1.0, + "content": "couver, Canada, July 2017. Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 506, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 133, + 505, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 105, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Emmeleia Panagiota Mastoropoulou. Enhancing deep active learning using selective self-training", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 115, + 144, + 505, + 158 + ], + "score": 1.0, + "content": "for image classification. Master’s thesis, KTH, School of Electrical Engineering and Computer", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 156, + 211, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 156, + 211, + 168 + ], + "score": 1.0, + "content": "Science (EECS), 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 133, + 505, + 168 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 174, + 504, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 505, + 187 + ], + "score": 1.0, + "content": "Matthew E Peters, Waleed Ammar, Chandra Bhagavatula, and Russell Power. Semi-supervised", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 186, + 496, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 496, + 198 + ], + "score": 1.0, + "content": "sequence tagging with bidirectional language models. arXiv preprint arXiv:1705.00108, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 174, + 505, + 198 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 486, + 216 + ], + "lines": [ + { + "bbox": [ + 105, + 202, + 486, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 202, + 486, + 218 + ], + "score": 1.0, + "content": "Slav Petrov and Ryan McDonald. Overview of the 2012 shared task on parsing the web. 2012.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 202, + 486, + 218 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 222, + 504, + 245 + ], + "lines": [ + { + "bbox": [ + 105, + 220, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 506, + 237 + ], + "score": 1.0, + "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 234, + 320, + 245 + ], + "spans": [ + { + "bbox": [ + 115, + 234, + 320, + 245 + ], + "score": 1.0, + "content": "models are unsupervised multitask learners. 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 220, + 506, + 245 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 252, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 504, + 264 + ], + "score": 1.0, + "content": "Antti Rasmus, Mathias Berglund, Mikko Honkala, Harri Valpola, and Tapani Raiko. Semi-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 262, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 115, + 262, + 506, + 277 + ], + "score": 1.0, + "content": "supervised learning with ladder networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 273, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 505, + 288 + ], + "score": 1.0, + "content": "Masashi Sugiyama, and Roman Garnett (eds.), Advances in Neural Information Processing Sys-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "tems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 296, + 345, + 309 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 345, + 309 + ], + "score": 1.0, + "content": "2015, Montreal, Quebec, Canada, pp. 3546–3554, 2015.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 252, + 506, + 309 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 314, + 504, + 338 + ], + "lines": [ + { + "bbox": [ + 106, + 315, + 505, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 505, + 328 + ], + "score": 1.0, + "content": "Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 326, + 498, + 339 + ], + "spans": [ + { + "bbox": [ + 116, + 326, + 498, + 339 + ], + "score": 1.0, + "content": "robust deep learning. In International Conference on Machine Learning, pp. 4334–4343, 2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 315, + 505, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 344, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 505, + 357 + ], + "score": 1.0, + "content": "Sebastian Ruder and Barbara Plank. Strong baselines for neural semi-supervised learning under", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 116, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "domain shift. In Proceedings of the 56th Annual Meeting of the Association for Computational", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 366, + 358, + 379 + ], + "spans": [ + { + "bbox": [ + 116, + 366, + 358, + 379 + ], + "score": 1.0, + "content": "Linguistics (Volume 1: Long Papers), pp. 1044–1054, 2018.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 106, + 344, + 505, + 379 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 385, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 505, + 399 + ], + "score": 1.0, + "content": "Erik F. Tjong Kim Sang and Fien De Meulder. Introduction to the conll-2003 shared task: Language-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 396, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 396, + 505, + 408 + ], + "score": 1.0, + "content": "independent named entity recognition. In Proceedings of the Seventh Conference on Natural", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 407, + 311, + 419 + ], + "spans": [ + { + "bbox": [ + 116, + 407, + 311, + 419 + ], + "score": 1.0, + "content": "Language Learning at HLT-NAACL 2003, 2003.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 384, + 505, + 419 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 116, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "with monolingual data. In Proceedings of the 54th Annual Meeting of the Association for Com-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 115, + 448, + 505, + 461 + ], + "score": 1.0, + "content": "putational Linguistics, ACL 2016, August 7-12, 2016, Berlin, Germany, Volume 1: Long Papers.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 459, + 431, + 470 + ], + "spans": [ + { + "bbox": [ + 116, + 459, + 431, + 470 + ], + "score": 1.0, + "content": "The Association for Computer Linguistics, 2016. doi: 10.18653/v1/p16-1009.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 425, + 505, + 470 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 476, + 504, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 490 + ], + "score": 1.0, + "content": "Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 116, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "with online hard example mining. In Proceedings of the IEEE conference on computer vision and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 500, + 277, + 511 + ], + "spans": [ + { + "bbox": [ + 115, + 500, + 277, + 511 + ], + "score": 1.0, + "content": "pattern recognition, pp. 761–769, 2016.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 476, + 505, + 511 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 517, + 504, + 551 + ], + "lines": [ + { + "bbox": [ + 105, + 517, + 504, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 504, + 531 + ], + "score": 1.0, + "content": "Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "net: Learning an explicit mapping for sample weighting. In Advances in Neural Information", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 540, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 116, + 540, + 289, + 552 + ], + "score": 1.0, + "content": "Processing Systems, pp. 1919–1930, 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 517, + 505, + 552 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 558, + 504, + 581 + ], + "lines": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 506, + 571 + ], + "score": 1.0, + "content": "Q. Sun, Y. Liu, T. Chua, and B. Schiele. Meta-transfer learning for few-shot learning. In 2019", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 115, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 403–412, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 558, + 506, + 582 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 632 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged con-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 599, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 115, + 599, + 505, + 611 + ], + "score": 1.0, + "content": "sistency targets improve semi-supervised deep learning results. In 5th International Conference", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 610, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 116, + 610, + 506, + 622 + ], + "score": 1.0, + "content": "on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Workshop Track", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 116, + 621, + 266, + 633 + ], + "spans": [ + { + "bbox": [ + 116, + 621, + 266, + 633 + ], + "score": 1.0, + "content": "Proceedings. OpenReview.net, 2017.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 587, + 506, + 633 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 639, + 503, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 653 + ], + "score": 1.0, + "content": "Erik F Tjong, Kim Sang, and Jorn Veenstra. Representing text chunks. In Ninth Conference of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 650, + 417, + 663 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 417, + 663 + ], + "score": 1.0, + "content": "European Chapter of the Association for Computational Linguistics, 1999.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 637, + 505, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 668, + 503, + 692 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 682 + ], + "score": 1.0, + "content": "Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V. Le. Unsupervised data", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 680, + 297, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 297, + 692 + ], + "score": 1.0, + "content": "augmentation for consistency training, 2019.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 668, + 505, + 692 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V. Le. Self-training with noisy student", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "improves imagenet classification. In Proceedings of the IEEE/CVF Conference on Computer", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 117, + 720, + 325, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 720, + 325, + 732 + ], + "score": 1.0, + "content": "Vision and Pattern Recognition (CVPR), June 2020.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Pengcheng Yin, Chunting Zhou, Junxian He, and Graham Neubig. Structvae: Tree-structured la-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tent variable models for semi-supervised semantic parsing. In Iryna Gurevych and Yusuke Miyao", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "(eds.), Proceedings of the 56th Annual Meeting of the Association for Computational Linguis-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "tics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pp. 754–765.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 433, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 433, + 138 + ], + "score": 1.0, + "content": "Association for Computational Linguistics, 2018. doi: 10.18653/v1/P18-1070.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 155, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 114, + 155, + 505, + 170 + ], + "score": 1.0, + "content": "deep learning requires rethinking generalization. In 5th International Conference on Learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 178, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 178, + 212, + 190 + ], + "score": 1.0, + "content": "OpenReview.net, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "Chunting Zhou and Graham Neubig. Multi-space variational encoder-decoders for semi-supervised", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 208, + 404, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 404, + 221 + ], + "score": 1.0, + "content": "labeled sequence transduction. arXiv preprint arXiv:1704.01691, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 504, + 240 + ], + "score": 1.0, + "content": "Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Rethinking pre-training and self-training. Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 248, + 157, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 157, + 261 + ], + "score": 1.0, + "content": "33, 2020.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "Pengcheng Yin, Chunting Zhou, Junxian He, and Graham Neubig. Structvae: Tree-structured la-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "tent variable models for semi-supervised semantic parsing. In Iryna Gurevych and Yusuke Miyao", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 505, + 118 + ], + "score": 1.0, + "content": "(eds.), Proceedings of the 56th Annual Meeting of the Association for Computational Linguis-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 115, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "tics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pp. 754–765.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 126, + 433, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 126, + 433, + 138 + ], + "score": 1.0, + "content": "Association for Computational Linguistics, 2018. doi: 10.18653/v1/P18-1070.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 83, + 505, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 144, + 505, + 189 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 505, + 158 + ], + "score": 1.0, + "content": "Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 155, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 114, + 155, + 505, + 170 + ], + "score": 1.0, + "content": "deep learning requires rethinking generalization. In 5th International Conference on Learning", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 167, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 505, + 180 + ], + "score": 1.0, + "content": "Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 178, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 116, + 178, + 212, + 190 + ], + "score": 1.0, + "content": "OpenReview.net, 2017.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 145, + 505, + 190 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 219 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "Chunting Zhou and Graham Neubig. Multi-space variational encoder-decoders for semi-supervised", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 208, + 404, + 221 + ], + "spans": [ + { + "bbox": [ + 115, + 208, + 404, + 221 + ], + "score": 1.0, + "content": "labeled sequence transduction. arXiv preprint arXiv:1704.01691, 2017.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 196, + 505, + 221 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 260 + ], + "lines": [ + { + "bbox": [ + 105, + 226, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 504, + 240 + ], + "score": 1.0, + "content": "Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 115, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Rethinking pre-training and self-training. Advances in Neural Information Processing Systems,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 248, + 157, + 261 + ], + "spans": [ + { + "bbox": [ + 115, + 248, + 157, + 261 + ], + "score": 1.0, + "content": "33, 2020.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 226, + 505, + 261 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 182, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 379, + 118 + ], + "lines": [ + { + "bbox": [ + 107, + 106, + 379, + 118 + ], + "spans": [ + { + "bbox": [ + 107, + 106, + 379, + 118 + ], + "score": 1.0, + "content": "A.1 EXPLORATIONS ON UNLABELED DATA AND MINI-BATCH S", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "Variation in model performance with unlabeled data. Table 12 shows the improvement in model", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 475, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 475, + 150 + ], + "score": 1.0, + "content": "performance as we inject more unlabeled data with diminishing returns after a certain point.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 412, + 167 + ], + "score": 1.0, + "content": "Variation in model performance with mini-batch S. We set the value of", + "type": "text" + }, + { + "bbox": [ + 412, + 155, + 420, + 165 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 153, + 469, + 167 + ], + "score": 1.0, + "content": "in Eq. 8 to", + "type": "text" + }, + { + "bbox": [ + 469, + 154, + 504, + 167 + ], + "score": 0.92, + "content": "\\{ 1 , 3 , 5 \\}", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "respectively to explore its impact on the re-weighting mechanism. From Figure 3 we observe that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 309, + 189 + ], + "score": 1.0, + "content": "the model is not super sensitive to hyper-parameter", + "type": "text" + }, + { + "bbox": [ + 309, + 177, + 317, + 186 + ], + "score": 0.73, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "but can achieve a better estimate of the weights", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 187, + 354, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 354, + 200 + ], + "score": 1.0, + "content": "of the pseudo-labeled data with increasing mini-batch values.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5 + }, + { + "type": "table", + "bbox": [ + 120, + 229, + 290, + 288 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 120, + 229, + 290, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 229, + 290, + 288 + ], + "spans": [ + { + "bbox": [ + 120, + 229, + 290, + 288 + ], + "score": 0.974, + "html": "
Ratio of Unlabeled DataDatasets
SNIPSCoNLL03
5%84.4772.92
25%87.1076.46
75%87.5076.56
", + "type": "table", + "image_path": "230f8633595e21b91bcb56d6a4b5450e066fa0bfd2a45bfdbb9e8e404f33f8c9.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 120, + 229, + 290, + 243.75 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 120, + 243.75, + 290, + 258.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 120, + 258.5, + 290, + 273.25 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 120, + 273.25, + 290, + 288.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 109, + 297, + 302, + 318 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 296, + 302, + 309 + ], + "spans": [ + { + "bbox": [ + 109, + 296, + 302, + 309 + ], + "score": 1.0, + "content": "Table 7: Varying proportion of unlabeled data", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 306, + 251, + 320 + ], + "spans": [ + { + "bbox": [ + 109, + 306, + 251, + 320 + ], + "score": 1.0, + "content": "for MetaST with 10 labels per slot.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 336, + 213, + 456, + 308 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 336, + 213, + 456, + 308 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 336, + 213, + 456, + 308 + ], + "spans": [ + { + "bbox": [ + 336, + 213, + 456, + 308 + ], + "score": 0.949, + "type": "image", + "image_path": "4acf5be292e2775588f2a99e982a0a9dfdc3fa3dd074d857771804a5f6d9f1ce.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 336, + 213, + 456, + 260.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 336, + 260.5, + 456, + 308.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 331, + 315, + 471, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 331, + 316, + 472, + 327 + ], + "spans": [ + { + "bbox": [ + 331, + 316, + 405, + 327 + ], + "score": 1.0, + "content": "Figure 3: Varying", + "type": "text" + }, + { + "bbox": [ + 405, + 316, + 413, + 326 + ], + "score": 0.69, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 316, + 472, + 327 + ], + "score": 1.0, + "content": "mini-batch la-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 325, + 444, + 340 + ], + "spans": [ + { + "bbox": [ + 330, + 325, + 444, + 340 + ], + "score": 1.0, + "content": "beled data for re-weighting.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "index": 15.0 + }, + { + "type": "title", + "bbox": [ + 107, + 358, + 374, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "score": 1.0, + "content": "A.2 ANALYSIS OF RE-WEIGHTING ON SNIPS AND CONLL03", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 393 + ], + "score": 1.0, + "content": "Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "mechanism, we illustrate token re-weighting of MetaST on CoNLL03 and SNIPS datasets with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 131, + 412 + ], + "score": 0.83, + "content": "{ \\mathrm { K } } { = } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "shot at step 100 in Fig. 4. Besides the observation in the experimental section, we observe that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 411, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 426 + ], + "score": 1.0, + "content": "many difficult and correct pseudo-labeled samples (low teacher confidence) are selected according", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 423, + 150, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 150, + 437 + ], + "score": 1.0, + "content": "to Fig. 4a.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 116, + 456, + 467, + 577 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 456, + 467, + 577 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 116, + 456, + 467, + 577 + ], + "spans": [ + { + "bbox": [ + 116, + 456, + 467, + 577 + ], + "score": 0.971, + "type": "image", + "image_path": "a1ce0ac42b92d98b8512da697f6121d4c41ab38cbc3fd9329203574c7580c82c.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 116, + 456, + 467, + 496.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 116, + 496.3333333333333, + 467, + 536.6666666666666 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 116, + 536.6666666666666, + 467, + 577.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 586, + 485, + 599 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 586, + 486, + 600 + ], + "spans": [ + { + "bbox": [ + 124, + 586, + 486, + 600 + ], + "score": 1.0, + "content": "Figure 4: Visualization of MetaST re-weighting examples on SNIPS and CoNLL03 (EN).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "index": 26.0 + }, + { + "type": "title", + "bbox": [ + 108, + 619, + 173, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 618, + 174, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 174, + 632 + ], + "score": 1.0, + "content": "A.3 K-SHOTS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "Effect of varying the number of few-shots K. We show the performance changes with respect", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 245, + 663 + ], + "score": 1.0, + "content": "to varying number of few-shots K", + "type": "text" + }, + { + "bbox": [ + 245, + 650, + 311, + 662 + ], + "score": 0.81, + "content": "\\{ 5 , 1 0 , 2 0 , 1 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "on Wikiann (en), MIT movie, MIT Restaurant,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "score": 1.0, + "content": "CoNLL2003 (En), Multilingual CoNLL and Multilingual Wikiann in Table 9-13. Since the number", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "of labeled examples for some slots in Email dataset is around 10, we only show 5 and 10 shots for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 683, + 208, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 208, + 695 + ], + "score": 1.0, + "content": "Email dataset in Table 8.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 182, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 106, + 379, + 118 + ], + "lines": [ + { + "bbox": [ + 107, + 106, + 379, + 118 + ], + "spans": [ + { + "bbox": [ + 107, + 106, + 379, + 118 + ], + "score": 1.0, + "content": "A.1 EXPLORATIONS ON UNLABELED DATA AND MINI-BATCH S", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 126, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "Variation in model performance with unlabeled data. Table 12 shows the improvement in model", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 475, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 475, + 150 + ], + "score": 1.0, + "content": "performance as we inject more unlabeled data with diminishing returns after a certain point.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 126, + 505, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 153, + 504, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 412, + 167 + ], + "score": 1.0, + "content": "Variation in model performance with mini-batch S. We set the value of", + "type": "text" + }, + { + "bbox": [ + 412, + 155, + 420, + 165 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 153, + 469, + 167 + ], + "score": 1.0, + "content": "in Eq. 8 to", + "type": "text" + }, + { + "bbox": [ + 469, + 154, + 504, + 167 + ], + "score": 0.92, + "content": "\\{ 1 , 3 , 5 \\}", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "respectively to explore its impact on the re-weighting mechanism. From Figure 3 we observe that", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 309, + 189 + ], + "score": 1.0, + "content": "the model is not super sensitive to hyper-parameter", + "type": "text" + }, + { + "bbox": [ + 309, + 177, + 317, + 186 + ], + "score": 0.73, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "but can achieve a better estimate of the weights", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 187, + 354, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 354, + 200 + ], + "score": 1.0, + "content": "of the pseudo-labeled data with increasing mini-batch values.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 153, + 506, + 200 + ] + }, + { + "type": "table", + "bbox": [ + 120, + 229, + 290, + 288 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 120, + 229, + 290, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 120, + 229, + 290, + 288 + ], + "spans": [ + { + "bbox": [ + 120, + 229, + 290, + 288 + ], + "score": 0.974, + "html": "
Ratio of Unlabeled DataDatasets
SNIPSCoNLL03
5%84.4772.92
25%87.1076.46
75%87.5076.56
", + "type": "table", + "image_path": "230f8633595e21b91bcb56d6a4b5450e066fa0bfd2a45bfdbb9e8e404f33f8c9.jpg" + } + ] + } + ], + "index": 9.5, + "virtual_lines": [ + { + "bbox": [ + 120, + 229, + 290, + 243.75 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 120, + 243.75, + 290, + 258.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 120, + 258.5, + 290, + 273.25 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 120, + 273.25, + 290, + 288.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 109, + 297, + 302, + 318 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 296, + 302, + 309 + ], + "spans": [ + { + "bbox": [ + 109, + 296, + 302, + 309 + ], + "score": 1.0, + "content": "Table 7: Varying proportion of unlabeled data", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 109, + 306, + 251, + 320 + ], + "spans": [ + { + "bbox": [ + 109, + 306, + 251, + 320 + ], + "score": 1.0, + "content": "for MetaST with 10 labels per slot.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 336, + 213, + 456, + 308 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 336, + 213, + 456, + 308 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 336, + 213, + 456, + 308 + ], + "spans": [ + { + "bbox": [ + 336, + 213, + 456, + 308 + ], + "score": 0.949, + "type": "image", + "image_path": "4acf5be292e2775588f2a99e982a0a9dfdc3fa3dd074d857771804a5f6d9f1ce.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 336, + 213, + 456, + 260.5 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 336, + 260.5, + 456, + 308.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 331, + 315, + 471, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 331, + 316, + 472, + 327 + ], + "spans": [ + { + "bbox": [ + 331, + 316, + 405, + 327 + ], + "score": 1.0, + "content": "Figure 3: Varying", + "type": "text" + }, + { + "bbox": [ + 405, + 316, + 413, + 326 + ], + "score": 0.69, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 316, + 472, + 327 + ], + "score": 1.0, + "content": "mini-batch la-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 330, + 325, + 444, + 340 + ], + "spans": [ + { + "bbox": [ + 330, + 325, + 444, + 340 + ], + "score": 1.0, + "content": "beled data for re-weighting.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + } + ], + "index": 15.0 + }, + { + "type": "title", + "bbox": [ + 107, + 358, + 374, + 370 + ], + "lines": [ + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 374, + 371 + ], + "score": 1.0, + "content": "A.2 ANALYSIS OF RE-WEIGHTING ON SNIPS AND CONLL03", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 505, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 378, + 505, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 393 + ], + "score": 1.0, + "content": "Analysis of pseudo-labeled data re-weighting. To visually explore the adaptive re-weighting", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "mechanism, we illustrate token re-weighting of MetaST on CoNLL03 and SNIPS datasets with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 401, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 131, + 412 + ], + "score": 0.83, + "content": "{ \\mathrm { K } } { = } 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 401, + 505, + 414 + ], + "score": 1.0, + "content": "shot at step 100 in Fig. 4. Besides the observation in the experimental section, we observe that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 411, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 505, + 426 + ], + "score": 1.0, + "content": "many difficult and correct pseudo-labeled samples (low teacher confidence) are selected according", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 423, + 150, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 150, + 437 + ], + "score": 1.0, + "content": "to Fig. 4a.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 378, + 505, + 437 + ] + }, + { + "type": "image", + "bbox": [ + 116, + 456, + 467, + 577 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 456, + 467, + 577 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 116, + 456, + 467, + 577 + ], + "spans": [ + { + "bbox": [ + 116, + 456, + 467, + 577 + ], + "score": 0.971, + "type": "image", + "image_path": "a1ce0ac42b92d98b8512da697f6121d4c41ab38cbc3fd9329203574c7580c82c.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 116, + 456, + 467, + 496.3333333333333 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 116, + 496.3333333333333, + 467, + 536.6666666666666 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 116, + 536.6666666666666, + 467, + 577.0 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 125, + 586, + 485, + 599 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 586, + 486, + 600 + ], + "spans": [ + { + "bbox": [ + 124, + 586, + 486, + 600 + ], + "score": 1.0, + "content": "Figure 4: Visualization of MetaST re-weighting examples on SNIPS and CoNLL03 (EN).", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + } + ], + "index": 26.0 + }, + { + "type": "title", + "bbox": [ + 108, + 619, + 173, + 630 + ], + "lines": [ + { + "bbox": [ + 106, + 618, + 174, + 632 + ], + "spans": [ + { + "bbox": [ + 106, + 618, + 174, + 632 + ], + "score": 1.0, + "content": "A.3 K-SHOTS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 639, + 505, + 695 + ], + "lines": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 505, + 651 + ], + "score": 1.0, + "content": "Effect of varying the number of few-shots K. We show the performance changes with respect", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 245, + 663 + ], + "score": 1.0, + "content": "to varying number of few-shots K", + "type": "text" + }, + { + "bbox": [ + 245, + 650, + 311, + 662 + ], + "score": 0.81, + "content": "\\{ 5 , 1 0 , 2 0 , 1 0 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "on Wikiann (en), MIT movie, MIT Restaurant,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "score": 1.0, + "content": "CoNLL2003 (En), Multilingual CoNLL and Multilingual Wikiann in Table 9-13. Since the number", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 505, + 685 + ], + "score": 1.0, + "content": "of labeled examples for some slots in Email dataset is around 10, we only show 5 and 10 shots for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 683, + 208, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 683, + 208, + 695 + ], + "score": 1.0, + "content": "Email dataset in Table 8.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 640, + 505, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 227, + 103, + 384, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 258, + 83, + 353, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 257, + 82, + 354, + 95 + ], + "spans": [ + { + "bbox": [ + 257, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "Table 8: Email Dataset.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 227, + 103, + 384, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 227, + 103, + 384, + 264 + ], + "spans": [ + { + "bbox": [ + 227, + 103, + 384, + 264 + ], + "score": 0.972, + "html": "
MethodShots
510
Full-supervision
BERT0.9444
Few-shot Supervision
BERT0.82110.8785
Few-shot Supervision + unlabeled data
CVT67.4478.24
SeqVAT64.6772.65
Mean Teacher84.1089.53
VAT83.2489.71
Classic ST86.8890.70
BOND84.9289.75
MetaST89.2192.18
", + "type": "table", + "image_path": "a0a70b768721fcc3b76448f14f149d4e4414c737bad75fd505c92a646c5b05ab.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 227, + 103, + 384, + 117.63636363636364 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 227, + 117.63636363636364, + 384, + 132.27272727272728 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 227, + 132.27272727272728, + 384, + 146.9090909090909 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 227, + 146.9090909090909, + 384, + 161.54545454545453 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 227, + 161.54545454545453, + 384, + 176.18181818181816 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 227, + 176.18181818181816, + 384, + 190.81818181818178 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 227, + 190.81818181818178, + 384, + 205.4545454545454 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 227, + 205.4545454545454, + 384, + 220.09090909090904 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 227, + 220.09090909090904, + 384, + 234.72727272727266 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 227, + 234.72727272727266, + 384, + 249.3636363636363 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 227, + 249.3636363636363, + 384, + 263.99999999999994 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 124, + 275, + 295, + 402 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 275, + 295, + 402 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 275, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 295, + 402 + ], + "score": 0.967, + "html": "
MethodShots (3 Slot Types)
51020100
Full-supervision
BERT84.04
Few-shot Supervision
BERT37.0145.6154.5367.87
Few-shot Supervision +unlabeled data
CVT16.0527.8946.4266.36
SeqVAT21.1135.1642.2662.37
Mean Teacher30.9241.4350.6167.16
VAT24.7238.8150.1566.31
Classic ST32.7246.1554.4168.64
BOND34.2248.7352.4568.89
MetaST55.0456.6160.3873.20
", + "type": "table", + "image_path": "95fcfbdbbf909669be074fbba1a4c6a7b11a739db36ef810a02b75c6dcd33573.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 124, + 275, + 295, + 289.1111111111111 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 124, + 289.1111111111111, + 295, + 303.2222222222222 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 124, + 303.2222222222222, + 295, + 317.33333333333326 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 124, + 317.33333333333326, + 295, + 331.44444444444434 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 124, + 331.44444444444434, + 295, + 345.55555555555543 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 124, + 345.55555555555543, + 295, + 359.6666666666665 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 124, + 359.6666666666665, + 295, + 373.7777777777776 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 124, + 373.7777777777776, + 295, + 387.8888888888887 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 124, + 387.8888888888887, + 295, + 401.9999999999998 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 146, + 411, + 272, + 422 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 410, + 272, + 424 + ], + "spans": [ + { + "bbox": [ + 145, + 410, + 272, + 424 + ], + "score": 1.0, + "content": "Table 9: Wikiann (En) Dataset.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + } + ], + "index": 23.0 + }, + { + "type": "table", + "bbox": [ + 317, + 275, + 487, + 402 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 317, + 275, + 487, + 402 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 317, + 275, + 487, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 487, + 402 + ], + "score": 0.968, + "html": "
MethodShots (12 Slot Types)
51020100
Full-supervision87.87
BERT
Few-shot Supervision
BERT62.8069.5075.8182.49
Few-shot Supervision +unlabeled data
CVT57.4862.7370.2081.82
SeqVAT60.9467.1074.1582.73
Mean Teacher58.9267.6275.2482.20
VAT60.7570.1775.4182.39
Classic ST63.3971.8876.5883.06
BOND62.5070.9175.5282.65
MetaST72.5777.6780.3384.35
", + "type": "table", + "image_path": "2e756955727d5af6ceff79d49d5a8c00dbdb2ebe481f26438fc9a3cd5983d3fb.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 317, + 275, + 487, + 289.1111111111111 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 317, + 289.1111111111111, + 487, + 303.2222222222222 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 317, + 303.2222222222222, + 487, + 317.33333333333326 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 317, + 317.33333333333326, + 487, + 331.44444444444434 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 317, + 331.44444444444434, + 487, + 345.55555555555543 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 317, + 345.55555555555543, + 487, + 359.6666666666665 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 317, + 359.6666666666665, + 487, + 373.7777777777776 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 317, + 373.7777777777776, + 487, + 387.8888888888887 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 317, + 387.8888888888887, + 487, + 401.9999999999998 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 340, + 411, + 462, + 422 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 339, + 410, + 463, + 424 + ], + "spans": [ + { + "bbox": [ + 339, + 410, + 463, + 424 + ], + "score": 1.0, + "content": "Figure 5: MIT Movie Dataset.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "index": 32.5 + }, + { + "type": "table", + "bbox": [ + 124, + 435, + 295, + 561 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 435, + 295, + 561 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 124, + 435, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 124, + 435, + 295, + 561 + ], + "score": 0.972, + "html": "
MethodShots (8 Slot Types)
51020100
Full-supervision
BERT78.95
Few-shot Supervision
BERT41.3954.0660.1272.24
Few-shot Supervision +unlabeled data
CVT33.7442.5751.3370.84
SeqVAT41.9451.5556.1571.39
Mean Teacher40.3751.7557.3472.40
VAT41.2953.3459.6872.65
Classic ST44.3556.8060.2873.13
BOND43.0155.7859.9673.60
MetaST53.0263.8367.8675.25
", + "type": "table", + "image_path": "f476962ad01d07f77ba58f1fdd395dfa4759197269c394736f1bb1c3e19b84bf.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 124, + 435, + 295, + 449.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 124, + 449.0, + 295, + 463.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 124, + 463.0, + 295, + 477.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 124, + 477.0, + 295, + 491.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 124, + 491.0, + 295, + 505.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 124, + 505.0, + 295, + 519.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 124, + 519.0, + 295, + 533.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 124, + 533.0, + 295, + 547.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 124, + 547.0, + 295, + 561.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 139, + 570, + 279, + 582 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 138, + 569, + 280, + 583 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 280, + 583 + ], + "score": 1.0, + "content": "Table 10: MIT Restaurant Dataset.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + } + ], + "index": 42.5 + }, + { + "type": "table", + "bbox": [ + 317, + 435, + 486, + 561 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 317, + 435, + 486, + 561 + ], + "group_id": 5, + "lines": [ + { + "bbox": [ + 317, + 435, + 486, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 435, + 486, + 561 + ], + "score": 0.97, + "html": "
MethodShots (4 Slot Types)
51020100
Full-supervision92.40
BERT
Few-shot Supervision
BERT63.8771.1573.5784.36
Few-shot Supervision +unlabeled data
CVT51.1554.3166.1181.99
SeqVAT58.0267.2174.1582.20
Mean Teacher59.0468.6772.6284.17
VAT57.0365.0372.6984.43
Classic ST64.0470.9974.6584.93
BOND62.5269.5674.1983.87
MetaST71.4976.6578.5485.77
", + "type": "table", + "image_path": "a5c5c774b1339fd4c329a18b099f6042e9d44fe22e679ea606230aedb95f867e.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 317, + 435, + 486, + 449.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 317, + 449.0, + 486, + 463.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 317, + 463.0, + 486, + 477.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 317, + 477.0, + 486, + 491.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 317, + 491.0, + 486, + 505.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 317, + 505.0, + 486, + 519.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 317, + 519.0, + 486, + 533.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 317, + 533.0, + 486, + 547.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 317, + 547.0, + 486, + 561.0 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 344, + 570, + 459, + 582 + ], + "group_id": 5, + "lines": [ + { + "bbox": [ + 343, + 568, + 460, + 584 + ], + "spans": [ + { + "bbox": [ + 343, + 568, + 460, + 584 + ], + "score": 1.0, + "content": "Table 11: CoNLL2003 (EN)", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + } + ], + "index": 48.0 + }, + { + "type": "table", + "bbox": [ + 124, + 594, + 295, + 706 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 594, + 295, + 706 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 124, + 594, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 124, + 594, + 295, + 706 + ], + "score": 0.848, + "html": "
MethodShots (4 Slot Types)
51020100
Full-supervision
BERT
Few-shot Supervision
BERT64.8070.7773.8980.61
Few-shot Supervision +unlabeled data
Mean Teacher64.5568.3473.8779.21
VAT64.9767.6374.2680.70
Classic ST67.9572.6973.7981.82
BOND69.4272.7976.0280.62
MetaST73.3476.6577.0182.11
", + "type": "table", + "image_path": "25284224f76b660bd08133cc6918cc8de0525a602474f8bf588087c496594e31.jpg" + } + ] + } + ], + "index": 55.5, + "virtual_lines": [ + { + "bbox": [ + 124, + 594, + 295, + 608.0 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 124, + 608.0, + 295, + 622.0 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 124, + 622.0, + 295, + 636.0 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 124, + 636.0, + 295, + 650.0 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 124, + 650.0, + 295, + 664.0 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 124, + 664.0, + 295, + 678.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 124, + 678.0, + 295, + 692.0 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 124, + 692.0, + 295, + 706.0 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 141, + 714, + 277, + 726 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 140, + 714, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 140, + 714, + 279, + 728 + ], + "score": 1.0, + "content": "Table 12: Multilingual CoNLL03.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 68 + } + ], + "index": 61.75 + }, + { + "type": "table", + "bbox": [ + 317, + 594, + 486, + 706 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 317, + 594, + 486, + 706 + ], + "group_id": 6, + "lines": [ + { + "bbox": [ + 317, + 594, + 486, + 706 + ], + "spans": [ + { + "bbox": [ + 317, + 594, + 486, + 706 + ], + "score": 0.976, + "html": "
MethodShots (3 Slot Types X 41 languages)
51020100
Full-supervision
BERT
Few-shot Supervision82.3385.70
BERT77.6879.67
Few-shot Supervision +unlabeled data
Mean Teacher77.0980.2382.1985.34
VAT74.7178.8282.6085.82
Classic ST76.7380.2482.3986.08
BOND78.8179.5782.1986.14
MetaST79.1081.6183.1485.57
", + "type": "table", + "image_path": "bdb9bde3a099c1cd2efc59fc41a63bf1059ccedbcf2cff1774df674b185df69c.jpg" + } + ] + } + ], + "index": 63.5, + "virtual_lines": [ + { + "bbox": [ + 317, + 594, + 486, + 608.0 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 317, + 608.0, + 486, + 622.0 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 317, + 622.0, + 486, + 636.0 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 317, + 636.0, + 486, + 650.0 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 317, + 650.0, + 486, + 664.0 + ], + "spans": [], + "index": 64 + }, + { + "bbox": [ + 317, + 664.0, + 486, + 678.0 + ], + "spans": [], + "index": 65 + }, + { + "bbox": [ + 317, + 678.0, + 486, + 692.0 + ], + "spans": [], + "index": 66 + }, + { + "bbox": [ + 317, + 692.0, + 486, + 706.0 + ], + "spans": [], + "index": 67 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 338, + 715, + 466, + 726 + ], + "group_id": 6, + "lines": [ + { + "bbox": [ + 336, + 713, + 466, + 727 + ], + "spans": [ + { + "bbox": [ + 336, + 713, + 466, + 727 + ], + "score": 1.0, + "content": "Table 13: Multilingual Wikiann", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 69 + } + ], + "index": 66.25 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 227, + 103, + 384, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 258, + 83, + 353, + 94 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 257, + 82, + 354, + 95 + ], + "spans": [ + { + "bbox": [ + 257, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "Table 8: Email Dataset.", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 227, + 103, + 384, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 227, + 103, + 384, + 264 + ], + "spans": [ + { + "bbox": [ + 227, + 103, + 384, + 264 + ], + "score": 0.972, + "html": "
MethodShots
510
Full-supervision
BERT0.9444
Few-shot Supervision
BERT0.82110.8785
Few-shot Supervision + unlabeled data
CVT67.4478.24
SeqVAT64.6772.65
Mean Teacher84.1089.53
VAT83.2489.71
Classic ST86.8890.70
BOND84.9289.75
MetaST89.2192.18
", + "type": "table", + "image_path": "a0a70b768721fcc3b76448f14f149d4e4414c737bad75fd505c92a646c5b05ab.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 227, + 103, + 384, + 117.63636363636364 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 227, + 117.63636363636364, + 384, + 132.27272727272728 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 227, + 132.27272727272728, + 384, + 146.9090909090909 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 227, + 146.9090909090909, + 384, + 161.54545454545453 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 227, + 161.54545454545453, + 384, + 176.18181818181816 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 227, + 176.18181818181816, + 384, + 190.81818181818178 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 227, + 190.81818181818178, + 384, + 205.4545454545454 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 227, + 205.4545454545454, + 384, + 220.09090909090904 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 227, + 220.09090909090904, + 384, + 234.72727272727266 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 227, + 234.72727272727266, + 384, + 249.3636363636363 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 227, + 249.3636363636363, + 384, + 263.99999999999994 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 3.0 + }, + { + "type": "table", + "bbox": [ + 124, + 275, + 295, + 402 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 275, + 295, + 402 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 275, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 124, + 275, + 295, + 402 + ], + "score": 0.967, + "html": "
MethodShots (3 Slot Types)
51020100
Full-supervision
BERT84.04
Few-shot Supervision
BERT37.0145.6154.5367.87
Few-shot Supervision +unlabeled data
CVT16.0527.8946.4266.36
SeqVAT21.1135.1642.2662.37
Mean Teacher30.9241.4350.6167.16
VAT24.7238.8150.1566.31
Classic ST32.7246.1554.4168.64
BOND34.2248.7352.4568.89
MetaST55.0456.6160.3873.20
", + "type": "table", + "image_path": "95fcfbdbbf909669be074fbba1a4c6a7b11a739db36ef810a02b75c6dcd33573.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 124, + 275, + 295, + 289.1111111111111 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 124, + 289.1111111111111, + 295, + 303.2222222222222 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 124, + 303.2222222222222, + 295, + 317.33333333333326 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 124, + 317.33333333333326, + 295, + 331.44444444444434 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 124, + 331.44444444444434, + 295, + 345.55555555555543 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 124, + 345.55555555555543, + 295, + 359.6666666666665 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 124, + 359.6666666666665, + 295, + 373.7777777777776 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 124, + 373.7777777777776, + 295, + 387.8888888888887 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 124, + 387.8888888888887, + 295, + 401.9999999999998 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 146, + 411, + 272, + 422 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 410, + 272, + 424 + ], + "spans": [ + { + "bbox": [ + 145, + 410, + 272, + 424 + ], + "score": 1.0, + "content": "Table 9: Wikiann (En) Dataset.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + } + ], + "index": 23.0 + }, + { + "type": "table", + "bbox": [ + 317, + 275, + 487, + 402 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 317, + 275, + 487, + 402 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 317, + 275, + 487, + 402 + ], + "spans": [ + { + "bbox": [ + 317, + 275, + 487, + 402 + ], + "score": 0.968, + "html": "
MethodShots (12 Slot Types)
51020100
Full-supervision87.87
BERT
Few-shot Supervision
BERT62.8069.5075.8182.49
Few-shot Supervision +unlabeled data
CVT57.4862.7370.2081.82
SeqVAT60.9467.1074.1582.73
Mean Teacher58.9267.6275.2482.20
VAT60.7570.1775.4182.39
Classic ST63.3971.8876.5883.06
BOND62.5070.9175.5282.65
MetaST72.5777.6780.3384.35
", + "type": "table", + "image_path": "2e756955727d5af6ceff79d49d5a8c00dbdb2ebe481f26438fc9a3cd5983d3fb.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 317, + 275, + 487, + 289.1111111111111 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 317, + 289.1111111111111, + 487, + 303.2222222222222 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 317, + 303.2222222222222, + 487, + 317.33333333333326 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 317, + 317.33333333333326, + 487, + 331.44444444444434 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 317, + 331.44444444444434, + 487, + 345.55555555555543 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 317, + 345.55555555555543, + 487, + 359.6666666666665 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 317, + 359.6666666666665, + 487, + 373.7777777777776 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 317, + 373.7777777777776, + 487, + 387.8888888888887 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 317, + 387.8888888888887, + 487, + 401.9999999999998 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 340, + 411, + 462, + 422 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 339, + 410, + 463, + 424 + ], + "spans": [ + { + "bbox": [ + 339, + 410, + 463, + 424 + ], + "score": 1.0, + "content": "Figure 5: MIT Movie Dataset.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + } + ], + "index": 32.5 + }, + { + "type": "table", + "bbox": [ + 124, + 435, + 295, + 561 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 435, + 295, + 561 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 124, + 435, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 124, + 435, + 295, + 561 + ], + "score": 0.972, + "html": "
MethodShots (8 Slot Types)
51020100
Full-supervision
BERT78.95
Few-shot Supervision
BERT41.3954.0660.1272.24
Few-shot Supervision +unlabeled data
CVT33.7442.5751.3370.84
SeqVAT41.9451.5556.1571.39
Mean Teacher40.3751.7557.3472.40
VAT41.2953.3459.6872.65
Classic ST44.3556.8060.2873.13
BOND43.0155.7859.9673.60
MetaST53.0263.8367.8675.25
", + "type": "table", + "image_path": "f476962ad01d07f77ba58f1fdd395dfa4759197269c394736f1bb1c3e19b84bf.jpg" + } + ] + } + ], + "index": 35, + "virtual_lines": [ + { + "bbox": [ + 124, + 435, + 295, + 449.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 124, + 449.0, + 295, + 463.0 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 124, + 463.0, + 295, + 477.0 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 124, + 477.0, + 295, + 491.0 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 124, + 491.0, + 295, + 505.0 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 124, + 505.0, + 295, + 519.0 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 124, + 519.0, + 295, + 533.0 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 124, + 533.0, + 295, + 547.0 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 124, + 547.0, + 295, + 561.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 139, + 570, + 279, + 582 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 138, + 569, + 280, + 583 + ], + "spans": [ + { + "bbox": [ + 138, + 569, + 280, + 583 + ], + "score": 1.0, + "content": "Table 10: MIT Restaurant Dataset.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 50 + } + ], + "index": 42.5 + }, + { + "type": "table", + "bbox": [ + 317, + 435, + 486, + 561 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 317, + 435, + 486, + 561 + ], + "group_id": 5, + "lines": [ + { + "bbox": [ + 317, + 435, + 486, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 435, + 486, + 561 + ], + "score": 0.97, + "html": "
MethodShots (4 Slot Types)
51020100
Full-supervision92.40
BERT
Few-shot Supervision
BERT63.8771.1573.5784.36
Few-shot Supervision +unlabeled data
CVT51.1554.3166.1181.99
SeqVAT58.0267.2174.1582.20
Mean Teacher59.0468.6772.6284.17
VAT57.0365.0372.6984.43
Classic ST64.0470.9974.6584.93
BOND62.5269.5674.1983.87
MetaST71.4976.6578.5485.77
", + "type": "table", + "image_path": "a5c5c774b1339fd4c329a18b099f6042e9d44fe22e679ea606230aedb95f867e.jpg" + } + ] + } + ], + "index": 45, + "virtual_lines": [ + { + "bbox": [ + 317, + 435, + 486, + 449.0 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 317, + 449.0, + 486, + 463.0 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 317, + 463.0, + 486, + 477.0 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 317, + 477.0, + 486, + 491.0 + ], + "spans": [], + "index": 44 + }, + { + "bbox": [ + 317, + 491.0, + 486, + 505.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 317, + 505.0, + 486, + 519.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 317, + 519.0, + 486, + 533.0 + ], + "spans": [], + "index": 47 + }, + { + "bbox": [ + 317, + 533.0, + 486, + 547.0 + ], + "spans": [], + "index": 48 + }, + { + "bbox": [ + 317, + 547.0, + 486, + 561.0 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 344, + 570, + 459, + 582 + ], + "group_id": 5, + "lines": [ + { + "bbox": [ + 343, + 568, + 460, + 584 + ], + "spans": [ + { + "bbox": [ + 343, + 568, + 460, + 584 + ], + "score": 1.0, + "content": "Table 11: CoNLL2003 (EN)", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + } + ], + "index": 48.0 + }, + { + "type": "table", + "bbox": [ + 124, + 594, + 295, + 706 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 124, + 594, + 295, + 706 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 124, + 594, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 124, + 594, + 295, + 706 + ], + "score": 0.848, + "html": "
MethodShots (4 Slot Types)
51020100
Full-supervision
BERT
Few-shot Supervision
BERT64.8070.7773.8980.61
Few-shot Supervision +unlabeled data
Mean Teacher64.5568.3473.8779.21
VAT64.9767.6374.2680.70
Classic ST67.9572.6973.7981.82
BOND69.4272.7976.0280.62
MetaST73.3476.6577.0182.11
", + "type": "table", + "image_path": "25284224f76b660bd08133cc6918cc8de0525a602474f8bf588087c496594e31.jpg" + } + ] + } + ], + "index": 55.5, + "virtual_lines": [ + { + "bbox": [ + 124, + 594, + 295, + 608.0 + ], + "spans": [], + "index": 52 + }, + { + "bbox": [ + 124, + 608.0, + 295, + 622.0 + ], + "spans": [], + "index": 53 + }, + { + "bbox": [ + 124, + 622.0, + 295, + 636.0 + ], + "spans": [], + "index": 54 + }, + { + "bbox": [ + 124, + 636.0, + 295, + 650.0 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 124, + 650.0, + 295, + 664.0 + ], + "spans": [], + "index": 56 + }, + { + "bbox": [ + 124, + 664.0, + 295, + 678.0 + ], + "spans": [], + "index": 57 + }, + { + "bbox": [ + 124, + 678.0, + 295, + 692.0 + ], + "spans": [], + "index": 58 + }, + { + "bbox": [ + 124, + 692.0, + 295, + 706.0 + ], + "spans": [], + "index": 59 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 141, + 714, + 277, + 726 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 140, + 714, + 279, + 728 + ], + "spans": [ + { + "bbox": [ + 140, + 714, + 279, + 728 + ], + "score": 1.0, + "content": "Table 12: Multilingual CoNLL03.", + "type": "text" + } + ], + "index": 68 + } + ], + "index": 68 + } + ], + "index": 61.75 + }, + { + "type": "table", + "bbox": [ + 317, + 594, + 486, + 706 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 317, + 594, + 486, + 706 + ], + "group_id": 6, + "lines": [ + { + "bbox": [ + 317, + 594, + 486, + 706 + ], + "spans": [ + { + "bbox": [ + 317, + 594, + 486, + 706 + ], + "score": 0.976, + "html": "
MethodShots (3 Slot Types X 41 languages)
51020100
Full-supervision
BERT
Few-shot Supervision82.3385.70
BERT77.6879.67
Few-shot Supervision +unlabeled data
Mean Teacher77.0980.2382.1985.34
VAT74.7178.8282.6085.82
Classic ST76.7380.2482.3986.08
BOND78.8179.5782.1986.14
MetaST79.1081.6183.1485.57
", + "type": "table", + "image_path": "bdb9bde3a099c1cd2efc59fc41a63bf1059ccedbcf2cff1774df674b185df69c.jpg" + } + ] + } + ], + "index": 63.5, + "virtual_lines": [ + { + "bbox": [ + 317, + 594, + 486, + 608.0 + ], + "spans": [], + "index": 60 + }, + { + "bbox": [ + 317, + 608.0, + 486, + 622.0 + ], + "spans": [], + "index": 61 + }, + { + "bbox": [ + 317, + 622.0, + 486, + 636.0 + ], + "spans": [], + "index": 62 + }, + { + "bbox": [ + 317, + 636.0, + 486, + 650.0 + ], + "spans": [], + "index": 63 + }, + { + "bbox": [ + 317, + 650.0, + 486, + 664.0 + ], + "spans": [], + "index": 64 + }, + { + "bbox": [ + 317, + 664.0, + 486, + 678.0 + ], + "spans": [], + "index": 65 + }, + { + "bbox": [ + 317, + 678.0, + 486, + 692.0 + ], + "spans": [], + "index": 66 + }, + { + "bbox": [ + 317, + 692.0, + 486, + 706.0 + ], + "spans": [], + "index": 67 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 338, + 715, + 466, + 726 + ], + "group_id": 6, + "lines": [ + { + "bbox": [ + 336, + 713, + 466, + 727 + ], + "spans": [ + { + "bbox": [ + 336, + 713, + 466, + 727 + ], + "score": 1.0, + "content": "Table 13: Multilingual Wikiann", + "type": "text" + } + ], + "index": 69 + } + ], + "index": 69 + } + ], + "index": 66.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 324, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 326, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 326, + 94 + ], + "score": 1.0, + "content": "A.4 IMPLEMENTATIONS AND HYPER-PARAMETER", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 503, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "We do not perform any hyper-parameter tuning for different datasets. The batch size and maxi-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "mum sequence length varies due to data characteristics and are as shown in Tbale 14. The hyper-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 255, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 255, + 136 + ], + "score": 1.0, + "content": "parameters are as shown in Table 14.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 141, + 502, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 504, + 155 + ], + "score": 1.0, + "content": "Also, we retain parameters from original BERT implementation from https://github.com/", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 153, + 255, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 255, + 165 + ], + "score": 1.0, + "content": "huggingface/transformers.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 108, + 169, + 503, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 182 + ], + "score": 1.0, + "content": "We implement SeqVAT based on https://github.com/jiesutd/NCRFpp and im-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "plement CVT following https://github.com/tensorflow/models/tree/master/", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 213, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 213, + 203 + ], + "score": 1.0, + "content": "research/cvt_text.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 108, + 212, + 502, + 285 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 212, + 502, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 212, + 502, + 285 + ], + "spans": [ + { + "bbox": [ + 108, + 212, + 502, + 285 + ], + "score": 0.973, + "html": "
DatasetSequence LengthBatch SizeLabeled data sample size B|Unlabeled Batch SizeBERT Encoder
SNIPS64163232BERT-base-uncased
Email64163232BERT-base-cased
Movie64163232BERT-base-uncased
Restaurant64161632BERT-base-uncased
CoNLL03 (EN)12816832BERT-base-cased
Wikiann (EN)12816832BERT-base-cased
CoNLLO3 (multilingual)128163232BERT-multilingual-base-cased
Wikiann (multilingaul)128163232BERT-multilingual-base-cased
", + "type": "table", + "image_path": "7d8a5ff09a19797def76c476a6247747396bbda92f80f3b10f8a54a82da45d85.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 108, + 212, + 502, + 236.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 236.33333333333334, + 502, + 260.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 260.6666666666667, + 502, + 285.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 142, + 294, + 467, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 293, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 469, + 307 + ], + "score": 1.0, + "content": "Table 14: Batch size, sequence length and BERT encoder choices across datasets", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "table", + "bbox": [ + 109, + 322, + 503, + 456 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 109, + 322, + 503, + 456 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 322, + 503, + 456 + ], + "spans": [ + { + "bbox": [ + 109, + 322, + 503, + 456 + ], + "score": 0.979, + "html": "
BERT attention dropout BERT hidden dropout Latest Iteration R in labeled data acquisition BERT output hidden size h0.3 0.3 5 768
Steps for fine-tuning teacher model on labeled data2000
Steps T for self-training model on unlabeled data3000
Mini-batch S Re-initialize Student5
Pseudo-label TypeY Hard
Warmup steps20
learning rate α5e-5
Weight_decay5e-6
", + "type": "table", + "image_path": "61aff8079c20ff5289ce3220cb1b1edc76f284a5836c724cdde2f73eb1a1b91e.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 109, + 322, + 503, + 366.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 109, + 366.6666666666667, + 503, + 411.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 109, + 411.33333333333337, + 503, + 456.00000000000006 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 248, + 464, + 363, + 476 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 246, + 462, + 365, + 478 + ], + "spans": [ + { + "bbox": [ + 246, + 462, + 365, + 478 + ], + "score": 1.0, + "content": "Table 15: Hyper-parameters.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + } + ], + "index": 15.0 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 307, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 752, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 324, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 326, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 326, + 94 + ], + "score": 1.0, + "content": "A.4 IMPLEMENTATIONS AND HYPER-PARAMETER", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 503, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "We do not perform any hyper-parameter tuning for different datasets. The batch size and maxi-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "mum sequence length varies due to data characteristics and are as shown in Tbale 14. The hyper-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 255, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 255, + 136 + ], + "score": 1.0, + "content": "parameters are as shown in Table 14.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 106, + 102, + 505, + 136 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 141, + 502, + 164 + ], + "lines": [ + { + "bbox": [ + 106, + 140, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 140, + 504, + 155 + ], + "score": 1.0, + "content": "Also, we retain parameters from original BERT implementation from https://github.com/", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 153, + 255, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 255, + 165 + ], + "score": 1.0, + "content": "huggingface/transformers.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 140, + 504, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 169, + 503, + 203 + ], + "lines": [ + { + "bbox": [ + 106, + 168, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 106, + 168, + 505, + 182 + ], + "score": 1.0, + "content": "We implement SeqVAT based on https://github.com/jiesutd/NCRFpp and im-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "plement CVT following https://github.com/tensorflow/models/tree/master/", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 192, + 213, + 203 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 213, + 203 + ], + "score": 1.0, + "content": "research/cvt_text.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 168, + 505, + 203 + ] + }, + { + "type": "table", + "bbox": [ + 108, + 212, + 502, + 285 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 108, + 212, + 502, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 212, + 502, + 285 + ], + "spans": [ + { + "bbox": [ + 108, + 212, + 502, + 285 + ], + "score": 0.973, + "html": "
DatasetSequence LengthBatch SizeLabeled data sample size B|Unlabeled Batch SizeBERT Encoder
SNIPS64163232BERT-base-uncased
Email64163232BERT-base-cased
Movie64163232BERT-base-uncased
Restaurant64161632BERT-base-uncased
CoNLL03 (EN)12816832BERT-base-cased
Wikiann (EN)12816832BERT-base-cased
CoNLLO3 (multilingual)128163232BERT-multilingual-base-cased
Wikiann (multilingaul)128163232BERT-multilingual-base-cased
", + "type": "table", + "image_path": "7d8a5ff09a19797def76c476a6247747396bbda92f80f3b10f8a54a82da45d85.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 108, + 212, + 502, + 236.33333333333334 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 236.33333333333334, + 502, + 260.6666666666667 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 260.6666666666667, + 502, + 285.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 142, + 294, + 467, + 306 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 293, + 469, + 307 + ], + "spans": [ + { + "bbox": [ + 142, + 293, + 469, + 307 + ], + "score": 1.0, + "content": "Table 14: Batch size, sequence length and BERT encoder choices across datasets", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + } + ], + "index": 11.0 + }, + { + "type": "table", + "bbox": [ + 109, + 322, + 503, + 456 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 109, + 322, + 503, + 456 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 322, + 503, + 456 + ], + "spans": [ + { + "bbox": [ + 109, + 322, + 503, + 456 + ], + "score": 0.979, + "html": "
BERT attention dropout BERT hidden dropout Latest Iteration R in labeled data acquisition BERT output hidden size h0.3 0.3 5 768
Steps for fine-tuning teacher model on labeled data2000
Steps T for self-training model on unlabeled data3000
Mini-batch S Re-initialize Student5
Pseudo-label TypeY Hard
Warmup steps20
learning rate α5e-5
Weight_decay5e-6
", + "type": "table", + "image_path": "61aff8079c20ff5289ce3220cb1b1edc76f284a5836c724cdde2f73eb1a1b91e.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 109, + 322, + 503, + 366.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 109, + 366.6666666666667, + 503, + 411.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 109, + 411.33333333333337, + 503, + 456.00000000000006 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 248, + 464, + 363, + 476 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 246, + 462, + 365, + 478 + ], + "spans": [ + { + "bbox": [ + 246, + 462, + 365, + 478 + ], + "score": 1.0, + "content": "Table 15: Hyper-parameters.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + } + ], + "index": 15.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/ARFshOO1Iu/ARFshOO1Iu_model.json b/parse/train/ARFshOO1Iu/ARFshOO1Iu_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8648abaac48a2464fa7487b58fc096f1d6e618d --- /dev/null +++ b/parse/train/ARFshOO1Iu/ARFshOO1Iu_model.json @@ -0,0 +1,19144 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1287, + 1402, + 1287, + 1402, + 1833, + 298, + 1833 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 398, + 583, + 1302, + 583, + 1302, + 1164, + 398, + 1164 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 299, + 1850, + 1402, + 1850, + 1402, + 2033, + 299, + 2033 + ], + "score": 0.977 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1401, + 219, + 1401, + 324, + 301, + 324 + ], + "score": 0.963 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 302, + 1217, + 573, + 1217, + 573, + 1252, + 302, + 1252 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 104, + 298, + 104 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 773, + 520, + 926, + 520, + 926, + 553, + 773, + 553 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 842, + 2089, + 857, + 2089, + 857, + 2112, + 842, + 2112 + ], + "score": 0.618 + }, + { + "category_id": 13, + "poly": [ + 773, + 1102, + 827, + 1102, + 827, + 1132, + 773, + 1132 + ], + "score": 0.88, + "latex": "1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 508, + 1468, + 563, + 1468, + 563, + 1498, + 508, + 1498 + ], + "score": 0.87, + "latex": "\\bar { 2 } 7 \\%" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 216.0, + 1404.0, + 216.0, + 1404.0, + 277.0, + 294.0, + 277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 273.0, + 924.0, + 273.0, + 924.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1214.0, + 579.0, + 1214.0, + 579.0, + 1261.0, + 294.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 519.0, + 932.0, + 519.0, + 932.0, + 556.0, + 769.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 860.0, + 2088.0, + 860.0, + 2118.0, + 841.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1284.0, + 1406.0, + 1284.0, + 1406.0, + 1320.0, + 293.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1351.0, + 293.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1345.0, + 1406.0, + 1345.0, + 1406.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1377.0, + 1404.0, + 1377.0, + 1404.0, + 1412.0, + 294.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1405.0, + 1406.0, + 1405.0, + 1406.0, + 1443.0, + 293.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1438.0, + 1407.0, + 1438.0, + 1407.0, + 1473.0, + 294.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1467.0, + 507.0, + 1467.0, + 507.0, + 1503.0, + 293.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1467.0, + 1406.0, + 1467.0, + 1406.0, + 1503.0, + 564.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1404.0, + 1498.0, + 1404.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1527.0, + 1406.0, + 1527.0, + 1406.0, + 1565.0, + 293.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 1406.0, + 1561.0, + 1406.0, + 1594.0, + 293.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1590.0, + 1407.0, + 1590.0, + 1407.0, + 1624.0, + 293.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1622.0, + 1404.0, + 1622.0, + 1404.0, + 1653.0, + 296.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 1406.0, + 1651.0, + 1406.0, + 1686.0, + 294.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1677.0, + 1407.0, + 1677.0, + 1407.0, + 1720.0, + 292.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1711.0, + 1406.0, + 1711.0, + 1406.0, + 1747.0, + 296.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 1406.0, + 1771.0, + 1406.0, + 1809.0, + 294.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1804.0, + 1323.0, + 1804.0, + 1323.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 586.0, + 1304.0, + 586.0, + 1304.0, + 619.0, + 395.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 615.0, + 1306.0, + 615.0, + 1306.0, + 651.0, + 393.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 646.0, + 1306.0, + 646.0, + 1306.0, + 684.0, + 393.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 674.0, + 1306.0, + 674.0, + 1306.0, + 713.0, + 393.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 709.0, + 1306.0, + 709.0, + 1306.0, + 740.0, + 393.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 739.0, + 1304.0, + 739.0, + 1304.0, + 769.0, + 395.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 770.0, + 1305.0, + 770.0, + 1305.0, + 800.0, + 395.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 799.0, + 1305.0, + 799.0, + 1305.0, + 832.0, + 396.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 830.0, + 1305.0, + 830.0, + 1305.0, + 861.0, + 394.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 860.0, + 1305.0, + 860.0, + 1305.0, + 896.0, + 393.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 889.0, + 1305.0, + 889.0, + 1305.0, + 926.0, + 394.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 919.0, + 1305.0, + 919.0, + 1305.0, + 954.0, + 394.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 950.0, + 1305.0, + 950.0, + 1305.0, + 984.0, + 394.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 982.0, + 1306.0, + 982.0, + 1306.0, + 1015.0, + 394.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1013.0, + 1306.0, + 1013.0, + 1306.0, + 1044.0, + 393.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1041.0, + 1306.0, + 1041.0, + 1306.0, + 1077.0, + 393.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1072.0, + 1305.0, + 1072.0, + 1305.0, + 1105.0, + 395.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1101.0, + 772.0, + 1101.0, + 772.0, + 1137.0, + 393.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1101.0, + 1305.0, + 1101.0, + 1305.0, + 1137.0, + 828.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1131.0, + 1062.0, + 1131.0, + 1062.0, + 1170.0, + 393.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 1407.0, + 1878.0, + 1407.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 1404.0, + 1909.0, + 1404.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1974.0, + 297.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1969.0, + 1404.0, + 1969.0, + 1404.0, + 2006.0, + 293.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 1406.0, + 2001.0, + 1406.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1607, + 1404, + 1607, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 859, + 1404, + 859, + 1404, + 1591, + 297, + 1591 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 599, + 1404, + 599, + 1404, + 844, + 298, + 844 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 523, + 193, + 1153, + 193, + 1153, + 516, + 523, + 516 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 681, + 528, + 1017, + 528, + 1017, + 559, + 681, + 559 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.724 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.713 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 854, + 76, + 854, + 104, + 298, + 104 + ], + "score": 0.122 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.096 + }, + { + "category_id": 13, + "poly": [ + 461, + 1637, + 715, + 1637, + 715, + 1671, + 461, + 1671 + ], + "score": 0.91, + "latex": "K = \\{ 5 , 1 0 , 2 0 , 1 0 0 \\} )" + }, + { + "category_id": 13, + "poly": [ + 1261, + 660, + 1315, + 660, + 1315, + 690, + 1261, + 690 + ], + "score": 0.88, + "latex": "1 0 \\%" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 202.0, + 1055.0, + 202.0, + 1055.0, + 227.0, + 883.0, + 227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 220.0, + 724.0, + 220.0, + 724.0, + 243.0, + 634.0, + 243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 224.0, + 1007.0, + 224.0, + 1007.0, + 241.0, + 930.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 268.0, + 706.0, + 268.0, + 706.0, + 285.0, + 652.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 265.0, + 979.0, + 265.0, + 979.0, + 282.0, + 929.0, + 282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 280.0, + 715.0, + 280.0, + 715.0, + 300.0, + 641.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 279.0, + 998.0, + 279.0, + 998.0, + 299.0, + 908.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 292.0, + 833.0, + 292.0, + 833.0, + 319.0, + 760.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 324.0, + 710.0, + 324.0, + 710.0, + 352.0, + 644.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 310.0, + 825.0, + 310.0, + 825.0, + 351.0, + 769.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 339.0, + 955.0, + 339.0, + 955.0, + 348.0, + 947.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 325.0, + 1136.0, + 325.0, + 1136.0, + 348.0, + 1068.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 349.0, + 1009.0, + 349.0, + 1009.0, + 375.0, + 893.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 359.0, + 1079.0, + 359.0, + 1079.0, + 379.0, + 1013.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 391.0, + 591.0, + 391.0, + 591.0, + 408.0, + 534.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 389.0, + 836.0, + 389.0, + 836.0, + 408.0, + 753.0, + 408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 384.0, + 986.0, + 384.0, + 986.0, + 403.0, + 916.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 406.0, + 578.0, + 406.0, + 578.0, + 425.0, + 546.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 405.0, + 809.0, + 405.0, + 809.0, + 424.0, + 778.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 399.0, + 993.0, + 399.0, + 993.0, + 419.0, + 909.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 414.0, + 967.0, + 414.0, + 967.0, + 433.0, + 935.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 454.0, + 881.0, + 454.0, + 881.0, + 480.0, + 741.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 477.0, + 869.0, + 477.0, + 869.0, + 499.0, + 755.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 480.0, + 1087.0, + 480.0, + 1087.0, + 503.0, + 1009.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.25, + 459.5, + 1114.25, + 459.5, + 1114.25, + 482.5, + 982.25, + 482.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 524.0, + 1019.0, + 524.0, + 1019.0, + 562.0, + 678.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1642.0, + 294.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1636.0, + 460.0, + 1636.0, + 460.0, + 1674.0, + 292.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1636.0, + 1406.0, + 1636.0, + 1406.0, + 1674.0, + 716.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1666.0, + 1405.0, + 1666.0, + 1405.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 1406.0, + 1698.0, + 1406.0, + 1732.0, + 294.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1731.0, + 1405.0, + 1731.0, + 1405.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1405.0, + 1760.0, + 1405.0, + 1795.0, + 292.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1788.0, + 1407.0, + 1788.0, + 1407.0, + 1824.0, + 290.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1885.0, + 292.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1907.0, + 1404.0, + 1907.0, + 1404.0, + 1950.0, + 292.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2007.0, + 291.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2037.0, + 292.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 860.0, + 1405.0, + 860.0, + 1405.0, + 891.0, + 296.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 889.0, + 1404.0, + 889.0, + 1404.0, + 923.0, + 293.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 919.0, + 1404.0, + 919.0, + 1404.0, + 954.0, + 294.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 951.0, + 1404.0, + 951.0, + 1404.0, + 984.0, + 293.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 976.0, + 1404.0, + 976.0, + 1404.0, + 1018.0, + 294.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1012.0, + 1404.0, + 1012.0, + 1404.0, + 1045.0, + 294.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1042.0, + 1404.0, + 1042.0, + 1404.0, + 1076.0, + 295.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1073.0, + 1405.0, + 1073.0, + 1405.0, + 1107.0, + 296.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1101.0, + 1405.0, + 1101.0, + 1405.0, + 1138.0, + 292.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1131.0, + 1405.0, + 1131.0, + 1405.0, + 1168.0, + 295.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1162.0, + 1404.0, + 1162.0, + 1404.0, + 1198.0, + 294.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1193.0, + 1405.0, + 1193.0, + 1405.0, + 1227.0, + 296.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1224.0, + 1404.0, + 1224.0, + 1404.0, + 1258.0, + 295.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1289.0, + 296.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1286.0, + 1404.0, + 1286.0, + 1404.0, + 1320.0, + 295.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1310.0, + 1404.0, + 1310.0, + 1404.0, + 1354.0, + 292.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1343.0, + 1404.0, + 1343.0, + 1404.0, + 1385.0, + 292.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1372.0, + 1405.0, + 1372.0, + 1405.0, + 1414.0, + 294.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1442.0, + 294.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1439.0, + 1402.0, + 1439.0, + 1402.0, + 1470.0, + 297.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 1405.0, + 1468.0, + 1405.0, + 1502.0, + 295.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1496.0, + 1408.0, + 1496.0, + 1408.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1564.0, + 294.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1562.0, + 988.0, + 1562.0, + 988.0, + 1593.0, + 296.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 601.0, + 1404.0, + 601.0, + 1404.0, + 630.0, + 296.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 627.0, + 1405.0, + 627.0, + 1405.0, + 665.0, + 293.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 659.0, + 1260.0, + 659.0, + 1260.0, + 692.0, + 294.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 659.0, + 1404.0, + 659.0, + 1404.0, + 692.0, + 1316.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 693.0, + 1404.0, + 693.0, + 1404.0, + 723.0, + 297.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 721.0, + 1401.0, + 721.0, + 1401.0, + 751.0, + 297.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 754.0, + 1402.0, + 754.0, + 1402.0, + 784.0, + 296.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 780.0, + 1405.0, + 780.0, + 1405.0, + 816.0, + 295.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 809.0, + 526.0, + 809.0, + 526.0, + 847.0, + 294.0, + 847.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 399, + 1404, + 399, + 1404, + 797, + 297, + 797 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 1110, + 1404, + 1110, + 1404, + 1404, + 297, + 1404 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 880, + 1403, + 880, + 1403, + 1098, + 297, + 1098 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1818, + 1404, + 1818, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 384, + 298, + 384 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 577, + 1595, + 1120, + 1595, + 1120, + 1686, + 577, + 1686 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 655, + 1424, + 1040, + 1424, + 1040, + 1486, + 655, + 1486 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 298, + 1507, + 1400, + 1507, + 1400, + 1575, + 298, + 1575 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 298, + 1701, + 900, + 1701, + 900, + 1735, + 298, + 1735 + ], + "score": 0.929 + }, + { + "category_id": 0, + "poly": [ + 301, + 1768, + 737, + 1768, + 737, + 1805, + 301, + 1805 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 299, + 822, + 558, + 822, + 558, + 857, + 299, + 857 + ], + "score": 0.909 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1433, + 1400, + 1433, + 1400, + 1463, + 1366, + 1463 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1623, + 1400, + 1623, + 1400, + 1653, + 1366, + 1653 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.633 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.63 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.399 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 854, + 76, + 854, + 104, + 299, + 104 + ], + "score": 0.371 + }, + { + "category_id": 14, + "poly": [ + 579, + 1590, + 1118, + 1590, + 1118, + 1687, + 579, + 1687 + ], + "score": 0.95, + "latex": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\underset { \\theta } { \\arg \\operatorname* { m i n } } \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) ," + }, + { + "category_id": 13, + "poly": [ + 677, + 943, + 895, + 943, + 895, + 977, + 677, + 977 + ], + "score": 0.93, + "latex": "\\mathbf { \\bar { \\cal X } } ~ = ~ \\{ x _ { 1 } , . . . , x _ { N } \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1360, + 563, + 1360, + 563, + 1407, + 297, + 1407 + ], + "score": 0.92, + "latex": "\\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ]" + }, + { + "category_id": 14, + "poly": [ + 655, + 1422, + 1043, + 1422, + 1043, + 1486, + 655, + 1486 + ], + "score": 0.92, + "latex": "\\hat { y } _ { m , n } ^ { ( t ) } = \\underset { c \\in C } { \\arg \\operatorname* { m a x } } f _ { n , c } ( x _ { m , n } ^ { u } ; \\theta _ { t e a } ^ { ( t ) } ) ." + }, + { + "category_id": 13, + "poly": [ + 736, + 1112, + 842, + 1112, + 842, + 1146, + 736, + 1146 + ], + "score": 0.92, + "latex": "f ( \\cdot ; \\theta _ { s t u } )" + }, + { + "category_id": 13, + "poly": [ + 749, + 1505, + 1093, + 1505, + 1093, + 1547, + 749, + 1547 + ], + "score": 0.92, + "latex": "( X ^ { u } , \\hat { Y } ^ { ( t ) } ) = \\{ ( X _ { m } ^ { u } , \\hat { Y } _ { m } ^ { ( t ) } ) \\} _ { m } ^ { M }" + }, + { + "category_id": 13, + "poly": [ + 533, + 1325, + 810, + 1325, + 810, + 1363, + 533, + 1363 + ], + "score": 0.92, + "latex": "X _ { m } ^ { u } \\doteq \\{ x _ { 1 , m } ^ { u } , . . . , x _ { N , m } ^ { u } \\}" + }, + { + "category_id": 13, + "poly": [ + 576, + 1112, + 681, + 1112, + 681, + 1146, + 576, + 1146 + ], + "score": 0.92, + "latex": "f ( \\cdot ; \\theta _ { t e a } )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1702, + 433, + 1702, + 433, + 1736, + 372, + 1736 + ], + "score": 0.91, + "latex": "l ( \\cdot , \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 523, + 521, + 578, + 521, + 578, + 551, + 523, + 551 + ], + "score": 0.89, + "latex": "1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 974, + 678, + 974, + 678, + 1009, + 297, + 1009 + ], + "score": 0.83, + "latex": "s = [ x _ { i } , . . . , x _ { j } ] ( 0 \\leq i \\leq j \\leq N )" + }, + { + "category_id": 13, + "poly": [ + 420, + 1326, + 449, + 1326, + 449, + 1354, + 420, + 1354 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 558, + 945, + 587, + 945, + 587, + 971, + 558, + 971 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 861, + 1326, + 886, + 1326, + 886, + 1353, + 861, + 1353 + ], + "score": 0.81, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1299, + 1196, + 1299, + 1196, + 1324, + 1168, + 1324 + ], + "score": 0.77, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 987, + 1372, + 1000, + 1372, + 1000, + 1396, + 987, + 1396 + ], + "score": 0.77, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 620, + 1039, + 643, + 1039, + 643, + 1064, + 620, + 1064 + ], + "score": 0.42, + "latex": "\\bigcirc" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1766.0, + 740.0, + 1766.0, + 740.0, + 1810.0, + 292.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 818.0, + 561.0, + 818.0, + 561.0, + 865.0, + 291.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 400.0, + 1401.0, + 400.0, + 1401.0, + 432.0, + 296.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 463.0, + 295.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 459.0, + 1406.0, + 459.0, + 1406.0, + 497.0, + 291.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 489.0, + 1404.0, + 489.0, + 1404.0, + 527.0, + 294.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 522.0, + 520.0, + 522.0, + 560.0, + 294.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 520.0, + 1405.0, + 520.0, + 1405.0, + 560.0, + 579.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 547.0, + 1405.0, + 547.0, + 1405.0, + 590.0, + 291.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 582.0, + 1405.0, + 582.0, + 1405.0, + 619.0, + 294.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 612.0, + 1405.0, + 612.0, + 1405.0, + 647.0, + 292.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 641.0, + 1405.0, + 641.0, + 1405.0, + 678.0, + 292.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 671.0, + 1405.0, + 671.0, + 1405.0, + 712.0, + 291.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 1404.0, + 703.0, + 1404.0, + 738.0, + 294.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 770.0, + 294.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 763.0, + 703.0, + 763.0, + 703.0, + 802.0, + 292.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1108.0, + 575.0, + 1108.0, + 575.0, + 1150.0, + 292.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1108.0, + 735.0, + 1108.0, + 735.0, + 1150.0, + 682.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1108.0, + 1406.0, + 1108.0, + 1406.0, + 1150.0, + 843.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1140.0, + 1405.0, + 1140.0, + 1405.0, + 1178.0, + 294.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1175.0, + 1404.0, + 1175.0, + 1404.0, + 1205.0, + 295.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1205.0, + 1404.0, + 1205.0, + 1404.0, + 1236.0, + 295.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1236.0, + 1404.0, + 1236.0, + 1404.0, + 1267.0, + 295.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1262.0, + 1406.0, + 1262.0, + 1406.0, + 1299.0, + 294.0, + 1299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1295.0, + 1167.0, + 1295.0, + 1167.0, + 1330.0, + 295.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1295.0, + 1404.0, + 1295.0, + 1404.0, + 1330.0, + 1197.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1315.0, + 419.0, + 1315.0, + 419.0, + 1373.0, + 287.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1315.0, + 532.0, + 1315.0, + 532.0, + 1373.0, + 450.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1315.0, + 860.0, + 1315.0, + 860.0, + 1373.0, + 811.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1315.0, + 1411.0, + 1315.0, + 1411.0, + 1373.0, + 887.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1356.0, + 296.0, + 1356.0, + 296.0, + 1416.0, + 291.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1356.0, + 986.0, + 1356.0, + 986.0, + 1416.0, + 564.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1356.0, + 1218.0, + 1356.0, + 1218.0, + 1416.0, + 1001.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 882.0, + 1405.0, + 882.0, + 1405.0, + 917.0, + 295.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 914.0, + 1404.0, + 914.0, + 1404.0, + 949.0, + 295.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 942.0, + 557.0, + 942.0, + 557.0, + 979.0, + 292.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 942.0, + 676.0, + 942.0, + 676.0, + 979.0, + 588.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 942.0, + 1406.0, + 942.0, + 1406.0, + 979.0, + 896.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 970.0, + 296.0, + 970.0, + 296.0, + 1015.0, + 290.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 970.0, + 1407.0, + 970.0, + 1407.0, + 1015.0, + 679.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1005.0, + 1405.0, + 1005.0, + 1405.0, + 1040.0, + 295.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1034.0, + 619.0, + 1034.0, + 619.0, + 1070.0, + 294.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1034.0, + 1405.0, + 1034.0, + 1405.0, + 1070.0, + 644.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 1341.0, + 1065.0, + 1341.0, + 1101.0, + 292.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1820.0, + 1404.0, + 1820.0, + 1404.0, + 1855.0, + 296.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1405.0, + 1850.0, + 1405.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1405.0, + 231.0, + 1405.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 296.0, + 294.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 357.0, + 294.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 354.0, + 1036.0, + 354.0, + 1036.0, + 388.0, + 296.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1488.0, + 748.0, + 1488.0, + 748.0, + 1562.0, + 286.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1488.0, + 1413.0, + 1488.0, + 1413.0, + 1562.0, + 1094.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1541.0, + 682.0, + 1541.0, + 682.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1698.0, + 371.0, + 1698.0, + 371.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 1698.0, + 902.0, + 1698.0, + 902.0, + 1740.0, + 434.0, + 1740.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 667, + 1404, + 667, + 1404, + 1033, + 297, + 1033 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1581, + 1404, + 1581, + 1404, + 1768, + 298, + 1768 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1820, + 1403, + 1820, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 408, + 1403, + 408, + 1403, + 653, + 298, + 653 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1042, + 1403, + 1042, + 1403, + 1211, + 297, + 1211 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1319, + 1403, + 1319, + 1403, + 1446, + 298, + 1446 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1402, + 229, + 1402, + 353, + 298, + 353 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 647, + 1224, + 1048, + 1224, + 1048, + 1312, + 647, + 1312 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 299, + 1505, + 1410, + 1505, + 1410, + 1568, + 299, + 1568 + ], + "score": 0.946 + }, + { + "category_id": 8, + "poly": [ + 587, + 1455, + 1113, + 1455, + 1113, + 1499, + 587, + 1499 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 298, + 1782, + 855, + 1782, + 855, + 1814, + 298, + 1814 + ], + "score": 0.92 + }, + { + "category_id": 0, + "poly": [ + 300, + 370, + 852, + 370, + 852, + 401, + 300, + 401 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1253, + 1400, + 1253, + 1400, + 1283, + 1366, + 1283 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1464, + 1399, + 1464, + 1399, + 1492, + 1366, + 1492 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.761 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 854, + 75, + 854, + 104, + 299, + 104 + ], + "score": 0.691 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.41 + }, + { + "category_id": 14, + "poly": [ + 648, + 1220, + 1051, + 1220, + 1051, + 1315, + 648, + 1315 + ], + "score": 0.96, + "latex": "l _ { b } ^ { m } = \\frac { 1 } { R } \\sum _ { r = 1 } ^ { R } l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t - r ) } ) ) ." + }, + { + "category_id": 13, + "poly": [ + 479, + 1081, + 700, + 1081, + 700, + 1122, + 479, + 1122 + ], + "score": 0.93, + "latex": "l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) )" + }, + { + "category_id": 13, + "poly": [ + 931, + 1041, + 980, + 1041, + 980, + 1083, + 931, + 1083 + ], + "score": 0.93, + "latex": "\\theta _ { s t u } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 994, + 1642, + 1054, + 1642, + 1054, + 1677, + 994, + 1677 + ], + "score": 0.92, + "latex": "\\{ B _ { s } ^ { l } \\}" + }, + { + "category_id": 13, + "poly": [ + 1211, + 1047, + 1327, + 1047, + 1327, + 1084, + 1211, + 1084 + ], + "score": 0.92, + "latex": "( X _ { m } ^ { l } , Y _ { m } )" + }, + { + "category_id": 13, + "poly": [ + 701, + 1411, + 816, + 1411, + 816, + 1447, + 701, + 1447 + ], + "score": 0.92, + "latex": "( X _ { m } ^ { l } , Y _ { m } ^ { l } )" + }, + { + "category_id": 14, + "poly": [ + 586, + 1452, + 1113, + 1452, + 1113, + 1498, + 586, + 1498 + ], + "score": 0.92, + "latex": "W _ { m } \\propto \\mathrm { m a x } ( l _ { b } ^ { m } - l ( Y _ { m } , f ( X _ { m } ^ { l } ; \\theta _ { s t u } ^ { ( t ) } ) ) , 0 ) + \\delta ." + }, + { + "category_id": 13, + "poly": [ + 1107, + 1147, + 1222, + 1147, + 1222, + 1182, + 1107, + 1182 + ], + "score": 0.92, + "latex": "( X _ { m } ^ { l } , Y _ { m } )" + }, + { + "category_id": 13, + "poly": [ + 554, + 1180, + 587, + 1180, + 587, + 1213, + 554, + 1213 + ], + "score": 0.9, + "latex": "l _ { b } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1181, + 1116, + 1181, + 1116, + 1213, + 1084, + 1213 + ], + "score": 0.9, + "latex": "l _ { b } ^ { m }" + }, + { + "category_id": 13, + "poly": [ + 751, + 1353, + 767, + 1353, + 767, + 1379, + 751, + 1379 + ], + "score": 0.82, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1537, + 1079, + 1537, + 1079, + 1563, + 1062, + 1563 + ], + "score": 0.81, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 563, + 1507, + 580, + 1507, + 580, + 1533, + 563, + 1533 + ], + "score": 0.8, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 1377, + 1149, + 1402, + 1149, + 1402, + 1176, + 1377, + 1176 + ], + "score": 0.79, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 297, + 1090, + 311, + 1090, + 311, + 1116, + 297, + 1116 + ], + "score": 0.72, + "latex": "t" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1780.0, + 855.0, + 1780.0, + 855.0, + 1818.0, + 293.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 369.0, + 855.0, + 369.0, + 855.0, + 405.0, + 295.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 667.0, + 1402.0, + 667.0, + 1402.0, + 701.0, + 295.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 697.0, + 1405.0, + 697.0, + 1405.0, + 731.0, + 295.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 727.0, + 1402.0, + 727.0, + 1402.0, + 761.0, + 295.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 759.0, + 1402.0, + 759.0, + 1402.0, + 789.0, + 295.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 788.0, + 1404.0, + 788.0, + 1404.0, + 822.0, + 295.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 820.0, + 1402.0, + 820.0, + 1402.0, + 851.0, + 293.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 850.0, + 1402.0, + 850.0, + 1402.0, + 880.0, + 296.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 878.0, + 1402.0, + 878.0, + 1402.0, + 913.0, + 295.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 911.0, + 1402.0, + 911.0, + 1402.0, + 943.0, + 294.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 943.0, + 1402.0, + 943.0, + 1402.0, + 974.0, + 297.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 970.0, + 1404.0, + 970.0, + 1404.0, + 1005.0, + 295.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 999.0, + 603.0, + 999.0, + 603.0, + 1036.0, + 296.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1581.0, + 1406.0, + 1581.0, + 1406.0, + 1617.0, + 296.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1612.0, + 1406.0, + 1612.0, + 1406.0, + 1649.0, + 293.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1642.0, + 993.0, + 1642.0, + 993.0, + 1678.0, + 294.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1642.0, + 1406.0, + 1642.0, + 1406.0, + 1678.0, + 1055.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1673.0, + 1406.0, + 1673.0, + 1406.0, + 1709.0, + 294.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1704.0, + 1406.0, + 1704.0, + 1406.0, + 1740.0, + 294.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1735.0, + 1402.0, + 1735.0, + 1402.0, + 1771.0, + 294.0, + 1771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1857.0, + 293.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1879.0, + 1405.0, + 1879.0, + 1405.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1943.0, + 1403.0, + 1943.0, + 1403.0, + 1976.0, + 293.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 598.0, + 2002.0, + 598.0, + 2036.0, + 294.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 407.0, + 1407.0, + 407.0, + 1407.0, + 441.0, + 294.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 437.0, + 1404.0, + 437.0, + 1404.0, + 471.0, + 294.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 467.0, + 1404.0, + 467.0, + 1404.0, + 498.0, + 294.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 501.0, + 1404.0, + 501.0, + 1404.0, + 530.0, + 297.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 527.0, + 1404.0, + 527.0, + 1404.0, + 563.0, + 296.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 556.0, + 1408.0, + 556.0, + 1408.0, + 595.0, + 293.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 589.0, + 1403.0, + 589.0, + 1403.0, + 623.0, + 294.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 618.0, + 1398.0, + 618.0, + 1398.0, + 655.0, + 296.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1030.0, + 930.0, + 1030.0, + 930.0, + 1098.0, + 286.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1030.0, + 1210.0, + 1030.0, + 1210.0, + 1098.0, + 981.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 1030.0, + 1414.0, + 1030.0, + 1414.0, + 1098.0, + 1328.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1072.0, + 296.0, + 1072.0, + 296.0, + 1136.0, + 285.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 1072.0, + 478.0, + 1072.0, + 478.0, + 1136.0, + 312.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 701.0, + 1072.0, + 1414.0, + 1072.0, + 1414.0, + 1136.0, + 701.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1117.0, + 1405.0, + 1117.0, + 1405.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1145.0, + 1106.0, + 1145.0, + 1106.0, + 1185.0, + 292.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 1145.0, + 1376.0, + 1145.0, + 1376.0, + 1185.0, + 1223.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1145.0, + 1408.0, + 1145.0, + 1408.0, + 1185.0, + 1403.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1175.0, + 553.0, + 1175.0, + 553.0, + 1214.0, + 291.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1175.0, + 1083.0, + 1175.0, + 1083.0, + 1214.0, + 588.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1175.0, + 1395.0, + 1175.0, + 1395.0, + 1214.0, + 1117.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1321.0, + 1403.0, + 1321.0, + 1403.0, + 1354.0, + 296.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1351.0, + 750.0, + 1351.0, + 750.0, + 1387.0, + 294.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 1351.0, + 1405.0, + 1351.0, + 1405.0, + 1387.0, + 768.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1383.0, + 1405.0, + 1383.0, + 1405.0, + 1416.0, + 293.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1410.0, + 700.0, + 1410.0, + 700.0, + 1451.0, + 292.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1410.0, + 950.0, + 1410.0, + 950.0, + 1451.0, + 817.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 229.0, + 1406.0, + 229.0, + 1406.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 292.0, + 296.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 326.0, + 296.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 324.0, + 756.0, + 324.0, + 756.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1505.0, + 562.0, + 1505.0, + 562.0, + 1540.0, + 297.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1505.0, + 1404.0, + 1505.0, + 1404.0, + 1540.0, + 581.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1535.0, + 1061.0, + 1535.0, + 1061.0, + 1571.0, + 296.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1535.0, + 1369.0, + 1535.0, + 1369.0, + 1571.0, + 1080.0, + 1571.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1117, + 1405, + 1117, + 1405, + 1428, + 296, + 1428 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 296, + 478, + 1406, + 478, + 1406, + 665, + 296, + 665 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 816, + 1404, + 816, + 1404, + 943, + 298, + 943 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1932, + 1404, + 1932, + 1404, + 2036, + 298, + 2036 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 223, + 1406, + 223, + 1406, + 329, + 297, + 329 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 689, + 1464, + 1010, + 1464, + 1010, + 1554, + 689, + 1554 + ], + "score": 0.962 + }, + { + "category_id": 8, + "poly": [ + 541, + 359, + 1154, + 359, + 1154, + 451, + 541, + 451 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 446, + 1676, + 1251, + 1676, + 1251, + 1770, + 446, + 1770 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 434, + 975, + 1266, + 975, + 1266, + 1075, + 434, + 1075 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 293, + 1581, + 1403, + 1581, + 1403, + 1646, + 293, + 1646 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 434, + 697, + 1263, + 697, + 1263, + 789, + 434, + 789 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 295, + 1813, + 1385, + 1813, + 1385, + 1847, + 295, + 1847 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 298, + 1879, + 830, + 1879, + 830, + 1913, + 298, + 1913 + ], + "score": 0.922 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1492, + 1400, + 1492, + 1400, + 1522, + 1365, + 1522 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1012, + 1400, + 1012, + 1400, + 1043, + 1366, + 1043 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 727, + 1400, + 727, + 1400, + 758, + 1366, + 758 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1365, + 390, + 1400, + 390, + 1400, + 421, + 1365, + 421 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1707, + 1400, + 1707, + 1400, + 1737, + 1365, + 1737 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 855, + 73, + 855, + 106, + 297, + 106 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2113, + 841, + 2113 + ], + "score": 0.752 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 855, + 74, + 855, + 106, + 297, + 106 + ], + "score": 0.107 + }, + { + "category_id": 14, + "poly": [ + 688, + 1459, + 1012, + 1459, + 1012, + 1554, + 688, + 1554 + ], + "score": 0.94, + "latex": "w _ { m , n } \\propto \\mathrm { m a x } ( \\sum _ { s = 1 } ^ { S } u _ { m , n , s } , 0 )" + }, + { + "category_id": 13, + "poly": [ + 1200, + 1213, + 1342, + 1213, + 1342, + 1246, + 1200, + 1246 + ], + "score": 0.94, + "latex": "\\{ B _ { 1 } ^ { l } , . . . , B _ { S } ^ { l } \\}" + }, + { + "category_id": 13, + "poly": [ + 864, + 223, + 1208, + 223, + 1208, + 269, + 864, + 269 + ], + "score": 0.94, + "latex": "\\{ \\hat { Y } _ { m } ^ { ( t ) } = [ \\hat { y } _ { m , 1 } ^ { ( t ) } , . . . , \\hat { y } _ { m , N } ^ { ( t ) } ] \\} _ { m = 1 } ^ { M }" + }, + { + "category_id": 14, + "poly": [ + 544, + 357, + 1155, + 357, + 1155, + 453, + 544, + 453 + ], + "score": 0.94, + "latex": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\sum _ { m = 1 } ^ { M } l ( \\hat { Y } _ { m } ^ { ( t ) } , f ( X _ { m } ^ { u } ; \\theta _ { s t u } ^ { ( t - 1 ) } ) ) \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 437, + 694, + 1263, + 694, + 1263, + 791, + 437, + 791 + ], + "score": 0.94, + "latex": "\\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ \\epsilon _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 449, + 1674, + 1251, + 1674, + 1251, + 1771, + 449, + 1771 + ], + "score": 0.94, + "latex": "\\hat { \\theta } _ { s t u } ^ { ( t ) } = \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } - \\alpha \\nabla \\big ( \\frac { 1 } { M } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { M } \\sum _ { n = 1 } ^ { N } [ w _ { m , n } \\cdot l ( \\hat { y } _ { m , n } ^ { ( t ) } , f ( x _ { m , n } ^ { u } ; \\hat { \\theta } _ { s t u } ^ { ( t - 1 ) } ) ) ] \\big ) ." + }, + { + "category_id": 14, + "poly": [ + 436, + 973, + 1264, + 973, + 1264, + 1075, + 436, + 1075 + ], + "score": 0.93, + "latex": "u _ { m , n , s } = - \\frac { \\partial } { \\partial \\epsilon _ { m , n , s } } \\big ( \\frac { 1 } { | \\mathcal { B } _ { s } ^ { l } | } \\frac { 1 } { N } \\sum _ { m = 1 } ^ { | \\mathcal { B } _ { s } ^ { l } | } \\sum _ { n = 1 } ^ { N } [ l ( y _ { m , n } , f ( x _ { m , n } ^ { l } ; \\hat { \\theta } _ { s t u } ^ { ( t ) } ( \\epsilon ) ) ] \\big ) | _ { \\epsilon _ { m , n , s } = 0 }" + }, + { + "category_id": 13, + "poly": [ + 904, + 2004, + 995, + 2004, + 995, + 2036, + 904, + 2036 + ], + "score": 0.93, + "latex": "( X ^ { l } , Y )" + }, + { + "category_id": 13, + "poly": [ + 355, + 1965, + 498, + 1965, + 498, + 2007, + 355, + 2007 + ], + "score": 0.92, + "latex": "\\theta _ { t e a } = \\theta _ { s t u } ^ { ( T ) } )" + }, + { + "category_id": 13, + "poly": [ + 367, + 1396, + 513, + 1396, + 513, + 1431, + 367, + 1431 + ], + "score": 0.91, + "latex": "\\left( x _ { m , n } ^ { u } , \\hat { y } _ { m , n } \\right)" + }, + { + "category_id": 13, + "poly": [ + 468, + 878, + 500, + 878, + 500, + 912, + 468, + 912 + ], + "score": 0.9, + "latex": "{ \\dot { B } } _ { s } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1352, + 1339, + 1372, + 1339, + 1372, + 1360, + 1352, + 1360 + ], + "score": 0.89, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1022, + 908, + 1055, + 908, + 1055, + 942, + 1022, + 942 + ], + "score": 0.89, + "latex": "B _ { s } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 1330, + 1587, + 1393, + 1587, + 1393, + 1617, + 1330, + 1617 + ], + "score": 0.87, + "latex": "w _ { m , n }" + }, + { + "category_id": 13, + "poly": [ + 515, + 1276, + 536, + 1276, + 536, + 1299, + 515, + 1299 + ], + "score": 0.84, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 842, + 1215, + 863, + 1215, + 863, + 1238, + 842, + 1238 + ], + "score": 0.81, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 650, + 271, + 670, + 271, + 670, + 293, + 650, + 293 + ], + "score": 0.77, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 367, + 268, + 381, + 268, + 381, + 293, + 367, + 293 + ], + "score": 0.76, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 571, + 270, + 599, + 270, + 599, + 293, + 571, + 293 + ], + "score": 0.75, + "latex": "m" + }, + { + "category_id": 13, + "poly": [ + 1289, + 884, + 1301, + 884, + 1301, + 906, + 1289, + 906 + ], + "score": 0.73, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 666, + 883, + 679, + 883, + 679, + 906, + 666, + 906 + ], + "score": 0.66, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1046, + 552, + 1059, + 552, + 1059, + 569, + 1046, + 569 + ], + "score": 0.64, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1878.0, + 833.0, + 1878.0, + 833.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 855.0, + 72.0, + 855.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 856.0, + 72.0, + 856.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1116.0, + 1407.0, + 1116.0, + 1407.0, + 1153.0, + 295.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1148.0, + 1405.0, + 1148.0, + 1405.0, + 1184.0, + 295.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1215.0, + 295.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1208.0, + 841.0, + 1208.0, + 841.0, + 1244.0, + 295.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1208.0, + 1199.0, + 1208.0, + 1199.0, + 1244.0, + 864.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1208.0, + 1403.0, + 1208.0, + 1403.0, + 1244.0, + 1343.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1238.0, + 1406.0, + 1238.0, + 1406.0, + 1275.0, + 295.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1270.0, + 514.0, + 1270.0, + 514.0, + 1305.0, + 295.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1270.0, + 1402.0, + 1270.0, + 1402.0, + 1305.0, + 537.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1302.0, + 1402.0, + 1302.0, + 1402.0, + 1334.0, + 295.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1333.0, + 1351.0, + 1333.0, + 1351.0, + 1365.0, + 298.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1333.0, + 1403.0, + 1333.0, + 1403.0, + 1365.0, + 1373.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1362.0, + 1405.0, + 1362.0, + 1405.0, + 1395.0, + 298.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1391.0, + 366.0, + 1391.0, + 366.0, + 1432.0, + 295.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1391.0, + 687.0, + 1391.0, + 687.0, + 1432.0, + 514.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1403.0, + 478.0, + 1403.0, + 514.0, + 294.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 507.0, + 1399.0, + 507.0, + 1399.0, + 545.0, + 293.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 539.0, + 1045.0, + 539.0, + 1045.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 539.0, + 1404.0, + 539.0, + 1404.0, + 575.0, + 1060.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 572.0, + 1404.0, + 572.0, + 1404.0, + 604.0, + 295.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 601.0, + 1406.0, + 601.0, + 1406.0, + 637.0, + 296.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 630.0, + 1290.0, + 630.0, + 1290.0, + 668.0, + 294.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 817.0, + 1402.0, + 817.0, + 1402.0, + 850.0, + 296.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 849.0, + 1405.0, + 849.0, + 1405.0, + 882.0, + 294.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 876.0, + 467.0, + 876.0, + 467.0, + 914.0, + 292.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 876.0, + 665.0, + 876.0, + 665.0, + 914.0, + 501.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 876.0, + 1288.0, + 876.0, + 1288.0, + 914.0, + 680.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 876.0, + 1405.0, + 876.0, + 1405.0, + 914.0, + 1302.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 908.0, + 1021.0, + 908.0, + 1021.0, + 945.0, + 292.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 908.0, + 1075.0, + 908.0, + 1075.0, + 945.0, + 1056.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1933.0, + 1405.0, + 1933.0, + 1405.0, + 1969.0, + 296.0, + 1969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1950.0, + 354.0, + 1950.0, + 354.0, + 2026.0, + 283.0, + 2026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 1950.0, + 1419.0, + 1950.0, + 1419.0, + 2026.0, + 499.0, + 2026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 903.0, + 2001.0, + 903.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2036.0, + 996.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 218.0, + 863.0, + 218.0, + 863.0, + 281.0, + 283.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 221.0, + 1410.0, + 221.0, + 1410.0, + 273.0, + 1209.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 263.0, + 366.0, + 263.0, + 366.0, + 300.0, + 292.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 263.0, + 570.0, + 263.0, + 570.0, + 300.0, + 382.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 263.0, + 649.0, + 263.0, + 649.0, + 300.0, + 600.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 263.0, + 1406.0, + 263.0, + 1406.0, + 300.0, + 671.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 297.0, + 1311.0, + 297.0, + 1311.0, + 329.0, + 296.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1573.0, + 1329.0, + 1573.0, + 1329.0, + 1625.0, + 292.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1573.0, + 1404.0, + 1573.0, + 1404.0, + 1625.0, + 1394.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1613.0, + 1391.0, + 1613.0, + 1391.0, + 1650.0, + 295.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1809.0, + 1388.0, + 1809.0, + 1388.0, + 1850.0, + 296.0, + 1850.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1668, + 1404, + 1668, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 828, + 1027, + 1382, + 1027, + 1382, + 1281, + 828, + 1281 + ], + "score": 0.974, + "html": "
Dataset# Slots#Train#Test#Lang
Email202.5K1kEN
SNIPS3913K0.7KEN
MIT Movie128.8K2.4KEN
MIT Restaurant86.9K1.5KEN
Wikiann (EN)320K10KEN
CoNLL03 (EN)415K3.6KEN
CoNLL031638K15K4
Wikiann123705K329K41
" + }, + { + "category_id": 1, + "poly": [ + 297, + 873, + 1403, + 873, + 1403, + 997, + 297, + 997 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 298, + 1318, + 1405, + 1318, + 1405, + 1651, + 298, + 1651 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 299, + 1013, + 807, + 1013, + 807, + 1316, + 299, + 1316 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1403, + 229, + 1403, + 295, + 293, + 295 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 821, + 559, + 821, + 559, + 855, + 299, + 855 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.896 + }, + { + "category_id": 7, + "poly": [ + 968, + 1286, + 1266, + 1286, + 1266, + 1315, + 968, + 1315 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 300, + 322, + 680, + 322, + 680, + 355, + 300, + 355 + ], + "score": 0.782 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.771 + }, + { + "category_id": 5, + "poly": [ + 296, + 363, + 1271, + 363, + 1271, + 797, + 296, + 797 + ], + "score": 0.323, + "html": "
Input:Labeled sequences (Xl,Y); Unlabeled sequences (X");Pre-trained BERT model with randomly initialized token classification layer f(;((O)); Batches S; Number of self-training iterations T.
Initialize teacher model 0tea =0(0) while not converged do
Fine-tune teacher model on small labeled data (Xl, Y);
Initialize the student model θ(0)=0(0);
Generate hard pseudo-labels Y(t）for unlabeled samples Xu with model f(·,0tea);
fort←1toTdo
Compute labeled data acquisition function according to Eq. 4;
function;Sample S mini-batches of labeled examples {Bl,.,B's} from (Xl,Y）based on labeled data acquisition
Randomly sample abatch of pseudo-labeled examples B from(Xu,Y(t));
Compute token-level weights in B based on the loss on {Bl,..,B's} according to Eq.8;
endstu
Update the teacher:θtea
end
" + }, + { + "category_id": 1, + "poly": [ + 296, + 363, + 1271, + 363, + 1271, + 797, + 296, + 797 + ], + "score": 0.278 + }, + { + "category_id": 13, + "poly": [ + 746, + 585, + 866, + 585, + 866, + 614, + 746, + 614 + ], + "score": 0.93, + "latex": "\\{ \\boldsymbol { B } _ { 1 } ^ { l } , . . . , \\boldsymbol { B } _ { S } ^ { l } \\}" + }, + { + "category_id": 13, + "poly": [ + 830, + 658, + 947, + 658, + 947, + 685, + 830, + 685 + ], + "score": 0.92, + "latex": "\\{ B _ { 1 } ^ { l } , . . . , B _ { S } ^ { l } \\}" + }, + { + "category_id": 13, + "poly": [ + 298, + 225, + 403, + 225, + 403, + 265, + 298, + 265 + ], + "score": 0.92, + "latex": "f ( \\cdot , \\theta _ { t e a } ^ { ( t ) } )" + }, + { + "category_id": 13, + "poly": [ + 483, + 684, + 572, + 684, + 572, + 714, + 483, + 714 + ], + "score": 0.91, + "latex": "f ( \\cdot , \\theta _ { s t u } ^ { ( t ) } )" + }, + { + "category_id": 13, + "poly": [ + 917, + 585, + 991, + 585, + 991, + 613, + 917, + 613 + ], + "score": 0.91, + "latex": "( X ^ { l } , Y )" + }, + { + "category_id": 13, + "poly": [ + 704, + 464, + 778, + 464, + 778, + 491, + 704, + 491 + ], + "score": 0.91, + "latex": "( X ^ { l } , Y )" + }, + { + "category_id": 13, + "poly": [ + 517, + 1532, + 757, + 1532, + 757, + 1563, + 517, + 1563 + ], + "score": 0.91, + "latex": "K \\in \\{ 5 , 1 0 , 2 0 , 1 0 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 918, + 520, + 1006, + 520, + 1006, + 546, + 918, + 546 + ], + "score": 0.9, + "latex": "f ( \\cdot , \\theta _ { t e a } )" + }, + { + "category_id": 13, + "poly": [ + 507, + 390, + 591, + 390, + 591, + 417, + 507, + 417 + ], + "score": 0.89, + "latex": "f ( \\cdot ; \\theta ^ { ( 0 ) } )" + }, + { + "category_id": 13, + "poly": [ + 486, + 417, + 593, + 417, + 593, + 442, + 486, + 442 + ], + "score": 0.89, + "latex": "\\theta _ { t e a } = \\theta ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 557, + 489, + 665, + 489, + 665, + 519, + 557, + 519 + ], + "score": 0.88, + "latex": "\\theta _ { s t u } ^ { ( 0 ) } = \\theta ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 1093, + 684, + 1134, + 684, + 1134, + 714, + 1093, + 714 + ], + "score": 0.88, + "latex": "\\theta _ { s t u } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 881, + 689, + 910, + 689, + 910, + 712, + 881, + 712 + ], + "score": 0.87, + "latex": "\\scriptstyle { B _ { u } }" + }, + { + "category_id": 13, + "poly": [ + 635, + 661, + 664, + 661, + 664, + 683, + 635, + 683 + ], + "score": 0.86, + "latex": "\\scriptstyle { B _ { u } }" + }, + { + "category_id": 13, + "poly": [ + 508, + 364, + 581, + 364, + 581, + 390, + 508, + 390 + ], + "score": 0.86, + "latex": "( X ^ { l } , Y )" + }, + { + "category_id": 13, + "poly": [ + 497, + 736, + 609, + 736, + 609, + 768, + 497, + 768 + ], + "score": 0.86, + "latex": "\\theta _ { t e a } = \\theta _ { s t u } ^ { ( T ) }" + }, + { + "category_id": 13, + "poly": [ + 370, + 544, + 426, + 544, + 426, + 563, + 370, + 563 + ], + "score": 0.85, + "latex": "t \\gets 1" + }, + { + "category_id": 13, + "poly": [ + 886, + 631, + 991, + 631, + 991, + 658, + 886, + 658 + ], + "score": 0.85, + "latex": "( X ^ { u } , \\hat { Y } ^ { ( t ) } )" + }, + { + "category_id": 13, + "poly": [ + 567, + 519, + 609, + 519, + 609, + 541, + 567, + 541 + ], + "score": 0.84, + "latex": "\\hat { Y } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 810, + 634, + 839, + 634, + 839, + 657, + 810, + 657 + ], + "score": 0.84, + "latex": "\\scriptstyle { B _ { u } }" + }, + { + "category_id": 13, + "poly": [ + 788, + 521, + 822, + 521, + 822, + 542, + 788, + 542 + ], + "score": 0.82, + "latex": "X ^ { u }" + }, + { + "category_id": 13, + "poly": [ + 760, + 366, + 810, + 366, + 810, + 391, + 760, + 391 + ], + "score": 0.82, + "latex": "( X ^ { u } )" + }, + { + "category_id": 13, + "poly": [ + 466, + 1731, + 495, + 1731, + 495, + 1756, + 466, + 1756 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 958, + 394, + 975, + 394, + 975, + 413, + 958, + 413 + ], + "score": 0.79, + "latex": "_ T" + }, + { + "category_id": 13, + "poly": [ + 897, + 1593, + 926, + 1593, + 926, + 1618, + 897, + 1618 + ], + "score": 0.79, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 450, + 545, + 468, + 545, + 468, + 563, + 450, + 563 + ], + "score": 0.78, + "latex": "_ T" + }, + { + "category_id": 13, + "poly": [ + 667, + 394, + 683, + 394, + 683, + 414, + 667, + 414 + ], + "score": 0.77, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 449, + 589, + 465, + 589, + 465, + 608, + 449, + 608 + ], + "score": 0.75, + "latex": "S" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 818.0, + 561.0, + 818.0, + 561.0, + 861.0, + 292.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1281.0, + 1268.0, + 1281.0, + 1268.0, + 1321.0, + 964.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 680.0, + 317.0, + 680.0, + 361.0, + 295.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1405.0, + 1665.0, + 1405.0, + 1703.0, + 294.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1699.0, + 1404.0, + 1699.0, + 1404.0, + 1730.0, + 296.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1728.0, + 465.0, + 1728.0, + 465.0, + 1762.0, + 295.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1762.0, + 496.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1793.0, + 295.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1789.0, + 1404.0, + 1789.0, + 1404.0, + 1823.0, + 295.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 295.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1883.0, + 1402.0, + 1883.0, + 1402.0, + 1913.0, + 296.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1944.0, + 297.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2004.0, + 294.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1397.0, + 2001.0, + 1397.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 873.0, + 1404.0, + 873.0, + 1404.0, + 906.0, + 295.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 902.0, + 1404.0, + 902.0, + 1404.0, + 939.0, + 294.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 936.0, + 1403.0, + 936.0, + 1403.0, + 969.0, + 295.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 964.0, + 1360.0, + 964.0, + 1360.0, + 1002.0, + 292.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1314.0, + 1406.0, + 1314.0, + 1406.0, + 1349.0, + 295.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1345.0, + 1406.0, + 1345.0, + 1406.0, + 1381.0, + 295.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1375.0, + 1405.0, + 1375.0, + 1405.0, + 1411.0, + 293.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1405.0, + 1407.0, + 1405.0, + 1407.0, + 1441.0, + 293.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1434.0, + 1406.0, + 1434.0, + 1406.0, + 1475.0, + 293.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1467.0, + 1403.0, + 1467.0, + 1403.0, + 1502.0, + 296.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1497.0, + 1403.0, + 1497.0, + 1403.0, + 1534.0, + 293.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1529.0, + 516.0, + 1529.0, + 516.0, + 1563.0, + 296.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1529.0, + 1403.0, + 1529.0, + 1403.0, + 1563.0, + 758.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1594.0, + 295.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1590.0, + 896.0, + 1590.0, + 896.0, + 1625.0, + 296.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1590.0, + 1403.0, + 1590.0, + 1403.0, + 1625.0, + 927.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1137.0, + 1620.0, + 1137.0, + 1654.0, + 295.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1009.0, + 808.0, + 1009.0, + 808.0, + 1045.0, + 295.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1041.0, + 809.0, + 1041.0, + 809.0, + 1076.0, + 293.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1072.0, + 809.0, + 1072.0, + 809.0, + 1107.0, + 294.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1099.0, + 810.0, + 1099.0, + 810.0, + 1139.0, + 293.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1131.0, + 810.0, + 1131.0, + 810.0, + 1166.0, + 294.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1162.0, + 810.0, + 1162.0, + 810.0, + 1198.0, + 293.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1194.0, + 809.0, + 1194.0, + 809.0, + 1226.0, + 294.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1222.0, + 810.0, + 1222.0, + 810.0, + 1259.0, + 293.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1253.0, + 810.0, + 1253.0, + 810.0, + 1290.0, + 295.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 811.0, + 1285.0, + 811.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 218.0, + 297.0, + 218.0, + 297.0, + 279.0, + 291.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 218.0, + 1411.0, + 218.0, + 1411.0, + 279.0, + 404.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 262.0, + 1361.0, + 262.0, + 1361.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 362.0, + 507.0, + 362.0, + 507.0, + 393.0, + 295.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 362.0, + 759.0, + 362.0, + 759.0, + 393.0, + 582.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 362.0, + 1266.0, + 362.0, + 1266.0, + 393.0, + 811.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 388.0, + 506.0, + 388.0, + 506.0, + 419.0, + 354.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 388.0, + 666.0, + 388.0, + 666.0, + 419.0, + 592.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 388.0, + 957.0, + 388.0, + 957.0, + 419.0, + 684.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 388.0, + 985.0, + 388.0, + 985.0, + 419.0, + 976.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 414.0, + 485.0, + 414.0, + 485.0, + 444.0, + 292.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 414.0, + 598.0, + 414.0, + 598.0, + 444.0, + 594.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 441.0, + 488.0, + 441.0, + 488.0, + 465.0, + 296.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 462.0, + 703.0, + 462.0, + 703.0, + 493.0, + 334.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 462.0, + 787.0, + 462.0, + 787.0, + 493.0, + 779.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 489.0, + 556.0, + 489.0, + 556.0, + 521.0, + 334.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 489.0, + 676.0, + 489.0, + 676.0, + 521.0, + 666.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.0, + 513.0, + 566.0, + 513.0, + 566.0, + 550.0, + 335.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 513.0, + 787.0, + 513.0, + 787.0, + 550.0, + 610.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 513.0, + 917.0, + 513.0, + 917.0, + 550.0, + 823.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 513.0, + 1016.0, + 513.0, + 1016.0, + 550.0, + 1007.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 542.0, + 369.0, + 542.0, + 369.0, + 567.0, + 338.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 542.0, + 449.0, + 542.0, + 449.0, + 567.0, + 427.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 542.0, + 498.0, + 542.0, + 498.0, + 567.0, + 469.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 563.0, + 871.0, + 563.0, + 871.0, + 588.0, + 381.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 379.0, + 584.0, + 448.0, + 584.0, + 448.0, + 616.0, + 379.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 584.0, + 745.0, + 584.0, + 745.0, + 616.0, + 466.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 584.0, + 916.0, + 584.0, + 916.0, + 616.0, + 867.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 584.0, + 1271.0, + 584.0, + 1271.0, + 616.0, + 992.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 608.0, + 460.0, + 608.0, + 460.0, + 636.0, + 380.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 630.0, + 809.0, + 630.0, + 809.0, + 660.0, + 377.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 630.0, + 885.0, + 630.0, + 885.0, + 660.0, + 840.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 630.0, + 1005.0, + 630.0, + 1005.0, + 660.0, + 992.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 657.0, + 634.0, + 657.0, + 634.0, + 688.0, + 380.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 657.0, + 829.0, + 657.0, + 829.0, + 688.0, + 665.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 657.0, + 1107.0, + 657.0, + 1107.0, + 688.0, + 948.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 675.0, + 482.0, + 675.0, + 482.0, + 727.0, + 373.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 675.0, + 880.0, + 675.0, + 880.0, + 727.0, + 573.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 675.0, + 1092.0, + 675.0, + 1092.0, + 727.0, + 911.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 675.0, + 1154.0, + 675.0, + 1154.0, + 727.0, + 1135.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 712.0, + 496.0, + 712.0, + 496.0, + 790.0, + 295.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 712.0, + 626.0, + 712.0, + 626.0, + 790.0, + 610.0, + 790.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 295, + 466, + 1399, + 466, + 1399, + 907, + 295, + 907 + ], + "score": 0.984, + "html": "
MethodSNIPSEmailMovieRestaurantCoNLL03 (EN)Wikiann (EN)
# Slots392012843
Full-supervision
BERT95.8094.4487.8778.9592.4084.04
Few-shot supervision (10 labels per slot)69.5054.0671.1545.61
BERT 79.01 87.85
Few-shot supervision (10 labels per slot) + unlabeled data
CVT78.2378.2462.7342.5754.3127.89
SeqVAT78.6772.6567.1051.5567.2135.16
MT79.4889.5367.6251.7568.6741.43
VAT79.0889.7170.1753.3465.0338.81
Classic ST83.2690.7071.8856.8070.9946.15
BOND83.5489.7570.9155.7869.5648.73
MetaST88.2392.1877.6763.8376.65
(0.04;↑12%)(0.47;↑4.93%)(0.10;↑11.76%)(1.62;↑18.07%)(0.73;17.73%)56.61 (0.4;↑24.12%)
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1032, + 1404, + 1032, + 1404, + 1460, + 297, + 1460 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 301, + 1473, + 1399, + 1473, + 1399, + 1610, + 301, + 1610 + ], + "score": 0.971, + "html": "
Dataset#Lang#SlotsFull Sup.Few-shot Sup.Few-shot supervision + unlabeled data
BERTBERTMTVATClassic STBONDMetaST
CoNLL0341687.6770.7768.3467.6372.6972.7976.41 (0.47) (↑ 7.97%)
Wikiann4112387.1779.6780.2378.8280.2479.5781.61 (0.14) (↑ 2.42%)
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1399, + 1972, + 1399, + 2034, + 297, + 2034 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1741, + 1404, + 1741, + 1404, + 1958, + 298, + 1958 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 446, + 297, + 446 + ], + "score": 0.777 + }, + { + "category_id": 1, + "poly": [ + 301, + 1632, + 1399, + 1632, + 1399, + 1726, + 301, + 1726 + ], + "score": 0.756 + }, + { + "category_id": 1, + "poly": [ + 297, + 922, + 1405, + 922, + 1405, + 1016, + 297, + 1016 + ], + "score": 0.729 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.647 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.63 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 859, + 2087, + 859, + 2111, + 842, + 2111 + ], + "score": 0.343 + }, + { + "category_id": 6, + "poly": [ + 297, + 229, + 1403, + 229, + 1403, + 446, + 297, + 446 + ], + "score": 0.333 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 854, + 75, + 854, + 104, + 299, + 104 + ], + "score": 0.317 + }, + { + "category_id": 6, + "poly": [ + 301, + 1632, + 1399, + 1632, + 1399, + 1726, + 301, + 1726 + ], + "score": 0.25 + }, + { + "category_id": 6, + "poly": [ + 297, + 922, + 1405, + 922, + 1405, + 1016, + 297, + 1016 + ], + "score": 0.168 + }, + { + "category_id": 13, + "poly": [ + 980, + 1186, + 1114, + 1186, + 1114, + 1215, + 980, + 1215 + ], + "score": 0.89, + "latex": "K \\times \\# S l o t s" + }, + { + "category_id": 13, + "poly": [ + 1064, + 382, + 1154, + 382, + 1154, + 413, + 1064, + 413 + ], + "score": 0.88, + "latex": "1 4 . 2 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1169, + 382, + 1259, + 382, + 1259, + 413, + 1169, + 413 + ], + "score": 0.88, + "latex": "1 4 . 9 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1273, + 382, + 1351, + 382, + 1351, + 413, + 1273, + 413 + ], + "score": 0.87, + "latex": "8 . 4 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 296, + 413, + 373, + 413, + 373, + 444, + 296, + 444 + ], + "score": 0.86, + "latex": "8 . 8 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 755, + 262, + 823, + 262, + 823, + 291, + 755, + 291 + ], + "score": 0.85, + "latex": "\\mathrm { K } { = } 1 0" + }, + { + "category_id": 13, + "poly": [ + 766, + 1974, + 795, + 1974, + 795, + 2000, + 766, + 2000 + ], + "score": 0.77, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 760, + 414, + 856, + 414, + 856, + 445, + 760, + 445 + ], + "score": 0.62, + "latex": "{ \\tt S e q V A T }" + }, + { + "category_id": 13, + "poly": [ + 693, + 669, + 712, + 669, + 712, + 688, + 693, + 688 + ], + "score": 0.43, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 457, + 953, + 555, + 953, + 555, + 984, + 457, + 984 + ], + "score": 0.32, + "latex": "{ \\tt S e q V A T } ," + }, + { + "category_id": 13, + "poly": [ + 1129, + 1484, + 1147, + 1484, + 1147, + 1502, + 1129, + 1502 + ], + "score": 0.31, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 754.0, + 262.0, + 754.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 297.0, + 824.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 326.0, + 292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 354.0, + 1404.0, + 354.0, + 1404.0, + 385.0, + 295.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1063.0, + 381.0, + 1063.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 381.0, + 1168.0, + 381.0, + 1168.0, + 419.0, + 1155.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 381.0, + 1272.0, + 381.0, + 1272.0, + 419.0, + 1260.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 381.0, + 1407.0, + 381.0, + 1407.0, + 419.0, + 1352.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 411.0, + 295.0, + 411.0, + 295.0, + 449.0, + 292.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 411.0, + 759.0, + 411.0, + 759.0, + 449.0, + 374.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 411.0, + 1345.0, + 411.0, + 1345.0, + 449.0, + 857.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1669.0, + 295.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1695.0, + 1367.0, + 1695.0, + 1367.0, + 1729.0, + 297.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 919.0, + 1405.0, + 919.0, + 1405.0, + 958.0, + 292.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 953.0, + 456.0, + 953.0, + 456.0, + 987.0, + 294.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 953.0, + 1406.0, + 953.0, + 1406.0, + 987.0, + 556.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 985.0, + 1338.0, + 985.0, + 1338.0, + 1019.0, + 295.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1031.0, + 1404.0, + 1031.0, + 1404.0, + 1068.0, + 294.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1099.0, + 294.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1091.0, + 1406.0, + 1091.0, + 1406.0, + 1128.0, + 292.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1124.0, + 1405.0, + 1124.0, + 1405.0, + 1160.0, + 294.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1155.0, + 1406.0, + 1155.0, + 1406.0, + 1190.0, + 294.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1185.0, + 979.0, + 1185.0, + 979.0, + 1221.0, + 294.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1185.0, + 1405.0, + 1185.0, + 1405.0, + 1221.0, + 1115.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1216.0, + 1406.0, + 1216.0, + 1406.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1246.0, + 1406.0, + 1246.0, + 1406.0, + 1282.0, + 295.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1273.0, + 1407.0, + 1273.0, + 1407.0, + 1313.0, + 292.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1304.0, + 1406.0, + 1304.0, + 1406.0, + 1346.0, + 292.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1337.0, + 1405.0, + 1337.0, + 1405.0, + 1373.0, + 294.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1365.0, + 1406.0, + 1365.0, + 1406.0, + 1402.0, + 294.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1434.0, + 295.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1425.0, + 1298.0, + 1425.0, + 1298.0, + 1467.0, + 292.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1966.0, + 765.0, + 1966.0, + 765.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1966.0, + 1404.0, + 1966.0, + 1404.0, + 2011.0, + 796.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1999.0, + 1404.0, + 1999.0, + 1404.0, + 2038.0, + 295.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1774.0, + 298.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1807.0, + 294.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1801.0, + 1406.0, + 1801.0, + 1406.0, + 1837.0, + 293.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1836.0, + 1405.0, + 1836.0, + 1405.0, + 1867.0, + 296.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1859.0, + 1405.0, + 1859.0, + 1405.0, + 1903.0, + 292.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1890.0, + 1404.0, + 1890.0, + 1404.0, + 1931.0, + 293.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1926.0, + 558.0, + 1926.0, + 558.0, + 1957.0, + 296.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 754.0, + 262.0, + 754.0, + 297.0, + 295.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 297.0, + 824.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 326.0, + 292.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 321.0, + 1407.0, + 321.0, + 1407.0, + 356.0, + 293.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 354.0, + 1404.0, + 354.0, + 1404.0, + 385.0, + 295.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1063.0, + 381.0, + 1063.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 381.0, + 1168.0, + 381.0, + 1168.0, + 419.0, + 1155.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 381.0, + 1272.0, + 381.0, + 1272.0, + 419.0, + 1260.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 381.0, + 1407.0, + 381.0, + 1407.0, + 419.0, + 1352.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 411.0, + 295.0, + 411.0, + 295.0, + 449.0, + 292.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 411.0, + 759.0, + 411.0, + 759.0, + 449.0, + 374.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 411.0, + 1345.0, + 411.0, + 1345.0, + 449.0, + 857.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1669.0, + 295.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 1405.0, + 1662.0, + 1405.0, + 1697.0, + 292.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1695.0, + 1367.0, + 1695.0, + 1367.0, + 1729.0, + 297.0, + 1729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 919.0, + 1405.0, + 919.0, + 1405.0, + 958.0, + 292.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 953.0, + 456.0, + 953.0, + 456.0, + 987.0, + 294.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 953.0, + 1406.0, + 953.0, + 1406.0, + 987.0, + 556.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 985.0, + 1338.0, + 985.0, + 1338.0, + 1019.0, + 295.0, + 1019.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1732, + 1404, + 1732, + 1404, + 2038, + 297, + 2038 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 482, + 1404, + 482, + 1404, + 790, + 298, + 790 + ], + "score": 0.98 + }, + { + "category_id": 5, + "poly": [ + 356, + 806, + 1349, + 806, + 1349, + 980, + 356, + 980 + ], + "score": 0.98, + "html": "
#SlotsFew-shot SupervisionFew-shot supervision + unlabeled data
BERTCVTSeqVATMTVATClassic STBONDMetaST(%Improvement)
570.6369.8269.3470.8571.3472.5972.8581.56 (↑15%)
1079.0178.2378.6779.4879.0883.2683.5488.22 (↑12%)
2086.8188.0485.0587.3188.1988.3288.9391.99 (↑6%)
10093.9094.6191.4694.2694.5393.9294.2295.39 (12%)
" + }, + { + "category_id": 5, + "poly": [ + 383, + 204, + 1316, + 204, + 1316, + 382, + 383, + 382 + ], + "score": 0.98, + "html": "
DatasetBERT (Full Supervision)BERT (Few-shot Supervision)MetaST（%Improvement）
MIT Movie87.8775.8180.33 (↑5.96%)
MIT Restaurant78.9560.1267.86 (↑12.87%)
CoNLL03 (EN)92.4077.4881.61 (↑ 5.33%)
Wikiann (EN)84.0462.0471.27 (个14.88%)
Average85.8268.8675.27 (↑ 9.31%)
" + }, + { + "category_id": 5, + "poly": [ + 354, + 1229, + 912, + 1229, + 912, + 1634, + 354, + 1634 + ], + "score": 0.973, + "html": "
MethodDatasets
SNIPSCoNLL03
BERT w/ Continued Pre-training +
Few-shot Supervision83.9669.84
Classic ST83.2670.99
Classic ST w/ Soft Pseudo-Labels81.1771.87
MetaST(ours) w/Hard Pseudo-Labels88.2376.65
MetaST w/ Soft Pseudo-Labels86.1675.84
MetaST w/o Iterative Teacher Fine-tune85.6472.74
MetaST w/o Labeled Data Acq.86.6375.02
Pseudo-labeled Data Re-weighting
MetaST w/o Re-weighting85.4873.02
MetaST (Easy)85.5674.53
MetaST (Difficult)86.3468.06
" + }, + { + "category_id": 3, + "poly": [ + 971, + 1285, + 1403, + 1285, + 1403, + 1578, + 971, + 1578 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 298, + 1066, + 1402, + 1066, + 1402, + 1221, + 298, + 1221 + ], + "score": 0.929 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.915 + }, + { + "category_id": 4, + "poly": [ + 970, + 1605, + 1391, + 1605, + 1391, + 1667, + 970, + 1667 + ], + "score": 0.897 + }, + { + "category_id": 6, + "poly": [ + 300, + 1657, + 965, + 1657, + 965, + 1720, + 300, + 1720 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.801 + }, + { + "category_id": 6, + "poly": [ + 296, + 403, + 1398, + 403, + 1398, + 466, + 296, + 466 + ], + "score": 0.455 + }, + { + "category_id": 1, + "poly": [ + 296, + 403, + 1398, + 403, + 1398, + 466, + 296, + 466 + ], + "score": 0.385 + }, + { + "category_id": 7, + "poly": [ + 299, + 992, + 1400, + 992, + 1400, + 1054, + 299, + 1054 + ], + "score": 0.333 + }, + { + "category_id": 13, + "poly": [ + 297, + 697, + 386, + 697, + 386, + 726, + 297, + 726 + ], + "score": 0.88, + "latex": "8 8 . 2 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 436, + 697, + 524, + 697, + 524, + 726, + 436, + 726 + ], + "score": 0.88, + "latex": "9 5 . 3 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 1122, + 697, + 1198, + 697, + 1198, + 726, + 1122, + 726 + ], + "score": 0.87, + "latex": "9 5 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1255, + 636, + 1309, + 636, + 1309, + 665, + 1255, + 665 + ], + "score": 0.86, + "latex": "12 \\%" + }, + { + "category_id": 13, + "poly": [ + 1360, + 636, + 1401, + 636, + 1401, + 664, + 1360, + 664 + ], + "score": 0.84, + "latex": "2 \\%" + }, + { + "category_id": 13, + "poly": [ + 881, + 994, + 911, + 994, + 911, + 1021, + 881, + 1021 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 888, + 1797, + 913, + 1797, + 913, + 1822, + 888, + 1822 + ], + "score": 0.72, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 628, + 1024, + 662, + 1024, + 662, + 1055, + 628, + 1055 + ], + "score": 0.46, + "latex": "( \\uparrow )" + }, + { + "category_id": 13, + "poly": [ + 1007, + 818, + 1023, + 818, + 1023, + 834, + 1007, + 834 + ], + "score": 0.42, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 647, + 1312, + 666, + 1312, + 666, + 1330, + 647, + 1330 + ], + "score": 0.35, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 452, + 437, + 484, + 437, + 484, + 466, + 452, + 466 + ], + "score": 0.29, + "latex": "( \\uparrow )" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1287.0, + 1033.0, + 1287.0, + 1033.0, + 1314.0, + 997.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1295.0, + 1259.0, + 1295.0, + 1259.0, + 1321.0, + 1082.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 1295.0, + 1396.0, + 1295.0, + 1396.0, + 1444.0, + 1368.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1307.0, + 1001.0, + 1307.0, + 1001.0, + 1515.0, + 970.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1318.0, + 1248.0, + 1318.0, + 1248.0, + 1343.0, + 1081.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1332.0, + 1032.0, + 1332.0, + 1032.0, + 1358.0, + 997.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1347.0, + 1069.0, + 1347.0, + 1069.0, + 1359.0, + 1055.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1340.0, + 1298.0, + 1340.0, + 1298.0, + 1363.0, + 1081.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1365.0, + 1070.0, + 1365.0, + 1070.0, + 1381.0, + 1054.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1362.0, + 1277.0, + 1362.0, + 1277.0, + 1384.0, + 1081.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1375.0, + 1032.0, + 1375.0, + 1032.0, + 1402.0, + 996.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1413.0, + 1364.0, + 1413.0, + 1364.0, + 1421.0, + 1355.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1419.0, + 1033.0, + 1419.0, + 1033.0, + 1445.0, + 997.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 1439.0, + 1392.0, + 1439.0, + 1392.0, + 1457.0, + 1378.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1366.0, + 1453.0, + 1378.0, + 1453.0, + 1378.0, + 1466.0, + 1366.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1454.0, + 1392.0, + 1454.0, + 1392.0, + 1471.0, + 1377.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1461.0, + 1032.0, + 1461.0, + 1032.0, + 1489.0, + 997.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1462.0, + 1365.0, + 1462.0, + 1365.0, + 1471.0, + 1355.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1471.0, + 1240.0, + 1471.0, + 1240.0, + 1488.0, + 1224.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 1471.0, + 1374.0, + 1471.0, + 1374.0, + 1484.0, + 1361.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1506.0, + 1033.0, + 1506.0, + 1033.0, + 1533.0, + 997.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1493.0, + 1058.0, + 1493.0, + 1058.0, + 1506.0, + 1046.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1509.0, + 1112.0, + 1509.0, + 1112.0, + 1526.0, + 1095.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1501.0, + 1147.0, + 1501.0, + 1147.0, + 1525.0, + 1121.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1499.0, + 1174.0, + 1499.0, + 1174.0, + 1516.0, + 1158.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1478.0, + 1201.0, + 1478.0, + 1201.0, + 1493.0, + 1184.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1477.0, + 1368.0, + 1477.0, + 1368.0, + 1518.0, + 1319.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1480.0, + 1393.0, + 1480.0, + 1393.0, + 1495.0, + 1377.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1530.0, + 1087.0, + 1530.0, + 1087.0, + 1557.0, + 1052.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1530.0, + 1166.0, + 1530.0, + 1166.0, + 1556.0, + 1130.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1530.0, + 1246.0, + 1530.0, + 1246.0, + 1556.0, + 1209.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1530.0, + 1324.0, + 1530.0, + 1324.0, + 1557.0, + 1289.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1367.0, + 1530.0, + 1403.0, + 1530.0, + 1403.0, + 1557.0, + 1367.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1549.0, + 1323.0, + 1549.0, + 1323.0, + 1581.0, + 1114.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1605.0, + 1392.0, + 1605.0, + 1392.0, + 1638.0, + 965.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1636.0, + 1310.0, + 1636.0, + 1310.0, + 1669.0, + 967.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 1656.0, + 968.0, + 1656.0, + 968.0, + 1691.0, + 300.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1686.0, + 968.0, + 1686.0, + 968.0, + 1722.0, + 302.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 396.0, + 1403.0, + 396.0, + 1403.0, + 442.0, + 294.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 433.0, + 451.0, + 433.0, + 451.0, + 469.0, + 296.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 433.0, + 1064.0, + 433.0, + 1064.0, + 469.0, + 485.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 991.0, + 880.0, + 991.0, + 880.0, + 1027.0, + 297.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 991.0, + 1404.0, + 991.0, + 1404.0, + 1027.0, + 912.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1022.0, + 627.0, + 1022.0, + 627.0, + 1057.0, + 295.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1022.0, + 1302.0, + 1022.0, + 1302.0, + 1057.0, + 663.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1728.0, + 1404.0, + 1728.0, + 1404.0, + 1768.0, + 295.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1762.0, + 1402.0, + 1762.0, + 1402.0, + 1796.0, + 295.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1790.0, + 887.0, + 1790.0, + 887.0, + 1830.0, + 292.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1790.0, + 1405.0, + 1790.0, + 1405.0, + 1830.0, + 914.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1823.0, + 1404.0, + 1823.0, + 1404.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1853.0, + 1405.0, + 1853.0, + 1405.0, + 1889.0, + 295.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1884.0, + 1404.0, + 1884.0, + 1404.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1915.0, + 1404.0, + 1915.0, + 1404.0, + 1948.0, + 292.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1946.0, + 1404.0, + 1946.0, + 1404.0, + 1980.0, + 294.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1976.0, + 1406.0, + 1976.0, + 1406.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2006.0, + 1356.0, + 2006.0, + 1356.0, + 2041.0, + 294.0, + 2041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 483.0, + 1405.0, + 483.0, + 1405.0, + 515.0, + 297.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 512.0, + 1403.0, + 512.0, + 1403.0, + 547.0, + 295.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 542.0, + 1404.0, + 542.0, + 1404.0, + 579.0, + 293.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 574.0, + 1404.0, + 574.0, + 1404.0, + 609.0, + 294.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 604.0, + 1406.0, + 604.0, + 1406.0, + 640.0, + 294.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 635.0, + 1254.0, + 635.0, + 1254.0, + 667.0, + 296.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 635.0, + 1359.0, + 635.0, + 1359.0, + 667.0, + 1310.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 664.0, + 1406.0, + 664.0, + 1406.0, + 698.0, + 293.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 695.0, + 435.0, + 695.0, + 435.0, + 730.0, + 387.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 695.0, + 1121.0, + 695.0, + 1121.0, + 730.0, + 525.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 695.0, + 1404.0, + 695.0, + 1404.0, + 730.0, + 1199.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 727.0, + 1402.0, + 727.0, + 1402.0, + 758.0, + 297.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 758.0, + 1029.0, + 758.0, + 1029.0, + 790.0, + 296.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1068.0, + 1404.0, + 1068.0, + 1404.0, + 1101.0, + 296.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1097.0, + 1404.0, + 1097.0, + 1404.0, + 1134.0, + 294.0, + 1134.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1129.0, + 1404.0, + 1129.0, + 1404.0, + 1162.0, + 296.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1161.0, + 1404.0, + 1161.0, + 1404.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1187.0, + 1017.0, + 1187.0, + 1017.0, + 1227.0, + 293.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 396.0, + 1403.0, + 396.0, + 1403.0, + 442.0, + 294.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 433.0, + 451.0, + 433.0, + 451.0, + 469.0, + 296.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 433.0, + 1064.0, + 433.0, + 1064.0, + 469.0, + 485.0, + 469.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 452, + 1404, + 452, + 1404, + 907, + 298, + 907 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1151, + 1403, + 1151, + 1403, + 1335, + 298, + 1335 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1351, + 1403, + 1351, + 1403, + 1566, + 298, + 1566 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1580, + 1403, + 1580, + 1403, + 1765, + 298, + 1765 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 923, + 1403, + 923, + 1403, + 1077, + 299, + 1077 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 230, + 1403, + 230, + 1403, + 445, + 299, + 445 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 300, + 1094, + 589, + 1094, + 589, + 1130, + 300, + 1130 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 300, + 1794, + 561, + 1794, + 561, + 1829, + 300, + 1829 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.786 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.776 + }, + { + "category_id": 13, + "poly": [ + 551, + 292, + 591, + 292, + 591, + 321, + 551, + 321 + ], + "score": 0.85, + "latex": "2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1334, + 954, + 1402, + 954, + 1402, + 983, + 1334, + 983 + ], + "score": 0.79, + "latex": "{ \\mathrm { K } } { = } 1 0" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1090.0, + 593.0, + 1090.0, + 593.0, + 1136.0, + 292.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1786.0, + 566.0, + 1786.0, + 566.0, + 1838.0, + 292.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 452.0, + 1402.0, + 452.0, + 1402.0, + 483.0, + 297.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 479.0, + 1405.0, + 479.0, + 1405.0, + 516.0, + 293.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 507.0, + 1404.0, + 507.0, + 1404.0, + 549.0, + 292.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 541.0, + 1403.0, + 541.0, + 1403.0, + 577.0, + 293.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 570.0, + 1404.0, + 570.0, + 1404.0, + 608.0, + 295.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 601.0, + 1404.0, + 601.0, + 1404.0, + 636.0, + 294.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 629.0, + 1405.0, + 629.0, + 1405.0, + 668.0, + 293.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 1405.0, + 662.0, + 1405.0, + 697.0, + 293.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 693.0, + 1405.0, + 693.0, + 1405.0, + 728.0, + 294.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 724.0, + 1405.0, + 724.0, + 1405.0, + 759.0, + 294.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 755.0, + 1402.0, + 755.0, + 1402.0, + 790.0, + 294.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1405.0, + 782.0, + 1405.0, + 822.0, + 293.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 817.0, + 1401.0, + 817.0, + 1401.0, + 848.0, + 296.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 843.0, + 1406.0, + 843.0, + 1406.0, + 883.0, + 292.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 873.0, + 648.0, + 873.0, + 648.0, + 913.0, + 293.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1150.0, + 1405.0, + 1150.0, + 1405.0, + 1187.0, + 294.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1183.0, + 1404.0, + 1183.0, + 1404.0, + 1215.0, + 296.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 1407.0, + 1211.0, + 1407.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1242.0, + 1404.0, + 1242.0, + 1404.0, + 1280.0, + 292.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1275.0, + 1405.0, + 1275.0, + 1405.0, + 1310.0, + 294.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 1049.0, + 1305.0, + 1049.0, + 1341.0, + 294.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1349.0, + 1408.0, + 1349.0, + 1408.0, + 1386.0, + 293.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1381.0, + 1407.0, + 1381.0, + 1407.0, + 1420.0, + 292.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1413.0, + 1404.0, + 1413.0, + 1404.0, + 1447.0, + 296.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1440.0, + 1407.0, + 1440.0, + 1407.0, + 1481.0, + 292.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1508.0, + 294.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1504.0, + 1405.0, + 1504.0, + 1405.0, + 1538.0, + 293.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1535.0, + 1397.0, + 1535.0, + 1397.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1403.0, + 1849.0, + 1403.0, + 1881.0, + 294.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 1321.0, + 1999.0, + 1321.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1577.0, + 1407.0, + 1577.0, + 1407.0, + 1620.0, + 292.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1611.0, + 1407.0, + 1611.0, + 1407.0, + 1647.0, + 293.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1637.0, + 1407.0, + 1637.0, + 1407.0, + 1680.0, + 292.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1674.0, + 1404.0, + 1674.0, + 1404.0, + 1706.0, + 294.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1702.0, + 1404.0, + 1702.0, + 1404.0, + 1737.0, + 294.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1735.0, + 808.0, + 1735.0, + 808.0, + 1768.0, + 293.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 919.0, + 1407.0, + 919.0, + 1407.0, + 963.0, + 293.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 952.0, + 1333.0, + 952.0, + 1333.0, + 988.0, + 293.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 1403.0, + 984.0, + 1403.0, + 1017.0, + 295.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1013.0, + 1404.0, + 1013.0, + 1404.0, + 1049.0, + 295.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1045.0, + 963.0, + 1045.0, + 963.0, + 1081.0, + 293.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 550.0, + 291.0, + 550.0, + 324.0, + 294.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 324.0, + 592.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 1403.0, + 322.0, + 1403.0, + 352.0, + 294.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 349.0, + 1403.0, + 349.0, + 1403.0, + 387.0, + 294.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 384.0, + 1404.0, + 384.0, + 1404.0, + 415.0, + 298.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 415.0, + 850.0, + 415.0, + 850.0, + 446.0, + 297.0, + 446.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 278, + 1404, + 278, + 1404, + 430, + 298, + 430 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 297, + 845, + 1404, + 845, + 1404, + 1030, + 297, + 1030 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 297, + 1484, + 1404, + 1484, + 1404, + 1639, + 297, + 1639 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 296, + 703, + 1404, + 703, + 1404, + 829, + 296, + 829 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 298, + 1656, + 1405, + 1656, + 1405, + 1812, + 298, + 1812 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 299, + 531, + 1404, + 531, + 1404, + 685, + 299, + 685 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 293, + 450, + 1402, + 450, + 1402, + 513, + 293, + 513 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 299, + 1829, + 1403, + 1829, + 1403, + 1954, + 299, + 1954 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 296, + 1971, + 1399, + 1971, + 1399, + 2034, + 296, + 2034 + ], + "score": 0.874 + }, + { + "category_id": 1, + "poly": [ + 295, + 1403, + 1400, + 1403, + 1400, + 1467, + 295, + 1467 + ], + "score": 0.865 + }, + { + "category_id": 1, + "poly": [ + 296, + 1322, + 1400, + 1322, + 1400, + 1386, + 296, + 1386 + ], + "score": 0.854 + }, + { + "category_id": 1, + "poly": [ + 296, + 1048, + 1402, + 1048, + 1402, + 1143, + 296, + 1143 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 296, + 1210, + 1407, + 1210, + 1407, + 1305, + 296, + 1305 + ], + "score": 0.847 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.833 + }, + { + "category_id": 1, + "poly": [ + 292, + 1159, + 1319, + 1159, + 1319, + 1194, + 292, + 1194 + ], + "score": 0.822 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.815 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 1404.0, + 279.0, + 1404.0, + 312.0, + 294.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 307.0, + 1405.0, + 307.0, + 1405.0, + 346.0, + 322.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 337.0, + 1404.0, + 337.0, + 1404.0, + 376.0, + 320.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 370.0, + 1405.0, + 370.0, + 1405.0, + 406.0, + 321.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 397.0, + 397.0, + 397.0, + 397.0, + 435.0, + 321.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 843.0, + 1405.0, + 843.0, + 1405.0, + 883.0, + 294.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 878.0, + 1405.0, + 878.0, + 1405.0, + 913.0, + 322.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 906.0, + 1405.0, + 906.0, + 1405.0, + 943.0, + 321.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 936.0, + 1405.0, + 936.0, + 1405.0, + 974.0, + 322.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 964.0, + 1406.0, + 964.0, + 1406.0, + 1008.0, + 319.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 999.0, + 545.0, + 999.0, + 545.0, + 1030.0, + 324.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1486.0, + 1405.0, + 1486.0, + 1405.0, + 1519.0, + 296.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1516.0, + 1405.0, + 1516.0, + 1405.0, + 1550.0, + 324.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1544.0, + 1405.0, + 1544.0, + 1405.0, + 1583.0, + 321.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1575.0, + 1406.0, + 1575.0, + 1406.0, + 1616.0, + 322.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1605.0, + 651.0, + 1605.0, + 651.0, + 1643.0, + 321.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 702.0, + 1405.0, + 702.0, + 1405.0, + 740.0, + 294.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 732.0, + 1405.0, + 732.0, + 1405.0, + 774.0, + 320.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 764.0, + 1406.0, + 764.0, + 1406.0, + 801.0, + 320.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 798.0, + 1261.0, + 798.0, + 1261.0, + 831.0, + 324.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1654.0, + 1405.0, + 1654.0, + 1405.0, + 1695.0, + 292.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1725.0, + 322.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1715.0, + 1406.0, + 1715.0, + 1406.0, + 1757.0, + 322.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1746.0, + 1406.0, + 1746.0, + 1406.0, + 1786.0, + 322.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1778.0, + 740.0, + 1778.0, + 740.0, + 1815.0, + 322.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 530.0, + 1405.0, + 530.0, + 1405.0, + 568.0, + 295.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 563.0, + 1404.0, + 563.0, + 1404.0, + 596.0, + 323.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 595.0, + 1401.0, + 595.0, + 1401.0, + 628.0, + 322.0, + 628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 621.0, + 1401.0, + 621.0, + 1401.0, + 660.0, + 322.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 654.0, + 835.0, + 654.0, + 835.0, + 686.0, + 322.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 447.0, + 1407.0, + 447.0, + 1407.0, + 488.0, + 292.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 482.0, + 803.0, + 482.0, + 803.0, + 513.0, + 322.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1827.0, + 1403.0, + 1827.0, + 1403.0, + 1866.0, + 294.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1859.0, + 1407.0, + 1859.0, + 1407.0, + 1896.0, + 319.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1893.0, + 1407.0, + 1893.0, + 1407.0, + 1925.0, + 323.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1921.0, + 1165.0, + 1921.0, + 1165.0, + 1956.0, + 322.0, + 1956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1969.0, + 1403.0, + 1969.0, + 1403.0, + 2005.0, + 295.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2004.0, + 634.0, + 2004.0, + 634.0, + 2034.0, + 321.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1402.0, + 1402.0, + 1402.0, + 1402.0, + 1440.0, + 295.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1434.0, + 1259.0, + 1434.0, + 1259.0, + 1470.0, + 321.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1321.0, + 1404.0, + 1321.0, + 1404.0, + 1359.0, + 294.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1353.0, + 1109.0, + 1353.0, + 1109.0, + 1387.0, + 321.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1048.0, + 1406.0, + 1048.0, + 1406.0, + 1086.0, + 294.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1081.0, + 1400.0, + 1081.0, + 1400.0, + 1115.0, + 324.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1113.0, + 975.0, + 1113.0, + 975.0, + 1144.0, + 321.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1211.0, + 1404.0, + 1211.0, + 1404.0, + 1245.0, + 295.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1242.0, + 1404.0, + 1242.0, + 1404.0, + 1276.0, + 323.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1272.0, + 920.0, + 1272.0, + 920.0, + 1306.0, + 319.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1159.0, + 1322.0, + 1159.0, + 1322.0, + 1199.0, + 295.0, + 1199.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 291, + 229, + 1399, + 229, + 1399, + 293, + 291, + 293 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.825 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1402, + 1942, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.816 + }, + { + "category_id": 1, + "poly": [ + 298, + 1795, + 1405, + 1795, + 1405, + 1917, + 298, + 1917 + ], + "score": 0.771 + }, + { + "category_id": 1, + "poly": [ + 299, + 1355, + 1401, + 1355, + 1401, + 1450, + 299, + 1450 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 296, + 1240, + 1403, + 1240, + 1403, + 1334, + 296, + 1334 + ], + "score": 0.74 + }, + { + "category_id": 1, + "poly": [ + 298, + 862, + 1402, + 862, + 1402, + 957, + 298, + 957 + ], + "score": 0.737 + }, + { + "category_id": 1, + "poly": [ + 298, + 746, + 1402, + 746, + 1402, + 843, + 298, + 843 + ], + "score": 0.735 + }, + { + "category_id": 1, + "poly": [ + 296, + 1093, + 1405, + 1093, + 1405, + 1219, + 296, + 1219 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 297, + 1471, + 1404, + 1471, + 1404, + 1597, + 297, + 1597 + ], + "score": 0.712 + }, + { + "category_id": 1, + "poly": [ + 299, + 1618, + 1404, + 1618, + 1404, + 1773, + 299, + 1773 + ], + "score": 0.711 + }, + { + "category_id": 1, + "poly": [ + 296, + 491, + 1400, + 491, + 1400, + 554, + 296, + 554 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 295, + 661, + 1400, + 661, + 1400, + 726, + 295, + 726 + ], + "score": 0.679 + }, + { + "category_id": 1, + "poly": [ + 297, + 977, + 1403, + 977, + 1403, + 1073, + 297, + 1073 + ], + "score": 0.655 + }, + { + "category_id": 1, + "poly": [ + 297, + 576, + 1401, + 576, + 1401, + 642, + 297, + 642 + ], + "score": 0.651 + }, + { + "category_id": 1, + "poly": [ + 296, + 314, + 1404, + 314, + 1404, + 470, + 296, + 470 + ], + "score": 0.572 + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1402.0, + 228.0, + 1402.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 261.0, + 1081.0, + 261.0, + 1081.0, + 294.0, + 321.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2008.0, + 320.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2002.0, + 990.0, + 2002.0, + 990.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1792.0, + 1403.0, + 1792.0, + 1403.0, + 1829.0, + 295.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1826.0, + 1406.0, + 1826.0, + 1406.0, + 1862.0, + 322.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1855.0, + 1406.0, + 1855.0, + 1406.0, + 1891.0, + 318.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1886.0, + 450.0, + 1886.0, + 450.0, + 1920.0, + 321.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1355.0, + 1405.0, + 1355.0, + 1405.0, + 1393.0, + 295.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1384.0, + 1405.0, + 1384.0, + 1405.0, + 1424.0, + 319.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1420.0, + 782.0, + 1420.0, + 782.0, + 1451.0, + 323.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1277.0, + 293.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1272.0, + 1405.0, + 1272.0, + 1405.0, + 1306.0, + 324.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1302.0, + 788.0, + 1302.0, + 788.0, + 1335.0, + 323.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 862.0, + 1402.0, + 862.0, + 1402.0, + 900.0, + 294.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 895.0, + 1403.0, + 895.0, + 1403.0, + 929.0, + 323.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 923.0, + 945.0, + 923.0, + 945.0, + 961.0, + 325.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 746.0, + 1406.0, + 746.0, + 1406.0, + 785.0, + 293.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 778.0, + 1406.0, + 778.0, + 1406.0, + 813.0, + 322.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 808.0, + 1362.0, + 808.0, + 1362.0, + 846.0, + 322.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1095.0, + 1405.0, + 1095.0, + 1405.0, + 1131.0, + 295.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1123.0, + 1408.0, + 1123.0, + 1408.0, + 1162.0, + 319.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1152.0, + 1405.0, + 1152.0, + 1405.0, + 1195.0, + 320.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1188.0, + 531.0, + 1188.0, + 531.0, + 1218.0, + 325.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1470.0, + 1404.0, + 1470.0, + 1404.0, + 1510.0, + 292.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1502.0, + 1404.0, + 1502.0, + 1404.0, + 1539.0, + 322.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1533.0, + 1405.0, + 1533.0, + 1405.0, + 1568.0, + 321.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1566.0, + 1048.0, + 1566.0, + 1048.0, + 1599.0, + 324.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1653.0, + 298.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1681.0, + 322.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1677.0, + 1405.0, + 1677.0, + 1405.0, + 1718.0, + 323.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1709.0, + 1404.0, + 1709.0, + 1404.0, + 1745.0, + 323.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1741.0, + 648.0, + 1741.0, + 648.0, + 1774.0, + 323.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 491.0, + 1405.0, + 491.0, + 1405.0, + 526.0, + 295.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 523.0, + 707.0, + 523.0, + 707.0, + 556.0, + 320.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 660.0, + 1405.0, + 660.0, + 1405.0, + 700.0, + 295.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 694.0, + 906.0, + 694.0, + 906.0, + 727.0, + 322.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 977.0, + 1404.0, + 977.0, + 1404.0, + 1015.0, + 293.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1010.0, + 1403.0, + 1010.0, + 1403.0, + 1045.0, + 324.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1041.0, + 1223.0, + 1041.0, + 1223.0, + 1074.0, + 322.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 577.0, + 1405.0, + 577.0, + 1405.0, + 613.0, + 296.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 606.0, + 1169.0, + 606.0, + 1169.0, + 643.0, + 321.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 315.0, + 1401.0, + 315.0, + 1401.0, + 348.0, + 295.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 345.0, + 1406.0, + 345.0, + 1406.0, + 383.0, + 323.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 376.0, + 1406.0, + 376.0, + 1406.0, + 413.0, + 324.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 407.0, + 1404.0, + 407.0, + 1404.0, + 444.0, + 323.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 436.0, + 1138.0, + 436.0, + 1138.0, + 473.0, + 322.0, + 473.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 372, + 1405, + 372, + 1405, + 466, + 297, + 466 + ], + "score": 0.911 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1404, + 229, + 1404, + 353, + 299, + 353 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 295, + 485, + 1400, + 485, + 1400, + 549, + 295, + 549 + ], + "score": 0.907 + }, + { + "category_id": 1, + "poly": [ + 298, + 700, + 1404, + 700, + 1404, + 857, + 298, + 857 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 297, + 1438, + 1401, + 1438, + 1401, + 1533, + 297, + 1533 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 294, + 619, + 1402, + 619, + 1402, + 682, + 294, + 682 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 290, + 1551, + 1402, + 1551, + 1402, + 1615, + 290, + 1615 + ], + "score": 0.88 + }, + { + "category_id": 1, + "poly": [ + 295, + 567, + 1350, + 567, + 1350, + 601, + 295, + 601 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 298, + 1181, + 1404, + 1181, + 1404, + 1308, + 298, + 1308 + ], + "score": 0.878 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1403, + 1941, + 1403, + 2034, + 300, + 2034 + ], + "score": 0.856 + }, + { + "category_id": 1, + "poly": [ + 298, + 1633, + 1404, + 1633, + 1404, + 1757, + 298, + 1757 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 298, + 1324, + 1400, + 1324, + 1400, + 1420, + 298, + 1420 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 298, + 1070, + 1405, + 1070, + 1405, + 1164, + 298, + 1164 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 296, + 874, + 1401, + 874, + 1401, + 940, + 296, + 940 + ], + "score": 0.839 + }, + { + "category_id": 1, + "poly": [ + 297, + 1857, + 1399, + 1857, + 1399, + 1923, + 297, + 1923 + ], + "score": 0.838 + }, + { + "category_id": 1, + "poly": [ + 299, + 956, + 1405, + 956, + 1405, + 1052, + 299, + 1052 + ], + "score": 0.836 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 854, + 75, + 854, + 105, + 299, + 105 + ], + "score": 0.77 + }, + { + "category_id": 1, + "poly": [ + 297, + 1776, + 1398, + 1776, + 1398, + 1840, + 297, + 1840 + ], + "score": 0.716 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 852, + 76, + 852, + 104, + 300, + 104 + ], + "score": 0.268 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 370.0, + 1405.0, + 370.0, + 1405.0, + 411.0, + 294.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 402.0, + 1405.0, + 402.0, + 1405.0, + 441.0, + 320.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 434.0, + 587.0, + 434.0, + 587.0, + 468.0, + 323.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1406.0, + 259.0, + 1406.0, + 299.0, + 321.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 326.0, + 324.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 321.0, + 1116.0, + 321.0, + 1116.0, + 357.0, + 323.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 485.0, + 1405.0, + 485.0, + 1405.0, + 521.0, + 294.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 517.0, + 1380.0, + 517.0, + 1380.0, + 550.0, + 322.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 701.0, + 1401.0, + 701.0, + 1401.0, + 734.0, + 297.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 729.0, + 1406.0, + 729.0, + 1406.0, + 770.0, + 322.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 761.0, + 1405.0, + 761.0, + 1405.0, + 800.0, + 321.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 792.0, + 1404.0, + 792.0, + 1404.0, + 829.0, + 322.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 823.0, + 959.0, + 823.0, + 959.0, + 860.0, + 322.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1438.0, + 1402.0, + 1438.0, + 1402.0, + 1476.0, + 293.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1471.0, + 1405.0, + 1471.0, + 1405.0, + 1505.0, + 322.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1500.0, + 804.0, + 1500.0, + 804.0, + 1534.0, + 323.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 612.0, + 1407.0, + 612.0, + 1407.0, + 661.0, + 293.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 651.0, + 891.0, + 651.0, + 891.0, + 682.0, + 322.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1550.0, + 1406.0, + 1550.0, + 1406.0, + 1587.0, + 293.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1582.0, + 1405.0, + 1582.0, + 1405.0, + 1618.0, + 322.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 563.0, + 1352.0, + 563.0, + 1352.0, + 606.0, + 294.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1218.0, + 293.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1214.0, + 1404.0, + 1214.0, + 1404.0, + 1250.0, + 323.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1245.0, + 1405.0, + 1245.0, + 1405.0, + 1281.0, + 320.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1275.0, + 1198.0, + 1275.0, + 1198.0, + 1308.0, + 323.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1940.0, + 1405.0, + 1940.0, + 1405.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 320.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 2002.0, + 904.0, + 2002.0, + 904.0, + 2036.0, + 326.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1632.0, + 1404.0, + 1632.0, + 1404.0, + 1669.0, + 293.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1666.0, + 1404.0, + 1666.0, + 1404.0, + 1698.0, + 322.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1697.0, + 1406.0, + 1697.0, + 1406.0, + 1730.0, + 323.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1727.0, + 740.0, + 1727.0, + 740.0, + 1759.0, + 323.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1323.0, + 1404.0, + 1323.0, + 1404.0, + 1362.0, + 294.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1356.0, + 1405.0, + 1356.0, + 1405.0, + 1392.0, + 323.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1391.0, + 772.0, + 1391.0, + 772.0, + 1420.0, + 320.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1067.0, + 1403.0, + 1067.0, + 1403.0, + 1109.0, + 293.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1102.0, + 1405.0, + 1102.0, + 1405.0, + 1136.0, + 322.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1132.0, + 865.0, + 1132.0, + 865.0, + 1166.0, + 323.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 876.0, + 1405.0, + 876.0, + 1405.0, + 912.0, + 295.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 906.0, + 1386.0, + 906.0, + 1386.0, + 942.0, + 323.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1856.0, + 1405.0, + 1856.0, + 1405.0, + 1896.0, + 296.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1890.0, + 827.0, + 1890.0, + 827.0, + 1924.0, + 321.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 1403.0, + 957.0, + 1403.0, + 992.0, + 295.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1024.0, + 323.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1019.0, + 996.0, + 1019.0, + 996.0, + 1053.0, + 323.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1772.0, + 1404.0, + 1772.0, + 1404.0, + 1814.0, + 293.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1807.0, + 1160.0, + 1807.0, + 1160.0, + 1843.0, + 323.0, + 1843.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 629, + 1404, + 629, + 1404, + 724, + 299, + 724 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 299, + 402, + 1403, + 402, + 1403, + 527, + 299, + 527 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 299, + 229, + 1404, + 229, + 1404, + 384, + 299, + 384 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 295, + 547, + 1404, + 547, + 1404, + 611, + 295, + 611 + ], + "score": 0.902 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2112, + 836, + 2112 + ], + "score": 0.832 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 854, + 75, + 854, + 104, + 299, + 104 + ], + "score": 0.561 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 852, + 75, + 852, + 104, + 299, + 104 + ], + "score": 0.453 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 856.0, + 72.0, + 856.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 629.0, + 1401.0, + 629.0, + 1401.0, + 667.0, + 294.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 659.0, + 1405.0, + 659.0, + 1405.0, + 700.0, + 321.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 690.0, + 437.0, + 690.0, + 437.0, + 725.0, + 321.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 404.0, + 1404.0, + 404.0, + 1404.0, + 440.0, + 295.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 432.0, + 1404.0, + 432.0, + 1404.0, + 474.0, + 319.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 466.0, + 1404.0, + 466.0, + 1404.0, + 502.0, + 322.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 496.0, + 591.0, + 496.0, + 591.0, + 530.0, + 323.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 1404.0, + 231.0, + 1404.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 259.0, + 1403.0, + 259.0, + 1403.0, + 297.0, + 321.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 328.0, + 322.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 320.0, + 1405.0, + 320.0, + 1405.0, + 357.0, + 322.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 350.0, + 1203.0, + 350.0, + 1203.0, + 386.0, + 321.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 547.0, + 1404.0, + 547.0, + 1404.0, + 583.0, + 297.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 578.0, + 1124.0, + 578.0, + 1124.0, + 614.0, + 322.0, + 614.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 336, + 637, + 806, + 637, + 806, + 802, + 336, + 802 + ], + "score": 0.974, + "html": "
Ratio of Unlabeled DataDatasets
SNIPSCoNLL03
5%84.4772.92
25%87.1076.46
75%87.5076.56
" + }, + { + "category_id": 1, + "poly": [ + 298, + 429, + 1403, + 429, + 1403, + 554, + 298, + 554 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 297, + 1054, + 1404, + 1054, + 1404, + 1208, + 297, + 1208 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 1776, + 1404, + 1776, + 1404, + 1931, + 298, + 1931 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 324, + 1267, + 1299, + 1267, + 1299, + 1605, + 324, + 1605 + ], + "score": 0.971 + }, + { + "category_id": 3, + "poly": [ + 935, + 594, + 1267, + 594, + 1267, + 856, + 935, + 856 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 295, + 352, + 1403, + 352, + 1403, + 415, + 295, + 415 + ], + "score": 0.946 + }, + { + "category_id": 4, + "poly": [ + 921, + 877, + 1311, + 877, + 1311, + 942, + 921, + 942 + ], + "score": 0.938 + }, + { + "category_id": 4, + "poly": [ + 349, + 1630, + 1348, + 1630, + 1348, + 1665, + 349, + 1665 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 1720, + 481, + 1720, + 481, + 1751, + 300, + 1751 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 298, + 997, + 1039, + 997, + 1039, + 1030, + 298, + 1030 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 301, + 227, + 508, + 227, + 508, + 261, + 301, + 261 + ], + "score": 0.884 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.853 + }, + { + "category_id": 0, + "poly": [ + 298, + 297, + 1055, + 297, + 1055, + 328, + 298, + 328 + ], + "score": 0.771 + }, + { + "category_id": 6, + "poly": [ + 303, + 825, + 839, + 825, + 839, + 886, + 303, + 886 + ], + "score": 0.736 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.693 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 854, + 75, + 854, + 105, + 297, + 105 + ], + "score": 0.221 + }, + { + "category_id": 1, + "poly": [ + 298, + 297, + 1055, + 297, + 1055, + 328, + 298, + 328 + ], + "score": 0.147 + }, + { + "category_id": 1, + "poly": [ + 303, + 825, + 839, + 825, + 839, + 886, + 303, + 886 + ], + "score": 0.094 + }, + { + "category_id": 13, + "poly": [ + 1304, + 429, + 1401, + 429, + 1401, + 464, + 1304, + 464 + ], + "score": 0.92, + "latex": "\\{ 1 , 3 , 5 \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1116, + 365, + 1116, + 365, + 1145, + 297, + 1145 + ], + "score": 0.83, + "latex": "{ \\mathrm { K } } { = } 1 0" + }, + { + "category_id": 13, + "poly": [ + 683, + 1808, + 866, + 1808, + 866, + 1841, + 683, + 1841 + ], + "score": 0.81, + "latex": "\\{ 5 , 1 0 , 2 0 , 1 0 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 1147, + 431, + 1169, + 431, + 1169, + 459, + 1147, + 459 + ], + "score": 0.81, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 860, + 493, + 882, + 493, + 882, + 519, + 860, + 519 + ], + "score": 0.73, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 1126, + 880, + 1148, + 880, + 1148, + 907, + 1126, + 907 + ], + "score": 0.69, + "latex": "S" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1272.0, + 371.0, + 1272.0, + 371.0, + 1293.0, + 341.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1288.0, + 402.0, + 1288.0, + 402.0, + 1313.0, + 390.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1280.0, + 554.0, + 1280.0, + 554.0, + 1319.0, + 409.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1275.0, + 903.0, + 1275.0, + 903.0, + 1296.0, + 875.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 1287.0, + 937.0, + 1287.0, + 937.0, + 1315.0, + 922.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1279.0, + 1087.0, + 1279.0, + 1087.0, + 1317.0, + 942.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1288.0, + 1284.0, + 1288.0, + 1284.0, + 1376.0, + 1263.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1317.0, + 373.0, + 1317.0, + 373.0, + 1339.0, + 341.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1319.0, + 403.0, + 1319.0, + 403.0, + 1332.0, + 387.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 1314.0, + 577.0, + 1314.0, + 577.0, + 1336.0, + 409.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1315.0, + 733.0, + 1315.0, + 733.0, + 1336.0, + 719.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1321.0, + 906.0, + 1321.0, + 906.0, + 1342.0, + 876.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1316.0, + 939.0, + 1316.0, + 939.0, + 1334.0, + 920.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1314.0, + 1114.0, + 1314.0, + 1114.0, + 1336.0, + 944.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 1322.0, + 1250.0, + 1322.0, + 1250.0, + 1328.0, + 1245.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1338.0, + 345.0, + 1338.0, + 345.0, + 1465.0, + 323.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1334.0, + 403.0, + 1334.0, + 403.0, + 1349.0, + 387.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 1333.0, + 569.0, + 1333.0, + 569.0, + 1352.0, + 411.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1337.0, + 659.0, + 1337.0, + 659.0, + 1366.0, + 639.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1337.0, + 881.0, + 1337.0, + 881.0, + 1465.0, + 856.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1332.0, + 939.0, + 1332.0, + 939.0, + 1352.0, + 919.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1333.0, + 1102.0, + 1333.0, + 1102.0, + 1352.0, + 945.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1336.0, + 1183.0, + 1336.0, + 1183.0, + 1350.0, + 1169.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1363.0, + 371.0, + 1363.0, + 371.0, + 1385.0, + 342.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1379.0, + 673.0, + 1379.0, + 673.0, + 1389.0, + 661.0, + 1389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1368.0, + 904.0, + 1368.0, + 904.0, + 1388.0, + 876.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1364.0, + 1286.0, + 1364.0, + 1286.0, + 1435.0, + 1256.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1385.0, + 594.0, + 1385.0, + 594.0, + 1399.0, + 581.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1393.0, + 1218.0, + 1393.0, + 1218.0, + 1404.0, + 1208.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1402.0, + 1252.0, + 1402.0, + 1252.0, + 1413.0, + 1241.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1411.0, + 370.0, + 1411.0, + 370.0, + 1431.0, + 341.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1411.0, + 504.0, + 1411.0, + 504.0, + 1422.0, + 495.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1422.0, + 633.0, + 1422.0, + 633.0, + 1437.0, + 619.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1413.0, + 904.0, + 1413.0, + 904.0, + 1434.0, + 876.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1428.0, + 744.0, + 1428.0, + 744.0, + 1444.0, + 732.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1267.0, + 1433.0, + 1281.0, + 1433.0, + 1281.0, + 1449.0, + 1267.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 1457.0, + 369.0, + 1457.0, + 369.0, + 1478.0, + 342.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1474.0, + 448.0, + 1474.0, + 448.0, + 1492.0, + 414.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1443.0, + 471.0, + 1443.0, + 471.0, + 1458.0, + 457.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1465.0, + 491.0, + 1465.0, + 491.0, + 1478.0, + 476.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1474.0, + 564.0, + 1474.0, + 564.0, + 1488.0, + 549.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1479.0, + 670.0, + 1479.0, + 670.0, + 1490.0, + 659.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1442.0, + 752.0, + 1442.0, + 752.0, + 1482.0, + 725.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1459.0, + 904.0, + 1459.0, + 904.0, + 1480.0, + 876.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1470.0, + 1081.0, + 1470.0, + 1081.0, + 1488.0, + 1063.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1466.0, + 1122.0, + 1466.0, + 1122.0, + 1480.0, + 1107.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1469.0, + 1233.0, + 1469.0, + 1233.0, + 1490.0, + 1205.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1447.0, + 1281.0, + 1447.0, + 1281.0, + 1493.0, + 1241.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1503.0, + 373.0, + 1503.0, + 373.0, + 1525.0, + 341.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1506.0, + 398.0, + 1506.0, + 398.0, + 1519.0, + 385.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1508.0, + 430.0, + 1508.0, + 430.0, + 1519.0, + 420.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1507.0, + 465.0, + 1507.0, + 465.0, + 1522.0, + 450.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 1493.0, + 492.0, + 1493.0, + 492.0, + 1508.0, + 478.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1493.0, + 574.0, + 1493.0, + 574.0, + 1507.0, + 560.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1501.0, + 611.0, + 1501.0, + 611.0, + 1519.0, + 578.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1507.0, + 666.0, + 1507.0, + 666.0, + 1517.0, + 657.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 1487.0, + 720.0, + 1487.0, + 720.0, + 1510.0, + 696.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1503.0, + 736.0, + 1503.0, + 736.0, + 1517.0, + 722.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1506.0, + 906.0, + 1506.0, + 906.0, + 1527.0, + 876.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1488.0, + 930.0, + 1488.0, + 930.0, + 1500.0, + 920.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1507.0, + 987.0, + 1507.0, + 987.0, + 1522.0, + 972.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1499.0, + 1021.0, + 1499.0, + 1021.0, + 1518.0, + 1000.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1494.0, + 1053.0, + 1494.0, + 1053.0, + 1512.0, + 1035.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1502.0, + 1219.0, + 1502.0, + 1219.0, + 1512.0, + 1210.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 1492.0, + 1235.0, + 1492.0, + 1235.0, + 1501.0, + 1226.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1490.0, + 1256.0, + 1490.0, + 1256.0, + 1512.0, + 1238.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1527.0, + 404.0, + 1527.0, + 404.0, + 1549.0, + 376.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1525.0, + 476.0, + 1525.0, + 476.0, + 1551.0, + 445.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1526.0, + 544.0, + 1526.0, + 544.0, + 1549.0, + 516.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1525.0, + 616.0, + 1525.0, + 616.0, + 1550.0, + 584.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1527.0, + 684.0, + 1527.0, + 684.0, + 1549.0, + 656.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1527.0, + 752.0, + 1527.0, + 752.0, + 1549.0, + 725.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1527.0, + 956.0, + 1527.0, + 956.0, + 1549.0, + 929.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1526.0, + 1039.0, + 1526.0, + 1039.0, + 1549.0, + 1012.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1526.0, + 1122.0, + 1526.0, + 1122.0, + 1549.0, + 1094.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1526.0, + 1205.0, + 1526.0, + 1205.0, + 1549.0, + 1177.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1527.0, + 1287.0, + 1527.0, + 1287.0, + 1550.0, + 1261.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1544.0, + 627.0, + 1544.0, + 627.0, + 1566.0, + 505.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1545.0, + 1161.0, + 1545.0, + 1161.0, + 1564.0, + 1041.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 1574.0, + 637.0, + 1574.0, + 637.0, + 1606.0, + 525.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1575.0, + 1191.0, + 1575.0, + 1191.0, + 1606.0, + 1042.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 590.0, + 992.0, + 590.0, + 992.0, + 622.0, + 957.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 608.0, + 1057.0, + 608.0, + 1057.0, + 637.0, + 1011.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 630.0, + 1058.0, + 630.0, + 1058.0, + 659.0, + 1012.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 645.0, + 993.0, + 645.0, + 993.0, + 677.0, + 956.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 652.0, + 1058.0, + 652.0, + 1058.0, + 682.0, + 1011.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 697.0, + 993.0, + 697.0, + 993.0, + 734.0, + 933.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 753.0, + 993.0, + 753.0, + 993.0, + 792.0, + 955.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 807.0, + 991.0, + 807.0, + 991.0, + 840.0, + 957.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 828.0, + 1204.0, + 828.0, + 1204.0, + 855.0, + 1034.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 878.0, + 1125.0, + 878.0, + 1125.0, + 911.0, + 921.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 878.0, + 1312.0, + 878.0, + 1312.0, + 911.0, + 1149.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 904.0, + 1236.0, + 904.0, + 1236.0, + 946.0, + 919.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 1629.0, + 1351.0, + 1629.0, + 1351.0, + 1669.0, + 346.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1717.0, + 484.0, + 1717.0, + 484.0, + 1756.0, + 295.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 997.0, + 1041.0, + 997.0, + 1041.0, + 1032.0, + 294.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 221.0, + 514.0, + 221.0, + 514.0, + 270.0, + 294.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 297.0, + 1055.0, + 297.0, + 1055.0, + 329.0, + 298.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 823.0, + 840.0, + 823.0, + 840.0, + 859.0, + 304.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 850.0, + 698.0, + 850.0, + 698.0, + 889.0, + 303.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 856.0, + 73.0, + 856.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 426.0, + 1146.0, + 426.0, + 1146.0, + 466.0, + 293.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 426.0, + 1303.0, + 426.0, + 1303.0, + 466.0, + 1170.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 462.0, + 1407.0, + 462.0, + 1407.0, + 495.0, + 294.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 493.0, + 859.0, + 493.0, + 859.0, + 525.0, + 294.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 493.0, + 1403.0, + 493.0, + 1403.0, + 525.0, + 883.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 522.0, + 984.0, + 522.0, + 984.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1052.0, + 1405.0, + 1052.0, + 1405.0, + 1093.0, + 295.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1085.0, + 1404.0, + 1085.0, + 1404.0, + 1118.0, + 294.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1116.0, + 1405.0, + 1116.0, + 1405.0, + 1150.0, + 366.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1144.0, + 1405.0, + 1144.0, + 1405.0, + 1185.0, + 292.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1175.0, + 418.0, + 1175.0, + 418.0, + 1214.0, + 294.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1811.0, + 296.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1806.0, + 682.0, + 1806.0, + 682.0, + 1843.0, + 294.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1806.0, + 1404.0, + 1806.0, + 1404.0, + 1843.0, + 867.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1839.0, + 1405.0, + 1839.0, + 1405.0, + 1872.0, + 296.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1867.0, + 1405.0, + 1867.0, + 1405.0, + 1904.0, + 293.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1899.0, + 578.0, + 1899.0, + 578.0, + 1932.0, + 294.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 352.0, + 1404.0, + 352.0, + 1404.0, + 387.0, + 297.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 381.0, + 1322.0, + 381.0, + 1322.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 297.0, + 1055.0, + 297.0, + 1055.0, + 329.0, + 298.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 823.0, + 840.0, + 823.0, + 840.0, + 859.0, + 304.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 850.0, + 698.0, + 850.0, + 698.0, + 889.0, + 303.0, + 889.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 881, + 1651, + 1352, + 1651, + 1352, + 1963, + 881, + 1963 + ], + "score": 0.976, + "html": "
MethodShots (3 Slot Types X 41 languages)
51020100
Full-supervision
BERT
Few-shot Supervision82.3385.70
BERT77.6879.67
Few-shot Supervision +unlabeled data
Mean Teacher77.0980.2382.1985.34
VAT74.7178.8282.6085.82
Classic ST76.7380.2482.3986.08
BOND78.8179.5782.1986.14
MetaST79.1081.6183.1485.57
" + }, + { + "category_id": 5, + "poly": [ + 346, + 1209, + 820, + 1209, + 820, + 1561, + 346, + 1561 + ], + "score": 0.972, + "html": "
MethodShots (8 Slot Types)
51020100
Full-supervision
BERT78.95
Few-shot Supervision
BERT41.3954.0660.1272.24
Few-shot Supervision +unlabeled data
CVT33.7442.5751.3370.84
SeqVAT41.9451.5556.1571.39
Mean Teacher40.3751.7557.3472.40
VAT41.2953.3459.6872.65
Classic ST44.3556.8060.2873.13
BOND43.0155.7859.9673.60
MetaST53.0263.8367.8675.25
" + }, + { + "category_id": 5, + "poly": [ + 631, + 288, + 1067, + 288, + 1067, + 735, + 631, + 735 + ], + "score": 0.972, + "html": "
MethodShots
510
Full-supervision
BERT0.9444
Few-shot Supervision
BERT0.82110.8785
Few-shot Supervision + unlabeled data
CVT67.4478.24
SeqVAT64.6772.65
Mean Teacher84.1089.53
VAT83.2489.71
Classic ST86.8890.70
BOND84.9289.75
MetaST89.2192.18
" + }, + { + "category_id": 5, + "poly": [ + 881, + 1209, + 1352, + 1209, + 1352, + 1561, + 881, + 1561 + ], + "score": 0.97, + "html": "
MethodShots (4 Slot Types)
51020100
Full-supervision92.40
BERT
Few-shot Supervision
BERT63.8771.1573.5784.36
Few-shot Supervision +unlabeled data
CVT51.1554.3166.1181.99
SeqVAT58.0267.2174.1582.20
Mean Teacher59.0468.6772.6284.17
VAT57.0365.0372.6984.43
Classic ST64.0470.9974.6584.93
BOND62.5269.5674.1983.87
MetaST71.4976.6578.5485.77
" + }, + { + "category_id": 5, + "poly": [ + 881, + 766, + 1353, + 766, + 1353, + 1118, + 881, + 1118 + ], + "score": 0.968, + "html": "
MethodShots (12 Slot Types)
51020100
Full-supervision87.87
BERT
Few-shot Supervision
BERT62.8069.5075.8182.49
Few-shot Supervision +unlabeled data
CVT57.4862.7370.2081.82
SeqVAT60.9467.1074.1582.73
Mean Teacher58.9267.6275.2482.20
VAT60.7570.1775.4182.39
Classic ST63.3971.8876.5883.06
BOND62.5070.9175.5282.65
MetaST72.5777.6780.3384.35
" + }, + { + "category_id": 5, + "poly": [ + 346, + 766, + 820, + 766, + 820, + 1118, + 346, + 1118 + ], + "score": 0.967, + "html": "
MethodShots (3 Slot Types)
51020100
Full-supervision
BERT84.04
Few-shot Supervision
BERT37.0145.6154.5367.87
Few-shot Supervision +unlabeled data
CVT16.0527.8946.4266.36
SeqVAT21.1135.1642.2662.37
Mean Teacher30.9241.4350.6167.16
VAT24.7238.8150.1566.31
Classic ST32.7246.1554.4168.64
BOND34.2248.7352.4568.89
MetaST55.0456.6160.3873.20
" + }, + { + "category_id": 6, + "poly": [ + 408, + 1143, + 756, + 1143, + 756, + 1174, + 408, + 1174 + ], + "score": 0.912 + }, + { + "category_id": 6, + "poly": [ + 958, + 1585, + 1277, + 1585, + 1277, + 1617, + 958, + 1617 + ], + "score": 0.909 + }, + { + "category_id": 6, + "poly": [ + 939, + 1987, + 1295, + 1987, + 1295, + 2018, + 939, + 2018 + ], + "score": 0.904 + }, + { + "category_id": 6, + "poly": [ + 392, + 1986, + 772, + 1986, + 772, + 2018, + 392, + 2018 + ], + "score": 0.901 + }, + { + "category_id": 6, + "poly": [ + 946, + 1143, + 1284, + 1143, + 1284, + 1174, + 946, + 1174 + ], + "score": 0.897 + }, + { + "category_id": 6, + "poly": [ + 717, + 231, + 981, + 231, + 981, + 263, + 717, + 263 + ], + "score": 0.896 + }, + { + "category_id": 6, + "poly": [ + 388, + 1586, + 775, + 1586, + 775, + 1617, + 388, + 1617 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.868 + }, + { + "category_id": 6, + "poly": [ + 388, + 1586, + 776, + 1586, + 776, + 1617, + 388, + 1617 + ], + "score": 0.108 + }, + { + "category_id": 13, + "poly": [ + 534, + 1823, + 549, + 1823, + 549, + 1839, + 534, + 1839 + ], + "score": 0.57, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 534, + 1380, + 549, + 1380, + 549, + 1395, + 534, + 1395 + ], + "score": 0.48, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1824, + 1084, + 1824, + 1084, + 1838, + 1069, + 1838 + ], + "score": 0.41, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 534, + 937, + 550, + 937, + 550, + 953, + 534, + 953 + ], + "score": 0.38, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1069, + 1380, + 1084, + 1380, + 1084, + 1396, + 1069, + 1396 + ], + "score": 0.35, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 874, + 503, + 894, + 503, + 894, + 524, + 874, + 524 + ], + "score": 0.33, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1069, + 938, + 1084, + 938, + 1084, + 952, + 1069, + 952 + ], + "score": 0.3, + "latex": "^ +" + }, + { + "category_id": 5, + "poly": [ + 347, + 1651, + 820, + 1651, + 820, + 1963, + 347, + 1963 + ], + "score": 0.848, + "html": "
MethodShots (4 Slot Types)
51020100
Full-supervision
BERT
Few-shot Supervision
BERT64.8070.7773.8980.61
Few-shot Supervision +unlabeled data
Mean Teacher64.5568.3473.8779.21
VAT64.9767.6374.2680.70
Classic ST67.9572.6973.7981.82
BOND69.4272.7976.0280.62
MetaST73.3476.6577.0182.11
" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 1139.0, + 758.0, + 1139.0, + 758.0, + 1178.0, + 405.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1579.0, + 1279.0, + 1579.0, + 1279.0, + 1623.0, + 954.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1983.0, + 1296.0, + 1983.0, + 1296.0, + 2022.0, + 936.0, + 2022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1984.0, + 775.0, + 1984.0, + 775.0, + 2024.0, + 389.0, + 2024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 1140.0, + 1288.0, + 1140.0, + 1288.0, + 1179.0, + 944.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 228.0, + 984.0, + 228.0, + 984.0, + 265.0, + 715.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1581.0, + 778.0, + 1581.0, + 778.0, + 1620.0, + 385.0, + 1620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1581.0, + 778.0, + 1581.0, + 778.0, + 1620.0, + 385.0, + 1620.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 304, + 897, + 1398, + 897, + 1398, + 1267, + 304, + 1267 + ], + "score": 0.979, + "html": "
BERT attention dropout BERT hidden dropout Latest Iteration R in labeled data acquisition BERT output hidden size h0.3 0.3 5 768
Steps for fine-tuning teacher model on labeled data2000
Steps T for self-training model on unlabeled data3000
Mini-batch S Re-initialize Student5
Pseudo-label TypeY Hard
Warmup steps20
learning rate α5e-5
Weight_decay5e-6
" + }, + { + "category_id": 5, + "poly": [ + 302, + 591, + 1397, + 591, + 1397, + 794, + 302, + 794 + ], + "score": 0.973, + "html": "
DatasetSequence LengthBatch SizeLabeled data sample size B|Unlabeled Batch SizeBERT Encoder
SNIPS64163232BERT-base-uncased
Email64163232BERT-base-cased
Movie64163232BERT-base-uncased
Restaurant64161632BERT-base-uncased
CoNLL03 (EN)12816832BERT-base-cased
Wikiann (EN)12816832BERT-base-cased
CoNLLO3 (multilingual)128163232BERT-multilingual-base-cased
Wikiann (multilingaul)128163232BERT-multilingual-base-cased
" + }, + { + "category_id": 1, + "poly": [ + 301, + 286, + 1399, + 286, + 1399, + 379, + 301, + 379 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 300, + 472, + 1399, + 472, + 1399, + 565, + 300, + 565 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 296, + 394, + 1397, + 394, + 1397, + 457, + 296, + 457 + ], + "score": 0.949 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 854, + 75, + 854, + 105, + 298, + 105 + ], + "score": 0.898 + }, + { + "category_id": 6, + "poly": [ + 397, + 818, + 1299, + 818, + 1299, + 851, + 397, + 851 + ], + "score": 0.894 + }, + { + "category_id": 6, + "poly": [ + 689, + 1290, + 1009, + 1290, + 1009, + 1323, + 689, + 1323 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2089, + 865, + 2089, + 865, + 2112, + 836, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 0, + "poly": [ + 300, + 230, + 901, + 230, + 901, + 261, + 300, + 261 + ], + "score": 0.732 + }, + { + "category_id": 1, + "poly": [ + 300, + 230, + 901, + 230, + 901, + 261, + 300, + 261 + ], + "score": 0.131 + }, + { + "category_id": 13, + "poly": [ + 451, + 1206, + 471, + 1206, + 471, + 1225, + 451, + 1225 + ], + "score": 0.35, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 815.0, + 1303.0, + 815.0, + 1303.0, + 853.0, + 396.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1286.0, + 1014.0, + 1286.0, + 1014.0, + 1330.0, + 685.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 907.0, + 231.0, + 907.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 285.0, + 1404.0, + 285.0, + 1404.0, + 323.0, + 296.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 319.0, + 1403.0, + 319.0, + 1403.0, + 353.0, + 296.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 711.0, + 352.0, + 711.0, + 378.0, + 295.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 469.0, + 1403.0, + 469.0, + 1403.0, + 508.0, + 295.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 502.0, + 1404.0, + 502.0, + 1404.0, + 539.0, + 295.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 535.0, + 593.0, + 535.0, + 593.0, + 564.0, + 295.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 391.0, + 1402.0, + 391.0, + 1402.0, + 432.0, + 295.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 426.0, + 711.0, + 426.0, + 711.0, + 461.0, + 296.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 907.0, + 231.0, + 907.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/B1l8BtlCb/B1l8BtlCb.md b/parse/train/B1l8BtlCb/B1l8BtlCb.md new file mode 100644 index 0000000000000000000000000000000000000000..059313d6fb37b5a67c45237ee9dc4bc5750479fa --- /dev/null +++ b/parse/train/B1l8BtlCb/B1l8BtlCb.md @@ -0,0 +1,318 @@ +# NON-AUTOREGRESSIVE NEURAL MACHINE TRANSLATION + +Jiatao $\mathbf { G u } ^ { \dagger }$ ∗, James Bradbury‡, Caiming Xiong‡, Victor O.K. Li†& Richard Socher‡ + +‡Salesforce Research +{james.bradbury,cxiong,rsocher}@salesforce.com +†The University of Hong Kong +{jiataogu, vli}@eee.hku.hk + +# ABSTRACT + +Existing approaches to neural machine translation condition each output word on previously generated outputs. We introduce a model that avoids this autoregressive property and produces its outputs in parallel, allowing an order of magnitude lower latency during inference. Through knowledge distillation, the use of input token fertilities as a latent variable, and policy gradient fine-tuning, we achieve this at a cost of as little as 2.0 BLEU points relative to the autoregressive Transformer network used as a teacher. We demonstrate substantial cumulative improvements associated with each of the three aspects of our training strategy, and validate our approach on IWSLT 2016 English–German and two WMT language pairs. By sampling fertilities in parallel at inference time, our non-autoregressive model achieves near-state-of-the-art performance of 29.8 BLEU on WMT 2016 English– Romanian. + +# 1 INTRODUCTION + +Neural network based models outperform traditional statistical models for machine translation (MT) (Bahdanau et al., 2015; Luong et al., 2015). However, state-of-the-art neural models are much slower than statistical MT approaches at inference time (Wu et al., 2016). Both model families use autoregressive decoders that operate one step at a time: they generate each token conditioned on the sequence of tokens previously generated. This process is not parallelizable, and, in the case of neural MT models, it is particularly slow because a computationally intensive neural network is used to generate each token. + +While several recently proposed models avoid recurrence at train time by leveraging convolutions (Kalchbrenner et al., 2016; Gehring et al., 2017; Kaiser et al., 2017) or self-attention (Vaswani et al., 2017) as more-parallelizable alternatives to recurrent neural networks (RNNs), use of autoregressive decoding makes it impossible to take full advantage of parallelism during inference. + +We introduce a non-autoregressive translation model based on the Transformer network (Vaswani et al., 2017). We modify the encoder of the original Transformer network by adding a module that predicts fertilities, sequences of numbers that form an important component of many traditional machine translation models (Brown et al., 1993). These fertilities are supervised during training and provide the decoder at inference time with a globally consistent plan on which to condition its simultaneously computed outputs. + +# 2 BACKGROUND + +# 2.1 AUTOREGRESSIVE NEURAL MACHINE TRANSLATION + +Given a source sentence $X = \{ x _ { 1 } , . . . , x _ { T ^ { \prime } } \}$ , a neural machine translation model factors the distribution over possible output sentences $Y = \{ y _ { 1 } , . . . , y _ { T } \}$ into a chain of conditional probabilities with a + +left-to-right causal structure: + +$$ +p _ { \mathcal { A } \mathcal { R } } ( \boldsymbol { Y } | \boldsymbol { X } ; \theta ) = \prod _ { t = 1 } ^ { T + 1 } p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \prime } } ; \theta ) , +$$ + +where the special tokens $y _ { 0 }$ (e.g. $\left. \mathrm { b o s } \right.$ ) and $y _ { T + 1 }$ (e.g. $\langle \cos \rangle$ ) are used to represent the beginning and end of all target sentences. These conditional probabilities are parameterized using a neural network. Typically, an encoder-decoder architecture (Sutskever et al., 2014) with a unidirectional RNN-based decoder is used to capture the causal structure of the output distribution. + +Maximum Likelihood training Choosing to factorize the machine translation output distribution autoregressively enables straightforward maximum likelihood training with a cross-entropy loss applied at each decoding step: + +$$ +\mathcal { L } _ { \mathrm { M L } } = \log p _ { \mathcal { A R } } ( Y | X ; \theta ) = \sum _ { t = 1 } ^ { T + 1 } \log p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \prime } } ; \theta ) . +$$ + +This loss provides direct supervision for each conditional probability prediction. + +Autoregressive NMT without RNNs Since the entire target translation is known at training time, the calculation of later conditional probabilities (and their corresponding losses) does not depend on the output words chosen during earlier decoding steps. Even though decoding must remain entirely sequential during inference, models can take advantage of this parallelism during training. One such approach replaces recurrent layers in the decoder with masked convolution layers (Kalchbrenner et al., 2016; Gehring et al., 2017) that provide the causal structure required by the autoregressive factorization. + +A recently introduced option which reduces sequential computation still further is to construct the decoder layers out of self-attention computations that have been causally masked in an analogous way. The state-of-the-art Transformer network takes this approach, which allows information to flow in the decoder across arbitrarily long distances in a constant number of operations, asymptotically fewer than required by convolutional architectures (Vaswani et al., 2017). + +# 2.2 NON-AUTOREGRESSIVE DECODING + +Pros and cons of autoregressive decoding The autoregressive factorization used by conventional NMT models has several benefits. It corresponds to the word-by-word nature of human language production and effectively captures the distribution of real translations. Autoregressive models achieve state-of-the-art performance on large-scale corpora and are easy to train, while beam search provides an effective local search method for finding approximately-optimal output translations. + +But there are also drawbacks. As the individual steps of the decoder must be run sequentially rather than in parallel, autoregressive decoding prevents architectures like the Transformer from fully realizing their train-time performance advantage during inference. Meanwhile, beam search suffers from diminishing returns with respect to beam size (Koehn & Knowles, 2017) and exhibits limited search parallelism because it introduces computational dependence between beams. + +Towards non-autoregressive decoding A na¨ıve solution is to remove the autoregressive connection directly from an existing encoder-decoder model. Assuming that the target sequence length $T$ can be modeled with a separate conditional distribution $p _ { L }$ , this becomes + +![](images/fb73e32da81d3465f857ae37043ab008349971a74c732394fc409a62ee78d60d.jpg) +Figure 1: Translating “A B C” to $^ { 6 6 } \mathrm { X }$ Y” using autoregressive and non-autoregressive neural MT architectures. The latter generates all output tokens in parallel. + +$$ +p _ { \mathcal N A } ( Y | X ; \theta ) = p _ { L } ( T | x _ { 1 : T ^ { \prime } } ; \theta ) \cdot \prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 : T ^ { \prime } } ; \theta ) . +$$ + +This model still has an explicit likelihood function, and it can still be trained using independent cross-entropy losses on each output distribution. Now, however, these distributions can be computed in parallel at inference time. + +![](images/899dc15129b43e72de49a617f813130a25104e3d34171cb067e8d5c905671e3e.jpg) +Figure 2: The architecture of the NAT, where the black solid arrows represent differentiable connections and the purple dashed arrows are non-differentiable operations. Each sublayer inside the encoder and decoder stacks also includes layer normalization and a residual connection. + +# 2.3 THE MULTIMODALITY PROBLEM + +However, this na¨ıve approach does not yield good results, because such a model exhibits complete conditional independence. Each token’s distribution $p ( y _ { t } )$ depends only on the source sentence $X$ . This makes it a poor approximation to the true target distribution, which exhibits strong correlation across time. Intuitively, such a decoder is akin to a panel of human translators each asked to provide a single word of a translation independently of the words their colleagues choose. + +In particular, consider an English source sentence like “Thank you.” This can be accurately translated into German as any one of “Danke.”, “Danke schon.”, or “Vielen Dank.”, all of which may ¨ occur in a given training corpus. This target distribution cannot be represented as a product of independent probability distributions for each of the first, second, and third words, because a conditionally independent distribution cannot allow “Danke schon.” and “Vielen Dank.” without also ¨ licensing “Danke Dank.” and “Vielen schon.”¨ + +The conditional independence assumption prevents a model from properly capturing the highly multimodal distribution of target translations. We call this the “multimodality problem” and introduce both a modified model and new training techniques to tackle this issue. + +# 3 THE NON-AUTOREGRESSIVE TRANSFORMER (NAT) + +We introduce a novel NMT model—the Non-Autoregressive Transformer (NAT)—that can produce an entire output translation in parallel. As shown in Fig. 2, the model is composed of the following four modules: an encoder stack, a decoder stack, a newly added fertility predictor (details in 3.3), and a translation predictor for token decoding. + +# 3.1 ENCODER STACK + +Similar to the autoregressive Transformer, both the encoder and decoder stacks are composed entirely of feed-forward networks (MLPs) and multi-head attention modules. Since no RNNs are used, there is no inherent requirement for sequential execution, making non-autoregressive decoding possible. For our proposed NAT, the encoder stays unchanged from the original Transformer network. + +# 3.2 DECODER STACK + +In order to translate non-autoregressively and parallelize the decoding process, we modify the decoder stack as follows. + +Decoder Inputs Before decoding starts, the NAT needs to know how long the target sentence will be in order to generate all words in parallel. More crucially, we cannot use time-shifted target outputs (during training) or previously predicted outputs (during inference) as the inputs to the first decoder layer. Omitting inputs to the first decoder layer entirely, or using only positional embeddings, resulted in very poor performance. Instead, we initialize the decoding process using copied source inputs from the encoder side. As the source and target sentences are often of different lengths, we propose two methods: + +• Copy source inputs uniformly: Each decoder input $t$ is a copy of the Round $( T ^ { \prime } t / T )$ -th encoder input. This is equivalent to “scanning” source inputs from left to right with a constant “speed,” and results in a decoding process that is deterministic given a (predicted) target length. • Copy source inputs using fertilities: A more powerful way, depicted in Fig. 2 and discussed in more detail below, is to copy each encoder input as a decoder input zero or more times, with the number of times each input is copied referred to as that input word’s “fertility.” In this case the source inputs are scanned from left to right at a “speed” that varies inversely with the fertility of each input; the decoding process is now conditioned on the sequence of fertilities, while the resulting output length is determined by the sum of all fertility values. + +Non-causal self-attention Without the constraint of an autoregressive factorization of the output distribution, we no longer need to prevent earlier decoding steps from accessing information from later steps. Thus we can avoid the causal mask used in the self-attention module of the conventional Transformer’s decoder. Instead, we mask out each query position only from attending to itself, which we found to improve decoder performance relative to unmasked self-attention. + +Positional attention We also include an additional positional attention module in each decoder layer, which is a multi-head attention module with the same general attention mechanism used in other parts of the Transformer network, i.e. + +$$ +{ \mathrm { A t t e n t i o n } } ( Q , K , V ) = { \mathrm { s o f t m a x } } \left( { \frac { Q K ^ { T } } { \sqrt { d _ { \mathrm { m o d e l } } } } } \right) \cdot V , +$$ + +where $d _ { \mathrm { m o d e l } }$ is the model hidden size, but with the positional encoding1 as both query and key and the decoder states as the value. This incorporates positional information directly into the attention process and provides a stronger positional signal than the embedding layer alone. We also hypothesize that this additional information improves the decoder’s ability to perform local reordering. + +# 3.3 MODELING FERTILITY TO TACKLE THE MULTIMODALITY PROBLEM + +The multimodality problem can be attacked by introducing a latent variable $z$ to directly model the nondeterminism in the translation process: we first sample $z$ from a prior distribution and then condition on $z$ to non-autoregressively generate a translation. + +One way to interpret this latent variable is as a sentence-level “plan” akin to those discussed in the language production literature (Martin et al., 2010). There are several desirable properties for this latent variable: + +• It should be simple to infer a value for the latent variable given a particular input-output pair, as this is needed to train the model end-to-end. • Adding $z$ to the conditioning context should account as much as possible for the correlations across time between different outputs, so that the remaining marginal probabilities at each output location are as close as possible to satisfying conditional independence. • It should not account for the variation in output translations so directly that $p ( y | x , z )$ becomes trivial to learn, since that is the function our decoder neural network will approximate. + +The factorization by length introduced in Eq. 3 provides a very weak example of a latent variable model, satisfying the first and third property but not the first. We propose the use of fertilities instead. These are integers for each word in the source sentence that correspond to the number of words in the target sentence that can be aligned to that source word using a hard alignment algorithm like IBM Model 2 (Brown et al., 1993). + +One of the most important properties of the proposed NAT is that it naturally introduces an informative latent variable when we choose to copy the encoder inputs based on predicted fertilities. More precisely, given a source sentence $X$ , the conditional probability of a target translation $Y$ is: + +$$ +p _ { \mathcal N \mathcal M } ( Y | X ; \theta ) = \sum _ { f _ { 1 } , . . . , f _ { T ^ { \prime } } \in \mathcal F } \left( \prod _ { t ^ { \prime } = 1 } ^ { T ^ { \prime } } p _ { F } ( f _ { t ^ { \prime } } | x _ { 1 : T ^ { \prime } } ; \theta ) \cdot \prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 } \{ f _ { 1 } \} , . . , x _ { T ^ { \prime } } \{ f _ { T ^ { \prime } } \} ; \theta ) \right) +$$ + +where value $\begin{array} { r } { \mathcal { F } = \{ f _ { 1 } , . . . , f _ { T ^ { \prime } } | \sum _ { t ^ { \prime } = 1 } ^ { T ^ { \prime } } f _ { t ^ { \prime } } = T , f _ { t ^ { \prime } } \in \mathbb { Z } ^ { * } \} } \end{array}$ is thand of all fertility sequdenotes the token ces—onerepeated ertilitytimes. $Y$ $x \{ f \}$ $x$ $f$ + +Fertility prediction As shown in Fig. 2, we model the fertility $p _ { F } \big ( f _ { t ^ { \prime } } | x _ { 1 : T ^ { \prime } } \big )$ at each position independently using a one-layer neural network with a softmax classifier $L = 5 0$ in our experiments) on top of the output of the last encoder layer. This models the way that fertility values are a property of each input word but depend on information and context from the entire sentence. + +Benefits of fertility Fertilities possess all three of the properties listed earlier as desired of a latent variable for non-autoregressive machine translation: + +• An external aligner provides a simple and fast approximate inference model that effectively reduces the unsupervised training problem to two supervised ones. +• Using fertilities as a latent variable makes significant progress towards solving the multimodality problem by providing a natural factorization of the output space. Given a source sentence, restricting the output distribution to those target sentences consistent with a particular fertility sequence dramatically reduces the mode space. Furthermore, the global choice of mode is factored into a set of local mode choices: namely, how to translate each input word. These local mode choices can be effectively supervised because the fertilities provide a fixed “scaffold.” +Including both fertilities and reordering in the latent variable would provide complete alignment statistics. This would make the decoding function trivially easy to approximate given the latent variable and force all of the modeling complexity into the encoder. Using fertilities alone allows the decoder to take some of this burden off of the encoder. + +Our use of fertilities as a latent variable also means that there is no need to have a separate means of explicitly modeling the length of the translation, which is simply the sum of fertilities. And fertilities provide a powerful way to condition the decoding process, allowing the model to generate diverse translations by sampling over the fertility space. + +# 3.4 TRANSLATION PREDICTOR AND THE DECODING PROCESS + +At inference time, the model can identify the translation with the highest conditional probability (see Eq. 5) by marginalizing over all possible latent fertility sequences. Given a fertility sequence, however, identifying the optimal translation only requires independently maximizing the local probability for each output position. We define $Y \overset { \cdot } { = } G ( x _ { 1 : T ^ { \prime } } , f _ { 1 : T ^ { \prime } } ; \theta )$ to represent the optimal translation given a source sentence and a sequence of fertility values. + +But searching and marginalizing over the whole fertility space is still intractable. We propose three heuristic decoding algorithms to reduce the search space of the NAT model: + +Argmax decoding Since the fertility sequence is also modeled with a conditionally independent factorization, we can simply estimate the best translation by choosing the highest-probability fertility for each input word: + +$$ +\hat { Y } _ { \mathrm { a r g m a x } } = G ( x _ { 1 : T ^ { \prime } } , \hat { f } _ { 1 : T ^ { \prime } } ; \theta ) , \mathrm { w h e r e \ } \hat { f } _ { t ^ { \prime } } = \underset { f } { \operatorname { a r g m a x } } p _ { F } ( f _ { t ^ { \prime } } | x _ { 1 : T ^ { \prime } } ; \theta ) +$$ + +Average decoding We can also estimate each fertility as the expectation of its corresponding softmax distribution: + +$$ +\hat { Y } _ { \mathrm { a v e r a g e } } = G ( x _ { 1 : T ^ { \prime } } , \hat { f } _ { 1 : T ^ { \prime } } ; \theta ) , \mathrm { w h e r e } \hat { f } _ { t ^ { \prime } } = \mathrm { R o u n d } \left( \sum _ { f _ { t ^ { \prime } } = 1 } ^ { L } p _ { F } \big ( f _ { t ^ { \prime } } | x _ { 1 : T ^ { \prime } } ; \theta \big ) f _ { t ^ { \prime } } \right) +$$ + +Noisy parallel decoding (NPD) A more accurate approximation of the true optimum of the target distribution, inspired by Cho (2016), is to draw samples from the fertility space and compute the best translation for each fertility sequence. We can then use the autoregressive teacher to identify the best overall translation: + +$$ +\hat { Y } _ { \mathrm { N P D } } = G ( x _ { 1 : T ^ { \prime } } , \underset { f _ { t ^ { \prime } } \sim p _ { F } } { \mathrm { a r g m a x } } p _ { \mathcal { A R } } ( G ( x _ { 1 : T ^ { \prime } } , f _ { 1 : T ^ { \prime } } ; \theta ) | X ; \theta ) ; \theta ) +$$ + +Note that, when using an autoregressive model as a scoring function for a set of decoded translations, it can run as fast as it does at train time because it can be provided with all decoder inputs in parallel. + +NPD is a stochastic search method, and it also increases the computational resources required linearly by the sample size. However, because all the search samples can be computed and scored entirely independently, the process only doubles the latency compared to computing a single translation if sufficient parallelism is available. + +# 4 TRAINING + +The proposed NAT contains a discrete sequential latent variable $f _ { 1 : T ^ { \prime } }$ , whose conditional posterior distribution $p ( f _ { 1 : T ^ { \prime } } | x _ { 1 : T ^ { \prime } } , y _ { 1 : T } ; \theta )$ we can approximate using a proposal distribution $q ( f _ { 1 : T ^ { \prime } } | x _ { 1 : T ^ { \prime } } , y _ { 1 : T } )$ . This provides a variational bound for the overall maximum likelihood loss: + +$$ +\begin{array} { r l } & { \mathcal { L } _ { \mathrm { M L } } = \log p _ { \mathcal { N A } } ( Y | X ; \theta ) = \log \displaystyle \sum _ { f _ { 1 : T ^ { \prime } } \in \mathcal { F } } p _ { F } ( f _ { 1 : T ^ { \prime } } | x _ { 1 : T ^ { \prime } } ; \theta ) \cdot p ( y _ { 1 : T } | x _ { 1 : T ^ { \prime } } , f _ { 1 : T ^ { \prime } } ; \theta ) } \\ & { \geq \displaystyle \sum _ { f _ { 1 : T ^ { \prime } \sim \mathcal { I } } } \left( \underbrace { T } _ { \mathrm { t - 1 } } \log p ( y _ { t } | x _ { 1 } \{ f _ { 1 } \} , . . . , x _ { T ^ { \prime } } \{ f _ { T ^ { \prime } } \} ; \theta ) \right)} _ { \mathrm { T r a n s l a t i o n ~ L o s s } } + \underbrace { \sum _ { t ^ { \prime } = 1 } ^ { T ^ { \prime } } \log p _ { F } ( f _ { t ^ { \prime } } | x _ { 1 : T ^ { \prime } } ; \theta ) } _ { \mathrm { F e r t i l i t y ~ L o s s } } + \mathcal { H } ( q ) \end{array} +$$ + +We choose a proposal distribution $q$ defined by a separate, fixed fertility model. Possible options include the output of an external aligner, which produces a deterministic sequence of integer fertilities for each (source, target) pair in a training corpus, or fertilities computed from the attention weights used in our fixed autoregressive teacher model. This simplifies the inference process considerably, as the expectation over $q$ is deterministic. + +The resulting loss function, consisting of the two bracketed terms in Eq. 9, allows us to train the entire model in a supervised fashion, using the inferred fertilities to simultaneously train the translation model $p$ and supervise the fertility neural network model $p _ { F }$ . + +# 4.1 SEQUENCE-LEVEL KNOWLEDGE DISTILLATION + +While the latent fertility model substantially improves the ability of the non-autoregressive output distribution to approximate the multimodal target distribution, it does not completely solve the problem of nondeterminism in the training data. In many cases, there are multiple correct translations consistent with a single sequence of fertilities—for instance, both “Danke schon.” and “Vielen ¨ dank.” are consistent with the English input “Thank you.” and the fertility sequence [2, 0, 1], because “you” is not directly translated in either German sentence. + +Thus we additionally apply sequence-level knowledge distillation (Kim & Rush, 2016) to construct a new corpus by training an autoregressive machine translation model, known as the teacher, on an existing training corpus, then using that model’s greedy outputs as the targets for training the nonautoregressive student. The resulting targets are less noisy and more deterministic, as the trained model will consistently translate a sentence like “Thank you.” into the same German translation every time; on the other hand, they are also lower in quality than the original dataset. + +# 4.2 FINE-TUNING + +Our supervised fertility model enables a decomposition of the overall maximum likelihood loss into translation and fertility terms, but it has some drawbacks compared to variational training. In particular, it heavily relies on the deterministic, approximate inference model provided by the external alignment system, while it would be desirable to train the entire model, including the fertility predictor, end to end. + +Thus we propose a fine-tuning step after training the NAT to convergence. We introduce an additional loss term consisting of the reverse K-L divergence with the teacher output distribution, a form of word-level knowledge distillation: + +$$ +\mathcal { L } _ { \mathrm { R K L } } \left( f _ { 1 : T ^ { \prime } } ; \theta \right) = \sum _ { t = 1 } ^ { T } \sum _ { y _ { t } } \left[ \log p _ { \mathcal { A R } } \left( y _ { t } | \hat { y } _ { 1 : t - 1 } , x _ { 1 : T ^ { \prime } } \right) \cdot p _ { \mathcal { N A } } \left( y _ { t } | x _ { 1 : T ^ { \prime } } , f _ { 1 : T ^ { \prime } } ; \theta \right) \right] , +$$ + +where $\hat { y } _ { 1 : T } = G ( x _ { 1 : T ^ { \prime } } , f _ { 1 : T ^ { \prime } } ; \theta )$ . Such a loss is more favorable towards highly peaked student output distributions than a standard cross-entropy error would be. + +Then we train the whole model jointly with a weighted sum of the original distillation loss and two such terms, one an expectation over the predicted fertility distribution, normalized with a baseline, and the other based on the external fertility inference model: + +$$ +\mathcal { L } _ { \mathrm { F T } } = \lambda \left( \underbrace { \mathbb { E } } _ { \int _ { \mathrm { I : T : } } r \sim p _ { F } } \left( \mathcal { L } _ { \mathrm { R K L } } \left( f _ { 1 : T ^ { \prime } } \right) - \mathcal { L } _ { \mathrm { R K L } } \left( \bar { f } _ { 1 : T ^ { \prime } } \right) \right) + \underbrace { \mathbb { E } } _ { f _ { 1 : T ^ { \prime } \sim q } } \left( \mathcal { L } _ { \mathrm { R K L } } \left( f _ { 1 : T ^ { \prime } } \right) \right) _ { \mathcal { L } _ { \mathrm { R P } } } \right) + ( 1 - \lambda ) \mathcal { L } _ { \mathrm { K D } } , +$$ + +where ${ \bar { f } } _ { 1 : T ^ { \prime } }$ is the average fertility computed by Eq. 7. The gradient with respect to the nondifferentiable ${ \mathcal { L } } _ { \mathrm { R L } }$ term can be estimated with REINFORCE (Williams, 1992), while the $\mathcal { L } _ { \mathrm { B P } }$ term can be trained using ordinary backpropagation. + +# 5 EXPERIMENTS + +# 5.1 EXPERIMENTAL SETTINGS + +Dataset We evaluate the proposed NAT on three widely used public machine translation corpora: IWSLT16 En–De2, WMT14 En–De,3 and WMT16 En–Ro4. We use IWSLT—which is smaller than the other two datasets—as the development dataset for ablation experiments, and additionally train and test our primary models on both directions of both WMT datasets. All the data are tokenized and segmented into subword symbols using byte-pair encoding (BPE) (Sennrich et al., 2015) to restrict the size of the vocabulary. For both WMT datasets, we use shared BPE vocabulary and additionally share encoder and decoder word embeddings; for IWSLT, we use separate English and German vocabulary and embeddings. + +Teacher Sequence-level knowledge distillation is applied to alleviate multimodality in the training dataset, using autoregressive models as the teachers. The same teacher model used for distillation is also used as a scoring function for fine-tuning and noisy parallel decoding. + +To enable a fair comparison, and benefit from its high translation quality, we implemented the autoregressive teachers using the state-of-the-art Transformer architecture. In addition, we use the same sizes and hyperparameters for each student and its respective teacher, with the exception of the newly added positional self-attention and fertility prediction modules. + +
ModelsWMT14WMT16IWSLT16
En→DeDe→EnEn→RoRo→EnEn→DeLatency / Speedup
NAT17.3520.6226.2227.8325.2039 ms15.6×
NAT (+FT)17.6921.4727.2929.0626.5239 ms15.6×
NAT (+FT + NPD s = 10)18.6622.4129.0230.7627.4479 ms7.68×
NAT (+FT + NPD s = 100)19.1723.2029.7931.4428.16257 ms2.36×
Autoregressive (b = 1)22.7126.3931.3531.0328.89408 ms1.49×
Autoregressive (b = 4)23.4527.0231.9131.7629.70607ms1.00×
+ +Table 1: BLEU scores on official test sets (newstest2014 for WMT En-De and newstest2016 for WMT En-Ro) or the development set for IWSLT. NAT models without NPD use argmax decoding. Latency is computed as the time to decode a single sentence without minibatching, averaged over the whole test set; decoding is implemented in PyTorch on a single NVIDIA Tesla P100. + +Preparation for knowledge distillation We first train all teacher models using maximum likelihood, then freeze their parameters. To avoid the redundancy of running fixed teacher models repeatedly on the same data, we decode the entire training set once using each teacher to create a new training dataset for its respective student. + +Encoder initialization We find it helpful to initialize the weights in the NAT student’s encoder with the encoder weights from its teacher, as the autoregressive and non-autoregressive models share the same encoder input and architecture. + +Fertility supervision during training As described above, we supervise the fertility predictions at train time by using a fixed aligner as a fertility inference function. We use the fast align5 implementation of IBM Model 2 for this purpose, with default parameters (Dyer et al., 2013). + +![](images/ff2cfd4d2f25ef6e98d968d65066c8e02ea0f947a91d09952aa8c8ffbab3302a.jpg) +Figure 3: BLEU scores on IWSLT development set as a function of sample size for noisy parallel decoding. NPD matches the performance of the other two decoding strategies after two samples, and exceeds the performance of the autoregressive teacher with around 1000. + +Hyperparameters For experiments on WMT datasets, we use the hyperparameter settings of the base Transformer model described in Vaswani et al. (2017), though without label smoothing. As IWSLT is a smaller corpus, and to reduce training time, we use a set of smaller hyperparameters $( d _ { \mathrm { m o d e l } } = 2 8 7 , d _ { \mathrm { h i d d e n } } = 5 0 7 , n _ { \mathrm { l a y e r } } = 5 , n _ { \mathrm { h e a d } } = \mathrm { \Omega } ^ { \ast }$ , and $t _ { \mathrm { w a r m u p } } = 7 4 6 )$ for all experiments on that dataset. For fine-tuning we use $\lambda = 0 . 2 5$ . + +Evaluation metrics We evaluate using tokenized and cased BLEU scores (Papineni et al., 2002). + +Implementation We have open-sourced our PyTorch implementation of the $\mathrm { N A T ^ { 6 } }$ . + +# 5.2 RESULTS + +Across the three datasets we used, the NAT performs between 2-5 BLEU points worse than its autoregressive teacher, with part or all of this gap addressed by the use of noisy parallel decoding. In the case of WMT16 English–Romanian, NPD improves the performance of our non-autoregressive model to within 0.2 BLEU points of the previous overall state of the art (Gehring et al., 2017). + +Comparing latencies on the development model shows a speedup of more than a factor of 10 over greedy autoregressive decoding, or a factor of 15 over beam search. Latencies for decoding with NPD, regardless of sample size, could be reduced to about $8 0 \mathrm { m s }$ by parallelizing across multiple GPUs because each sample can be generated, then scored, independently from the others. + +# 5.3 ABLATION STUDY + +We also conduct an extensive ablation study with the proposed NAT on the IWSLT dataset. First, we note that the model fails to train when provided with only positional embeddings as input to the decoder. Second, we see that training on the distillation corpus rather than the ground truth provides a fairly consistent improvement of around 5 BLEU points. Third, switching from uniform copying of source inputs to fertility-based copying improves performance by four BLEU points when using ground-truth training or two when using distillation. + +
Distillation b=1 b=4Decoder Inputs +uniform+fertility+PosAttFine-tuning +LKD +LBPBLEU +CRLBLEU (T)
~2
√16.51
√18.87
√√√20.72
√√21.12
<√24.0243.91
√25.2045.41
√√22.44 厂
兴 √√××
√√××
√√√√25.7646.11
√√√√26.5247.38
+ +Table 2: Ablation performance on the IWSLT development set. BLEU (T) refers to the BLEU score on a version of the development set that has been translated by the teacher model. An $\times$ indicates that fine-tuning caused that model to get worse. When uniform copying is used as the decoder inputs, the ground-truth target lengths are provided. All models use argmax decoding. + +Fine-tuning does not converge with reinforcement learning alone, or with the $\mathcal { L } _ { \mathrm { B P } }$ term alone, but use of all three fine-tuning terms together leads to an improvement of around 1.5 BLEU points. Training the student model from a distillation corpus produced using beam search is similar to training from the greedily-distilled corpus. + +Figure 4: Two examples comparing translations produced by an autoregressive (AR) and nonautoregressive Transformer as well as the result of noisy parallel decoding with sample size 100. Repeated words are highlighted in gray. + +
Source:politicians try to pick wordsand use words to shape realityand control reality,butin fact,reality changes words
far more than words can ever change reality.
Target: AR:Politiker versuchen Worte zu benutzen,um die Realitat zu formen und die Realitat zu kontrolieren,aber
tatsächlich verandert die Realitat Worte viel mehr,als Worte die Realitat jemals verändern konnten.
Politikerversuchen Worter zu wahlen und Worter zur Realitat zu gestalten und Realitätzu steuern,aber in
NAT:Wirklichkeit verändert sich die Realitat viel mehrals Worte,die die Realität verändern konnen.
Politikerversuchen,Worter wahlen und zu verwenden,um Realitatzu formen und Realitatzu formen,aber
NAT+NPD:tatsächlich ändert Realitat Realitat viel mehrals Wortedie Realität Realitat verändern.
Politikerversuchen,Worter wahlenund zu verwenden,um Realitat Realitat formenund die Realitatzu formen， aber tatsächlich ändert die Realität Worte viel mehrals Worte jemals die Realität verändern konnen.
Source:Isee wheelchairs bought and sold like used cars.
Target:ich erlebe,dass Rollstuhle gekauft und verkauft werden wie Gebrauchtwagen
AR: NAT:ich sehe Rollstuhlen,die wie Autos verkauft und verkauft werden.
NAT+NPD:ich sehe,dass Stuhle Stuhle und verkauftwie Autos verkauft.
ich sehe Rollühle kauften und verkaufte wie Autos.
+ +We include two examples of translations from the IWSLT development set in Fig. 4. Instances of repeated words or phrases, highlighted in gray, are most prevalent in the non-autoregressive output for the relatively complex first example sentence. Two pairs of repeated words in the first example, as + +se lucreaza la soluti de genul acesta . + +se la solutii de genul acesta . +se lucreaza la solutii de acesta . +se lucreaza solutii de genul acesta . +se se lucreaza la solutii de acesta . +se lucreaza lucreaza la solutii de acesta . +se se lucreaza lucreaza la solutii de acesta . +se se lucreaza lucreaza la solutii de de acesta . +se se lucreaza lucreaza la solutii de genul acesta . solutions on this kind are done. +work done on solutions like this . +solutions on this kind is done . +work is done on solutions like this . +work is done on solutions like this . +work is being done on solutions like this . +work is being done on solutions such as this. +work is being done on solutions such this kind . + +Figure 5: A Romanian–English example translated with noisy parallel decoding. At left are eight sampled fertility sequences from the encoder, represented with their corresponding decoder input sequences. Each of these values for the latent variable leads to a different possible output translation, shown at right. The autoregressive Transformer then picks the best translation, shown in red, a process which is much faster than directly using it to generate output. + +well as a pair in the second, are not present in the versions with noisy parallel decoding, suggesting that NPD scoring using the teacher model can filter out such mistakes. The translations produced by the NAT with NPD, while of a similar quality to those produced by the autoregressive model, are also noticeably more literal. + +We also show an example of the noisy parallel decoding process in Fig. 5, demonstrating the diversity of translations that can be found by sampling from the fertility space. + +# 6 CONCLUSION + +We introduce a latent variable model for non-autoregressive machine translation that enables a decoder based on Vaswani et al. (2017) to take full advantage of its exceptional degree of internal parallelism even at inference time. As a result, we measure translation latencies of one-tenth that of an equal-sized autoregressive model, while maintaining competitive BLEU scores. + +# REFERENCES + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015. + +Peter Brown, Vincent della Pietra, Stephen della Pietra, and Robert Mercer. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263–311, 1993. + +Kyunghyun Cho. Noisy parallel approximate decoding for conditional recurrent language model. arXiv preprint arXiv:1605.03835, 2016. + +Chris Dyer, Victor Chahuneau, and Noah Smith. A simple, fast, and effective reparameterization of IBM Model 2. In NAACL, 2013. + +Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann Dauphin. Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122, 2017. + +Łukasz Kaiser, Aidan Gomez, and Franc¸ois Chollet. Depthwise separable convolutions for neural machine translation. arXiv preprint arXiv:1706.03059, 2017. + +Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukc¸uoglu. Neural machine translation in linear time. ˇ arXiv preprint arXiv:1610.10099, 2016. + +Yoon Kim and Alexander Rush. Sequence-level knowledge distillation. In EMNLP, 2016. + +Philipp Koehn and Rebecca Knowles. Six challenges for neural machine translation. arXiv preprint arXiv:1706.03872, 2017. + +Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attentionbased neural machine translation. In EMNLP, 2015. + +Randi Martin, Jason Crowther, Meredith Knight, Franklin Tamborello, and Chin-Lung Yang. Planning in sentence production: Evidence for the phrase as a default planning scope. Cognition, 116 (2):177–192, 2010. + +Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In ACL, pp. 311–318, 2002. + +Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. + +Ilya Sutskever, Oriol Vinyals, and Quoc L ˆ e. Sequence to sequence learning with neural networks. ˆ In NIPS, 2014. + +Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. + +Ronald Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229–256, 1992. + +Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey, J. Klingner, A. Shah, M. Johnson, X. Liu, Ł. Kaiser, S. Gouws, Y. Kato, T. Kudo, H. Kazawa, K. Stevens, G. Kurian, N. Patil, W. Wang, C. Young, J. Smith, J. Riesa, A. Rudnick, O. Vinyals, G. Corrado, M. Hughes, and J. Dean. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. + +![](images/8397d58044f6d9c8a19dcbb03ac376664021ffaa63d7eac01c5e0d5c288a1df7.jpg) +Figure 6: The schematic structure of training and inference for the NAT. The “distilled data” contains target sentences decoded by the autoregressive model and ground-truth source sentences. + +![](images/65e35fd7ac2e233ef8cdae03751ebce2756e57bc1378b23f1a59c83c62823be8.jpg) +Figure 7: The translation latency, computed as the time to decode a single sentence without minibatching, for each sentence in the IWSLT development set as a function of its length. The autoregressive model has latency linear in the decoding length, while the latency of the NAT is nearly constant for typical lengths, even with NPD with sample size 10. When using NPD with sample size 100, the level of parallelism is enough to more than saturate the GPU, leading again to linear latencies. + +![](images/19326d4ac73a662073b026f5ada8aa0be903a2801d0ffdb32756ff0c9f97cdcc.jpg) +Figure 8: Learning curves for training and fine-tuning of the NAT on IWSLT. BLEU scores are on the development set. \ No newline at end of file diff --git a/parse/train/B1l8BtlCb/B1l8BtlCb_content_list.json b/parse/train/B1l8BtlCb/B1l8BtlCb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b2e307427d3111f5d949d04f0e85adaaa5ad1465 --- /dev/null +++ b/parse/train/B1l8BtlCb/B1l8BtlCb_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "NON-AUTOREGRESSIVE NEURAL MACHINE TRANSLATION ", + "text_level": 1, + "bbox": [ + 174, + 101, + 586, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiatao $\\mathbf { G u } ^ { \\dagger }$ ∗, James Bradbury‡, Caiming Xiong‡, Victor O.K. Li†& Richard Socher‡ ", + "bbox": [ + 184, + 169, + 758, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "‡Salesforce Research \n{james.bradbury,cxiong,rsocher}@salesforce.com \n†The University of Hong Kong \n{jiataogu, vli}@eee.hku.hk ", + "bbox": [ + 184, + 186, + 633, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 279, + 544, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing approaches to neural machine translation condition each output word on previously generated outputs. We introduce a model that avoids this autoregressive property and produces its outputs in parallel, allowing an order of magnitude lower latency during inference. Through knowledge distillation, the use of input token fertilities as a latent variable, and policy gradient fine-tuning, we achieve this at a cost of as little as 2.0 BLEU points relative to the autoregressive Transformer network used as a teacher. We demonstrate substantial cumulative improvements associated with each of the three aspects of our training strategy, and validate our approach on IWSLT 2016 English–German and two WMT language pairs. By sampling fertilities in parallel at inference time, our non-autoregressive model achieves near-state-of-the-art performance of 29.8 BLEU on WMT 2016 English– Romanian. ", + "bbox": [ + 233, + 310, + 764, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 506, + 336, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural network based models outperform traditional statistical models for machine translation (MT) (Bahdanau et al., 2015; Luong et al., 2015). However, state-of-the-art neural models are much slower than statistical MT approaches at inference time (Wu et al., 2016). Both model families use autoregressive decoders that operate one step at a time: they generate each token conditioned on the sequence of tokens previously generated. This process is not parallelizable, and, in the case of neural MT models, it is particularly slow because a computationally intensive neural network is used to generate each token. ", + "bbox": [ + 174, + 539, + 825, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While several recently proposed models avoid recurrence at train time by leveraging convolutions (Kalchbrenner et al., 2016; Gehring et al., 2017; Kaiser et al., 2017) or self-attention (Vaswani et al., 2017) as more-parallelizable alternatives to recurrent neural networks (RNNs), use of autoregressive decoding makes it impossible to take full advantage of parallelism during inference. ", + "bbox": [ + 174, + 643, + 823, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce a non-autoregressive translation model based on the Transformer network (Vaswani et al., 2017). We modify the encoder of the original Transformer network by adding a module that predicts fertilities, sequences of numbers that form an important component of many traditional machine translation models (Brown et al., 1993). These fertilities are supervised during training and provide the decoder at inference time with a globally consistent plan on which to condition its simultaneously computed outputs. ", + "bbox": [ + 174, + 707, + 825, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 176, + 810, + 326, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2.1 AUTOREGRESSIVE NEURAL MACHINE TRANSLATION ", + "text_level": 1, + "bbox": [ + 174, + 843, + 583, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a source sentence $X = \\{ x _ { 1 } , . . . , x _ { T ^ { \\prime } } \\}$ , a neural machine translation model factors the distribution over possible output sentences $Y = \\{ y _ { 1 } , . . . , y _ { T } \\}$ into a chain of conditional probabilities with a ", + "bbox": [ + 176, + 869, + 823, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "left-to-right causal structure: ", + "bbox": [ + 174, + 103, + 362, + 118 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/633a748a4002cb496ac4d94d9bde49381f4173fa71b235023f3e7a76c5656215.jpg", + "text": "$$\np _ { \\mathcal { A } \\mathcal { R } } ( \\boldsymbol { Y } | \\boldsymbol { X } ; \\theta ) = \\prod _ { t = 1 } ^ { T + 1 } p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) ,\n$$", + "text_format": "latex", + "bbox": [ + 354, + 122, + 642, + 166 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the special tokens $y _ { 0 }$ (e.g. $\\left. \\mathrm { b o s } \\right.$ ) and $y _ { T + 1 }$ (e.g. $\\langle \\cos \\rangle$ ) are used to represent the beginning and end of all target sentences. These conditional probabilities are parameterized using a neural network. Typically, an encoder-decoder architecture (Sutskever et al., 2014) with a unidirectional RNN-based decoder is used to capture the causal structure of the output distribution. ", + "bbox": [ + 174, + 170, + 825, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Maximum Likelihood training Choosing to factorize the machine translation output distribution autoregressively enables straightforward maximum likelihood training with a cross-entropy loss applied at each decoding step: ", + "bbox": [ + 174, + 234, + 823, + 279 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/975cc03de217f9263a661785207cd37c0ad59d69289d6f9743d2345d95881151.jpg", + "text": "$$\n\\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { A R } } ( Y | X ; \\theta ) = \\sum _ { t = 1 } ^ { T + 1 } \\log p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) .\n$$", + "text_format": "latex", + "bbox": [ + 305, + 282, + 691, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This loss provides direct supervision for each conditional probability prediction. ", + "bbox": [ + 176, + 329, + 699, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive NMT without RNNs Since the entire target translation is known at training time, the calculation of later conditional probabilities (and their corresponding losses) does not depend on the output words chosen during earlier decoding steps. Even though decoding must remain entirely sequential during inference, models can take advantage of this parallelism during training. One such approach replaces recurrent layers in the decoder with masked convolution layers (Kalchbrenner et al., 2016; Gehring et al., 2017) that provide the causal structure required by the autoregressive factorization. ", + "bbox": [ + 173, + 352, + 825, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A recently introduced option which reduces sequential computation still further is to construct the decoder layers out of self-attention computations that have been causally masked in an analogous way. The state-of-the-art Transformer network takes this approach, which allows information to flow in the decoder across arbitrarily long distances in a constant number of operations, asymptotically fewer than required by convolutional architectures (Vaswani et al., 2017). ", + "bbox": [ + 174, + 457, + 825, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 NON-AUTOREGRESSIVE DECODING ", + "text_level": 1, + "bbox": [ + 176, + 544, + 462, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pros and cons of autoregressive decoding The autoregressive factorization used by conventional NMT models has several benefits. It corresponds to the word-by-word nature of human language production and effectively captures the distribution of real translations. Autoregressive models achieve state-of-the-art performance on large-scale corpora and are easy to train, while beam search provides an effective local search method for finding approximately-optimal output translations. ", + "bbox": [ + 173, + 563, + 823, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "But there are also drawbacks. As the individual steps of the decoder must be run sequentially rather than in parallel, autoregressive decoding prevents architectures like the Transformer from fully realizing their train-time performance advantage during inference. Meanwhile, beam search suffers from diminishing returns with respect to beam size (Koehn & Knowles, 2017) and exhibits limited search parallelism because it introduces computational dependence between beams. ", + "bbox": [ + 174, + 640, + 581, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards non-autoregressive decoding A na¨ıve solution is to remove the autoregressive connection directly from an existing encoder-decoder model. Assuming that the target sequence length $T$ can be modeled with a separate conditional distribution $p _ { L }$ , this becomes ", + "bbox": [ + 173, + 760, + 581, + 830 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fb73e32da81d3465f857ae37043ab008349971a74c732394fc409a62ee78d60d.jpg", + "image_caption": [ + "Figure 1: Translating “A B C” to $^ { 6 6 } \\mathrm { X }$ Y” using autoregressive and non-autoregressive neural MT architectures. The latter generates all output tokens in parallel. " + ], + "image_footnote": [], + "bbox": [ + 598, + 643, + 825, + 786 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/d97e3003546373ed25378b0bc946fe1615293facf85f3967deae841210d37fb1.jpg", + "text": "$$\np _ { \\mathcal N A } ( Y | X ; \\theta ) = p _ { L } ( T | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 : T ^ { \\prime } } ; \\theta ) .\n$$", + "text_format": "latex", + "bbox": [ + 194, + 834, + 539, + 877 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This model still has an explicit likelihood function, and it can still be trained using independent cross-entropy losses on each output distribution. Now, however, these distributions can be computed in parallel at inference time. ", + "bbox": [ + 174, + 881, + 823, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/899dc15129b43e72de49a617f813130a25104e3d34171cb067e8d5c905671e3e.jpg", + "image_caption": [ + "Figure 2: The architecture of the NAT, where the black solid arrows represent differentiable connections and the purple dashed arrows are non-differentiable operations. Each sublayer inside the encoder and decoder stacks also includes layer normalization and a residual connection. " + ], + "image_footnote": [], + "bbox": [ + 189, + 99, + 797, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 THE MULTIMODALITY PROBLEM ", + "text_level": 1, + "bbox": [ + 176, + 434, + 444, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, this na¨ıve approach does not yield good results, because such a model exhibits complete conditional independence. Each token’s distribution $p ( y _ { t } )$ depends only on the source sentence $X$ . This makes it a poor approximation to the true target distribution, which exhibits strong correlation across time. Intuitively, such a decoder is akin to a panel of human translators each asked to provide a single word of a translation independently of the words their colleagues choose. ", + "bbox": [ + 174, + 465, + 825, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In particular, consider an English source sentence like “Thank you.” This can be accurately translated into German as any one of “Danke.”, “Danke schon.”, or “Vielen Dank.”, all of which may ¨ occur in a given training corpus. This target distribution cannot be represented as a product of independent probability distributions for each of the first, second, and third words, because a conditionally independent distribution cannot allow “Danke schon.” and “Vielen Dank.” without also ¨ licensing “Danke Dank.” and “Vielen schon.”¨ ", + "bbox": [ + 174, + 542, + 825, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The conditional independence assumption prevents a model from properly capturing the highly multimodal distribution of target translations. We call this the “multimodality problem” and introduce both a modified model and new training techniques to tackle this issue. ", + "bbox": [ + 174, + 633, + 825, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 THE NON-AUTOREGRESSIVE TRANSFORMER (NAT) ", + "text_level": 1, + "bbox": [ + 176, + 709, + 640, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a novel NMT model—the Non-Autoregressive Transformer (NAT)—that can produce an entire output translation in parallel. As shown in Fig. 2, the model is composed of the following four modules: an encoder stack, a decoder stack, a newly added fertility predictor (details in 3.3), and a translation predictor for token decoding. ", + "bbox": [ + 174, + 750, + 825, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 ENCODER STACK ", + "text_level": 1, + "bbox": [ + 176, + 835, + 334, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to the autoregressive Transformer, both the encoder and decoder stacks are composed entirely of feed-forward networks (MLPs) and multi-head attention modules. Since no RNNs are used, there is no inherent requirement for sequential execution, making non-autoregressive decoding possible. For our proposed NAT, the encoder stays unchanged from the original Transformer network. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2 DECODER STACK ", + "text_level": 1, + "bbox": [ + 176, + 103, + 334, + 117 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to translate non-autoregressively and parallelize the decoding process, we modify the decoder stack as follows. ", + "bbox": [ + 173, + 128, + 823, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoder Inputs Before decoding starts, the NAT needs to know how long the target sentence will be in order to generate all words in parallel. More crucially, we cannot use time-shifted target outputs (during training) or previously predicted outputs (during inference) as the inputs to the first decoder layer. Omitting inputs to the first decoder layer entirely, or using only positional embeddings, resulted in very poor performance. Instead, we initialize the decoding process using copied source inputs from the encoder side. As the source and target sentences are often of different lengths, we propose two methods: ", + "bbox": [ + 173, + 166, + 825, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• Copy source inputs uniformly: Each decoder input $t$ is a copy of the Round $( T ^ { \\prime } t / T )$ -th encoder input. This is equivalent to “scanning” source inputs from left to right with a constant “speed,” and results in a decoding process that is deterministic given a (predicted) target length. • Copy source inputs using fertilities: A more powerful way, depicted in Fig. 2 and discussed in more detail below, is to copy each encoder input as a decoder input zero or more times, with the number of times each input is copied referred to as that input word’s “fertility.” In this case the source inputs are scanned from left to right at a “speed” that varies inversely with the fertility of each input; the decoding process is now conditioned on the sequence of fertilities, while the resulting output length is determined by the sum of all fertility values. ", + "bbox": [ + 174, + 273, + 825, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Non-causal self-attention Without the constraint of an autoregressive factorization of the output distribution, we no longer need to prevent earlier decoding steps from accessing information from later steps. Thus we can avoid the causal mask used in the self-attention module of the conventional Transformer’s decoder. Instead, we mask out each query position only from attending to itself, which we found to improve decoder performance relative to unmasked self-attention. ", + "bbox": [ + 174, + 416, + 825, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positional attention We also include an additional positional attention module in each decoder layer, which is a multi-head attention module with the same general attention mechanism used in other parts of the Transformer network, i.e. ", + "bbox": [ + 174, + 494, + 825, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/7591f42425b5f5d9a239ca6fc7711621177920ae7db10daccd0a298639ff6dee.jpg", + "text": "$$\n{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { \\mathrm { m o d e l } } } } } \\right) \\cdot V ,\n$$", + "text_format": "latex", + "bbox": [ + 339, + 541, + 656, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d _ { \\mathrm { m o d e l } }$ is the model hidden size, but with the positional encoding1 as both query and key and the decoder states as the value. This incorporates positional information directly into the attention process and provides a stronger positional signal than the embedding layer alone. We also hypothesize that this additional information improves the decoder’s ability to perform local reordering. ", + "bbox": [ + 174, + 580, + 823, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3 MODELING FERTILITY TO TACKLE THE MULTIMODALITY PROBLEM ", + "text_level": 1, + "bbox": [ + 174, + 652, + 687, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multimodality problem can be attacked by introducing a latent variable $z$ to directly model the nondeterminism in the translation process: we first sample $z$ from a prior distribution and then condition on $z$ to non-autoregressively generate a translation. ", + "bbox": [ + 176, + 679, + 825, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One way to interpret this latent variable is as a sentence-level “plan” akin to those discussed in the language production literature (Martin et al., 2010). There are several desirable properties for this latent variable: ", + "bbox": [ + 176, + 728, + 823, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• It should be simple to infer a value for the latent variable given a particular input-output pair, as this is needed to train the model end-to-end. • Adding $z$ to the conditioning context should account as much as possible for the correlations across time between different outputs, so that the remaining marginal probabilities at each output location are as close as possible to satisfying conditional independence. • It should not account for the variation in output translations so directly that $p ( y | x , z )$ becomes trivial to learn, since that is the function our decoder neural network will approximate. ", + "bbox": [ + 173, + 781, + 826, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The factorization by length introduced in Eq. 3 provides a very weak example of a latent variable model, satisfying the first and third property but not the first. We propose the use of fertilities instead. These are integers for each word in the source sentence that correspond to the number of words in the target sentence that can be aligned to that source word using a hard alignment algorithm like IBM Model 2 (Brown et al., 1993). ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One of the most important properties of the proposed NAT is that it naturally introduces an informative latent variable when we choose to copy the encoder inputs based on predicted fertilities. More precisely, given a source sentence $X$ , the conditional probability of a target translation $Y$ is: ", + "bbox": [ + 174, + 180, + 823, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/addbc095e30bd0f2ec3e327b7ac2d196aff125984ec04b5d2c19836176794d07.jpg", + "text": "$$\np _ { \\mathcal N \\mathcal M } ( Y | X ; \\theta ) = \\sum _ { f _ { 1 } , . . . , f _ { T ^ { \\prime } } \\in \\mathcal F } \\left( \\prod _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)\n$$", + "text_format": "latex", + "bbox": [ + 199, + 229, + 777, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where value $\\begin{array} { r } { \\mathcal { F } = \\{ f _ { 1 } , . . . , f _ { T ^ { \\prime } } | \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } f _ { t ^ { \\prime } } = T , f _ { t ^ { \\prime } } \\in \\mathbb { Z } ^ { * } \\} } \\end{array}$ is thand of all fertility sequdenotes the token ces—onerepeated ertilitytimes. $Y$ $x \\{ f \\}$ $x$ $f$ ", + "bbox": [ + 171, + 287, + 825, + 320 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fertility prediction As shown in Fig. 2, we model the fertility $p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } \\big )$ at each position independently using a one-layer neural network with a softmax classifier $L = 5 0$ in our experiments) on top of the output of the last encoder layer. This models the way that fertility values are a property of each input word but depend on information and context from the entire sentence. ", + "bbox": [ + 173, + 328, + 825, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benefits of fertility Fertilities possess all three of the properties listed earlier as desired of a latent variable for non-autoregressive machine translation: ", + "bbox": [ + 174, + 393, + 823, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "• An external aligner provides a simple and fast approximate inference model that effectively reduces the unsupervised training problem to two supervised ones. \n• Using fertilities as a latent variable makes significant progress towards solving the multimodality problem by providing a natural factorization of the output space. Given a source sentence, restricting the output distribution to those target sentences consistent with a particular fertility sequence dramatically reduces the mode space. Furthermore, the global choice of mode is factored into a set of local mode choices: namely, how to translate each input word. These local mode choices can be effectively supervised because the fertilities provide a fixed “scaffold.” \nIncluding both fertilities and reordering in the latent variable would provide complete alignment statistics. This would make the decoding function trivially easy to approximate given the latent variable and force all of the modeling complexity into the encoder. Using fertilities alone allows the decoder to take some of this burden off of the encoder. ", + "bbox": [ + 173, + 435, + 825, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our use of fertilities as a latent variable also means that there is no need to have a separate means of explicitly modeling the length of the translation, which is simply the sum of fertilities. And fertilities provide a powerful way to condition the decoding process, allowing the model to generate diverse translations by sampling over the fertility space. ", + "bbox": [ + 173, + 625, + 825, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4 TRANSLATION PREDICTOR AND THE DECODING PROCESS ", + "text_level": 1, + "bbox": [ + 174, + 698, + 616, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At inference time, the model can identify the translation with the highest conditional probability (see Eq. 5) by marginalizing over all possible latent fertility sequences. Given a fertility sequence, however, identifying the optimal translation only requires independently maximizing the local probability for each output position. We define $Y \\overset { \\cdot } { = } G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )$ to represent the optimal translation given a source sentence and a sequence of fertility values. ", + "bbox": [ + 173, + 724, + 825, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "But searching and marginalizing over the whole fertility space is still intractable. We propose three heuristic decoding algorithms to reduce the search space of the NAT model: ", + "bbox": [ + 171, + 801, + 823, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Argmax decoding Since the fertility sequence is also modeled with a conditionally independent factorization, we can simply estimate the best translation by choosing the highest-probability fertility for each input word: ", + "bbox": [ + 174, + 839, + 823, + 882 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/c87b82f443ad023fe764cc137a25c0b5f30813f1855f5398ec17018f945487d0.jpg", + "text": "$$\n\\hat { Y } _ { \\mathrm { a r g m a x } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e \\ } \\hat { f } _ { t ^ { \\prime } } = \\underset { f } { \\operatorname { a r g m a x } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta )\n$$", + "text_format": "latex", + "bbox": [ + 279, + 888, + 718, + 917 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Average decoding We can also estimate each fertility as the expectation of its corresponding softmax distribution: ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/cddc8316aa514a2040ceed5ae75c4f96052fd7ea9df38f637de3023bdc4c2c2c.jpg", + "text": "$$\n\\hat { Y } _ { \\mathrm { a v e r a g e } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e } \\hat { f } _ { t ^ { \\prime } } = \\mathrm { R o u n d } \\left( \\sum _ { f _ { t ^ { \\prime } } = 1 } ^ { L } p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta \\big ) f _ { t ^ { \\prime } } \\right)\n$$", + "text_format": "latex", + "bbox": [ + 241, + 136, + 754, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Noisy parallel decoding (NPD) A more accurate approximation of the true optimum of the target distribution, inspired by Cho (2016), is to draw samples from the fertility space and compute the best translation for each fertility sequence. We can then use the autoregressive teacher to identify the best overall translation: ", + "bbox": [ + 174, + 204, + 823, + 260 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/ddb83a2f54effd61e8585241161c5ecfe255d9b7958c6a9e1d1be0a28272f922.jpg", + "text": "$$\n\\hat { Y } _ { \\mathrm { N P D } } = G ( x _ { 1 : T ^ { \\prime } } , \\underset { f _ { t ^ { \\prime } } \\sim p _ { F } } { \\mathrm { a r g m a x } } p _ { \\mathcal { A R } } ( G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) | X ; \\theta ) ; \\theta )\n$$", + "text_format": "latex", + "bbox": [ + 303, + 266, + 694, + 295 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that, when using an autoregressive model as a scoring function for a set of decoded translations, it can run as fast as it does at train time because it can be provided with all decoder inputs in parallel. ", + "bbox": [ + 171, + 303, + 823, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NPD is a stochastic search method, and it also increases the computational resources required linearly by the sample size. However, because all the search samples can be computed and scored entirely independently, the process only doubles the latency compared to computing a single translation if sufficient parallelism is available. ", + "bbox": [ + 174, + 338, + 825, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 TRAINING ", + "text_level": 1, + "bbox": [ + 174, + 415, + 292, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed NAT contains a discrete sequential latent variable $f _ { 1 : T ^ { \\prime } }$ , whose conditional posterior distribution $p ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } ; \\theta )$ we can approximate using a proposal distribution $q ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } )$ . This provides a variational bound for the overall maximum likelihood loss: ", + "bbox": [ + 174, + 446, + 825, + 491 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/98d5a44ed0ad3d2976698667125e9e76d3327aac64ae197046cd812c046fa407.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { N A } } ( Y | X ; \\theta ) = \\log \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } } \\in \\mathcal { F } } p _ { F } ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot p ( y _ { 1 : T } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) } \\\\ & { \\geq \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } \\sim \\mathcal { I } } } \\left( \\underbrace { T } _ { \\mathrm { t - 1 } } \\log p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)} _ { \\mathrm { T r a n s l a t i o n ~ L o s s } } + \\underbrace { \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } \\log p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) } _ { \\mathrm { F e r t i l i t y ~ L o s s } } + \\mathcal { H } ( q ) \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 187, + 507, + 790, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We choose a proposal distribution $q$ defined by a separate, fixed fertility model. Possible options include the output of an external aligner, which produces a deterministic sequence of integer fertilities for each (source, target) pair in a training corpus, or fertilities computed from the attention weights used in our fixed autoregressive teacher model. This simplifies the inference process considerably, as the expectation over $q$ is deterministic. ", + "bbox": [ + 174, + 641, + 825, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The resulting loss function, consisting of the two bracketed terms in Eq. 9, allows us to train the entire model in a supervised fashion, using the inferred fertilities to simultaneously train the translation model $p$ and supervise the fertility neural network model $p _ { F }$ . ", + "bbox": [ + 174, + 718, + 825, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 SEQUENCE-LEVEL KNOWLEDGE DISTILLATION ", + "text_level": 1, + "bbox": [ + 173, + 777, + 545, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While the latent fertility model substantially improves the ability of the non-autoregressive output distribution to approximate the multimodal target distribution, it does not completely solve the problem of nondeterminism in the training data. In many cases, there are multiple correct translations consistent with a single sequence of fertilities—for instance, both “Danke schon.” and “Vielen ¨ dank.” are consistent with the English input “Thank you.” and the fertility sequence [2, 0, 1], because “you” is not directly translated in either German sentence. ", + "bbox": [ + 174, + 804, + 825, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus we additionally apply sequence-level knowledge distillation (Kim & Rush, 2016) to construct a new corpus by training an autoregressive machine translation model, known as the teacher, on an existing training corpus, then using that model’s greedy outputs as the targets for training the nonautoregressive student. The resulting targets are less noisy and more deterministic, as the trained model will consistently translate a sentence like “Thank you.” into the same German translation every time; on the other hand, they are also lower in quality than the original dataset. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 160 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 FINE-TUNING ", + "text_level": 1, + "bbox": [ + 174, + 176, + 312, + 190 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our supervised fertility model enables a decomposition of the overall maximum likelihood loss into translation and fertility terms, but it has some drawbacks compared to variational training. In particular, it heavily relies on the deterministic, approximate inference model provided by the external alignment system, while it would be desirable to train the entire model, including the fertility predictor, end to end. ", + "bbox": [ + 173, + 202, + 825, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thus we propose a fine-tuning step after training the NAT to convergence. We introduce an additional loss term consisting of the reverse K-L divergence with the teacher output distribution, a form of word-level knowledge distillation: ", + "bbox": [ + 174, + 279, + 823, + 320 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/e65993392162bd75d4d5f03dbe09c11504275ad6f8a4d1ecdf2eaf6fe3d240bf.jpg", + "text": "$$\n\\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } ; \\theta \\right) = \\sum _ { t = 1 } ^ { T } \\sum _ { y _ { t } } \\left[ \\log p _ { \\mathcal { A R } } \\left( y _ { t } | \\hat { y } _ { 1 : t - 1 } , x _ { 1 : T ^ { \\prime } } \\right) \\cdot p _ { \\mathcal { N A } } \\left( y _ { t } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta \\right) \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 233, + 324, + 763, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\hat { y } _ { 1 : T } = G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )$ . Such a loss is more favorable towards highly peaked student output distributions than a standard cross-entropy error would be. ", + "bbox": [ + 173, + 373, + 826, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we train the whole model jointly with a weighted sum of the original distillation loss and two such terms, one an expectation over the predicted fertility distribution, normalized with a baseline, and the other based on the external fertility inference model: ", + "bbox": [ + 173, + 409, + 825, + 450 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/57b8b0401c44a9da9dffb4a771b588d3f01db84a0e94c59950fcea341c8459e4.jpg", + "text": "$$\n\\mathcal { L } _ { \\mathrm { F T } } = \\lambda \\left( \\underbrace { \\mathbb { E } } _ { \\int _ { \\mathrm { I : T : } } r \\sim p _ { F } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) - \\mathcal { L } _ { \\mathrm { R K L } } \\left( \\bar { f } _ { 1 : T ^ { \\prime } } \\right) \\right) + \\underbrace { \\mathbb { E } } _ { f _ { 1 : T ^ { \\prime } \\sim q } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) \\right) _ { \\mathcal { L } _ { \\mathrm { R P } } } \\right) + ( 1 - \\lambda ) \\mathcal { L } _ { \\mathrm { K D } } ,\n$$", + "text_format": "latex", + "bbox": [ + 179, + 454, + 789, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where ${ \\bar { f } } _ { 1 : T ^ { \\prime } }$ is the average fertility computed by Eq. 7. The gradient with respect to the nondifferentiable ${ \\mathcal { L } } _ { \\mathrm { R L } }$ term can be estimated with REINFORCE (Williams, 1992), while the $\\mathcal { L } _ { \\mathrm { B P } }$ term can be trained using ordinary backpropagation. ", + "bbox": [ + 174, + 534, + 823, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 595, + 326, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1 EXPERIMENTAL SETTINGS ", + "text_level": 1, + "bbox": [ + 174, + 627, + 400, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset We evaluate the proposed NAT on three widely used public machine translation corpora: IWSLT16 En–De2, WMT14 En–De,3 and WMT16 En–Ro4. We use IWSLT—which is smaller than the other two datasets—as the development dataset for ablation experiments, and additionally train and test our primary models on both directions of both WMT datasets. All the data are tokenized and segmented into subword symbols using byte-pair encoding (BPE) (Sennrich et al., 2015) to restrict the size of the vocabulary. For both WMT datasets, we use shared BPE vocabulary and additionally share encoder and decoder word embeddings; for IWSLT, we use separate English and German vocabulary and embeddings. ", + "bbox": [ + 173, + 646, + 825, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Teacher Sequence-level knowledge distillation is applied to alleviate multimodality in the training dataset, using autoregressive models as the teachers. The same teacher model used for distillation is also used as a scoring function for fine-tuning and noisy parallel decoding. ", + "bbox": [ + 174, + 766, + 823, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To enable a fair comparison, and benefit from its high translation quality, we implemented the autoregressive teachers using the state-of-the-art Transformer architecture. In addition, we use the same sizes and hyperparameters for each student and its respective teacher, with the exception of the newly added positional self-attention and fertility prediction modules. ", + "bbox": [ + 174, + 815, + 825, + 872 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b27d229f92b1dc501a1101859c66ddaef0b786b8513588c7b5b795ee0a526bb9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsWMT14WMT16IWSLT16
En→DeDe→EnEn→RoRo→EnEn→DeLatency / Speedup
NAT17.3520.6226.2227.8325.2039 ms15.6×
NAT (+FT)17.6921.4727.2929.0626.5239 ms15.6×
NAT (+FT + NPD s = 10)18.6622.4129.0230.7627.4479 ms7.68×
NAT (+FT + NPD s = 100)19.1723.2029.7931.4428.16257 ms2.36×
Autoregressive (b = 1)22.7126.3931.3531.0328.89408 ms1.49×
Autoregressive (b = 4)23.4527.0231.9131.7629.70607ms1.00×
", + "bbox": [ + 173, + 101, + 828, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1: BLEU scores on official test sets (newstest2014 for WMT En-De and newstest2016 for WMT En-Ro) or the development set for IWSLT. NAT models without NPD use argmax decoding. Latency is computed as the time to decode a single sentence without minibatching, averaged over the whole test set; decoding is implemented in PyTorch on a single NVIDIA Tesla P100. ", + "bbox": [ + 173, + 234, + 825, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Preparation for knowledge distillation We first train all teacher models using maximum likelihood, then freeze their parameters. To avoid the redundancy of running fixed teacher models repeatedly on the same data, we decode the entire training set once using each teacher to create a new training dataset for its respective student. ", + "bbox": [ + 173, + 306, + 516, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Encoder initialization We find it helpful to initialize the weights in the NAT student’s encoder with the encoder weights from its teacher, as the autoregressive and non-autoregressive models share the same encoder input and architecture. ", + "bbox": [ + 173, + 416, + 516, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fertility supervision during training As described above, we supervise the fertility predictions at train time by using a fixed aligner as a fertility inference function. We use the fast align5 implementation of IBM Model 2 for this purpose, with default parameters (Dyer et al., 2013). ", + "bbox": [ + 174, + 497, + 516, + 580 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff2cfd4d2f25ef6e98d968d65066c8e02ea0f947a91d09952aa8c8ffbab3302a.jpg", + "image_caption": [ + "Figure 3: BLEU scores on IWSLT development set as a function of sample size for noisy parallel decoding. NPD matches the performance of the other two decoding strategies after two samples, and exceeds the performance of the autoregressive teacher with around 1000. " + ], + "image_footnote": [], + "bbox": [ + 535, + 300, + 795, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameters For experiments on WMT datasets, we use the hyperparameter settings of the base Transformer model described in Vaswani et al. (2017), though without label smoothing. As IWSLT is a smaller corpus, and to reduce training time, we use a set of smaller hyperparameters $( d _ { \\mathrm { m o d e l } } = 2 8 7 , d _ { \\mathrm { h i d d e n } } = 5 0 7 , n _ { \\mathrm { l a y e r } } = 5 , n _ { \\mathrm { h e a d } } = \\mathrm { \\Omega } ^ { \\ast }$ , and $t _ { \\mathrm { w a r m u p } } = 7 4 6 )$ for all experiments on that dataset. For fine-tuning we use $\\lambda = 0 . 2 5$ . ", + "bbox": [ + 174, + 593, + 825, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation metrics We evaluate using tokenized and cased BLEU scores (Papineni et al., 2002). ", + "bbox": [ + 176, + 674, + 820, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation We have open-sourced our PyTorch implementation of the $\\mathrm { N A T ^ { 6 } }$ . ", + "bbox": [ + 173, + 700, + 730, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 734, + 277, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Across the three datasets we used, the NAT performs between 2-5 BLEU points worse than its autoregressive teacher, with part or all of this gap addressed by the use of noisy parallel decoding. In the case of WMT16 English–Romanian, NPD improves the performance of our non-autoregressive model to within 0.2 BLEU points of the previous overall state of the art (Gehring et al., 2017). ", + "bbox": [ + 176, + 762, + 825, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing latencies on the development model shows a speedup of more than a factor of 10 over greedy autoregressive decoding, or a factor of 15 over beam search. Latencies for decoding with NPD, regardless of sample size, could be reduced to about $8 0 \\mathrm { m s }$ by parallelizing across multiple GPUs because each sample can be generated, then scored, independently from the others. ", + "bbox": [ + 174, + 825, + 825, + 881 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3 ABLATION STUDY ", + "text_level": 1, + "bbox": [ + 174, + 103, + 341, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We also conduct an extensive ablation study with the proposed NAT on the IWSLT dataset. First, we note that the model fails to train when provided with only positional embeddings as input to the decoder. Second, we see that training on the distillation corpus rather than the ground truth provides a fairly consistent improvement of around 5 BLEU points. Third, switching from uniform copying of source inputs to fertility-based copying improves performance by four BLEU points when using ground-truth training or two when using distillation. ", + "bbox": [ + 174, + 130, + 825, + 214 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/e54d8b7021cfac76f841f708bf7c607fe1294c5aa3739137af72c28cad0aff7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Distillation b=1 b=4Decoder Inputs +uniform+fertility+PosAttFine-tuning +LKD +LBPBLEU +CRLBLEU (T)
~2
√16.51
√18.87
√√√20.72
√√21.12
<√24.0243.91
√25.2045.41
√√22.44 厂
兴 √√××
√√××
√√√√25.7646.11
√√√√26.5247.38
", + "bbox": [ + 199, + 228, + 797, + 433 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Table 2: Ablation performance on the IWSLT development set. BLEU (T) refers to the BLEU score on a version of the development set that has been translated by the teacher model. An $\\times$ indicates that fine-tuning caused that model to get worse. When uniform copying is used as the decoder inputs, the ground-truth target lengths are provided. All models use argmax decoding. ", + "bbox": [ + 173, + 443, + 825, + 500 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fine-tuning does not converge with reinforcement learning alone, or with the $\\mathcal { L } _ { \\mathrm { B P } }$ term alone, but use of all three fine-tuning terms together leads to an improvement of around 1.5 BLEU points. Training the student model from a distillation corpus produced using beam search is similar to training from the greedily-distilled corpus. ", + "bbox": [ + 173, + 513, + 825, + 570 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/f8421f544ec54c247fda2ccf5acde08a1d896cc7e4be518f611e266b0beaaf82.jpg", + "table_caption": [ + "Figure 4: Two examples comparing translations produced by an autoregressive (AR) and nonautoregressive Transformer as well as the result of noisy parallel decoding with sample size 100. Repeated words are highlighted in gray. " + ], + "table_footnote": [], + "table_body": "
Source:politicians try to pick wordsand use words to shape realityand control reality,butin fact,reality changes words
far more than words can ever change reality.
Target: AR:Politiker versuchen Worte zu benutzen,um die Realitat zu formen und die Realitat zu kontrolieren,aber
tatsächlich verandert die Realitat Worte viel mehr,als Worte die Realitat jemals verändern konnten.
Politikerversuchen Worter zu wahlen und Worter zur Realitat zu gestalten und Realitätzu steuern,aber in
NAT:Wirklichkeit verändert sich die Realitat viel mehrals Worte,die die Realität verändern konnen.
Politikerversuchen,Worter wahlen und zu verwenden,um Realitatzu formen und Realitatzu formen,aber
NAT+NPD:tatsächlich ändert Realitat Realitat viel mehrals Wortedie Realität Realitat verändern.
Politikerversuchen,Worter wahlenund zu verwenden,um Realitat Realitat formenund die Realitatzu formen， aber tatsächlich ändert die Realität Worte viel mehrals Worte jemals die Realität verändern konnen.
Source:Isee wheelchairs bought and sold like used cars.
Target:ich erlebe,dass Rollstuhle gekauft und verkauft werden wie Gebrauchtwagen
AR: NAT:ich sehe Rollstuhlen,die wie Autos verkauft und verkauft werden.
NAT+NPD:ich sehe,dass Stuhle Stuhle und verkauftwie Autos verkauft.
ich sehe Rollühle kauften und verkaufte wie Autos.
", + "bbox": [ + 184, + 597, + 805, + 801 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We include two examples of translations from the IWSLT development set in Fig. 4. Instances of repeated words or phrases, highlighted in gray, are most prevalent in the non-autoregressive output for the relatively complex first example sentence. Two pairs of repeated words in the first example, as ", + "bbox": [ + 176, + 882, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "se lucreaza la soluti de genul acesta . ", + "bbox": [ + 210, + 107, + 428, + 119 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "se la solutii de genul acesta . \nse lucreaza la solutii de acesta . \nse lucreaza solutii de genul acesta . \nse se lucreaza la solutii de acesta . \nse lucreaza lucreaza la solutii de acesta . \nse se lucreaza lucreaza la solutii de acesta . \nse se lucreaza lucreaza la solutii de de acesta . \nse se lucreaza lucreaza la solutii de genul acesta . solutions on this kind are done. \nwork done on solutions like this . \nsolutions on this kind is done . \nwork is done on solutions like this . \nwork is done on solutions like this . \nwork is being done on solutions like this . \nwork is being done on solutions such as this. \nwork is being done on solutions such this kind . ", + "bbox": [ + 207, + 132, + 493, + 255 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 521, + 132, + 790, + 256 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Figure 5: A Romanian–English example translated with noisy parallel decoding. At left are eight sampled fertility sequences from the encoder, represented with their corresponding decoder input sequences. Each of these values for the latent variable leads to a different possible output translation, shown at right. The autoregressive Transformer then picks the best translation, shown in red, a process which is much faster than directly using it to generate output. ", + "bbox": [ + 173, + 276, + 826, + 345 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "well as a pair in the second, are not present in the versions with noisy parallel decoding, suggesting that NPD scoring using the teacher model can filter out such mistakes. The translations produced by the NAT with NPD, while of a similar quality to those produced by the autoregressive model, are also noticeably more literal. ", + "bbox": [ + 174, + 372, + 825, + 428 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We also show an example of the noisy parallel decoding process in Fig. 5, demonstrating the diversity of translations that can be found by sampling from the fertility space. ", + "bbox": [ + 174, + 435, + 823, + 463 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 484, + 318, + 501 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We introduce a latent variable model for non-autoregressive machine translation that enables a decoder based on Vaswani et al. (2017) to take full advantage of its exceptional degree of internal parallelism even at inference time. As a result, we measure translation latencies of one-tenth that of an equal-sized autoregressive model, while maintaining competitive BLEU scores. ", + "bbox": [ + 174, + 517, + 825, + 573 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 595, + 285, + 611 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. In ICLR, 2015. ", + "bbox": [ + 176, + 619, + 823, + 647 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Peter Brown, Vincent della Pietra, Stephen della Pietra, and Robert Mercer. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263–311, 1993. ", + "bbox": [ + 173, + 657, + 826, + 700 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Kyunghyun Cho. Noisy parallel approximate decoding for conditional recurrent language model. arXiv preprint arXiv:1605.03835, 2016. ", + "bbox": [ + 174, + 712, + 821, + 741 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chris Dyer, Victor Chahuneau, and Noah Smith. A simple, fast, and effective reparameterization of IBM Model 2. In NAACL, 2013. ", + "bbox": [ + 176, + 751, + 823, + 780 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann Dauphin. Convolutional sequence to sequence learning. arXiv preprint arXiv:1705.03122, 2017. ", + "bbox": [ + 173, + 790, + 823, + 820 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Łukasz Kaiser, Aidan Gomez, and Franc¸ois Chollet. Depthwise separable convolutions for neural machine translation. arXiv preprint arXiv:1706.03059, 2017. ", + "bbox": [ + 171, + 830, + 823, + 859 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray Kavukc¸uoglu. Neural machine translation in linear time. ˇ arXiv preprint arXiv:1610.10099, 2016. ", + "bbox": [ + 174, + 869, + 826, + 898 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yoon Kim and Alexander Rush. Sequence-level knowledge distillation. In EMNLP, 2016. ", + "bbox": [ + 171, + 909, + 764, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Philipp Koehn and Rebecca Knowles. Six challenges for neural machine translation. arXiv preprint arXiv:1706.03872, 2017. ", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attentionbased neural machine translation. In EMNLP, 2015. ", + "bbox": [ + 173, + 140, + 823, + 170 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Randi Martin, Jason Crowther, Meredith Knight, Franklin Tamborello, and Chin-Lung Yang. Planning in sentence production: Evidence for the phrase as a default planning scope. Cognition, 116 (2):177–192, 2010. ", + "bbox": [ + 174, + 178, + 823, + 222 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. In ACL, pp. 311–318, 2002. ", + "bbox": [ + 171, + 229, + 823, + 260 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909, 2015. ", + "bbox": [ + 171, + 268, + 823, + 297 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ilya Sutskever, Oriol Vinyals, and Quoc L ˆ e. Sequence to sequence learning with neural networks. ˆ In NIPS, 2014. ", + "bbox": [ + 171, + 305, + 823, + 334 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. ", + "bbox": [ + 173, + 343, + 821, + 373 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Ronald Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine Learning, 8(3-4):229–256, 1992. ", + "bbox": [ + 169, + 381, + 823, + 410 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey, J. Klingner, A. Shah, M. Johnson, X. Liu, Ł. Kaiser, S. Gouws, Y. Kato, T. Kudo, H. Kazawa, K. Stevens, G. Kurian, N. Patil, W. Wang, C. Young, J. Smith, J. Riesa, A. Rudnick, O. Vinyals, G. Corrado, M. Hughes, and J. Dean. Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144, 2016. ", + "bbox": [ + 176, + 419, + 826, + 502 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/8397d58044f6d9c8a19dcbb03ac376664021ffaa63d7eac01c5e0d5c288a1df7.jpg", + "image_caption": [ + "Figure 6: The schematic structure of training and inference for the NAT. The “distilled data” contains target sentences decoded by the autoregressive model and ground-truth source sentences. " + ], + "image_footnote": [], + "bbox": [ + 179, + 135, + 815, + 446 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/65e35fd7ac2e233ef8cdae03751ebce2756e57bc1378b23f1a59c83c62823be8.jpg", + "image_caption": [ + "Figure 7: The translation latency, computed as the time to decode a single sentence without minibatching, for each sentence in the IWSLT development set as a function of its length. The autoregressive model has latency linear in the decoding length, while the latency of the NAT is nearly constant for typical lengths, even with NPD with sample size 10. When using NPD with sample size 100, the level of parallelism is enough to more than saturate the GPU, leading again to linear latencies. " + ], + "image_footnote": [], + "bbox": [ + 230, + 541, + 733, + 790 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/19326d4ac73a662073b026f5ada8aa0be903a2801d0ffdb32756ff0c9f97cdcc.jpg", + "image_caption": [ + "Figure 8: Learning curves for training and fine-tuning of the NAT on IWSLT. BLEU scores are on the development set. " + ], + "image_footnote": [], + "bbox": [ + 246, + 381, + 720, + 627 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/parse/train/B1l8BtlCb/B1l8BtlCb_middle.json b/parse/train/B1l8BtlCb/B1l8BtlCb_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..3948899bcfad7fa837330973d00915aaf92376ef --- /dev/null +++ b/parse/train/B1l8BtlCb/B1l8BtlCb_middle.json @@ -0,0 +1,28562 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 359, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 286, + 96 + ], + "score": 1.0, + "content": "NON-AUTOREGRESSIVE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 360, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 360, + 117 + ], + "score": 1.0, + "content": "NEURAL MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 464, + 147 + ], + "lines": [ + { + "bbox": [ + 110, + 134, + 466, + 149 + ], + "spans": [ + { + "bbox": [ + 110, + 134, + 140, + 149 + ], + "score": 1.0, + "content": "Jiatao", + "type": "text" + }, + { + "bbox": [ + 141, + 135, + 159, + 146 + ], + "score": 0.82, + "content": "\\mathbf { G u } ^ { \\dagger }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 134, + 466, + 149 + ], + "score": 1.0, + "content": "∗, James Bradbury‡, Caiming Xiong‡, Victor O.K. Li†& Richard Socher‡", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 113, + 148, + 388, + 192 + ], + "lines": [ + { + "bbox": [ + 111, + 145, + 200, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 200, + 158 + ], + "score": 1.0, + "content": "‡Salesforce Research", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 158, + 388, + 170 + ], + "spans": [ + { + "bbox": [ + 113, + 158, + 388, + 170 + ], + "score": 1.0, + "content": "{james.bradbury,cxiong,rsocher}@salesforce.com", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 168, + 240, + 182 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 240, + 182 + ], + "score": 1.0, + "content": "†The University of Hong Kong", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 181, + 268, + 192 + ], + "spans": [ + { + "bbox": [ + 112, + 181, + 268, + 192 + ], + "score": 1.0, + "content": "{jiataogu, vli}@eee.hku.hk", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "title", + "bbox": [ + 278, + 221, + 333, + 233 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 336, + 235 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 336, + 235 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 246, + 468, + 378 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "score": 1.0, + "content": "Existing approaches to neural machine translation condition each output word on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "score": 1.0, + "content": "previously generated outputs. We introduce a model that avoids this autoregressive", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 140, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "property and produces its outputs in parallel, allowing an order of magnitude lower", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "score": 1.0, + "content": "latency during inference. Through knowledge distillation, the use of input token", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "score": 1.0, + "content": "fertilities as a latent variable, and policy gradient fine-tuning, we achieve this at", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "score": 1.0, + "content": "a cost of as little as 2.0 BLEU points relative to the autoregressive Transformer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 313, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 324 + ], + "score": 1.0, + "content": "network used as a teacher. We demonstrate substantial cumulative improvements", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 470, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 337 + ], + "score": 1.0, + "content": "associated with each of the three aspects of our training strategy, and validate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "score": 1.0, + "content": "our approach on IWSLT 2016 English–German and two WMT language pairs.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 346, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 346, + 469, + 357 + ], + "score": 1.0, + "content": "By sampling fertilities in parallel at inference time, our non-autoregressive model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "score": 1.0, + "content": "achieves near-state-of-the-art performance of 29.8 BLEU on WMT 2016 English–", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 367, + 188, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 188, + 379 + ], + "score": 1.0, + "content": "Romanian.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 401, + 206, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 208, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 208, + 417 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 504, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 504, + 439 + ], + "score": 1.0, + "content": "Neural network based models outperform traditional statistical models for machine translation (MT)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 439, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 450 + ], + "score": 1.0, + "content": "(Bahdanau et al., 2015; Luong et al., 2015). However, state-of-the-art neural models are much", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "slower than statistical MT approaches at inference time (Wu et al., 2016). Both model families use", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "score": 1.0, + "content": "autoregressive decoders that operate one step at a time: they generate each token conditioned on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 471, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 483 + ], + "score": 1.0, + "content": "the sequence of tokens previously generated. This process is not parallelizable, and, in the case of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "neural MT models, it is particularly slow because a computationally intensive neural network is used", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 494, + 200, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 200, + 504 + ], + "score": 1.0, + "content": "to generate each token.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 504, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "While several recently proposed models avoid recurrence at train time by leveraging convolu-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "tions (Kalchbrenner et al., 2016; Gehring et al., 2017; Kaiser et al., 2017) or self-attention (Vaswani", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "score": 1.0, + "content": "et al., 2017) as more-parallelizable alternatives to recurrent neural networks (RNNs), use of autore-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 542, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 479, + 556 + ], + "score": 1.0, + "content": "gressive decoding makes it impossible to take full advantage of parallelism during inference.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "score": 1.0, + "content": "We introduce a non-autoregressive translation model based on the Transformer network (Vaswani", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "et al., 2017). We modify the encoder of the original Transformer network by adding a module that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "predicts fertilities, sequences of numbers that form an important component of many traditional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 591, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 606 + ], + "score": 1.0, + "content": "machine translation models (Brown et al., 1993). These fertilities are supervised during training", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "and provide the decoder at inference time with a globally consistent plan on which to condition its", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 614, + 245, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 245, + 627 + ], + "score": 1.0, + "content": "simultaneously computed outputs.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 642, + 200, + 656 + ], + "lines": [ + { + "bbox": [ + 104, + 641, + 202, + 659 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 202, + 659 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 668, + 357, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 359, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 359, + 681 + ], + "score": 1.0, + "content": "2.1 AUTOREGRESSIVE NEURAL MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 204, + 703 + ], + "score": 1.0, + "content": "Given a source sentence", + "type": "text" + }, + { + "bbox": [ + 205, + 690, + 280, + 702 + ], + "score": 0.92, + "content": "X = \\{ x _ { 1 } , . . . , x _ { T ^ { \\prime } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 689, + 506, + 703 + ], + "score": 1.0, + "content": ", a neural machine translation model factors the distribu-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 248, + 714 + ], + "score": 1.0, + "content": "tion over possible output sentences", + "type": "text" + }, + { + "bbox": [ + 248, + 701, + 318, + 713 + ], + "score": 0.93, + "content": "Y = \\{ y _ { 1 } , . . . , y _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "into a chain of conditional probabilities with a", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 721, + 430, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 719, + 430, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 430, + 734 + ], + "score": 1.0, + "content": "∗This work was completed while the first author was interning at Salesforce Research.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 359, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 286, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 286, + 96 + ], + "score": 1.0, + "content": "NON-AUTOREGRESSIVE", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 360, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 360, + 117 + ], + "score": 1.0, + "content": "NEURAL MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 464, + 147 + ], + "lines": [ + { + "bbox": [ + 110, + 134, + 466, + 149 + ], + "spans": [ + { + "bbox": [ + 110, + 134, + 140, + 149 + ], + "score": 1.0, + "content": "Jiatao", + "type": "text" + }, + { + "bbox": [ + 141, + 135, + 159, + 146 + ], + "score": 0.82, + "content": "\\mathbf { G u } ^ { \\dagger }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 134, + 466, + 149 + ], + "score": 1.0, + "content": "∗, James Bradbury‡, Caiming Xiong‡, Victor O.K. Li†& Richard Socher‡", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 110, + 134, + 466, + 149 + ] + }, + { + "type": "list", + "bbox": [ + 113, + 148, + 388, + 192 + ], + "lines": [ + { + "bbox": [ + 111, + 145, + 200, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 145, + 200, + 158 + ], + "score": 1.0, + "content": "‡Salesforce Research", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 158, + 388, + 170 + ], + "spans": [ + { + "bbox": [ + 113, + 158, + 388, + 170 + ], + "score": 1.0, + "content": "{james.bradbury,cxiong,rsocher}@salesforce.com", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 168, + 240, + 182 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 240, + 182 + ], + "score": 1.0, + "content": "†The University of Hong Kong", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 181, + 268, + 192 + ], + "spans": [ + { + "bbox": [ + 112, + 181, + 268, + 192 + ], + "score": 1.0, + "content": "{jiataogu, vli}@eee.hku.hk", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + } + ], + "index": 4.5, + "bbox_fs": [ + 111, + 145, + 388, + 192 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 221, + 333, + 233 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 336, + 235 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 336, + 235 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 246, + 468, + 378 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 260 + ], + "score": 1.0, + "content": "Existing approaches to neural machine translation condition each output word on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 470, + 271 + ], + "score": 1.0, + "content": "previously generated outputs. We introduce a model that avoids this autoregressive", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 140, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 140, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "property and produces its outputs in parallel, allowing an order of magnitude lower", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "score": 1.0, + "content": "latency during inference. Through knowledge distillation, the use of input token", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 470, + 304 + ], + "score": 1.0, + "content": "fertilities as a latent variable, and policy gradient fine-tuning, we achieve this at", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "score": 1.0, + "content": "a cost of as little as 2.0 BLEU points relative to the autoregressive Transformer", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 313, + 469, + 324 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 324 + ], + "score": 1.0, + "content": "network used as a teacher. We demonstrate substantial cumulative improvements", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 470, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 337 + ], + "score": 1.0, + "content": "associated with each of the three aspects of our training strategy, and validate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 470, + 347 + ], + "score": 1.0, + "content": "our approach on IWSLT 2016 English–German and two WMT language pairs.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 346, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 346, + 469, + 357 + ], + "score": 1.0, + "content": "By sampling fertilities in parallel at inference time, our non-autoregressive model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "score": 1.0, + "content": "achieves near-state-of-the-art performance of 29.8 BLEU on WMT 2016 English–", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 367, + 188, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 188, + 379 + ], + "score": 1.0, + "content": "Romanian.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5, + "bbox_fs": [ + 140, + 246, + 470, + 379 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 401, + 206, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 208, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 208, + 417 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 427, + 505, + 504 + ], + "lines": [ + { + "bbox": [ + 106, + 427, + 504, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 504, + 439 + ], + "score": 1.0, + "content": "Neural network based models outperform traditional statistical models for machine translation (MT)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 439, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 505, + 450 + ], + "score": 1.0, + "content": "(Bahdanau et al., 2015; Luong et al., 2015). However, state-of-the-art neural models are much", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "slower than statistical MT approaches at inference time (Wu et al., 2016). Both model families use", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 472 + ], + "score": 1.0, + "content": "autoregressive decoders that operate one step at a time: they generate each token conditioned on", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 471, + 506, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 506, + 483 + ], + "score": 1.0, + "content": "the sequence of tokens previously generated. This process is not parallelizable, and, in the case of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "neural MT models, it is particularly slow because a computationally intensive neural network is used", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 494, + 200, + 504 + ], + "spans": [ + { + "bbox": [ + 106, + 494, + 200, + 504 + ], + "score": 1.0, + "content": "to generate each token.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 427, + 506, + 504 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 510, + 504, + 554 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 505, + 522 + ], + "score": 1.0, + "content": "While several recently proposed models avoid recurrence at train time by leveraging convolu-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 505, + 532 + ], + "score": 1.0, + "content": "tions (Kalchbrenner et al., 2016; Gehring et al., 2017; Kaiser et al., 2017) or self-attention (Vaswani", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 544 + ], + "score": 1.0, + "content": "et al., 2017) as more-parallelizable alternatives to recurrent neural networks (RNNs), use of autore-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 542, + 479, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 479, + 556 + ], + "score": 1.0, + "content": "gressive decoding makes it impossible to take full advantage of parallelism during inference.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 509, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 505, + 571 + ], + "score": 1.0, + "content": "We introduce a non-autoregressive translation model based on the Transformer network (Vaswani", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 505, + 583 + ], + "score": 1.0, + "content": "et al., 2017). We modify the encoder of the original Transformer network by adding a module that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "predicts fertilities, sequences of numbers that form an important component of many traditional", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 591, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 606 + ], + "score": 1.0, + "content": "machine translation models (Brown et al., 1993). These fertilities are supervised during training", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 505, + 615 + ], + "score": 1.0, + "content": "and provide the decoder at inference time with a globally consistent plan on which to condition its", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 614, + 245, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 245, + 627 + ], + "score": 1.0, + "content": "simultaneously computed outputs.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 560, + 505, + 627 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 642, + 200, + 656 + ], + "lines": [ + { + "bbox": [ + 104, + 641, + 202, + 659 + ], + "spans": [ + { + "bbox": [ + 104, + 641, + 202, + 659 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "title", + "bbox": [ + 107, + 668, + 357, + 680 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 359, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 359, + 681 + ], + "score": 1.0, + "content": "2.1 AUTOREGRESSIVE NEURAL MACHINE TRANSLATION", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 108, + 689, + 504, + 713 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 204, + 703 + ], + "score": 1.0, + "content": "Given a source sentence", + "type": "text" + }, + { + "bbox": [ + 205, + 690, + 280, + 702 + ], + "score": 0.92, + "content": "X = \\{ x _ { 1 } , . . . , x _ { T ^ { \\prime } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 689, + 506, + 703 + ], + "score": 1.0, + "content": ", a neural machine translation model factors the distribu-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 248, + 714 + ], + "score": 1.0, + "content": "tion over possible output sentences", + "type": "text" + }, + { + "bbox": [ + 248, + 701, + 318, + 713 + ], + "score": 0.93, + "content": "Y = \\{ y _ { 1 } , . . . , y _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "into a chain of conditional probabilities with a", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 689, + 506, + 714 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 222, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 223, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 223, + 95 + ], + "score": 1.0, + "content": "left-to-right causal structure:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 97, + 393, + 132 + ], + "lines": [ + { + "bbox": [ + 217, + 97, + 393, + 132 + ], + "spans": [ + { + "bbox": [ + 217, + 97, + 393, + 132 + ], + "score": 0.95, + "content": "p _ { \\mathcal { A } \\mathcal { R } } ( \\boldsymbol { Y } | \\boldsymbol { X } ; \\theta ) = \\prod _ { t = 1 } ^ { T + 1 } p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) ,", + "type": "interline_equation", + "image_path": "633a748a4002cb496ac4d94d9bde49381f4173fa71b235023f3e7a76c5656215.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 97, + 393, + 114.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 217, + 114.5, + 393, + 132.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 135, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 208, + 150 + ], + "score": 1.0, + "content": "where the special tokens", + "type": "text" + }, + { + "bbox": [ + 208, + 137, + 219, + 147 + ], + "score": 0.84, + "content": "y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 135, + 243, + 150 + ], + "score": 1.0, + "content": "(e.g.", + "type": "text" + }, + { + "bbox": [ + 243, + 136, + 266, + 148 + ], + "score": 0.89, + "content": "\\left. \\mathrm { b o s } \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 135, + 288, + 150 + ], + "score": 1.0, + "content": ") and", + "type": "text" + }, + { + "bbox": [ + 288, + 137, + 311, + 147 + ], + "score": 0.89, + "content": "y _ { T + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 135, + 334, + 150 + ], + "score": 1.0, + "content": "(e.g.", + "type": "text" + }, + { + "bbox": [ + 335, + 136, + 357, + 148 + ], + "score": 0.89, + "content": "\\langle \\cos \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 135, + 506, + 150 + ], + "score": 1.0, + "content": ") are used to represent the beginning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "and end of all target sentences. These conditional probabilities are parameterized using a neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 158, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 170 + ], + "score": 1.0, + "content": "network. Typically, an encoder-decoder architecture (Sutskever et al., 2014) with a unidirectional", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 446, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 446, + 181 + ], + "score": 1.0, + "content": "RNN-based decoder is used to capture the causal structure of the output distribution.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 504, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "Maximum Likelihood training Choosing to factorize the machine translation output distribution", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "score": 1.0, + "content": "autoregressively enables straightforward maximum likelihood training with a cross-entropy loss ap-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 207, + 220, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 220, + 223 + ], + "score": 1.0, + "content": "plied at each decoding step:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 224, + 423, + 258 + ], + "lines": [ + { + "bbox": [ + 187, + 224, + 423, + 258 + ], + "spans": [ + { + "bbox": [ + 187, + 224, + 423, + 258 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { A R } } ( Y | X ; \\theta ) = \\sum _ { t = 1 } ^ { T + 1 } \\log p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) .", + "type": "interline_equation", + "image_path": "975cc03de217f9263a661785207cd37c0ad59d69289d6f9743d2345d95881151.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 224, + 423, + 241.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 241.0, + 423, + 258.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 261, + 428, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 429, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 429, + 276 + ], + "score": 1.0, + "content": "This loss provides direct supervision for each conditional probability prediction.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "Autoregressive NMT without RNNs Since the entire target translation is known at training time,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "the calculation of later conditional probabilities (and their corresponding losses) does not depend on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 303, + 504, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 504, + 314 + ], + "score": 1.0, + "content": "the output words chosen during earlier decoding steps. Even though decoding must remain entirely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "sequential during inference, models can take advantage of this parallelism during training. One such", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 323, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 337 + ], + "score": 1.0, + "content": "approach replaces recurrent layers in the decoder with masked convolution layers (Kalchbrenner", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 333, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 348 + ], + "score": 1.0, + "content": "et al., 2016; Gehring et al., 2017) that provide the causal structure required by the autoregressive", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 346, + 161, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 161, + 357 + ], + "score": 1.0, + "content": "factorization.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "A recently introduced option which reduces sequential computation still further is to construct the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "score": 1.0, + "content": "decoder layers out of self-attention computations that have been causally masked in an analogous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "way. The state-of-the-art Transformer network takes this approach, which allows information to flow", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "in the decoder across arbitrarily long distances in a constant number of operations, asymptotically", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 406, + 401, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 401, + 419 + ], + "score": 1.0, + "content": "fewer than required by convolutional architectures (Vaswani et al., 2017).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 431, + 283, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 284, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 284, + 443 + ], + "score": 1.0, + "content": "2.2 NON-AUTOREGRESSIVE DECODING", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 504, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "Pros and cons of autoregressive decoding The autoregressive factorization used by conventional", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "NMT models has several benefits. It corresponds to the word-by-word nature of human language", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "production and effectively captures the distribution of real translations. Autoregressive models", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "achieve state-of-the-art performance on large-scale corpora and are easy to train, while beam search", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 491, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 491, + 503 + ], + "score": 1.0, + "content": "provides an effective local search method for finding approximately-optimal output translations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 507, + 356, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 356, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 356, + 520 + ], + "score": 1.0, + "content": "But there are also drawbacks. As the individual steps of the de-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 518, + 356, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 356, + 531 + ], + "score": 1.0, + "content": "coder must be run sequentially rather than in parallel, autore-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 529, + 357, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 357, + 541 + ], + "score": 1.0, + "content": "gressive decoding prevents architectures like the Transformer", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 540, + 357, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 357, + 553 + ], + "score": 1.0, + "content": "from fully realizing their train-time performance advantage", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 552, + 356, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 356, + 562 + ], + "score": 1.0, + "content": "during inference. Meanwhile, beam search suffers from dimin-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 562, + 356, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 356, + 574 + ], + "score": 1.0, + "content": "ishing returns with respect to beam size (Koehn & Knowles,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 573, + 356, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 356, + 585 + ], + "score": 1.0, + "content": "2017) and exhibits limited search parallelism because it intro-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 584, + 307, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 307, + 595 + ], + "score": 1.0, + "content": "duces computational dependence between beams.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 356, + 658 + ], + "lines": [ + { + "bbox": [ + 107, + 602, + 357, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 602, + 357, + 614 + ], + "score": 1.0, + "content": "Towards non-autoregressive decoding A na¨ıve solution is", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 613, + 356, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 356, + 626 + ], + "score": 1.0, + "content": "to remove the autoregressive connection directly from an exist-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 623, + 357, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 357, + 637 + ], + "score": 1.0, + "content": "ing encoder-decoder model. Assuming that the target sequence", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 635, + 356, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 134, + 648 + ], + "score": 1.0, + "content": "length", + "type": "text" + }, + { + "bbox": [ + 134, + 636, + 143, + 645 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 635, + 356, + 648 + ], + "score": 1.0, + "content": "can be modeled with a separate conditional distribu-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 646, + 195, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 124, + 658 + ], + "score": 1.0, + "content": "tion", + "type": "text" + }, + { + "bbox": [ + 124, + 648, + 136, + 658 + ], + "score": 0.87, + "content": "p _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 646, + 195, + 658 + ], + "score": 1.0, + "content": ", this becomes", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 366, + 510, + 505, + 623 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 366, + 510, + 505, + 623 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 366, + 510, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 366, + 510, + 505, + 623 + ], + "score": 0.964, + "type": "image", + "image_path": "fb73e32da81d3465f857ae37043ab008349971a74c732394fc409a62ee78d60d.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 366, + 510, + 505, + 566.5 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 366, + 566.5, + 505, + 623.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 363, + 636, + 505, + 692 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 362, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 362, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "Figure 1: Translating “A B C”", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 362, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 362, + 646, + 374, + 659 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 375, + 647, + 388, + 658 + ], + "score": 0.26, + "content": "^ { 6 6 } \\mathrm { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "Y” using autoregressive and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 362, + 658, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 362, + 658, + 505, + 669 + ], + "score": 1.0, + "content": "non-autoregressive neural MT ar-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 363, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 363, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "chitectures. The latter generates all", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 363, + 680, + 465, + 692 + ], + "spans": [ + { + "bbox": [ + 363, + 680, + 465, + 692 + ], + "score": 1.0, + "content": "output tokens in parallel.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50 + } + ], + "index": 44.75 + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 661, + 330, + 695 + ], + "lines": [ + { + "bbox": [ + 119, + 661, + 330, + 695 + ], + "spans": [ + { + "bbox": [ + 119, + 661, + 330, + 695 + ], + "score": 0.93, + "content": "p _ { \\mathcal N A } ( Y | X ; \\theta ) = p _ { L } ( T | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 : T ^ { \\prime } } ; \\theta ) .", + "type": "interline_equation", + "image_path": "d97e3003546373ed25378b0bc946fe1615293facf85f3967deae841210d37fb1.jpg" + } + ] + } + ], + "index": 47.5, + "virtual_lines": [ + { + "bbox": [ + 119, + 661, + 330, + 678.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 119, + 678.0, + 330, + 695.0 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "This model still has an explicit likelihood function, and it can still be trained using independent", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "cross-entropy losses on each output distribution. Now, however, these distributions can be computed", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 721, + 221, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 221, + 733 + ], + "score": 1.0, + "content": "in parallel at inference time.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 222, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 223, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 223, + 95 + ], + "score": 1.0, + "content": "left-to-right causal structure:", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 105, + 81, + 223, + 95 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 217, + 97, + 393, + 132 + ], + "lines": [ + { + "bbox": [ + 217, + 97, + 393, + 132 + ], + "spans": [ + { + "bbox": [ + 217, + 97, + 393, + 132 + ], + "score": 0.95, + "content": "p _ { \\mathcal { A } \\mathcal { R } } ( \\boldsymbol { Y } | \\boldsymbol { X } ; \\theta ) = \\prod _ { t = 1 } ^ { T + 1 } p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) ,", + "type": "interline_equation", + "image_path": "633a748a4002cb496ac4d94d9bde49381f4173fa71b235023f3e7a76c5656215.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 217, + 97, + 393, + 114.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 217, + 114.5, + 393, + 132.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 135, + 505, + 180 + ], + "lines": [ + { + "bbox": [ + 105, + 135, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 135, + 208, + 150 + ], + "score": 1.0, + "content": "where the special tokens", + "type": "text" + }, + { + "bbox": [ + 208, + 137, + 219, + 147 + ], + "score": 0.84, + "content": "y _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 135, + 243, + 150 + ], + "score": 1.0, + "content": "(e.g.", + "type": "text" + }, + { + "bbox": [ + 243, + 136, + 266, + 148 + ], + "score": 0.89, + "content": "\\left. \\mathrm { b o s } \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 135, + 288, + 150 + ], + "score": 1.0, + "content": ") and", + "type": "text" + }, + { + "bbox": [ + 288, + 137, + 311, + 147 + ], + "score": 0.89, + "content": "y _ { T + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 135, + 334, + 150 + ], + "score": 1.0, + "content": "(e.g.", + "type": "text" + }, + { + "bbox": [ + 335, + 136, + 357, + 148 + ], + "score": 0.89, + "content": "\\langle \\cos \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 135, + 506, + 150 + ], + "score": 1.0, + "content": ") are used to represent the beginning", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "spans": [ + { + "bbox": [ + 105, + 146, + 505, + 159 + ], + "score": 1.0, + "content": "and end of all target sentences. These conditional probabilities are parameterized using a neural", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 158, + 505, + 170 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 170 + ], + "score": 1.0, + "content": "network. Typically, an encoder-decoder architecture (Sutskever et al., 2014) with a unidirectional", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 446, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 446, + 181 + ], + "score": 1.0, + "content": "RNN-based decoder is used to capture the causal structure of the output distribution.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 135, + 506, + 181 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 186, + 504, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "Maximum Likelihood training Choosing to factorize the machine translation output distribution", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 197, + 506, + 212 + ], + "score": 1.0, + "content": "autoregressively enables straightforward maximum likelihood training with a cross-entropy loss ap-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 207, + 220, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 207, + 220, + 223 + ], + "score": 1.0, + "content": "plied at each decoding step:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 187, + 506, + 223 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 224, + 423, + 258 + ], + "lines": [ + { + "bbox": [ + 187, + 224, + 423, + 258 + ], + "spans": [ + { + "bbox": [ + 187, + 224, + 423, + 258 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { A R } } ( Y | X ; \\theta ) = \\sum _ { t = 1 } ^ { T + 1 } \\log p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) .", + "type": "interline_equation", + "image_path": "975cc03de217f9263a661785207cd37c0ad59d69289d6f9743d2345d95881151.jpg" + } + ] + } + ], + "index": 10.5, + "virtual_lines": [ + { + "bbox": [ + 187, + 224, + 423, + 241.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 187, + 241.0, + 423, + 258.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 261, + 428, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 259, + 429, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 429, + 276 + ], + "score": 1.0, + "content": "This loss provides direct supervision for each conditional probability prediction.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 259, + 429, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 279, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 279, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 505, + 293 + ], + "score": 1.0, + "content": "Autoregressive NMT without RNNs Since the entire target translation is known at training time,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 505, + 303 + ], + "score": 1.0, + "content": "the calculation of later conditional probabilities (and their corresponding losses) does not depend on", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 303, + 504, + 314 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 504, + 314 + ], + "score": 1.0, + "content": "the output words chosen during earlier decoding steps. Even though decoding must remain entirely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 325 + ], + "score": 1.0, + "content": "sequential during inference, models can take advantage of this parallelism during training. One such", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 323, + 506, + 337 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 337 + ], + "score": 1.0, + "content": "approach replaces recurrent layers in the decoder with masked convolution layers (Kalchbrenner", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 333, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 348 + ], + "score": 1.0, + "content": "et al., 2016; Gehring et al., 2017) that provide the causal structure required by the autoregressive", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 346, + 161, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 161, + 357 + ], + "score": 1.0, + "content": "factorization.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 279, + 506, + 357 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 375 + ], + "score": 1.0, + "content": "A recently introduced option which reduces sequential computation still further is to construct the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 372, + 505, + 386 + ], + "score": 1.0, + "content": "decoder layers out of self-attention computations that have been causally masked in an analogous", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 397 + ], + "score": 1.0, + "content": "way. The state-of-the-art Transformer network takes this approach, which allows information to flow", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "in the decoder across arbitrarily long distances in a constant number of operations, asymptotically", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 406, + 401, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 401, + 419 + ], + "score": 1.0, + "content": "fewer than required by convolutional architectures (Vaswani et al., 2017).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 363, + 505, + 419 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 431, + 283, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 284, + 443 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 284, + 443 + ], + "score": 1.0, + "content": "2.2 NON-AUTOREGRESSIVE DECODING", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 504, + 502 + ], + "lines": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "Pros and cons of autoregressive decoding The autoregressive factorization used by conventional", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 506, + 471 + ], + "score": 1.0, + "content": "NMT models has several benefits. It corresponds to the word-by-word nature of human language", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 506, + 481 + ], + "score": 1.0, + "content": "production and effectively captures the distribution of real translations. Autoregressive models", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 480, + 505, + 492 + ], + "score": 1.0, + "content": "achieve state-of-the-art performance on large-scale corpora and are easy to train, while beam search", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 491, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 491, + 503 + ], + "score": 1.0, + "content": "provides an effective local search method for finding approximately-optimal output translations.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 446, + 506, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 507, + 356, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 506, + 356, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 356, + 520 + ], + "score": 1.0, + "content": "But there are also drawbacks. As the individual steps of the de-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 518, + 356, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 356, + 531 + ], + "score": 1.0, + "content": "coder must be run sequentially rather than in parallel, autore-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 529, + 357, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 357, + 541 + ], + "score": 1.0, + "content": "gressive decoding prevents architectures like the Transformer", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 540, + 357, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 357, + 553 + ], + "score": 1.0, + "content": "from fully realizing their train-time performance advantage", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 552, + 356, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 356, + 562 + ], + "score": 1.0, + "content": "during inference. Meanwhile, beam search suffers from dimin-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 562, + 356, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 356, + 574 + ], + "score": 1.0, + "content": "ishing returns with respect to beam size (Koehn & Knowles,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 573, + 356, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 356, + 585 + ], + "score": 1.0, + "content": "2017) and exhibits limited search parallelism because it intro-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 584, + 307, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 307, + 595 + ], + "score": 1.0, + "content": "duces computational dependence between beams.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 506, + 357, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 356, + 658 + ], + "lines": [ + { + "bbox": [ + 107, + 602, + 357, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 602, + 357, + 614 + ], + "score": 1.0, + "content": "Towards non-autoregressive decoding A na¨ıve solution is", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 613, + 356, + 626 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 356, + 626 + ], + "score": 1.0, + "content": "to remove the autoregressive connection directly from an exist-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 623, + 357, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 357, + 637 + ], + "score": 1.0, + "content": "ing encoder-decoder model. Assuming that the target sequence", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 635, + 356, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 134, + 648 + ], + "score": 1.0, + "content": "length", + "type": "text" + }, + { + "bbox": [ + 134, + 636, + 143, + 645 + ], + "score": 0.81, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 635, + 356, + 648 + ], + "score": 1.0, + "content": "can be modeled with a separate conditional distribu-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 646, + 195, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 124, + 658 + ], + "score": 1.0, + "content": "tion", + "type": "text" + }, + { + "bbox": [ + 124, + 648, + 136, + 658 + ], + "score": 0.87, + "content": "p _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 646, + 195, + 658 + ], + "score": 1.0, + "content": ", this becomes", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 602, + 357, + 658 + ] + }, + { + "type": "image", + "bbox": [ + 366, + 510, + 505, + 623 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 366, + 510, + 505, + 623 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 366, + 510, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 366, + 510, + 505, + 623 + ], + "score": 0.964, + "type": "image", + "image_path": "fb73e32da81d3465f857ae37043ab008349971a74c732394fc409a62ee78d60d.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 366, + 510, + 505, + 566.5 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 366, + 566.5, + 505, + 623.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 363, + 636, + 505, + 692 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 362, + 635, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 362, + 635, + 505, + 648 + ], + "score": 1.0, + "content": "Figure 1: Translating “A B C”", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 362, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 362, + 646, + 374, + 659 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 375, + 647, + 388, + 658 + ], + "score": 0.26, + "content": "^ { 6 6 } \\mathrm { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "Y” using autoregressive and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 362, + 658, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 362, + 658, + 505, + 669 + ], + "score": 1.0, + "content": "non-autoregressive neural MT ar-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 363, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 363, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "chitectures. The latter generates all", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 363, + 680, + 465, + 692 + ], + "spans": [ + { + "bbox": [ + 363, + 680, + 465, + 692 + ], + "score": 1.0, + "content": "output tokens in parallel.", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 50 + } + ], + "index": 44.75 + }, + { + "type": "interline_equation", + "bbox": [ + 119, + 661, + 330, + 695 + ], + "lines": [ + { + "bbox": [ + 119, + 661, + 330, + 695 + ], + "spans": [ + { + "bbox": [ + 119, + 661, + 330, + 695 + ], + "score": 0.93, + "content": "p _ { \\mathcal N A } ( Y | X ; \\theta ) = p _ { L } ( T | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 : T ^ { \\prime } } ; \\theta ) .", + "type": "interline_equation", + "image_path": "d97e3003546373ed25378b0bc946fe1615293facf85f3967deae841210d37fb1.jpg" + } + ] + } + ], + "index": 47.5, + "virtual_lines": [ + { + "bbox": [ + 119, + 661, + 330, + 678.0 + ], + "spans": [], + "index": 46 + }, + { + "bbox": [ + 119, + 678.0, + 330, + 695.0 + ], + "spans": [], + "index": 49 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "This model still has an explicit likelihood function, and it can still be trained using independent", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "cross-entropy losses on each output distribution. Now, however, these distributions can be computed", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 721, + 221, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 221, + 733 + ], + "score": 1.0, + "content": "in parallel at inference time.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 54, + "bbox_fs": [ + 105, + 699, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 79, + 488, + 290 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 79, + 488, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 79, + 488, + 290 + ], + "spans": [ + { + "bbox": [ + 116, + 79, + 488, + 290 + ], + "score": 0.977, + "type": "image", + "image_path": "899dc15129b43e72de49a617f813130a25104e3d34171cb067e8d5c905671e3e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 79, + 488, + 149.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 149.33333333333331, + 488, + 219.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 219.66666666666663, + 488, + 289.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 300, + 504, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "score": 1.0, + "content": "Figure 2: The architecture of the NAT, where the black solid arrows represent differentiable con-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "nections and the purple dashed arrows are non-differentiable operations. Each sublayer inside the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 323, + 459, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 459, + 334 + ], + "score": 1.0, + "content": "encoder and decoder stacks also includes layer normalization and a residual connection.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 344, + 272, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 273, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 273, + 357 + ], + "score": 1.0, + "content": "2.3 THE MULTIMODALITY PROBLEM", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "However, this na¨ıve approach does not yield good results, because such a model exhibits complete", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 380, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 318, + 393 + ], + "score": 1.0, + "content": "conditional independence. Each token’s distribution", + "type": "text" + }, + { + "bbox": [ + 318, + 380, + 340, + 393 + ], + "score": 0.92, + "content": "p ( y _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 381, + 491, + 393 + ], + "score": 1.0, + "content": "depends only on the source sentence", + "type": "text" + }, + { + "bbox": [ + 491, + 381, + 501, + 390 + ], + "score": 0.79, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 381, + 504, + 393 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "This makes it a poor approximation to the true target distribution, which exhibits strong correlation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "across time. Intuitively, such a decoder is akin to a panel of human translators each asked to provide", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 413, + 434, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 434, + 426 + ], + "score": 1.0, + "content": "a single word of a translation independently of the words their colleagues choose.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "In particular, consider an English source sentence like “Thank you.” This can be accurately trans-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 439, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 455 + ], + "score": 1.0, + "content": "lated into German as any one of “Danke.”, “Danke schon.”, or “Vielen Dank.”, all of which may ¨", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "score": 1.0, + "content": "occur in a given training corpus. This target distribution cannot be represented as a product of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "independent probability distributions for each of the first, second, and third words, because a con-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "ditionally independent distribution cannot allow “Danke schon.” and “Vielen Dank.” without also ¨", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 485, + 289, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 289, + 497 + ], + "score": 1.0, + "content": "licensing “Danke Dank.” and “Vielen schon.”¨", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "score": 1.0, + "content": "The conditional independence assumption prevents a model from properly capturing the highly mul-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "timodal distribution of target translations. We call this the “multimodality problem” and introduce", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 523, + 392, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 392, + 537 + ], + "score": 1.0, + "content": "both a modified model and new training techniques to tackle this issue.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 562, + 392, + 575 + ], + "lines": [ + { + "bbox": [ + 104, + 561, + 393, + 577 + ], + "spans": [ + { + "bbox": [ + 104, + 561, + 393, + 577 + ], + "score": 1.0, + "content": "3 THE NON-AUTOREGRESSIVE TRANSFORMER (NAT)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "We introduce a novel NMT model—the Non-Autoregressive Transformer (NAT)—that can produce", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "an entire output translation in parallel. As shown in Fig. 2, the model is composed of the following", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "four modules: an encoder stack, a decoder stack, a newly added fertility predictor (details in 3.3),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 626, + 293, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 293, + 640 + ], + "score": 1.0, + "content": "and a translation predictor for token decoding.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 108, + 662, + 205, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 207, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 207, + 676 + ], + "score": 1.0, + "content": "3.1 ENCODER STACK", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "score": 1.0, + "content": "Similar to the autoregressive Transformer, both the encoder and decoder stacks are composed en-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "tirely of feed-forward networks (MLPs) and multi-head attention modules. Since no RNNs are used,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "there is no inherent requirement for sequential execution, making non-autoregressive decoding pos-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 720, + 503, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 503, + 732 + ], + "score": 1.0, + "content": "sible. For our proposed NAT, the encoder stays unchanged from the original Transformer network.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 79, + 488, + 290 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 116, + 79, + 488, + 290 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 116, + 79, + 488, + 290 + ], + "spans": [ + { + "bbox": [ + 116, + 79, + 488, + 290 + ], + "score": 0.977, + "type": "image", + "image_path": "899dc15129b43e72de49a617f813130a25104e3d34171cb067e8d5c905671e3e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 116, + 79, + 488, + 149.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 116, + 149.33333333333331, + 488, + 219.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 116, + 219.66666666666663, + 488, + 289.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 300, + 504, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "spans": [ + { + "bbox": [ + 106, + 300, + 504, + 313 + ], + "score": 1.0, + "content": "Figure 2: The architecture of the NAT, where the black solid arrows represent differentiable con-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 324 + ], + "score": 1.0, + "content": "nections and the purple dashed arrows are non-differentiable operations. Each sublayer inside the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 323, + 459, + 334 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 459, + 334 + ], + "score": 1.0, + "content": "encoder and decoder stacks also includes layer normalization and a residual connection.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 344, + 272, + 356 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 273, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 273, + 357 + ], + "score": 1.0, + "content": "2.3 THE MULTIMODALITY PROBLEM", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 369, + 505, + 424 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 382 + ], + "score": 1.0, + "content": "However, this na¨ıve approach does not yield good results, because such a model exhibits complete", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 380, + 504, + 393 + ], + "spans": [ + { + "bbox": [ + 106, + 381, + 318, + 393 + ], + "score": 1.0, + "content": "conditional independence. Each token’s distribution", + "type": "text" + }, + { + "bbox": [ + 318, + 380, + 340, + 393 + ], + "score": 0.92, + "content": "p ( y _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 381, + 491, + 393 + ], + "score": 1.0, + "content": "depends only on the source sentence", + "type": "text" + }, + { + "bbox": [ + 491, + 381, + 501, + 390 + ], + "score": 0.79, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 381, + 504, + 393 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "This makes it a poor approximation to the true target distribution, which exhibits strong correlation", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "across time. Intuitively, such a decoder is akin to a panel of human translators each asked to provide", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 413, + 434, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 434, + 426 + ], + "score": 1.0, + "content": "a single word of a translation independently of the words their colleagues choose.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 369, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 430, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 443 + ], + "score": 1.0, + "content": "In particular, consider an English source sentence like “Thank you.” This can be accurately trans-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 439, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 455 + ], + "score": 1.0, + "content": "lated into German as any one of “Danke.”, “Danke schon.”, or “Vielen Dank.”, all of which may ¨", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 505, + 464 + ], + "score": 1.0, + "content": "occur in a given training corpus. This target distribution cannot be represented as a product of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "independent probability distributions for each of the first, second, and third words, because a con-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "ditionally independent distribution cannot allow “Danke schon.” and “Vielen Dank.” without also ¨", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 485, + 289, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 289, + 497 + ], + "score": 1.0, + "content": "licensing “Danke Dank.” and “Vielen schon.”¨", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 429, + 506, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 535 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 515 + ], + "score": 1.0, + "content": "The conditional independence assumption prevents a model from properly capturing the highly mul-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "timodal distribution of target translations. We call this the “multimodality problem” and introduce", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 523, + 392, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 392, + 537 + ], + "score": 1.0, + "content": "both a modified model and new training techniques to tackle this issue.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 501, + 505, + 537 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 562, + 392, + 575 + ], + "lines": [ + { + "bbox": [ + 104, + 561, + 393, + 577 + ], + "spans": [ + { + "bbox": [ + 104, + 561, + 393, + 577 + ], + "score": 1.0, + "content": "3 THE NON-AUTOREGRESSIVE TRANSFORMER (NAT)", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 638 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "We introduce a novel NMT model—the Non-Autoregressive Transformer (NAT)—that can produce", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "an entire output translation in parallel. As shown in Fig. 2, the model is composed of the following", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "four modules: an encoder stack, a decoder stack, a newly added fertility predictor (details in 3.3),", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 626, + 293, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 293, + 640 + ], + "score": 1.0, + "content": "and a translation predictor for token decoding.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 594, + 505, + 640 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 662, + 205, + 674 + ], + "lines": [ + { + "bbox": [ + 105, + 661, + 207, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 207, + 676 + ], + "score": 1.0, + "content": "3.1 ENCODER STACK", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 504, + 700 + ], + "score": 1.0, + "content": "Similar to the autoregressive Transformer, both the encoder and decoder stacks are composed en-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 710 + ], + "score": 1.0, + "content": "tirely of feed-forward networks (MLPs) and multi-head attention modules. Since no RNNs are used,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "there is no inherent requirement for sequential execution, making non-autoregressive decoding pos-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 720, + 503, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 503, + 732 + ], + "score": 1.0, + "content": "sible. For our proposed NAT, the encoder stays unchanged from the original Transformer network.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 687, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 205, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 208, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 208, + 96 + ], + "score": 1.0, + "content": "3.2 DECODER STACK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 504, + 125 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 504, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 504, + 116 + ], + "score": 1.0, + "content": "In order to translate non-autoregressively and parallelize the decoding process, we modify the de-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 200, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 200, + 127 + ], + "score": 1.0, + "content": "coder stack as follows.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "Decoder Inputs Before decoding starts, the NAT needs to know how long the target sentence", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "will be in order to generate all words in parallel. More crucially, we cannot use time-shifted target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 155, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 166 + ], + "score": 1.0, + "content": "outputs (during training) or previously predicted outputs (during inference) as the inputs to the first", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "decoder layer. Omitting inputs to the first decoder layer entirely, or using only positional embed-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "dings, resulted in very poor performance. Instead, we initialize the decoding process using copied", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "source inputs from the encoder side. As the source and target sentences are often of different lengths,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 200, + 210, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 210, + 210 + ], + "score": 1.0, + "content": "we propose two methods:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 328, + 231 + ], + "score": 1.0, + "content": "• Copy source inputs uniformly: Each decoder input", + "type": "text" + }, + { + "bbox": [ + 328, + 219, + 333, + 228 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 217, + 425, + 231 + ], + "score": 1.0, + "content": "is a copy of the Round", + "type": "text" + }, + { + "bbox": [ + 425, + 218, + 458, + 230 + ], + "score": 0.78, + "content": "( T ^ { \\prime } t / T )", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 217, + 506, + 231 + ], + "score": 1.0, + "content": "-th encoder", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "input. This is equivalent to “scanning” source inputs from left to right with a constant “speed,”", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 240, + 463, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 463, + 253 + ], + "score": 1.0, + "content": "and results in a decoding process that is deterministic given a (predicted) target length.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 108, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "• Copy source inputs using fertilities: A more powerful way, depicted in Fig. 2 and discussed", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "in more detail below, is to copy each encoder input as a decoder input zero or more times, with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "the number of times each input is copied referred to as that input word’s “fertility.” In this case", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "the source inputs are scanned from left to right at a “speed” that varies inversely with the fertility", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "of each input; the decoding process is now conditioned on the sequence of fertilities, while the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 310, + 397, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 310, + 397, + 322 + ], + "score": 1.0, + "content": "resulting output length is determined by the sum of all fertility values.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "Non-causal self-attention Without the constraint of an autoregressive factorization of the output", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "distribution, we no longer need to prevent earlier decoding steps from accessing information from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "later steps. Thus we can avoid the causal mask used in the self-attention module of the conventional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Transformer’s decoder. Instead, we mask out each query position only from attending to itself,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 374, + 448, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 448, + 386 + ], + "score": 1.0, + "content": "which we found to improve decoder performance relative to unmasked self-attention.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "Positional attention We also include an additional positional attention module in each decoder", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "score": 1.0, + "content": "layer, which is a multi-head attention module with the same general attention mechanism used in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 414, + 281, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 281, + 426 + ], + "score": 1.0, + "content": "other parts of the Transformer network, i.e.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 429, + 402, + 457 + ], + "lines": [ + { + "bbox": [ + 208, + 429, + 402, + 457 + ], + "spans": [ + { + "bbox": [ + 208, + 429, + 402, + 457 + ], + "score": 0.93, + "content": "{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { \\mathrm { m o d e l } } } } } \\right) \\cdot V ,", + "type": "interline_equation", + "image_path": "7591f42425b5f5d9a239ca6fc7711621177920ae7db10daccd0a298639ff6dee.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 429, + 402, + 443.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 208, + 443.0, + 402, + 457.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 504, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 133, + 473 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 461, + 158, + 472 + ], + "score": 0.92, + "content": "d _ { \\mathrm { m o d e l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "is the model hidden size, but with the positional encoding1 as both query and key and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "the decoder states as the value. This incorporates positional information directly into the attention", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "process and provides a stronger positional signal than the embedding layer alone. We also hypothe-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 492, + 487, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 487, + 507 + ], + "score": 1.0, + "content": "size that this additional information improves the decoder’s ability to perform local reordering.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 517, + 421, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 421, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 421, + 530 + ], + "score": 1.0, + "content": "3.3 MODELING FERTILITY TO TACKLE THE MULTIMODALITY PROBLEM", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 538, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 423, + 551 + ], + "score": 1.0, + "content": "The multimodality problem can be attacked by introducing a latent variable", + "type": "text" + }, + { + "bbox": [ + 423, + 541, + 430, + 549 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "to directly model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 359, + 562 + ], + "score": 1.0, + "content": "the nondeterminism in the translation process: we first sample", + "type": "text" + }, + { + "bbox": [ + 360, + 552, + 366, + 560 + ], + "score": 0.74, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "from a prior distribution and then", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 561, + 353, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 159, + 573 + ], + "score": 1.0, + "content": "condition on", + "type": "text" + }, + { + "bbox": [ + 159, + 563, + 165, + 570 + ], + "score": 0.71, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 561, + 353, + 573 + ], + "score": 1.0, + "content": "to non-autoregressively generate a translation.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "One way to interpret this latent variable is as a sentence-level “plan” akin to those discussed in the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "language production literature (Martin et al., 2010). There are several desirable properties for this", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 599, + 168, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 168, + 611 + ], + "score": 1.0, + "content": "latent variable:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 619, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "score": 1.0, + "content": "• It should be simple to infer a value for the latent variable given a particular input-output pair, as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 631, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 631, + 294, + 642 + ], + "score": 1.0, + "content": "this is needed to train the model end-to-end.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 149, + 658 + ], + "score": 1.0, + "content": "• Adding", + "type": "text" + }, + { + "bbox": [ + 150, + 647, + 157, + 655 + ], + "score": 0.74, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "to the conditioning context should account as much as possible for the correlations", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "across time between different outputs, so that the remaining marginal probabilities at each output", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 666, + 404, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 666, + 404, + 680 + ], + "score": 1.0, + "content": "location are as close as possible to satisfying conditional independence.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 428, + 694 + ], + "score": 1.0, + "content": "• It should not account for the variation in output translations so directly that", + "type": "text" + }, + { + "bbox": [ + 428, + 681, + 465, + 693 + ], + "score": 0.93, + "content": "p ( y | x , z )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "becomes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 691, + 463, + 705 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 463, + 705 + ], + "score": 1.0, + "content": "trivial to learn, since that is the function our decoder neural network will approximate.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 708, + 502, + 723 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 212, + 723 + ], + "score": 1.0, + "content": "1The positional encoding", + "type": "text" + }, + { + "bbox": [ + 212, + 714, + 219, + 722 + ], + "score": 0.72, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 708, + 276, + 723 + ], + "score": 1.0, + "content": "is computed as", + "type": "text" + }, + { + "bbox": [ + 277, + 710, + 379, + 722 + ], + "score": 0.9, + "content": "p ( j , k ) = \\sin { ( j / 1 0 0 0 0 ^ { k / d } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 708, + 416, + 723 + ], + "score": 1.0, + "content": "(for even", + "type": "text" + }, + { + "bbox": [ + 417, + 712, + 423, + 721 + ], + "score": 0.67, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 708, + 437, + 723 + ], + "score": 1.0, + "content": ") or", + "type": "text" + }, + { + "bbox": [ + 438, + 710, + 502, + 723 + ], + "score": 0.89, + "content": "\\cos { \\left( j / 1 0 0 0 0 ^ { k / d } \\right) }", + "type": "inline_equation" + } + ] + }, + { + "bbox": [ + 106, + 721, + 356, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 138, + 732 + ], + "score": 1.0, + "content": "(for odd", + "type": "text" + }, + { + "bbox": [ + 138, + 722, + 144, + 731 + ], + "score": 0.74, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 721, + 174, + 732 + ], + "score": 1.0, + "content": "), where", + "type": "text" + }, + { + "bbox": [ + 174, + 722, + 180, + 732 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 721, + 272, + 732 + ], + "score": 1.0, + "content": "is the timestep index and", + "type": "text" + }, + { + "bbox": [ + 273, + 722, + 279, + 730 + ], + "score": 0.8, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 721, + 356, + 732 + ], + "score": 1.0, + "content": "is the channel index.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 205, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 208, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 208, + 96 + ], + "score": 1.0, + "content": "3.2 DECODER STACK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 102, + 504, + 125 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 504, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 504, + 116 + ], + "score": 1.0, + "content": "In order to translate non-autoregressively and parallelize the decoding process, we modify the de-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 200, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 200, + 127 + ], + "score": 1.0, + "content": "coder stack as follows.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 101, + 504, + 127 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 132, + 505, + 209 + ], + "lines": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "Decoder Inputs Before decoding starts, the NAT needs to know how long the target sentence", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 143, + 506, + 156 + ], + "score": 1.0, + "content": "will be in order to generate all words in parallel. More crucially, we cannot use time-shifted target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 155, + 506, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 506, + 166 + ], + "score": 1.0, + "content": "outputs (during training) or previously predicted outputs (during inference) as the inputs to the first", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "decoder layer. Omitting inputs to the first decoder layer entirely, or using only positional embed-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 506, + 189 + ], + "score": 1.0, + "content": "dings, resulted in very poor performance. Instead, we initialize the decoding process using copied", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 505, + 199 + ], + "score": 1.0, + "content": "source inputs from the encoder side. As the source and target sentences are often of different lengths,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 200, + 210, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 210, + 210 + ], + "score": 1.0, + "content": "we propose two methods:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 132, + 506, + 210 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 217, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 328, + 231 + ], + "score": 1.0, + "content": "• Copy source inputs uniformly: Each decoder input", + "type": "text" + }, + { + "bbox": [ + 328, + 219, + 333, + 228 + ], + "score": 0.78, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 217, + 425, + 231 + ], + "score": 1.0, + "content": "is a copy of the Round", + "type": "text" + }, + { + "bbox": [ + 425, + 218, + 458, + 230 + ], + "score": 0.78, + "content": "( T ^ { \\prime } t / T )", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 217, + 506, + 231 + ], + "score": 1.0, + "content": "-th encoder", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 229, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 115, + 229, + 505, + 241 + ], + "score": 1.0, + "content": "input. This is equivalent to “scanning” source inputs from left to right with a constant “speed,”", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 240, + 463, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 240, + 463, + 253 + ], + "score": 1.0, + "content": "and results in a decoding process that is deterministic given a (predicted) target length.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 108, + 254, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 108, + 254, + 505, + 267 + ], + "score": 1.0, + "content": "• Copy source inputs using fertilities: A more powerful way, depicted in Fig. 2 and discussed", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 115, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "in more detail below, is to copy each encoder input as a decoder input zero or more times, with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "the number of times each input is copied referred to as that input word’s “fertility.” In this case", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "the source inputs are scanned from left to right at a “speed” that varies inversely with the fertility", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "of each input; the decoding process is now conditioned on the sequence of fertilities, while the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 310, + 397, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 310, + 397, + 322 + ], + "score": 1.0, + "content": "resulting output length is determined by the sum of all fertility values.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 217, + 506, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 505, + 386 + ], + "lines": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "Non-causal self-attention Without the constraint of an autoregressive factorization of the output", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 505, + 353 + ], + "score": 1.0, + "content": "distribution, we no longer need to prevent earlier decoding steps from accessing information from", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 505, + 365 + ], + "score": 1.0, + "content": "later steps. Thus we can avoid the causal mask used in the self-attention module of the conventional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "Transformer’s decoder. Instead, we mask out each query position only from attending to itself,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 374, + 448, + 386 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 448, + 386 + ], + "score": 1.0, + "content": "which we found to improve decoder performance relative to unmasked self-attention.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 330, + 505, + 386 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 404 + ], + "score": 1.0, + "content": "Positional attention We also include an additional positional attention module in each decoder", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 505, + 415 + ], + "score": 1.0, + "content": "layer, which is a multi-head attention module with the same general attention mechanism used in", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 414, + 281, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 281, + 426 + ], + "score": 1.0, + "content": "other parts of the Transformer network, i.e.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 392, + 505, + 426 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 429, + 402, + 457 + ], + "lines": [ + { + "bbox": [ + 208, + 429, + 402, + 457 + ], + "spans": [ + { + "bbox": [ + 208, + 429, + 402, + 457 + ], + "score": 0.93, + "content": "{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { \\mathrm { m o d e l } } } } } \\right) \\cdot V ,", + "type": "interline_equation", + "image_path": "7591f42425b5f5d9a239ca6fc7711621177920ae7db10daccd0a298639ff6dee.jpg" + } + ] + } + ], + "index": 27.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 429, + 402, + 443.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 208, + 443.0, + 402, + 457.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 460, + 504, + 505 + ], + "lines": [ + { + "bbox": [ + 106, + 460, + 506, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 460, + 133, + 473 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 461, + 158, + 472 + ], + "score": 0.92, + "content": "d _ { \\mathrm { m o d e l } }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 460, + 506, + 473 + ], + "score": 1.0, + "content": "is the model hidden size, but with the positional encoding1 as both query and key and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "the decoder states as the value. This incorporates positional information directly into the attention", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 495 + ], + "score": 1.0, + "content": "process and provides a stronger positional signal than the embedding layer alone. We also hypothe-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 492, + 487, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 487, + 507 + ], + "score": 1.0, + "content": "size that this additional information improves the decoder’s ability to perform local reordering.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 460, + 506, + 507 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 517, + 421, + 529 + ], + "lines": [ + { + "bbox": [ + 106, + 517, + 421, + 530 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 421, + 530 + ], + "score": 1.0, + "content": "3.3 MODELING FERTILITY TO TACKLE THE MULTIMODALITY PROBLEM", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 538, + 505, + 572 + ], + "lines": [ + { + "bbox": [ + 106, + 538, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 423, + 551 + ], + "score": 1.0, + "content": "The multimodality problem can be attacked by introducing a latent variable", + "type": "text" + }, + { + "bbox": [ + 423, + 541, + 430, + 549 + ], + "score": 0.75, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 538, + 505, + 551 + ], + "score": 1.0, + "content": "to directly model", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 359, + 562 + ], + "score": 1.0, + "content": "the nondeterminism in the translation process: we first sample", + "type": "text" + }, + { + "bbox": [ + 360, + 552, + 366, + 560 + ], + "score": 0.74, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "from a prior distribution and then", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 561, + 353, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 159, + 573 + ], + "score": 1.0, + "content": "condition on", + "type": "text" + }, + { + "bbox": [ + 159, + 563, + 165, + 570 + ], + "score": 0.71, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 561, + 353, + 573 + ], + "score": 1.0, + "content": "to non-autoregressively generate a translation.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 106, + 538, + 505, + 573 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 577, + 504, + 610 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 589 + ], + "score": 1.0, + "content": "One way to interpret this latent variable is as a sentence-level “plan” akin to those discussed in the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 600 + ], + "score": 1.0, + "content": "language production literature (Martin et al., 2010). There are several desirable properties for this", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 599, + 168, + 611 + ], + "spans": [ + { + "bbox": [ + 106, + 599, + 168, + 611 + ], + "score": 1.0, + "content": "latent variable:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 577, + 505, + 611 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 619, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 632 + ], + "score": 1.0, + "content": "• It should be simple to infer a value for the latent variable given a particular input-output pair, as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 631, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 115, + 631, + 294, + 642 + ], + "score": 1.0, + "content": "this is needed to train the model end-to-end.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 644, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 149, + 658 + ], + "score": 1.0, + "content": "• Adding", + "type": "text" + }, + { + "bbox": [ + 150, + 647, + 157, + 655 + ], + "score": 0.74, + "content": "z", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 644, + 505, + 658 + ], + "score": 1.0, + "content": "to the conditioning context should account as much as possible for the correlations", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 655, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 655, + 505, + 669 + ], + "score": 1.0, + "content": "across time between different outputs, so that the remaining marginal probabilities at each output", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 666, + 404, + 680 + ], + "spans": [ + { + "bbox": [ + 115, + 666, + 404, + 680 + ], + "score": 1.0, + "content": "location are as close as possible to satisfying conditional independence.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 428, + 694 + ], + "score": 1.0, + "content": "• It should not account for the variation in output translations so directly that", + "type": "text" + }, + { + "bbox": [ + 428, + 681, + 465, + 693 + ], + "score": 0.93, + "content": "p ( y | x , z )", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "becomes", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 115, + 691, + 463, + 705 + ], + "spans": [ + { + "bbox": [ + 115, + 691, + 463, + 705 + ], + "score": 1.0, + "content": "trivial to learn, since that is the function our decoder neural network will approximate.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 618, + 505, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The factorization by length introduced in Eq. 3 provides a very weak example of a latent variable", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "model, satisfying the first and third property but not the first. We propose the use of fertilities instead.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "These are integers for each word in the source sentence that correspond to the number of words in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "the target sentence that can be aligned to that source word using a hard alignment algorithm like", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 248, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 248, + 138 + ], + "score": 1.0, + "content": "IBM Model 2 (Brown et al., 1993).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 504, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "One of the most important properties of the proposed NAT is that it naturally introduces an informa-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "tive latent variable when we choose to copy the encoder inputs based on predicted fertilities. More", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 475, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 243, + 178 + ], + "score": 1.0, + "content": "precisely, given a source sentence", + "type": "text" + }, + { + "bbox": [ + 243, + 166, + 253, + 175 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 165, + 453, + 178 + ], + "score": 1.0, + "content": ", the conditional probability of a target translation", + "type": "text" + }, + { + "bbox": [ + 453, + 167, + 462, + 175 + ], + "score": 0.84, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 165, + 475, + 178 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "interline_equation", + "bbox": [ + 122, + 182, + 476, + 222 + ], + "lines": [ + { + "bbox": [ + 122, + 182, + 476, + 222 + ], + "spans": [ + { + "bbox": [ + 122, + 182, + 476, + 222 + ], + "score": 0.95, + "content": "p _ { \\mathcal N \\mathcal M } ( Y | X ; \\theta ) = \\sum _ { f _ { 1 } , . . . , f _ { T ^ { \\prime } } \\in \\mathcal F } \\left( \\prod _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)", + "type": "interline_equation", + "image_path": "addbc095e30bd0f2ec3e327b7ac2d196aff125984ec04b5d2c19836176794d07.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 122, + 182, + 476, + 195.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 122, + 195.33333333333334, + 476, + 208.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 122, + 208.66666666666669, + 476, + 222.00000000000003 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 228, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 102, + 222, + 509, + 255 + ], + "spans": [ + { + "bbox": [ + 102, + 222, + 133, + 255 + ], + "score": 1.0, + "content": "where value", + "type": "text" + }, + { + "bbox": [ + 133, + 227, + 311, + 243 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { F } = \\{ f _ { 1 } , . . . , f _ { T ^ { \\prime } } | \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } f _ { t ^ { \\prime } } = T , f _ { t ^ { \\prime } } \\in \\mathbb { Z } ^ { * } \\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 222, + 330, + 255 + ], + "score": 1.0, + "content": "is thand", + "type": "text" + }, + { + "bbox": [ + 353, + 222, + 425, + 255 + ], + "score": 1.0, + "content": "of all fertility sequdenotes the token", + "type": "text" + }, + { + "bbox": [ + 433, + 222, + 470, + 255 + ], + "score": 1.0, + "content": "ces—onerepeated", + "type": "text" + }, + { + "bbox": [ + 478, + 222, + 509, + 255 + ], + "score": 1.0, + "content": "ertilitytimes.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 241, + 477, + 254 + ], + "spans": [ + { + "bbox": [ + 303, + 243, + 312, + 252 + ], + "score": 0.81, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 241, + 353, + 254 + ], + "score": 0.94, + "content": "x \\{ f \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 244, + 433, + 252 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 242, + 477, + 253 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 260, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 104, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 104, + 259, + 369, + 273 + ], + "score": 1.0, + "content": "Fertility prediction As shown in Fig. 2, we model the fertility", + "type": "text" + }, + { + "bbox": [ + 369, + 261, + 424, + 273 + ], + "score": 0.91, + "content": "p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 259, + 505, + 273 + ], + "score": 1.0, + "content": "at each position in-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 392, + 285 + ], + "score": 1.0, + "content": "dependently using a one-layer neural network with a softmax classifier", + "type": "text" + }, + { + "bbox": [ + 393, + 273, + 424, + 282 + ], + "score": 0.86, + "content": "L = 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "in our experiments)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "on top of the output of the last encoder layer. This models the way that fertility values are a property", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 294, + 442, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 442, + 306 + ], + "score": 1.0, + "content": "of each input word but depend on information and context from the entire sentence.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "Benefits of fertility Fertilities possess all three of the properties listed earlier as desired of a latent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 323, + 316, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 316, + 336 + ], + "score": 1.0, + "content": "variable for non-autoregressive machine translation:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 345, + 505, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "score": 1.0, + "content": "• An external aligner provides a simple and fast approximate inference model that effectively re-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 356, + 376, + 369 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 376, + 369 + ], + "score": 1.0, + "content": "duces the unsupervised training problem to two supervised ones.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "• Using fertilities as a latent variable makes significant progress towards solving the multimodality", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "problem by providing a natural factorization of the output space. Given a source sentence, restrict-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 116, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "ing the output distribution to those target sentences consistent with a particular fertility sequence", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 405, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 116, + 405, + 506, + 416 + ], + "score": 1.0, + "content": "dramatically reduces the mode space. Furthermore, the global choice of mode is factored into a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "set of local mode choices: namely, how to translate each input word. These local mode choices", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 425, + 428, + 438 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 428, + 438 + ], + "score": 1.0, + "content": "can be effectively supervised because the fertilities provide a fixed “scaffold.”", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 111, + 441, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 111, + 441, + 505, + 454 + ], + "score": 1.0, + "content": "Including both fertilities and reordering in the latent variable would provide complete alignment", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "statistics. This would make the decoding function trivially easy to approximate given the latent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 116, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "variable and force all of the modeling complexity into the encoder. Using fertilities alone allows", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 474, + 350, + 486 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 350, + 486 + ], + "score": 1.0, + "content": "the decoder to take some of this burden off of the encoder.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "Our use of fertilities as a latent variable also means that there is no need to have a separate means of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 507, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 518 + ], + "score": 1.0, + "content": "explicitly modeling the length of the translation, which is simply the sum of fertilities. And fertilities", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "provide a powerful way to condition the decoding process, allowing the model to generate diverse", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 300, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 300, + 542 + ], + "score": 1.0, + "content": "translations by sampling over the fertility space.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 107, + 553, + 377, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 378, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 378, + 567 + ], + "score": 1.0, + "content": "3.4 TRANSLATION PREDICTOR AND THE DECODING PROCESS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "score": 1.0, + "content": "At inference time, the model can identify the translation with the highest conditional probability (see", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 585, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 504, + 599 + ], + "score": 1.0, + "content": "Eq. 5) by marginalizing over all possible latent fertility sequences. Given a fertility sequence, how-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 596, + 504, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 504, + 610 + ], + "score": 1.0, + "content": "ever, identifying the optimal translation only requires independently maximizing the local probabil-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 607, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 266, + 621 + ], + "score": 1.0, + "content": "ity for each output position. We define", + "type": "text" + }, + { + "bbox": [ + 266, + 607, + 360, + 619 + ], + "score": 0.92, + "content": "Y \\overset { \\cdot } { = } G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 607, + 505, + 621 + ], + "score": 1.0, + "content": "to represent the optimal translation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 618, + 339, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 339, + 631 + ], + "score": 1.0, + "content": "given a source sentence and a sequence of fertility values.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 105, + 635, + 504, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 648 + ], + "score": 1.0, + "content": "But searching and marginalizing over the whole fertility space is still intractable. We propose three", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 646, + 411, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 411, + 658 + ], + "score": 1.0, + "content": "heuristic decoding algorithms to reduce the search space of the NAT model:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Argmax decoding Since the fertility sequence is also modeled with a conditionally independent", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 675, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 505, + 690 + ], + "score": 1.0, + "content": "factorization, we can simply estimate the best translation by choosing the highest-probability fertility", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 190, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 190, + 699 + ], + "score": 1.0, + "content": "for each input word:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + }, + { + "type": "interline_equation", + "bbox": [ + 171, + 704, + 440, + 727 + ], + "lines": [ + { + "bbox": [ + 171, + 704, + 440, + 727 + ], + "spans": [ + { + "bbox": [ + 171, + 704, + 440, + 727 + ], + "score": 0.89, + "content": "\\hat { Y } _ { \\mathrm { a r g m a x } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e \\ } \\hat { f } _ { t ^ { \\prime } } = \\underset { f } { \\operatorname { a r g m a x } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta )", + "type": "interline_equation", + "image_path": "c87b82f443ad023fe764cc137a25c0b5f30813f1855f5398ec17018f945487d0.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 171, + 704, + 440, + 727 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "The factorization by length introduced in Eq. 3 provides a very weak example of a latent variable", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "model, satisfying the first and third property but not the first. We propose the use of fertilities instead.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "These are integers for each word in the source sentence that correspond to the number of words in", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 128 + ], + "score": 1.0, + "content": "the target sentence that can be aligned to that source word using a hard alignment algorithm like", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 248, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 248, + 138 + ], + "score": 1.0, + "content": "IBM Model 2 (Brown et al., 1993).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 82, + 505, + 138 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 143, + 504, + 177 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "One of the most important properties of the proposed NAT is that it naturally introduces an informa-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 506, + 167 + ], + "score": 1.0, + "content": "tive latent variable when we choose to copy the encoder inputs based on predicted fertilities. More", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 475, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 243, + 178 + ], + "score": 1.0, + "content": "precisely, given a source sentence", + "type": "text" + }, + { + "bbox": [ + 243, + 166, + 253, + 175 + ], + "score": 0.82, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 165, + 453, + 178 + ], + "score": 1.0, + "content": ", the conditional probability of a target translation", + "type": "text" + }, + { + "bbox": [ + 453, + 167, + 462, + 175 + ], + "score": 0.84, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 165, + 475, + 178 + ], + "score": 1.0, + "content": "is:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 142, + 506, + 178 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 122, + 182, + 476, + 222 + ], + "lines": [ + { + "bbox": [ + 122, + 182, + 476, + 222 + ], + "spans": [ + { + "bbox": [ + 122, + 182, + 476, + 222 + ], + "score": 0.95, + "content": "p _ { \\mathcal N \\mathcal M } ( Y | X ; \\theta ) = \\sum _ { f _ { 1 } , . . . , f _ { T ^ { \\prime } } \\in \\mathcal F } \\left( \\prod _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)", + "type": "interline_equation", + "image_path": "addbc095e30bd0f2ec3e327b7ac2d196aff125984ec04b5d2c19836176794d07.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 122, + 182, + 476, + 195.33333333333334 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 122, + 195.33333333333334, + 476, + 208.66666666666669 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 122, + 208.66666666666669, + 476, + 222.00000000000003 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 228, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 102, + 222, + 509, + 255 + ], + "spans": [ + { + "bbox": [ + 102, + 222, + 133, + 255 + ], + "score": 1.0, + "content": "where value", + "type": "text" + }, + { + "bbox": [ + 133, + 227, + 311, + 243 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathcal { F } = \\{ f _ { 1 } , . . . , f _ { T ^ { \\prime } } | \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } f _ { t ^ { \\prime } } = T , f _ { t ^ { \\prime } } \\in \\mathbb { Z } ^ { * } \\} } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 222, + 330, + 255 + ], + "score": 1.0, + "content": "is thand", + "type": "text" + }, + { + "bbox": [ + 353, + 222, + 425, + 255 + ], + "score": 1.0, + "content": "of all fertility sequdenotes the token", + "type": "text" + }, + { + "bbox": [ + 433, + 222, + 470, + 255 + ], + "score": 1.0, + "content": "ces—onerepeated", + "type": "text" + }, + { + "bbox": [ + 478, + 222, + 509, + 255 + ], + "score": 1.0, + "content": "ertilitytimes.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 303, + 241, + 477, + 254 + ], + "spans": [ + { + "bbox": [ + 303, + 243, + 312, + 252 + ], + "score": 0.81, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 241, + 353, + 254 + ], + "score": 0.94, + "content": "x \\{ f \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 244, + 433, + 252 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 470, + 242, + 477, + 253 + ], + "score": 0.85, + "content": "f", + "type": "inline_equation" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 102, + 222, + 509, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 260, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 104, + 259, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 104, + 259, + 369, + 273 + ], + "score": 1.0, + "content": "Fertility prediction As shown in Fig. 2, we model the fertility", + "type": "text" + }, + { + "bbox": [ + 369, + 261, + 424, + 273 + ], + "score": 0.91, + "content": "p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 259, + 505, + 273 + ], + "score": 1.0, + "content": "at each position in-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 271, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 392, + 285 + ], + "score": 1.0, + "content": "dependently using a one-layer neural network with a softmax classifier", + "type": "text" + }, + { + "bbox": [ + 393, + 273, + 424, + 282 + ], + "score": 0.86, + "content": "L = 5 0", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 271, + 505, + 285 + ], + "score": 1.0, + "content": "in our experiments)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "on top of the output of the last encoder layer. This models the way that fertility values are a property", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 294, + 442, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 442, + 306 + ], + "score": 1.0, + "content": "of each input word but depend on information and context from the entire sentence.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 259, + 505, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 312, + 504, + 335 + ], + "lines": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 505, + 325 + ], + "score": 1.0, + "content": "Benefits of fertility Fertilities possess all three of the properties listed earlier as desired of a latent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 323, + 316, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 316, + 336 + ], + "score": 1.0, + "content": "variable for non-autoregressive machine translation:", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 312, + 505, + 336 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 345, + 505, + 485 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 358 + ], + "score": 1.0, + "content": "• An external aligner provides a simple and fast approximate inference model that effectively re-", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 356, + 376, + 369 + ], + "spans": [ + { + "bbox": [ + 115, + 356, + 376, + 369 + ], + "score": 1.0, + "content": "duces the unsupervised training problem to two supervised ones.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 370, + 505, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 384 + ], + "score": 1.0, + "content": "• Using fertilities as a latent variable makes significant progress towards solving the multimodality", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 382, + 505, + 395 + ], + "score": 1.0, + "content": "problem by providing a natural factorization of the output space. Given a source sentence, restrict-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 116, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "ing the output distribution to those target sentences consistent with a particular fertility sequence", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 405, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 116, + 405, + 506, + 416 + ], + "score": 1.0, + "content": "dramatically reduces the mode space. Furthermore, the global choice of mode is factored into a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 115, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "set of local mode choices: namely, how to translate each input word. These local mode choices", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 425, + 428, + 438 + ], + "spans": [ + { + "bbox": [ + 115, + 425, + 428, + 438 + ], + "score": 1.0, + "content": "can be effectively supervised because the fertilities provide a fixed “scaffold.”", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 111, + 441, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 111, + 441, + 505, + 454 + ], + "score": 1.0, + "content": "Including both fertilities and reordering in the latent variable would provide complete alignment", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 115, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "statistics. This would make the decoding function trivially easy to approximate given the latent", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 116, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "variable and force all of the modeling complexity into the encoder. Using fertilities alone allows", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 474, + 350, + 486 + ], + "spans": [ + { + "bbox": [ + 116, + 474, + 350, + 486 + ], + "score": 1.0, + "content": "the decoder to take some of this burden off of the encoder.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 344, + 506, + 486 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 495, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 506, + 507 + ], + "score": 1.0, + "content": "Our use of fertilities as a latent variable also means that there is no need to have a separate means of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 507, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 518 + ], + "score": 1.0, + "content": "explicitly modeling the length of the translation, which is simply the sum of fertilities. And fertilities", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 506, + 531 + ], + "score": 1.0, + "content": "provide a powerful way to condition the decoding process, allowing the model to generate diverse", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 300, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 300, + 542 + ], + "score": 1.0, + "content": "translations by sampling over the fertility space.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 495, + 506, + 542 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 553, + 377, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 378, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 378, + 567 + ], + "score": 1.0, + "content": "3.4 TRANSLATION PREDICTOR AND THE DECODING PROCESS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 574, + 505, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 588 + ], + "score": 1.0, + "content": "At inference time, the model can identify the translation with the highest conditional probability (see", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 585, + 504, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 504, + 599 + ], + "score": 1.0, + "content": "Eq. 5) by marginalizing over all possible latent fertility sequences. Given a fertility sequence, how-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 596, + 504, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 504, + 610 + ], + "score": 1.0, + "content": "ever, identifying the optimal translation only requires independently maximizing the local probabil-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 607, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 266, + 621 + ], + "score": 1.0, + "content": "ity for each output position. We define", + "type": "text" + }, + { + "bbox": [ + 266, + 607, + 360, + 619 + ], + "score": 0.92, + "content": "Y \\overset { \\cdot } { = } G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 607, + 505, + 621 + ], + "score": 1.0, + "content": "to represent the optimal translation", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 618, + 339, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 339, + 631 + ], + "score": 1.0, + "content": "given a source sentence and a sequence of fertility values.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 573, + 505, + 631 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 635, + 504, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 648 + ], + "score": 1.0, + "content": "But searching and marginalizing over the whole fertility space is still intractable. We propose three", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 646, + 411, + 658 + ], + "spans": [ + { + "bbox": [ + 106, + 646, + 411, + 658 + ], + "score": 1.0, + "content": "heuristic decoding algorithms to reduce the search space of the NAT model:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 633, + 505, + 658 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 665, + 504, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "Argmax decoding Since the fertility sequence is also modeled with a conditionally independent", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 675, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 505, + 690 + ], + "score": 1.0, + "content": "factorization, we can simply estimate the best translation by choosing the highest-probability fertility", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 687, + 190, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 190, + 699 + ], + "score": 1.0, + "content": "for each input word:", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 665, + 506, + 699 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 171, + 704, + 440, + 727 + ], + "lines": [ + { + "bbox": [ + 171, + 704, + 440, + 727 + ], + "spans": [ + { + "bbox": [ + 171, + 704, + 440, + 727 + ], + "score": 0.89, + "content": "\\hat { Y } _ { \\mathrm { a r g m a x } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e \\ } \\hat { f } _ { t ^ { \\prime } } = \\underset { f } { \\operatorname { a r g m a x } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta )", + "type": "interline_equation", + "image_path": "c87b82f443ad023fe764cc137a25c0b5f30813f1855f5398ec17018f945487d0.jpg" + } + ] + } + ], + "index": 46, + "virtual_lines": [ + { + "bbox": [ + 171, + 704, + 440, + 727 + ], + "spans": [], + "index": 46 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Average decoding We can also estimate each fertility as the expectation of its corresponding soft-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 177, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 177, + 106 + ], + "score": 1.0, + "content": "max distribution:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 108, + 462, + 149 + ], + "lines": [ + { + "bbox": [ + 148, + 108, + 462, + 149 + ], + "spans": [ + { + "bbox": [ + 148, + 108, + 462, + 149 + ], + "score": 0.92, + "content": "\\hat { Y } _ { \\mathrm { a v e r a g e } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e } \\hat { f } _ { t ^ { \\prime } } = \\mathrm { R o u n d } \\left( \\sum _ { f _ { t ^ { \\prime } } = 1 } ^ { L } p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta \\big ) f _ { t ^ { \\prime } } \\right)", + "type": "interline_equation", + "image_path": "cddc8316aa514a2040ceed5ae75c4f96052fd7ea9df38f637de3023bdc4c2c2c.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 148, + 108, + 462, + 121.66666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 148, + 121.66666666666667, + 462, + 135.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 148, + 135.33333333333334, + 462, + 149.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "Noisy parallel decoding (NPD) A more accurate approximation of the true optimum of the target", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "score": 1.0, + "content": "distribution, inspired by Cho (2016), is to draw samples from the fertility space and compute the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "score": 1.0, + "content": "best translation for each fertility sequence. We can then use the autoregressive teacher to identify", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 217, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 217, + 207 + ], + "score": 1.0, + "content": "the best overall translation:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 211, + 425, + 234 + ], + "lines": [ + { + "bbox": [ + 186, + 211, + 425, + 234 + ], + "spans": [ + { + "bbox": [ + 186, + 211, + 425, + 234 + ], + "score": 0.92, + "content": "\\hat { Y } _ { \\mathrm { N P D } } = G ( x _ { 1 : T ^ { \\prime } } , \\underset { f _ { t ^ { \\prime } } \\sim p _ { F } } { \\mathrm { a r g m a x } } p _ { \\mathcal { A R } } ( G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) | X ; \\theta ) ; \\theta )", + "type": "interline_equation", + "image_path": "ddb83a2f54effd61e8585241161c5ecfe255d9b7958c6a9e1d1be0a28272f922.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 186, + 211, + 425, + 234 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 240, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "Note that, when using an autoregressive model as a scoring function for a set of decoded translations,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 251, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 264 + ], + "score": 1.0, + "content": "it can run as fast as it does at train time because it can be provided with all decoder inputs in parallel.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "NPD is a stochastic search method, and it also increases the computational resources required lin-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "early by the sample size. However, because all the search samples can be computed and scored", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "entirely independently, the process only doubles the latency compared to computing a single trans-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 302, + 276, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 276, + 313 + ], + "score": 1.0, + "content": "lation if sufficient parallelism is available.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 107, + 329, + 179, + 342 + ], + "lines": [ + { + "bbox": [ + 104, + 328, + 180, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 328, + 180, + 344 + ], + "score": 1.0, + "content": "4 TRAINING", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 396, + 368 + ], + "score": 1.0, + "content": "The proposed NAT contains a discrete sequential latent variable", + "type": "text" + }, + { + "bbox": [ + 396, + 355, + 417, + 367 + ], + "score": 0.89, + "content": "f _ { 1 : T ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 354, + 505, + 368 + ], + "score": 1.0, + "content": ", whose conditional", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 198, + 379 + ], + "score": 1.0, + "content": "posterior distribution", + "type": "text" + }, + { + "bbox": [ + 198, + 366, + 287, + 378 + ], + "score": 0.92, + "content": "p ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "we can approximate using a proposal distribution", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 185, + 389 + ], + "score": 0.92, + "content": "q ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 376, + 505, + 389 + ], + "score": 1.0, + "content": ". This provides a variational bound for the overall maximum likelihood loss:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 402, + 484, + 498 + ], + "lines": [ + { + "bbox": [ + 115, + 402, + 484, + 498 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 484, + 498 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { N A } } ( Y | X ; \\theta ) = \\log \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } } \\in \\mathcal { F } } p _ { F } ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot p ( y _ { 1 : T } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) } \\\\ & { \\geq \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } \\sim \\mathcal { I } } } \\left( \\underbrace { T } _ { \\mathrm { t - 1 } } \\log p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)} _ { \\mathrm { T r a n s l a t i o n ~ L o s s } } + \\underbrace { \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } \\log p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) } _ { \\mathrm { F e r t i l i t y ~ L o s s } } + \\mathcal { H } ( q ) \\end{array}", + "type": "interline_equation", + "image_path": "98d5a44ed0ad3d2976698667125e9e76d3327aac64ae197046cd812c046fa407.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 115, + 402, + 484, + 434.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 115, + 434.0, + 484, + 466.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 115, + 466.0, + 484, + 498.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 243, + 522 + ], + "score": 1.0, + "content": "We choose a proposal distribution", + "type": "text" + }, + { + "bbox": [ + 244, + 511, + 250, + 520 + ], + "score": 0.75, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "defined by a separate, fixed fertility model. Possible options in-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "clude the output of an external aligner, which produces a deterministic sequence of integer fertilities", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 530, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 543 + ], + "score": 1.0, + "content": "for each (source, target) pair in a training corpus, or fertilities computed from the attention weights", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 540, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 555 + ], + "score": 1.0, + "content": "used in our fixed autoregressive teacher model. This simplifies the inference process considerably,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 553, + 273, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 200, + 565 + ], + "score": 1.0, + "content": "as the expectation over", + "type": "text" + }, + { + "bbox": [ + 200, + 554, + 207, + 564 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 553, + 273, + 565 + ], + "score": 1.0, + "content": "is deterministic.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "The resulting loss function, consisting of the two bracketed terms in Eq. 9, allows us to train the en-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "tire model in a supervised fashion, using the inferred fertilities to simultaneously train the translation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 590, + 352, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 133, + 605 + ], + "score": 1.0, + "content": "model", + "type": "text" + }, + { + "bbox": [ + 134, + 593, + 140, + 603 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 590, + 335, + 605 + ], + "score": 1.0, + "content": "and supervise the fertility neural network model", + "type": "text" + }, + { + "bbox": [ + 335, + 592, + 348, + 603 + ], + "score": 0.85, + "content": "p _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 590, + 352, + 605 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "title", + "bbox": [ + 106, + 616, + 334, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 335, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 335, + 630 + ], + "score": 1.0, + "content": "4.1 SEQUENCE-LEVEL KNOWLEDGE DISTILLATION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "While the latent fertility model substantially improves the ability of the non-autoregressive out-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "put distribution to approximate the multimodal target distribution, it does not completely solve the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "problem of nondeterminism in the training data. In many cases, there are multiple correct transla-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "tions consistent with a single sequence of fertilities—for instance, both “Danke schon.” and “Vielen ¨", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "dank.” are consistent with the English input “Thank you.” and the fertility sequence [2, 0, 1], because", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 693, + 340, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 693, + 340, + 704 + ], + "score": 1.0, + "content": "“you” is not directly translated in either German sentence.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Thus we additionally apply sequence-level knowledge distillation (Kim & Rush, 2016) to construct", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "a new corpus by training an autoregressive machine translation model, known as the teacher, on an", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 310, + 761 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Average decoding We can also estimate each fertility as the expectation of its corresponding soft-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 177, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 177, + 106 + ], + "score": 1.0, + "content": "max distribution:", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 81, + 505, + 106 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 148, + 108, + 462, + 149 + ], + "lines": [ + { + "bbox": [ + 148, + 108, + 462, + 149 + ], + "spans": [ + { + "bbox": [ + 148, + 108, + 462, + 149 + ], + "score": 0.92, + "content": "\\hat { Y } _ { \\mathrm { a v e r a g e } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e } \\hat { f } _ { t ^ { \\prime } } = \\mathrm { R o u n d } \\left( \\sum _ { f _ { t ^ { \\prime } } = 1 } ^ { L } p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta \\big ) f _ { t ^ { \\prime } } \\right)", + "type": "interline_equation", + "image_path": "cddc8316aa514a2040ceed5ae75c4f96052fd7ea9df38f637de3023bdc4c2c2c.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 148, + 108, + 462, + 121.66666666666667 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 148, + 121.66666666666667, + 462, + 135.33333333333334 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 148, + 135.33333333333334, + 462, + 149.0 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 175 + ], + "score": 1.0, + "content": "Noisy parallel decoding (NPD) A more accurate approximation of the true optimum of the target", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "score": 1.0, + "content": "distribution, inspired by Cho (2016), is to draw samples from the fertility space and compute the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "score": 1.0, + "content": "best translation for each fertility sequence. We can then use the autoregressive teacher to identify", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 217, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 217, + 207 + ], + "score": 1.0, + "content": "the best overall translation:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 161, + 506, + 207 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 186, + 211, + 425, + 234 + ], + "lines": [ + { + "bbox": [ + 186, + 211, + 425, + 234 + ], + "spans": [ + { + "bbox": [ + 186, + 211, + 425, + 234 + ], + "score": 0.92, + "content": "\\hat { Y } _ { \\mathrm { N P D } } = G ( x _ { 1 : T ^ { \\prime } } , \\underset { f _ { t ^ { \\prime } } \\sim p _ { F } } { \\mathrm { a r g m a x } } p _ { \\mathcal { A R } } ( G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) | X ; \\theta ) ; \\theta )", + "type": "interline_equation", + "image_path": "ddb83a2f54effd61e8585241161c5ecfe255d9b7958c6a9e1d1be0a28272f922.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 186, + 211, + 425, + 234 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 240, + 504, + 263 + ], + "lines": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 253 + ], + "score": 1.0, + "content": "Note that, when using an autoregressive model as a scoring function for a set of decoded translations,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 251, + 505, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 505, + 264 + ], + "score": 1.0, + "content": "it can run as fast as it does at train time because it can be provided with all decoder inputs in parallel.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 240, + 505, + 264 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 268, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "NPD is a stochastic search method, and it also increases the computational resources required lin-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 280, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 291 + ], + "score": 1.0, + "content": "early by the sample size. However, because all the search samples can be computed and scored", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "entirely independently, the process only doubles the latency compared to computing a single trans-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 302, + 276, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 276, + 313 + ], + "score": 1.0, + "content": "lation if sufficient parallelism is available.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 268, + 506, + 313 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 329, + 179, + 342 + ], + "lines": [ + { + "bbox": [ + 104, + 328, + 180, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 328, + 180, + 344 + ], + "score": 1.0, + "content": "4 TRAINING", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 505, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 396, + 368 + ], + "score": 1.0, + "content": "The proposed NAT contains a discrete sequential latent variable", + "type": "text" + }, + { + "bbox": [ + 396, + 355, + 417, + 367 + ], + "score": 0.89, + "content": "f _ { 1 : T ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 354, + 505, + 368 + ], + "score": 1.0, + "content": ", whose conditional", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 198, + 379 + ], + "score": 1.0, + "content": "posterior distribution", + "type": "text" + }, + { + "bbox": [ + 198, + 366, + 287, + 378 + ], + "score": 0.92, + "content": "p ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "we can approximate using a proposal distribution", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 185, + 389 + ], + "score": 0.92, + "content": "q ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } )", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 376, + 505, + 389 + ], + "score": 1.0, + "content": ". This provides a variational bound for the overall maximum likelihood loss:", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 354, + 506, + 389 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 115, + 402, + 484, + 498 + ], + "lines": [ + { + "bbox": [ + 115, + 402, + 484, + 498 + ], + "spans": [ + { + "bbox": [ + 115, + 402, + 484, + 498 + ], + "score": 0.95, + "content": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { N A } } ( Y | X ; \\theta ) = \\log \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } } \\in \\mathcal { F } } p _ { F } ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot p ( y _ { 1 : T } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) } \\\\ & { \\geq \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } \\sim \\mathcal { I } } } \\left( \\underbrace { T } _ { \\mathrm { t - 1 } } \\log p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)} _ { \\mathrm { T r a n s l a t i o n ~ L o s s } } + \\underbrace { \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } \\log p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) } _ { \\mathrm { F e r t i l i t y ~ L o s s } } + \\mathcal { H } ( q ) \\end{array}", + "type": "interline_equation", + "image_path": "98d5a44ed0ad3d2976698667125e9e76d3327aac64ae197046cd812c046fa407.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 115, + 402, + 484, + 434.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 115, + 434.0, + 484, + 466.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 115, + 466.0, + 484, + 498.0 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 508, + 505, + 564 + ], + "lines": [ + { + "bbox": [ + 106, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 243, + 522 + ], + "score": 1.0, + "content": "We choose a proposal distribution", + "type": "text" + }, + { + "bbox": [ + 244, + 511, + 250, + 520 + ], + "score": 0.75, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "defined by a separate, fixed fertility model. Possible options in-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "clude the output of an external aligner, which produces a deterministic sequence of integer fertilities", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 530, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 505, + 543 + ], + "score": 1.0, + "content": "for each (source, target) pair in a training corpus, or fertilities computed from the attention weights", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 540, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 555 + ], + "score": 1.0, + "content": "used in our fixed autoregressive teacher model. This simplifies the inference process considerably,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 553, + 273, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 200, + 565 + ], + "score": 1.0, + "content": "as the expectation over", + "type": "text" + }, + { + "bbox": [ + 200, + 554, + 207, + 564 + ], + "score": 0.78, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 553, + 273, + 565 + ], + "score": 1.0, + "content": "is deterministic.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 508, + 505, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 569, + 505, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "The resulting loss function, consisting of the two bracketed terms in Eq. 9, allows us to train the en-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 593 + ], + "score": 1.0, + "content": "tire model in a supervised fashion, using the inferred fertilities to simultaneously train the translation", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 590, + 352, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 133, + 605 + ], + "score": 1.0, + "content": "model", + "type": "text" + }, + { + "bbox": [ + 134, + 593, + 140, + 603 + ], + "score": 0.79, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 590, + 335, + 605 + ], + "score": 1.0, + "content": "and supervise the fertility neural network model", + "type": "text" + }, + { + "bbox": [ + 335, + 592, + 348, + 603 + ], + "score": 0.85, + "content": "p _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 590, + 352, + 605 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 568, + 505, + 605 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 616, + 334, + 628 + ], + "lines": [ + { + "bbox": [ + 106, + 617, + 335, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 335, + 630 + ], + "score": 1.0, + "content": "4.1 SEQUENCE-LEVEL KNOWLEDGE DISTILLATION", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 637, + 505, + 704 + ], + "lines": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "While the latent fertility model substantially improves the ability of the non-autoregressive out-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "put distribution to approximate the multimodal target distribution, it does not completely solve the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "problem of nondeterminism in the training data. In many cases, there are multiple correct transla-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 671, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 505, + 682 + ], + "score": 1.0, + "content": "tions consistent with a single sequence of fertilities—for instance, both “Danke schon.” and “Vielen ¨", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "dank.” are consistent with the English input “Thank you.” and the fertility sequence [2, 0, 1], because", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 693, + 340, + 704 + ], + "spans": [ + { + "bbox": [ + 107, + 693, + 340, + 704 + ], + "score": 1.0, + "content": "“you” is not directly translated in either German sentence.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 637, + 505, + 704 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Thus we additionally apply sequence-level knowledge distillation (Kim & Rush, 2016) to construct", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "score": 1.0, + "content": "a new corpus by training an autoregressive machine translation model, known as the teacher, on an", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "existing training corpus, then using that model’s greedy outputs as the targets for training the non-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "autoregressive student. The resulting targets are less noisy and more deterministic, as the trained", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "model will consistently translate a sentence like “Thank you.” into the same German translation", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 446, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 446, + 128 + ], + "score": 1.0, + "content": "every time; on the other hand, they are also lower in quality than the original dataset.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "existing training corpus, then using that model’s greedy outputs as the targets for training the non-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 94, + 505, + 105 + ], + "score": 1.0, + "content": "autoregressive student. The resulting targets are less noisy and more deterministic, as the trained", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "model will consistently translate a sentence like “Thank you.” into the same German translation", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 446, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 446, + 128 + ], + "score": 1.0, + "content": "every time; on the other hand, they are also lower in quality than the original dataset.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 191, + 151 + ], + "lines": [ + { + "bbox": [ + 105, + 138, + 193, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 193, + 153 + ], + "score": 1.0, + "content": "4.2 FINE-TUNING", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "score": 1.0, + "content": "Our supervised fertility model enables a decomposition of the overall maximum likelihood loss", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "score": 1.0, + "content": "into translation and fertility terms, but it has some drawbacks compared to variational training. In", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "particular, it heavily relies on the deterministic, approximate inference model provided by the ex-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "score": 1.0, + "content": "ternal alignment system, while it would be desirable to train the entire model, including the fertility", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 193, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 193, + 216 + ], + "score": 1.0, + "content": "predictor, end to end.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 221, + 504, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 234 + ], + "score": 1.0, + "content": "Thus we propose a fine-tuning step after training the NAT to convergence. We introduce an addi-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "tional loss term consisting of the reverse K-L divergence with the teacher output distribution, a form", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 256, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 256, + 255 + ], + "score": 1.0, + "content": "of word-level knowledge distillation:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 257, + 467, + 293 + ], + "lines": [ + { + "bbox": [ + 143, + 257, + 467, + 293 + ], + "spans": [ + { + "bbox": [ + 143, + 257, + 467, + 293 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } ; \\theta \\right) = \\sum _ { t = 1 } ^ { T } \\sum _ { y _ { t } } \\left[ \\log p _ { \\mathcal { A R } } \\left( y _ { t } | \\hat { y } _ { 1 : t - 1 } , x _ { 1 : T ^ { \\prime } } \\right) \\cdot p _ { \\mathcal { N A } } \\left( y _ { t } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta \\right) \\right] ,", + "type": "interline_equation", + "image_path": "e65993392162bd75d4d5f03dbe09c11504275ad6f8a4d1ecdf2eaf6fe3d240bf.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 143, + 257, + 467, + 269.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 143, + 269.0, + 467, + 281.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 143, + 281.0, + 467, + 293.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 506, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 133, + 309 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 296, + 235, + 308 + ], + "score": 0.92, + "content": "\\hat { y } _ { 1 : T } = G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 296, + 505, + 309 + ], + "score": 1.0, + "content": ". Such a loss is more favorable towards highly peaked student output", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 307, + 341, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 341, + 319 + ], + "score": 1.0, + "content": "distributions than a standard cross-entropy error would be.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "Then we train the whole model jointly with a weighted sum of the original distillation loss and two", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "such terms, one an expectation over the predicted fertility distribution, normalized with a baseline,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 347, + 349, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 349, + 358 + ], + "score": 1.0, + "content": "and the other based on the external fertility inference model:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 110, + 360, + 483, + 420 + ], + "lines": [ + { + "bbox": [ + 110, + 360, + 483, + 420 + ], + "spans": [ + { + "bbox": [ + 110, + 360, + 483, + 420 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { \\mathrm { F T } } = \\lambda \\left( \\underbrace { \\mathbb { E } } _ { \\int _ { \\mathrm { I : T : } } r \\sim p _ { F } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) - \\mathcal { L } _ { \\mathrm { R K L } } \\left( \\bar { f } _ { 1 : T ^ { \\prime } } \\right) \\right) + \\underbrace { \\mathbb { E } } _ { f _ { 1 : T ^ { \\prime } \\sim q } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) \\right) _ { \\mathcal { L } _ { \\mathrm { R P } } } \\right) + ( 1 - \\lambda ) \\mathcal { L } _ { \\mathrm { K D } } ,", + "type": "interline_equation", + "image_path": "57b8b0401c44a9da9dffb4a771b588d3f01db84a0e94c59950fcea341c8459e4.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 110, + 360, + 483, + 380.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 110, + 380.0, + 483, + 400.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 110, + 400.0, + 483, + 420.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 504, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 135, + 436 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 423, + 156, + 435 + ], + "score": 0.91, + "content": "{ \\bar { f } } _ { 1 : T ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "is the average fertility computed by Eq. 7. The gradient with respect to the non-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 163, + 447 + ], + "score": 1.0, + "content": "differentiable", + "type": "text" + }, + { + "bbox": [ + 163, + 435, + 180, + 446 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { R L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 433, + 465, + 447 + ], + "score": 1.0, + "content": "term can be estimated with REINFORCE (Williams, 1992), while the", + "type": "text" + }, + { + "bbox": [ + 466, + 434, + 483, + 446 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\mathrm { B P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "term", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 295, + 458 + ], + "score": 1.0, + "content": "can be trained using ordinary backpropagation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 200, + 485 + ], + "lines": [ + { + "bbox": [ + 104, + 471, + 202, + 487 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 202, + 487 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 497, + 245, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 246, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 246, + 510 + ], + "score": 1.0, + "content": "5.1 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Dataset We evaluate the proposed NAT on three widely used public machine translation corpora:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "IWSLT16 En–De2, WMT14 En–De,3 and WMT16 En–Ro4. We use IWSLT—which is smaller than", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "the other two datasets—as the development dataset for ablation experiments, and additionally train", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 557 + ], + "score": 1.0, + "content": "and test our primary models on both directions of both WMT datasets. All the data are tokenized and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "segmented into subword symbols using byte-pair encoding (BPE) (Sennrich et al., 2015) to restrict", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 567, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 504, + 579 + ], + "score": 1.0, + "content": "the size of the vocabulary. For both WMT datasets, we use shared BPE vocabulary and additionally", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "share encoder and decoder word embeddings; for IWSLT, we use separate English and German", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 589, + 223, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 223, + 602 + ], + "score": 1.0, + "content": "vocabulary and embeddings.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 504, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 621 + ], + "score": 1.0, + "content": "Teacher Sequence-level knowledge distillation is applied to alleviate multimodality in the training", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "dataset, using autoregressive models as the teachers. The same teacher model used for distillation is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 629, + 406, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 406, + 643 + ], + "score": 1.0, + "content": "also used as a scoring function for fine-tuning and noisy parallel decoding.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "To enable a fair comparison, and benefit from its high translation quality, we implemented the au-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 658, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 670 + ], + "score": 1.0, + "content": "toregressive teachers using the state-of-the-art Transformer architecture. In addition, we use the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "same sizes and hyperparameters for each student and its respective teacher, with the exception of the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 679, + 387, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 387, + 692 + ], + "score": 1.0, + "content": "newly added positional self-attention and fertility prediction modules.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 699, + 288, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 698, + 194, + 710 + ], + "spans": [ + { + "bbox": [ + 119, + 698, + 194, + 710 + ], + "score": 1.0, + "content": "2https://wit3.fbk.eu/", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 709, + 289, + 722 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 289, + 722 + ], + "score": 1.0, + "content": "3http://www.statmt.org/wmt14/translation-task", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 720, + 289, + 733 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 289, + 733 + ], + "score": 1.0, + "content": "4http://www.statmt.org/wmt16/translation-task", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 127 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 105, + 83, + 505, + 128 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 191, + 151 + ], + "lines": [ + { + "bbox": [ + 105, + 138, + 193, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 193, + 153 + ], + "score": 1.0, + "content": "4.2 FINE-TUNING", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 160, + 505, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 505, + 173 + ], + "score": 1.0, + "content": "Our supervised fertility model enables a decomposition of the overall maximum likelihood loss", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 505, + 185 + ], + "score": 1.0, + "content": "into translation and fertility terms, but it has some drawbacks compared to variational training. In", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "particular, it heavily relies on the deterministic, approximate inference model provided by the ex-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "score": 1.0, + "content": "ternal alignment system, while it would be desirable to train the entire model, including the fertility", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 193, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 193, + 216 + ], + "score": 1.0, + "content": "predictor, end to end.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 161, + 506, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 221, + 504, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 221, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 234 + ], + "score": 1.0, + "content": "Thus we propose a fine-tuning step after training the NAT to convergence. We introduce an addi-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 232, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 245 + ], + "score": 1.0, + "content": "tional loss term consisting of the reverse K-L divergence with the teacher output distribution, a form", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 243, + 256, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 256, + 255 + ], + "score": 1.0, + "content": "of word-level knowledge distillation:", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 221, + 505, + 255 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 257, + 467, + 293 + ], + "lines": [ + { + "bbox": [ + 143, + 257, + 467, + 293 + ], + "spans": [ + { + "bbox": [ + 143, + 257, + 467, + 293 + ], + "score": 0.93, + "content": "\\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } ; \\theta \\right) = \\sum _ { t = 1 } ^ { T } \\sum _ { y _ { t } } \\left[ \\log p _ { \\mathcal { A R } } \\left( y _ { t } | \\hat { y } _ { 1 : t - 1 } , x _ { 1 : T ^ { \\prime } } \\right) \\cdot p _ { \\mathcal { N A } } \\left( y _ { t } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta \\right) \\right] ,", + "type": "interline_equation", + "image_path": "e65993392162bd75d4d5f03dbe09c11504275ad6f8a4d1ecdf2eaf6fe3d240bf.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 143, + 257, + 467, + 269.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 143, + 269.0, + 467, + 281.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 143, + 281.0, + 467, + 293.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 506, + 319 + ], + "lines": [ + { + "bbox": [ + 106, + 296, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 296, + 133, + 309 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 296, + 235, + 308 + ], + "score": 0.92, + "content": "\\hat { y } _ { 1 : T } = G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 296, + 505, + 309 + ], + "score": 1.0, + "content": ". Such a loss is more favorable towards highly peaked student output", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 307, + 341, + 319 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 341, + 319 + ], + "score": 1.0, + "content": "distributions than a standard cross-entropy error would be.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 296, + 505, + 319 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 324, + 505, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 324, + 505, + 336 + ], + "score": 1.0, + "content": "Then we train the whole model jointly with a weighted sum of the original distillation loss and two", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 106, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "such terms, one an expectation over the predicted fertility distribution, normalized with a baseline,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 347, + 349, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 349, + 358 + ], + "score": 1.0, + "content": "and the other based on the external fertility inference model:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 324, + 506, + 358 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 110, + 360, + 483, + 420 + ], + "lines": [ + { + "bbox": [ + 110, + 360, + 483, + 420 + ], + "spans": [ + { + "bbox": [ + 110, + 360, + 483, + 420 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { \\mathrm { F T } } = \\lambda \\left( \\underbrace { \\mathbb { E } } _ { \\int _ { \\mathrm { I : T : } } r \\sim p _ { F } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) - \\mathcal { L } _ { \\mathrm { R K L } } \\left( \\bar { f } _ { 1 : T ^ { \\prime } } \\right) \\right) + \\underbrace { \\mathbb { E } } _ { f _ { 1 : T ^ { \\prime } \\sim q } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) \\right) _ { \\mathcal { L } _ { \\mathrm { R P } } } \\right) + ( 1 - \\lambda ) \\mathcal { L } _ { \\mathrm { K D } } ,", + "type": "interline_equation", + "image_path": "57b8b0401c44a9da9dffb4a771b588d3f01db84a0e94c59950fcea341c8459e4.jpg" + } + ] + } + ], + "index": 22, + "virtual_lines": [ + { + "bbox": [ + 110, + 360, + 483, + 380.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 110, + 380.0, + 483, + 400.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 110, + 400.0, + 483, + 420.0 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 504, + 457 + ], + "lines": [ + { + "bbox": [ + 106, + 423, + 505, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 135, + 436 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 423, + 156, + 435 + ], + "score": 0.91, + "content": "{ \\bar { f } } _ { 1 : T ^ { \\prime } }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 423, + 505, + 436 + ], + "score": 1.0, + "content": "is the average fertility computed by Eq. 7. The gradient with respect to the non-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 433, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 163, + 447 + ], + "score": 1.0, + "content": "differentiable", + "type": "text" + }, + { + "bbox": [ + 163, + 435, + 180, + 446 + ], + "score": 0.9, + "content": "{ \\mathcal { L } } _ { \\mathrm { R L } }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 433, + 465, + 447 + ], + "score": 1.0, + "content": "term can be estimated with REINFORCE (Williams, 1992), while the", + "type": "text" + }, + { + "bbox": [ + 466, + 434, + 483, + 446 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { \\mathrm { B P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 433, + 505, + 447 + ], + "score": 1.0, + "content": "term", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 295, + 458 + ], + "score": 1.0, + "content": "can be trained using ordinary backpropagation.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 423, + 505, + 458 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 200, + 485 + ], + "lines": [ + { + "bbox": [ + 104, + 471, + 202, + 487 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 202, + 487 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 497, + 245, + 509 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 246, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 246, + 510 + ], + "score": 1.0, + "content": "5.1 EXPERIMENTAL SETTINGS", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Dataset We evaluate the proposed NAT on three widely used public machine translation corpora:", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "IWSLT16 En–De2, WMT14 En–De,3 and WMT16 En–Ro4. We use IWSLT—which is smaller than", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 547 + ], + "score": 1.0, + "content": "the other two datasets—as the development dataset for ablation experiments, and additionally train", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 506, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 557 + ], + "score": 1.0, + "content": "and test our primary models on both directions of both WMT datasets. All the data are tokenized and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "segmented into subword symbols using byte-pair encoding (BPE) (Sennrich et al., 2015) to restrict", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 567, + 504, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 504, + 579 + ], + "score": 1.0, + "content": "the size of the vocabulary. For both WMT datasets, we use shared BPE vocabulary and additionally", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 506, + 591 + ], + "score": 1.0, + "content": "share encoder and decoder word embeddings; for IWSLT, we use separate English and German", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 589, + 223, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 223, + 602 + ], + "score": 1.0, + "content": "vocabulary and embeddings.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 512, + 506, + 602 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 504, + 641 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 621 + ], + "score": 1.0, + "content": "Teacher Sequence-level knowledge distillation is applied to alleviate multimodality in the training", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "dataset, using autoregressive models as the teachers. The same teacher model used for distillation is", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 629, + 406, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 406, + 643 + ], + "score": 1.0, + "content": "also used as a scoring function for fine-tuning and noisy parallel decoding.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 606, + 506, + 643 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 505, + 691 + ], + "lines": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 505, + 659 + ], + "score": 1.0, + "content": "To enable a fair comparison, and benefit from its high translation quality, we implemented the au-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 658, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 670 + ], + "score": 1.0, + "content": "toregressive teachers using the state-of-the-art Transformer architecture. In addition, we use the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 669, + 505, + 681 + ], + "score": 1.0, + "content": "same sizes and hyperparameters for each student and its respective teacher, with the exception of the", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 679, + 387, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 387, + 692 + ], + "score": 1.0, + "content": "newly added positional self-attention and fertility prediction modules.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 646, + 505, + 692 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 80, + 507, + 178 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 80, + 507, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 507, + 178 + ], + "score": 0.984, + "html": "
ModelsWMT14WMT16IWSLT16
En→DeDe→EnEn→RoRo→EnEn→DeLatency / Speedup
NAT17.3520.6226.2227.8325.2039 ms15.6×
NAT (+FT)17.6921.4727.2929.0626.5239 ms15.6×
NAT (+FT + NPD s = 10)18.6622.4129.0230.7627.4479 ms7.68×
NAT (+FT + NPD s = 100)19.1723.2029.7931.4428.16257 ms2.36×
Autoregressive (b = 1)22.7126.3931.3531.0328.89408 ms1.49×
Autoregressive (b = 4)23.4527.0231.9131.7629.70607ms1.00×
", + "type": "table", + "image_path": "b27d229f92b1dc501a1101859c66ddaef0b786b8513588c7b5b795ee0a526bb9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 80, + 507, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 112.66666666666666, + 507, + 145.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 145.33333333333331, + 507, + 177.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 186, + 505, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "Table 1: BLEU scores on official test sets (newstest2014 for WMT En-De and newstest2016", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "for WMT En-Ro) or the development set for IWSLT. NAT models without NPD use argmax decod-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "score": 1.0, + "content": "ing. Latency is computed as the time to decode a single sentence without minibatching, averaged", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 219, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 481, + 230 + ], + "score": 1.0, + "content": "over the whole test set; decoding is implemented in PyTorch on a single NVIDIA Tesla P100.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 316, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 317, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 317, + 254 + ], + "score": 1.0, + "content": "Preparation for knowledge distillation We first", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 253, + 316, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 316, + 265 + ], + "score": 1.0, + "content": "train all teacher models using maximum likelihood,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 265, + 317, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 317, + 277 + ], + "score": 1.0, + "content": "then freeze their parameters. To avoid the redun-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 275, + 317, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 317, + 288 + ], + "score": 1.0, + "content": "dancy of running fixed teacher models repeatedly on", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 286, + 317, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 317, + 299 + ], + "score": 1.0, + "content": "the same data, we decode the entire training set once", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 298, + 317, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 317, + 310 + ], + "score": 1.0, + "content": "using each teacher to create a new training dataset", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 309, + 208, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 208, + 321 + ], + "score": 1.0, + "content": "for its respective student.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 316, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 329, + 316, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 316, + 340 + ], + "score": 1.0, + "content": "Encoder initialization We find it helpful to initial-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 341, + 316, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 316, + 352 + ], + "score": 1.0, + "content": "ize the weights in the NAT student’s encoder with the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 351, + 317, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 317, + 363 + ], + "score": 1.0, + "content": "encoder weights from its teacher, as the autoregres-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 363, + 317, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 317, + 374 + ], + "score": 1.0, + "content": "sive and non-autoregressive models share the same", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 374, + 232, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 232, + 385 + ], + "score": 1.0, + "content": "encoder input and architecture.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 316, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 317, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 268, + 407 + ], + "score": 1.0, + "content": "Fertility supervision during training", + "type": "text" + }, + { + "bbox": [ + 285, + 394, + 317, + 406 + ], + "score": 1.0, + "content": "As de-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 406, + 317, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 317, + 417 + ], + "score": 1.0, + "content": "scribed above, we supervise the fertility predictions", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 416, + 316, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 316, + 428 + ], + "score": 1.0, + "content": "at train time by using a fixed aligner as a fertility", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 426, + 318, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 318, + 439 + ], + "score": 1.0, + "content": "inference function. We use the fast align5 im-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 438, + 317, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 317, + 450 + ], + "score": 1.0, + "content": "plementation of IBM Model 2 for this purpose, with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 449, + 261, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 261, + 461 + ], + "score": 1.0, + "content": "default parameters (Dyer et al., 2013).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5 + }, + { + "type": "image", + "bbox": [ + 328, + 238, + 487, + 344 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 328, + 238, + 487, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 328, + 238, + 487, + 344 + ], + "spans": [ + { + "bbox": [ + 328, + 238, + 487, + 344 + ], + "score": 0.97, + "type": "image", + "image_path": "ff2cfd4d2f25ef6e98d968d65066c8e02ea0f947a91d09952aa8c8ffbab3302a.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 328, + 238, + 487, + 251.25 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 328, + 251.25, + 487, + 264.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 328, + 264.5, + 487, + 277.75 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 328, + 277.75, + 487, + 291.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 328, + 291.0, + 487, + 304.25 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 328, + 304.25, + 487, + 317.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 328, + 317.5, + 487, + 330.75 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 328, + 330.75, + 487, + 344.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 356, + 505, + 433 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 322, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 322, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "Figure 3: BLEU scores on IWSLT de-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 324, + 368, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 324, + 368, + 505, + 379 + ], + "score": 1.0, + "content": "velopment set as a function of sample size", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 323, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 323, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "for noisy parallel decoding. NPD matches", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 323, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 323, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "the performance of the other two decoding", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 324, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 324, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "strategies after two samples, and exceeds the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 323, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 323, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "performance of the autoregressive teacher", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 324, + 422, + 399, + 433 + ], + "spans": [ + { + "bbox": [ + 324, + 422, + 399, + 433 + ], + "score": 1.0, + "content": "with around 1000.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + } + ], + "index": 32.25 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "Hyperparameters For experiments on WMT datasets, we use the hyperparameter settings of the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 480, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 494 + ], + "score": 1.0, + "content": "base Transformer model described in Vaswani et al. (2017), though without label smoothing. As", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "IWSLT is a smaller corpus, and to reduce training time, we use a set of smaller hyperparameters", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 108, + 501, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 108, + 503, + 309, + 514 + ], + "score": 0.45, + "content": "( d _ { \\mathrm { m o d e l } } = 2 8 7 , d _ { \\mathrm { h i d d e n } } = 5 0 7 , n _ { \\mathrm { l a y e r } } = 5 , n _ { \\mathrm { h e a d } } = \\mathrm { \\Omega } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 501, + 333, + 517 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 333, + 503, + 393, + 515 + ], + "score": 0.88, + "content": "t _ { \\mathrm { w a r m u p } } = 7 4 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 501, + 506, + 517 + ], + "score": 1.0, + "content": "for all experiments on that", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 513, + 274, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 232, + 526 + ], + "score": 1.0, + "content": "dataset. For fine-tuning we use", + "type": "text" + }, + { + "bbox": [ + 232, + 515, + 270, + 524 + ], + "score": 0.87, + "content": "\\lambda = 0 . 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 513, + 274, + 526 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 108, + 534, + 502, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 503, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 503, + 547 + ], + "score": 1.0, + "content": "Evaluation metrics We evaluate using tokenized and cased BLEU scores (Papineni et al., 2002).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 447, + 567 + ], + "lines": [ + { + "bbox": [ + 104, + 553, + 448, + 569 + ], + "spans": [ + { + "bbox": [ + 104, + 553, + 419, + 569 + ], + "score": 1.0, + "content": "Implementation We have open-sourced our PyTorch implementation of the", + "type": "text" + }, + { + "bbox": [ + 419, + 555, + 443, + 566 + ], + "score": 0.62, + "content": "\\mathrm { N A T ^ { 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 553, + 448, + 569 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46 + }, + { + "type": "title", + "bbox": [ + 108, + 582, + 170, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 172, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 172, + 596 + ], + "score": 1.0, + "content": "5.2 RESULTS", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 108, + 604, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "Across the three datasets we used, the NAT performs between 2-5 BLEU points worse than its", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "autoregressive teacher, with part or all of this gap addressed by the use of noisy parallel decoding. In", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "the case of WMT16 English–Romanian, NPD improves the performance of our non-autoregressive", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 637, + 483, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 483, + 649 + ], + "score": 1.0, + "content": "model to within 0.2 BLEU points of the previous overall state of the art (Gehring et al., 2017).", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "Comparing latencies on the development model shows a speedup of more than a factor of 10 over", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "greedy autoregressive decoding, or a factor of 15 over beam search. Latencies for decoding with", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 349, + 688 + ], + "score": 1.0, + "content": "NPD, regardless of sample size, could be reduced to about", + "type": "text" + }, + { + "bbox": [ + 350, + 676, + 373, + 686 + ], + "score": 0.39, + "content": "8 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "by parallelizing across multiple", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 686, + 466, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 466, + 700 + ], + "score": 1.0, + "content": "GPUs because each sample can be generated, then scored, independently from the others.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 710, + 276, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 709, + 244, + 721 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 244, + 721 + ], + "score": 1.0, + "content": "5https://github.com/clab/fast align", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 720, + 277, + 732 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 277, + 732 + ], + "score": 1.0, + "content": "6https://github.com/salesforce/nonauto-nmt", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 80, + 507, + 178 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 80, + 507, + 178 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 80, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 80, + 507, + 178 + ], + "score": 0.984, + "html": "
ModelsWMT14WMT16IWSLT16
En→DeDe→EnEn→RoRo→EnEn→DeLatency / Speedup
NAT17.3520.6226.2227.8325.2039 ms15.6×
NAT (+FT)17.6921.4727.2929.0626.5239 ms15.6×
NAT (+FT + NPD s = 10)18.6622.4129.0230.7627.4479 ms7.68×
NAT (+FT + NPD s = 100)19.1723.2029.7931.4428.16257 ms2.36×
Autoregressive (b = 1)22.7126.3931.3531.0328.89408 ms1.49×
Autoregressive (b = 4)23.4527.0231.9131.7629.70607ms1.00×
", + "type": "table", + "image_path": "b27d229f92b1dc501a1101859c66ddaef0b786b8513588c7b5b795ee0a526bb9.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 80, + 507, + 112.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 112.66666666666666, + 507, + 145.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 145.33333333333331, + 507, + 177.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 186, + 505, + 230 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 198 + ], + "score": 1.0, + "content": "Table 1: BLEU scores on official test sets (newstest2014 for WMT En-De and newstest2016", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 197, + 505, + 209 + ], + "score": 1.0, + "content": "for WMT En-Ro) or the development set for IWSLT. NAT models without NPD use argmax decod-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 506, + 221 + ], + "score": 1.0, + "content": "ing. Latency is computed as the time to decode a single sentence without minibatching, averaged", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 219, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 481, + 230 + ], + "score": 1.0, + "content": "over the whole test set; decoding is implemented in PyTorch on a single NVIDIA Tesla P100.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 186, + 506, + 230 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 243, + 316, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 317, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 317, + 254 + ], + "score": 1.0, + "content": "Preparation for knowledge distillation We first", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 253, + 316, + 265 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 316, + 265 + ], + "score": 1.0, + "content": "train all teacher models using maximum likelihood,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 265, + 317, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 317, + 277 + ], + "score": 1.0, + "content": "then freeze their parameters. To avoid the redun-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 275, + 317, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 317, + 288 + ], + "score": 1.0, + "content": "dancy of running fixed teacher models repeatedly on", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 286, + 317, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 317, + 299 + ], + "score": 1.0, + "content": "the same data, we decode the entire training set once", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 298, + 317, + 310 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 317, + 310 + ], + "score": 1.0, + "content": "using each teacher to create a new training dataset", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 309, + 208, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 208, + 321 + ], + "score": 1.0, + "content": "for its respective student.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 243, + 317, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 330, + 316, + 384 + ], + "lines": [ + { + "bbox": [ + 106, + 329, + 316, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 316, + 340 + ], + "score": 1.0, + "content": "Encoder initialization We find it helpful to initial-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 341, + 316, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 316, + 352 + ], + "score": 1.0, + "content": "ize the weights in the NAT student’s encoder with the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 351, + 317, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 317, + 363 + ], + "score": 1.0, + "content": "encoder weights from its teacher, as the autoregres-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 363, + 317, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 317, + 374 + ], + "score": 1.0, + "content": "sive and non-autoregressive models share the same", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 374, + 232, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 374, + 232, + 385 + ], + "score": 1.0, + "content": "encoder input and architecture.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 106, + 329, + 317, + 385 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 394, + 316, + 460 + ], + "lines": [ + { + "bbox": [ + 106, + 394, + 317, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 268, + 407 + ], + "score": 1.0, + "content": "Fertility supervision during training", + "type": "text" + }, + { + "bbox": [ + 285, + 394, + 317, + 406 + ], + "score": 1.0, + "content": "As de-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 406, + 317, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 317, + 417 + ], + "score": 1.0, + "content": "scribed above, we supervise the fertility predictions", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 416, + 316, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 316, + 428 + ], + "score": 1.0, + "content": "at train time by using a fixed aligner as a fertility", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 426, + 318, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 318, + 439 + ], + "score": 1.0, + "content": "inference function. We use the fast align5 im-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 438, + 317, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 317, + 450 + ], + "score": 1.0, + "content": "plementation of IBM Model 2 for this purpose, with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 449, + 261, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 449, + 261, + 461 + ], + "score": 1.0, + "content": "default parameters (Dyer et al., 2013).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 394, + 318, + 461 + ] + }, + { + "type": "image", + "bbox": [ + 328, + 238, + 487, + 344 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 328, + 238, + 487, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 328, + 238, + 487, + 344 + ], + "spans": [ + { + "bbox": [ + 328, + 238, + 487, + 344 + ], + "score": 0.97, + "type": "image", + "image_path": "ff2cfd4d2f25ef6e98d968d65066c8e02ea0f947a91d09952aa8c8ffbab3302a.jpg" + } + ] + } + ], + "index": 28.5, + "virtual_lines": [ + { + "bbox": [ + 328, + 238, + 487, + 251.25 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 328, + 251.25, + 487, + 264.5 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 328, + 264.5, + 487, + 277.75 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 328, + 277.75, + 487, + 291.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 328, + 291.0, + 487, + 304.25 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 328, + 304.25, + 487, + 317.5 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 328, + 317.5, + 487, + 330.75 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 328, + 330.75, + 487, + 344.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 324, + 356, + 505, + 433 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 322, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 322, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "Figure 3: BLEU scores on IWSLT de-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 324, + 368, + 505, + 379 + ], + "spans": [ + { + "bbox": [ + 324, + 368, + 505, + 379 + ], + "score": 1.0, + "content": "velopment set as a function of sample size", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 323, + 378, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 323, + 378, + 505, + 390 + ], + "score": 1.0, + "content": "for noisy parallel decoding. NPD matches", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 323, + 389, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 323, + 389, + 505, + 402 + ], + "score": 1.0, + "content": "the performance of the other two decoding", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 324, + 400, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 324, + 400, + 505, + 412 + ], + "score": 1.0, + "content": "strategies after two samples, and exceeds the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 323, + 411, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 323, + 411, + 505, + 423 + ], + "score": 1.0, + "content": "performance of the autoregressive teacher", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 324, + 422, + 399, + 433 + ], + "spans": [ + { + "bbox": [ + 324, + 422, + 399, + 433 + ], + "score": 1.0, + "content": "with around 1000.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 36 + } + ], + "index": 32.25 + }, + { + "type": "text", + "bbox": [ + 107, + 470, + 505, + 525 + ], + "lines": [ + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "Hyperparameters For experiments on WMT datasets, we use the hyperparameter settings of the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 480, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 505, + 494 + ], + "score": 1.0, + "content": "base Transformer model described in Vaswani et al. (2017), though without label smoothing. As", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 505, + 505 + ], + "score": 1.0, + "content": "IWSLT is a smaller corpus, and to reduce training time, we use a set of smaller hyperparameters", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 108, + 501, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 108, + 503, + 309, + 514 + ], + "score": 0.45, + "content": "( d _ { \\mathrm { m o d e l } } = 2 8 7 , d _ { \\mathrm { h i d d e n } } = 5 0 7 , n _ { \\mathrm { l a y e r } } = 5 , n _ { \\mathrm { h e a d } } = \\mathrm { \\Omega } ^ { \\ast }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 501, + 333, + 517 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 333, + 503, + 393, + 515 + ], + "score": 0.88, + "content": "t _ { \\mathrm { w a r m u p } } = 7 4 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 501, + 506, + 517 + ], + "score": 1.0, + "content": "for all experiments on that", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 513, + 274, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 232, + 526 + ], + "score": 1.0, + "content": "dataset. For fine-tuning we use", + "type": "text" + }, + { + "bbox": [ + 232, + 515, + 270, + 524 + ], + "score": 0.87, + "content": "\\lambda = 0 . 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 513, + 274, + 526 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 469, + 506, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 534, + 502, + 546 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 503, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 503, + 547 + ], + "score": 1.0, + "content": "Evaluation metrics We evaluate using tokenized and cased BLEU scores (Papineni et al., 2002).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 533, + 503, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 447, + 567 + ], + "lines": [ + { + "bbox": [ + 104, + 553, + 448, + 569 + ], + "spans": [ + { + "bbox": [ + 104, + 553, + 419, + 569 + ], + "score": 1.0, + "content": "Implementation We have open-sourced our PyTorch implementation of the", + "type": "text" + }, + { + "bbox": [ + 419, + 555, + 443, + 566 + ], + "score": 0.62, + "content": "\\mathrm { N A T ^ { 6 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 553, + 448, + 569 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 46, + "bbox_fs": [ + 104, + 553, + 448, + 569 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 582, + 170, + 594 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 172, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 172, + 596 + ], + "score": 1.0, + "content": "5.2 RESULTS", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 47 + }, + { + "type": "text", + "bbox": [ + 108, + 604, + 505, + 648 + ], + "lines": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "Across the three datasets we used, the NAT performs between 2-5 BLEU points worse than its", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 505, + 627 + ], + "score": 1.0, + "content": "autoregressive teacher, with part or all of this gap addressed by the use of noisy parallel decoding. In", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "the case of WMT16 English–Romanian, NPD improves the performance of our non-autoregressive", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 637, + 483, + 649 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 483, + 649 + ], + "score": 1.0, + "content": "model to within 0.2 BLEU points of the previous overall state of the art (Gehring et al., 2017).", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 49.5, + "bbox_fs": [ + 105, + 604, + 506, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 666 + ], + "score": 1.0, + "content": "Comparing latencies on the development model shows a speedup of more than a factor of 10 over", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "greedy autoregressive decoding, or a factor of 15 over beam search. Latencies for decoding with", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 676, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 349, + 688 + ], + "score": 1.0, + "content": "NPD, regardless of sample size, could be reduced to about", + "type": "text" + }, + { + "bbox": [ + 350, + 676, + 373, + 686 + ], + "score": 0.39, + "content": "8 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 676, + 505, + 688 + ], + "score": 1.0, + "content": "by parallelizing across multiple", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 686, + 466, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 466, + 700 + ], + "score": 1.0, + "content": "GPUs because each sample can be generated, then scored, independently from the others.", + "type": "text" + } + ], + "index": 55 + } + ], + "index": 53.5, + "bbox_fs": [ + 105, + 654, + 505, + 700 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 209, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 210, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 210, + 95 + ], + "score": 1.0, + "content": "5.3 ABLATION STUDY", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "We also conduct an extensive ablation study with the proposed NAT on the IWSLT dataset. First,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "we note that the model fails to train when provided with only positional embeddings as input to the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "decoder. Second, we see that training on the distillation corpus rather than the ground truth provides", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 505, + 151 + ], + "score": 1.0, + "content": "a fairly consistent improvement of around 5 BLEU points. Third, switching from uniform copying", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 506, + 162 + ], + "score": 1.0, + "content": "of source inputs to fertility-based copying improves performance by four BLEU points when using", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 317, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 317, + 171 + ], + "score": 1.0, + "content": "ground-truth training or two when using distillation.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 122, + 181, + 488, + 343 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 122, + 181, + 488, + 343 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 181, + 488, + 343 + ], + "spans": [ + { + "bbox": [ + 122, + 181, + 488, + 343 + ], + "score": 0.963, + "html": "
Distillation b=1 b=4Decoder Inputs +uniform+fertility+PosAttFine-tuning +LKD +LBPBLEU +CRLBLEU (T)
~2
√16.51
√18.87
√√√20.72
√√21.12
<√24.0243.91
√25.2045.41
√√22.44 厂
兴 √√××
√√××
√√√√25.7646.11
√√√√26.5247.38
", + "type": "table", + "image_path": "e54d8b7021cfac76f841f708bf7c607fe1294c5aa3739137af72c28cad0aff7f.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 122, + 181, + 488, + 235.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 122, + 235.0, + 488, + 289.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 122, + 289.0, + 488, + 343.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "Table 2: Ablation performance on the IWSLT development set. BLEU (T) refers to the BLEU score", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 456, + 375 + ], + "score": 1.0, + "content": "on a version of the development set that has been translated by the teacher model. An", + "type": "text" + }, + { + "bbox": [ + 456, + 363, + 466, + 373 + ], + "score": 0.74, + "content": "\\times", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "that fine-tuning caused that model to get worse. When uniform copying is used as the decoder", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 384, + 451, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 451, + 397 + ], + "score": 1.0, + "content": "inputs, the ground-truth target lengths are provided. All models use argmax decoding.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 407, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 410, + 420 + ], + "score": 1.0, + "content": "Fine-tuning does not converge with reinforcement learning alone, or with the", + "type": "text" + }, + { + "bbox": [ + 410, + 408, + 428, + 419 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { \\mathrm { B P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "term alone, but use", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 417, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 433 + ], + "score": 1.0, + "content": "of all three fine-tuning terms together leads to an improvement of around 1.5 BLEU points. Training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 429, + 504, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 504, + 442 + ], + "score": 1.0, + "content": "the student model from a distillation corpus produced using beam search is similar to training from", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 440, + 223, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 223, + 455 + ], + "score": 1.0, + "content": "the greedily-distilled corpus.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "table", + "bbox": [ + 113, + 473, + 493, + 635 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 113, + 473, + 493, + 635 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 473, + 493, + 635 + ], + "spans": [ + { + "bbox": [ + 113, + 473, + 493, + 635 + ], + "score": 0.602, + "html": "
Source:politicians try to pick wordsand use words to shape realityand control reality,butin fact,reality changes words
far more than words can ever change reality.
Target: AR:Politiker versuchen Worte zu benutzen,um die Realitat zu formen und die Realitat zu kontrolieren,aber
tatsächlich verandert die Realitat Worte viel mehr,als Worte die Realitat jemals verändern konnten.
Politikerversuchen Worter zu wahlen und Worter zur Realitat zu gestalten und Realitätzu steuern,aber in
NAT:Wirklichkeit verändert sich die Realitat viel mehrals Worte,die die Realität verändern konnen.
Politikerversuchen,Worter wahlen und zu verwenden,um Realitatzu formen und Realitatzu formen,aber
NAT+NPD:tatsächlich ändert Realitat Realitat viel mehrals Wortedie Realität Realitat verändern.
Politikerversuchen,Worter wahlenund zu verwenden,um Realitat Realitat formenund die Realitatzu formen， aber tatsächlich ändert die Realität Worte viel mehrals Worte jemals die Realität verändern konnen.
Source:Isee wheelchairs bought and sold like used cars.
Target:ich erlebe,dass Rollstuhle gekauft und verkauft werden wie Gebrauchtwagen
AR: NAT:ich sehe Rollstuhlen,die wie Autos verkauft und verkauft werden.
NAT+NPD:ich sehe,dass Stuhle Stuhle und verkauftwie Autos verkauft.
ich sehe Rollühle kauften und verkaufte wie Autos.
", + "type": "table", + "image_path": "f8421f544ec54c247fda2ccf5acde08a1d896cc7e4be518f611e266b0beaaf82.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 113, + 473, + 493, + 527.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 113, + 527.0, + 493, + 581.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 113, + 581.0, + 493, + 635.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 653, + 505, + 687 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 653, + 504, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 504, + 666 + ], + "score": 1.0, + "content": "Figure 4: Two examples comparing translations produced by an autoregressive (AR) and non-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 664, + 504, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 504, + 676 + ], + "score": 1.0, + "content": "autoregressive Transformer as well as the result of noisy parallel decoding with sample size 100.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 673, + 268, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 268, + 689 + ], + "score": 1.0, + "content": "Repeated words are highlighted in gray.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "We include two examples of translations from the IWSLT development set in Fig. 4. Instances of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "repeated words or phrases, highlighted in gray, are most prevalent in the non-autoregressive output", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "for the relatively complex first example sentence. Two pairs of repeated words in the first example, as", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 209, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 210, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 210, + 95 + ], + "score": 1.0, + "content": "5.3 ABLATION STUDY", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 505, + 170 + ], + "lines": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 505, + 116 + ], + "score": 1.0, + "content": "We also conduct an extensive ablation study with the proposed NAT on the IWSLT dataset. First,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 505, + 127 + ], + "score": 1.0, + "content": "we note that the model fails to train when provided with only positional embeddings as input to the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 137 + ], + "score": 1.0, + "content": "decoder. Second, we see that training on the distillation corpus rather than the ground truth provides", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 134, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 505, + 151 + ], + "score": 1.0, + "content": "a fairly consistent improvement of around 5 BLEU points. Third, switching from uniform copying", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 145, + 506, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 506, + 162 + ], + "score": 1.0, + "content": "of source inputs to fertility-based copying improves performance by four BLEU points when using", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 317, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 317, + 171 + ], + "score": 1.0, + "content": "ground-truth training or two when using distillation.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 102, + 506, + 171 + ] + }, + { + "type": "table", + "bbox": [ + 122, + 181, + 488, + 343 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 122, + 181, + 488, + 343 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 181, + 488, + 343 + ], + "spans": [ + { + "bbox": [ + 122, + 181, + 488, + 343 + ], + "score": 0.963, + "html": "
Distillation b=1 b=4Decoder Inputs +uniform+fertility+PosAttFine-tuning +LKD +LBPBLEU +CRLBLEU (T)
~2
√16.51
√18.87
√√√20.72
√√21.12
<√24.0243.91
√25.2045.41
√√22.44 厂
兴 √√××
√√××
√√√√25.7646.11
√√√√26.5247.38
", + "type": "table", + "image_path": "e54d8b7021cfac76f841f708bf7c607fe1294c5aa3739137af72c28cad0aff7f.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 122, + 181, + 488, + 235.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 122, + 235.0, + 488, + 289.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 122, + 289.0, + 488, + 343.0 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 364 + ], + "score": 1.0, + "content": "Table 2: Ablation performance on the IWSLT development set. BLEU (T) refers to the BLEU score", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 362, + 506, + 375 + ], + "spans": [ + { + "bbox": [ + 104, + 362, + 456, + 375 + ], + "score": 1.0, + "content": "on a version of the development set that has been translated by the teacher model. An", + "type": "text" + }, + { + "bbox": [ + 456, + 363, + 466, + 373 + ], + "score": 0.74, + "content": "\\times", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 362, + 506, + 375 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "that fine-tuning caused that model to get worse. When uniform copying is used as the decoder", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 384, + 451, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 451, + 397 + ], + "score": 1.0, + "content": "inputs, the ground-truth target lengths are provided. All models use argmax decoding.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 104, + 351, + 506, + 397 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 407, + 505, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 410, + 420 + ], + "score": 1.0, + "content": "Fine-tuning does not converge with reinforcement learning alone, or with the", + "type": "text" + }, + { + "bbox": [ + 410, + 408, + 428, + 419 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { \\mathrm { B P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "term alone, but use", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 417, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 417, + 505, + 433 + ], + "score": 1.0, + "content": "of all three fine-tuning terms together leads to an improvement of around 1.5 BLEU points. Training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 429, + 504, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 504, + 442 + ], + "score": 1.0, + "content": "the student model from a distillation corpus produced using beam search is similar to training from", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 440, + 223, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 223, + 455 + ], + "score": 1.0, + "content": "the greedily-distilled corpus.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 407, + 505, + 455 + ] + }, + { + "type": "table", + "bbox": [ + 113, + 473, + 493, + 635 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 113, + 473, + 493, + 635 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 113, + 473, + 493, + 635 + ], + "spans": [ + { + "bbox": [ + 113, + 473, + 493, + 635 + ], + "score": 0.602, + "html": "
Source:politicians try to pick wordsand use words to shape realityand control reality,butin fact,reality changes words
far more than words can ever change reality.
Target: AR:Politiker versuchen Worte zu benutzen,um die Realitat zu formen und die Realitat zu kontrolieren,aber
tatsächlich verandert die Realitat Worte viel mehr,als Worte die Realitat jemals verändern konnten.
Politikerversuchen Worter zu wahlen und Worter zur Realitat zu gestalten und Realitätzu steuern,aber in
NAT:Wirklichkeit verändert sich die Realitat viel mehrals Worte,die die Realität verändern konnen.
Politikerversuchen,Worter wahlen und zu verwenden,um Realitatzu formen und Realitatzu formen,aber
NAT+NPD:tatsächlich ändert Realitat Realitat viel mehrals Wortedie Realität Realitat verändern.
Politikerversuchen,Worter wahlenund zu verwenden,um Realitat Realitat formenund die Realitatzu formen， aber tatsächlich ändert die Realität Worte viel mehrals Worte jemals die Realität verändern konnen.
Source:Isee wheelchairs bought and sold like used cars.
Target:ich erlebe,dass Rollstuhle gekauft und verkauft werden wie Gebrauchtwagen
AR: NAT:ich sehe Rollstuhlen,die wie Autos verkauft und verkauft werden.
NAT+NPD:ich sehe,dass Stuhle Stuhle und verkauftwie Autos verkauft.
ich sehe Rollühle kauften und verkaufte wie Autos.
", + "type": "table", + "image_path": "f8421f544ec54c247fda2ccf5acde08a1d896cc7e4be518f611e266b0beaaf82.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 113, + 473, + 493, + 527.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 113, + 527.0, + 493, + 581.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 113, + 581.0, + 493, + 635.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 106, + 653, + 505, + 687 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 653, + 504, + 666 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 504, + 666 + ], + "score": 1.0, + "content": "Figure 4: Two examples comparing translations produced by an autoregressive (AR) and non-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 664, + 504, + 676 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 504, + 676 + ], + "score": 1.0, + "content": "autoregressive Transformer as well as the result of noisy parallel decoding with sample size 100.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 673, + 268, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 268, + 689 + ], + "score": 1.0, + "content": "Repeated words are highlighted in gray.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "We include two examples of translations from the IWSLT development set in Fig. 4. Instances of", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "repeated words or phrases, highlighted in gray, are most prevalent in the non-autoregressive output", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "for the relatively complex first example sentence. Two pairs of repeated words in the first example, as", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 699, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 129, + 85, + 262, + 95 + ], + "lines": [ + { + "bbox": [ + 128, + 84, + 263, + 95 + ], + "spans": [ + { + "bbox": [ + 128, + 84, + 263, + 95 + ], + "score": 0.96, + "content": " se lucreaza la soluti de genul acesta .", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 127, + 105, + 302, + 202 + ], + "lines": [ + { + "bbox": [ + 128, + 105, + 230, + 115 + ], + "spans": [ + { + "bbox": [ + 128, + 105, + 230, + 115 + ], + "score": 0.981, + "content": "se la solutii de genul acesta .", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 128, + 118, + 240, + 127 + ], + "spans": [ + { + "bbox": [ + 128, + 118, + 240, + 127 + ], + "score": 0.981, + "content": " se lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 130, + 253, + 140 + ], + "spans": [ + { + "bbox": [ + 128, + 130, + 253, + 140 + ], + "score": 0.977, + "content": " se lucreaza solutii de genul acesta .", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 143, + 251, + 152 + ], + "spans": [ + { + "bbox": [ + 129, + 143, + 251, + 152 + ], + "score": 0.983, + "content": "se se lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 128, + 155, + 272, + 164 + ], + "spans": [ + { + "bbox": [ + 128, + 155, + 272, + 164 + ], + "score": 0.972, + "content": " se lucreaza lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 127, + 168, + 282, + 177 + ], + "spans": [ + { + "bbox": [ + 127, + 168, + 282, + 177 + ], + "score": 0.976, + "content": " se se lucreaza lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 128, + 180, + 293, + 189 + ], + "spans": [ + { + "bbox": [ + 128, + 180, + 293, + 189 + ], + "score": 0.989, + "content": "se se lucreaza lucreaza la solutii de de acesta .", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 192, + 303, + 202 + ], + "spans": [ + { + "bbox": [ + 128, + 192, + 303, + 202 + ], + "score": 0.994, + "content": "se se lucreaza lucreaza la solutii de genul acesta .", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 8.0 + }, + { + "type": "text", + "bbox": [ + 319, + 105, + 484, + 203 + ], + "lines": [ + { + "bbox": [ + 320, + 106, + 431, + 114 + ], + "spans": [ + { + "bbox": [ + 320, + 106, + 431, + 114 + ], + "score": 0.98, + "content": "solutions on this kind are done.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 319, + 118, + 433, + 127 + ], + "spans": [ + { + "bbox": [ + 319, + 118, + 433, + 127 + ], + "score": 0.981, + "content": "work done on solutions like this .", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 319, + 130, + 426, + 140 + ], + "spans": [ + { + "bbox": [ + 319, + 130, + 426, + 140 + ], + "score": 0.986, + "content": "solutions on this kind is done .", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 319, + 143, + 441, + 152 + ], + "spans": [ + { + "bbox": [ + 319, + 143, + 441, + 152 + ], + "score": 0.971, + "content": "work is done on solutions like this .", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 319, + 155, + 441, + 164 + ], + "spans": [ + { + "bbox": [ + 319, + 155, + 441, + 164 + ], + "score": 0.985, + "content": "work is done on solutions like this .", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 168, + 476, + 177 + ], + "spans": [ + { + "bbox": [ + 320, + 168, + 476, + 177 + ], + "score": 0.98, + "content": "work is being done on solutions like this .", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 180, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 320, + 180, + 480, + 190 + ], + "score": 0.996, + "content": "work is being done on solutions such as this.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 192, + 485, + 202 + ], + "spans": [ + { + "bbox": [ + 320, + 192, + 485, + 202 + ], + "score": 0.993, + "content": "work is being done on solutions such this kind .", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 9.0 + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 506, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 219, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 231 + ], + "score": 1.0, + "content": "Figure 5: A Romanian–English example translated with noisy parallel decoding. At left are eight", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 229, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 506, + 243 + ], + "score": 1.0, + "content": "sampled fertility sequences from the encoder, represented with their corresponding decoder input", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 240, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 254 + ], + "score": 1.0, + "content": "sequences. Each of these values for the latent variable leads to a different possible output translation,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "shown at right. The autoregressive Transformer then picks the best translation, shown in red, a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 262, + 385, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 385, + 276 + ], + "score": 1.0, + "content": "process which is much faster than directly using it to generate output.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 309 + ], + "score": 1.0, + "content": "well as a pair in the second, are not present in the versions with noisy parallel decoding, suggesting", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 306, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 319 + ], + "score": 1.0, + "content": "that NPD scoring using the teacher model can filter out such mistakes. The translations produced", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "by the NAT with NPD, while of a similar quality to those produced by the autoregressive model, are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 329, + 221, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 221, + 340 + ], + "score": 1.0, + "content": "also noticeably more literal.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 504, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 504, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 504, + 358 + ], + "score": 1.0, + "content": "We also show an example of the noisy parallel decoding process in Fig. 5, demonstrating the diver-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 356, + 402, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 402, + 369 + ], + "score": 1.0, + "content": "sity of translations that can be found by sampling from the fertility space.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 108, + 384, + 195, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 197, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 197, + 400 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 410, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "We introduce a latent variable model for non-autoregressive machine translation that enables a de-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "coder based on Vaswani et al. (2017) to take full advantage of its exceptional degree of internal", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "parallelism even at inference time. As a result, we measure translation latencies of one-tenth that of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 443, + 438, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 438, + 456 + ], + "score": 1.0, + "content": "an equal-sized autoregressive model, while maintaining competitive BLEU scores.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5 + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 175, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 472, + 176, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 176, + 485 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 491, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 501, + 303, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 303, + 513 + ], + "score": 1.0, + "content": "learning to align and translate. In ICLR, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 506, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "score": 1.0, + "content": "Peter Brown, Vincent della Pietra, Stephen della Pietra, and Robert Mercer. The mathematics of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263–311,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 543, + 142, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 543, + 142, + 556 + ], + "score": 1.0, + "content": "1993.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 503, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 504, + 576 + ], + "score": 1.0, + "content": "Kyunghyun Cho. Noisy parallel approximate decoding for conditional recurrent language model.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 575, + 279, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 279, + 587 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1605.03835, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5 + }, + { + "type": "text", + "bbox": [ + 108, + 595, + 504, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "Chris Dyer, Victor Chahuneau, and Noah Smith. A simple, fast, and effective reparameterization of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 606, + 249, + 618 + ], + "spans": [ + { + "bbox": [ + 116, + 606, + 249, + 618 + ], + "score": 1.0, + "content": "IBM Model 2. In NAACL, 2013.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann Dauphin. Convolutional", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 638, + 406, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 406, + 649 + ], + "score": 1.0, + "content": "sequence to sequence learning. arXiv preprint arXiv:1705.03122, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 105, + 658, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Łukasz Kaiser, Aidan Gomez, and Franc¸ois Chollet. Depthwise separable convolutions for neural", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 669, + 363, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 363, + 681 + ], + "score": 1.0, + "content": "machine translation. arXiv preprint arXiv:1706.03059, 2017.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 506, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 703 + ], + "score": 1.0, + "content": "Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 116, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 116, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "Kavukc¸uoglu. Neural machine translation in linear time. ˇ arXiv preprint arXiv:1610.10099, 2016.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5 + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 468, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 469, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 469, + 733 + ], + "score": 1.0, + "content": "Yoon Kim and Alexander Rush. Sequence-level knowledge distillation. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 301, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 129, + 85, + 262, + 95 + ], + "lines": [ + { + "bbox": [ + 128, + 84, + 263, + 95 + ], + "spans": [ + { + "bbox": [ + 128, + 84, + 263, + 95 + ], + "score": 0.96, + "content": " se lucreaza la soluti de genul acesta .", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 128, + 84, + 263, + 95 + ] + }, + { + "type": "list", + "bbox": [ + 127, + 105, + 302, + 202 + ], + "lines": [ + { + "bbox": [ + 128, + 105, + 230, + 115 + ], + "spans": [ + { + "bbox": [ + 128, + 105, + 230, + 115 + ], + "score": 0.981, + "content": "se la solutii de genul acesta .", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 118, + 240, + 127 + ], + "spans": [ + { + "bbox": [ + 128, + 118, + 240, + 127 + ], + "score": 0.981, + "content": " se lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 130, + 253, + 140 + ], + "spans": [ + { + "bbox": [ + 128, + 130, + 253, + 140 + ], + "score": 0.977, + "content": " se lucreaza solutii de genul acesta .", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 143, + 251, + 152 + ], + "spans": [ + { + "bbox": [ + 129, + 143, + 251, + 152 + ], + "score": 0.983, + "content": "se se lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 155, + 272, + 164 + ], + "spans": [ + { + "bbox": [ + 128, + 155, + 272, + 164 + ], + "score": 0.972, + "content": " se lucreaza lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 127, + 168, + 282, + 177 + ], + "spans": [ + { + "bbox": [ + 127, + 168, + 282, + 177 + ], + "score": 0.976, + "content": " se se lucreaza lucreaza la solutii de acesta .", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 180, + 293, + 189 + ], + "spans": [ + { + "bbox": [ + 128, + 180, + 293, + 189 + ], + "score": 0.989, + "content": "se se lucreaza lucreaza la solutii de de acesta .", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 192, + 303, + 202 + ], + "spans": [ + { + "bbox": [ + 128, + 192, + 303, + 202 + ], + "score": 0.994, + "content": "se se lucreaza lucreaza la solutii de genul acesta .", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 320, + 106, + 431, + 114 + ], + "spans": [ + { + "bbox": [ + 320, + 106, + 431, + 114 + ], + "score": 0.98, + "content": "solutions on this kind are done.", + "type": "text" + } + ], + "index": 2, + "is_list_end_line": true + }, + { + "bbox": [ + 319, + 118, + 433, + 127 + ], + "spans": [ + { + "bbox": [ + 319, + 118, + 433, + 127 + ], + "score": 0.981, + "content": "work done on solutions like this .", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 319, + 130, + 426, + 140 + ], + "spans": [ + { + "bbox": [ + 319, + 130, + 426, + 140 + ], + "score": 0.986, + "content": "solutions on this kind is done .", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 319, + 143, + 441, + 152 + ], + "spans": [ + { + "bbox": [ + 319, + 143, + 441, + 152 + ], + "score": 0.971, + "content": "work is done on solutions like this .", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 319, + 155, + 441, + 164 + ], + "spans": [ + { + "bbox": [ + 319, + 155, + 441, + 164 + ], + "score": 0.985, + "content": "work is done on solutions like this .", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 320, + 168, + 476, + 177 + ], + "spans": [ + { + "bbox": [ + 320, + 168, + 476, + 177 + ], + "score": 0.98, + "content": "work is being done on solutions like this .", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 320, + 180, + 480, + 190 + ], + "spans": [ + { + "bbox": [ + 320, + 180, + 480, + 190 + ], + "score": 0.996, + "content": "work is being done on solutions such as this.", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 320, + 192, + 485, + 202 + ], + "spans": [ + { + "bbox": [ + 320, + 192, + 485, + 202 + ], + "score": 0.993, + "content": "work is being done on solutions such this kind .", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 8.0, + "bbox_fs": [ + 127, + 105, + 303, + 202 + ] + }, + { + "type": "list", + "bbox": [ + 319, + 105, + 484, + 203 + ], + "lines": [], + "index": 9.0, + "bbox_fs": [ + 319, + 106, + 485, + 202 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 219, + 506, + 274 + ], + "lines": [ + { + "bbox": [ + 106, + 219, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 231 + ], + "score": 1.0, + "content": "Figure 5: A Romanian–English example translated with noisy parallel decoding. At left are eight", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 229, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 506, + 243 + ], + "score": 1.0, + "content": "sampled fertility sequences from the encoder, represented with their corresponding decoder input", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 240, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 240, + 505, + 254 + ], + "score": 1.0, + "content": "sequences. Each of these values for the latent variable leads to a different possible output translation,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "spans": [ + { + "bbox": [ + 105, + 251, + 506, + 264 + ], + "score": 1.0, + "content": "shown at right. The autoregressive Transformer then picks the best translation, shown in red, a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 262, + 385, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 385, + 276 + ], + "score": 1.0, + "content": "process which is much faster than directly using it to generate output.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 219, + 506, + 276 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 295, + 505, + 339 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 309 + ], + "score": 1.0, + "content": "well as a pair in the second, are not present in the versions with noisy parallel decoding, suggesting", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 306, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 319 + ], + "score": 1.0, + "content": "that NPD scoring using the teacher model can filter out such mistakes. The translations produced", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "by the NAT with NPD, while of a similar quality to those produced by the autoregressive model, are", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 329, + 221, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 221, + 340 + ], + "score": 1.0, + "content": "also noticeably more literal.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 294, + 506, + 340 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 345, + 504, + 367 + ], + "lines": [ + { + "bbox": [ + 106, + 344, + 504, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 344, + 504, + 358 + ], + "score": 1.0, + "content": "We also show an example of the noisy parallel decoding process in Fig. 5, demonstrating the diver-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 356, + 402, + 369 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 402, + 369 + ], + "score": 1.0, + "content": "sity of translations that can be found by sampling from the fertility space.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 344, + 504, + 369 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 384, + 195, + 397 + ], + "lines": [ + { + "bbox": [ + 105, + 383, + 197, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 197, + 400 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 410, + 505, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "We introduce a latent variable model for non-autoregressive machine translation that enables a de-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 505, + 433 + ], + "score": 1.0, + "content": "coder based on Vaswani et al. (2017) to take full advantage of its exceptional degree of internal", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 506, + 444 + ], + "score": 1.0, + "content": "parallelism even at inference time. As a result, we measure translation latencies of one-tenth that of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 443, + 438, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 438, + 456 + ], + "score": 1.0, + "content": "an equal-sized autoregressive model, while maintaining competitive BLEU scores.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 410, + 506, + 456 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 472, + 175, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 472, + 176, + 485 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 176, + 485 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 108, + 491, + 504, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 504 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 501, + 303, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 303, + 513 + ], + "score": 1.0, + "content": "learning to align and translate. In ICLR, 2015.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 489, + 505, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 506, + 555 + ], + "lines": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "score": 1.0, + "content": "Peter Brown, Vincent della Pietra, Stephen della Pietra, and Robert Mercer. The mathematics of", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263–311,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 116, + 543, + 142, + 556 + ], + "spans": [ + { + "bbox": [ + 116, + 543, + 142, + 556 + ], + "score": 1.0, + "content": "1993.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 522, + 506, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 564, + 503, + 587 + ], + "lines": [ + { + "bbox": [ + 106, + 563, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 563, + 504, + 576 + ], + "score": 1.0, + "content": "Kyunghyun Cho. Noisy parallel approximate decoding for conditional recurrent language model.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 575, + 279, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 575, + 279, + 587 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1605.03835, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39.5, + "bbox_fs": [ + 106, + 563, + 504, + 587 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 595, + 504, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 506, + 608 + ], + "score": 1.0, + "content": "Chris Dyer, Victor Chahuneau, and Noah Smith. A simple, fast, and effective reparameterization of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 606, + 249, + 618 + ], + "spans": [ + { + "bbox": [ + 116, + 606, + 249, + 618 + ], + "score": 1.0, + "content": "IBM Model 2. In NAACL, 2013.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 595, + 506, + 618 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann Dauphin. Convolutional", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 638, + 406, + 649 + ], + "spans": [ + { + "bbox": [ + 115, + 638, + 406, + 649 + ], + "score": 1.0, + "content": "sequence to sequence learning. arXiv preprint arXiv:1705.03122, 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 626, + 505, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 658, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 658, + 505, + 671 + ], + "score": 1.0, + "content": "Łukasz Kaiser, Aidan Gomez, and Franc¸ois Chollet. Depthwise separable convolutions for neural", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 116, + 669, + 363, + 681 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 363, + 681 + ], + "score": 1.0, + "content": "machine translation. arXiv preprint arXiv:1706.03059, 2017.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5, + "bbox_fs": [ + 106, + 658, + 505, + 681 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 689, + 506, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 688, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 703 + ], + "score": 1.0, + "content": "Nal Kalchbrenner, Lasse Espeholt, Karen Simonyan, Aaron van den Oord, Alex Graves, and Koray", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 116, + 700, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 116, + 700, + 505, + 712 + ], + "score": 1.0, + "content": "Kavukc¸uoglu. Neural machine translation in linear time. ˇ arXiv preprint arXiv:1610.10099, 2016.", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 688, + 505, + 712 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 720, + 468, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 719, + 469, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 719, + 469, + 733 + ], + "score": 1.0, + "content": "Yoon Kim and Alexander Rush. Sequence-level knowledge distillation. In EMNLP, 2016.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49, + "bbox_fs": [ + 106, + 719, + 469, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Philipp Koehn and Rebecca Knowles. Six challenges for neural machine translation. arXiv preprint", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 219, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 219, + 105 + ], + "score": 1.0, + "content": "arXiv:1706.03872, 2017.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 326, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 326, + 135 + ], + "score": 1.0, + "content": "based neural machine translation. In EMNLP, 2015.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 504, + 155 + ], + "score": 1.0, + "content": "Randi Martin, Jason Crowther, Meredith Knight, Franklin Tamborello, and Chin-Lung Yang. Plan-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "ning in sentence production: Evidence for the phrase as a default planning scope. Cognition, 116", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 196, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 196, + 176 + ], + "score": 1.0, + "content": "(2):177–192, 2010.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 182, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "score": 1.0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 369, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 369, + 207 + ], + "score": 1.0, + "content": "evaluation of machine translation. In ACL, pp. 311–318, 2002.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 105, + 213, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 504, + 225 + ], + "score": 1.0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 224, + 341, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 224, + 341, + 236 + ], + "score": 1.0, + "content": "subword units. arXiv preprint arXiv:1508.07909, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 105, + 242, + 504, + 265 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "score": 1.0, + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc L ˆ e. Sequence to sequence learning with neural networks. ˆ", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 253, + 178, + 265 + ], + "spans": [ + { + "bbox": [ + 115, + 253, + 178, + 265 + ], + "score": 1.0, + "content": "In NIPS, 2014.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 272, + 503, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 284, + 501, + 297 + ], + "spans": [ + { + "bbox": [ + 116, + 284, + 501, + 297 + ], + "score": 1.0, + "content": "Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 104, + 302, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "Ronald Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 314, + 326, + 325 + ], + "spans": [ + { + "bbox": [ + 116, + 314, + 326, + 325 + ], + "score": 1.0, + "content": "learning. Machine Learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 332, + 506, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 346 + ], + "score": 1.0, + "content": "Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "K. Macherey, J. Klingner, A. Shah, M. Johnson, X. Liu, Ł. Kaiser, S. Gouws, Y. Kato, T. Kudo,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "H. Kazawa, K. Stevens, G. Kurian, N. Patil, W. Wang, C. Young, J. Smith, J. Riesa, A. Rudnick,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "O. Vinyals, G. Corrado, M. Hughes, and J. Dean. Google’s neural machine translation system:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 114, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 385, + 142, + 399 + ], + "spans": [ + { + "bbox": [ + 115, + 385, + 142, + 399 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "Philipp Koehn and Rebecca Knowles. Six challenges for neural machine translation. arXiv preprint", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 219, + 105 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 219, + 105 + ], + "score": 1.0, + "content": "arXiv:1706.03872, 2017.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 505, + 105 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 111, + 504, + 135 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. Effective approaches to attention-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 326, + 135 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 326, + 135 + ], + "score": 1.0, + "content": "based neural machine translation. In EMNLP, 2015.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 111, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 141, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 504, + 155 + ], + "score": 1.0, + "content": "Randi Martin, Jason Crowther, Meredith Knight, Franklin Tamborello, and Chin-Lung Yang. Plan-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 505, + 167 + ], + "score": 1.0, + "content": "ning in sentence production: Evidence for the phrase as a default planning scope. Cognition, 116", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 164, + 196, + 176 + ], + "spans": [ + { + "bbox": [ + 116, + 164, + 196, + 176 + ], + "score": 1.0, + "content": "(2):177–192, 2010.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 141, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 182, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 504, + 195 + ], + "score": 1.0, + "content": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 194, + 369, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 194, + 369, + 207 + ], + "score": 1.0, + "content": "evaluation of machine translation. In ACL, pp. 311–318, 2002.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 183, + 504, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 213, + 504, + 236 + ], + "lines": [ + { + "bbox": [ + 106, + 212, + 504, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 504, + 225 + ], + "score": 1.0, + "content": "Rico Sennrich, Barry Haddow, and Alexandra Birch. Neural machine translation of rare words with", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 224, + 341, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 224, + 341, + 236 + ], + "score": 1.0, + "content": "subword units. arXiv preprint arXiv:1508.07909, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 212, + 504, + 236 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 242, + 504, + 265 + ], + "lines": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 504, + 255 + ], + "score": 1.0, + "content": "Ilya Sutskever, Oriol Vinyals, and Quoc L ˆ e. Sequence to sequence learning with neural networks. ˆ", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 253, + 178, + 265 + ], + "spans": [ + { + "bbox": [ + 115, + 253, + 178, + 265 + ], + "score": 1.0, + "content": "In NIPS, 2014.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 106, + 243, + 504, + 265 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 272, + 503, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "spans": [ + { + "bbox": [ + 106, + 273, + 504, + 285 + ], + "score": 1.0, + "content": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan Gomez, Łukasz", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 284, + 501, + 297 + ], + "spans": [ + { + "bbox": [ + 116, + 284, + 501, + 297 + ], + "score": 1.0, + "content": "Kaiser, and Illia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 273, + 504, + 297 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 302, + 504, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "Ronald Williams. Simple statistical gradient-following algorithms for connectionist reinforcement", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 314, + 326, + 325 + ], + "spans": [ + { + "bbox": [ + 116, + 314, + 326, + 325 + ], + "score": 1.0, + "content": "learning. Machine Learning, 8(3-4):229–256, 1992.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 303, + 505, + 325 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 332, + 506, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 331, + 506, + 346 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 506, + 346 + ], + "score": 1.0, + "content": "Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 116, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 116, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "K. Macherey, J. Klingner, A. Shah, M. Johnson, X. Liu, Ł. Kaiser, S. Gouws, Y. Kato, T. Kudo,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 115, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "H. Kazawa, K. Stevens, G. Kurian, N. Patil, W. Wang, C. Young, J. Smith, J. Riesa, A. Rudnick,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 115, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "O. Vinyals, G. Corrado, M. Hughes, and J. Dean. Google’s neural machine translation system:", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 114, + 375, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 114, + 375, + 506, + 389 + ], + "score": 1.0, + "content": "Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 385, + 142, + 399 + ], + "spans": [ + { + "bbox": [ + 115, + 385, + 142, + 399 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 331, + 506, + 399 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 107, + 499, + 354 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 107, + 499, + 354 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 107, + 499, + 354 + ], + "spans": [ + { + "bbox": [ + 110, + 107, + 499, + 354 + ], + "score": 0.973, + "type": "image", + "image_path": "8397d58044f6d9c8a19dcbb03ac376664021ffaa63d7eac01c5e0d5c288a1df7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 107, + 499, + 189.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 189.33333333333331, + 499, + 271.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 271.66666666666663, + 499, + 353.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 365, + 505, + 388 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "Figure 6: The schematic structure of training and inference for the NAT. The “distilled data” contains", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 376, + 464, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 464, + 390 + ], + "score": 1.0, + "content": "target sentences decoded by the autoregressive model and ground-truth source sentences.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 141, + 429, + 449, + 626 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 141, + 429, + 449, + 626 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 141, + 429, + 449, + 626 + ], + "spans": [ + { + "bbox": [ + 141, + 429, + 449, + 626 + ], + "score": 0.97, + "type": "image", + "image_path": "65e35fd7ac2e233ef8cdae03751ebce2756e57bc1378b23f1a59c83c62823be8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 141, + 429, + 449, + 494.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 141, + 494.6666666666667, + 449, + 560.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 141, + 560.3333333333334, + 449, + 626.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 641, + 505, + 708 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "Figure 7: The translation latency, computed as the time to decode a single sentence without mini-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 653, + 504, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 504, + 664 + ], + "score": 1.0, + "content": "batching, for each sentence in the IWSLT development set as a function of its length. The autore-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "score": 1.0, + "content": "gressive model has latency linear in the decoding length, while the latency of the NAT is nearly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "constant for typical lengths, even with NPD with sample size 10. When using NPD with sample", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "score": 1.0, + "content": "size 100, the level of parallelism is enough to more than saturate the GPU, leading again to linear", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 696, + 147, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 147, + 709 + ], + "score": 1.0, + "content": "latencies.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + } + ], + "index": 8.25 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 82, + 271, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 272, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 272, + 96 + ], + "score": 1.0, + "content": "A SCHEMATIC AND ANALYSIS", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 107, + 499, + 354 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 107, + 499, + 354 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 107, + 499, + 354 + ], + "spans": [ + { + "bbox": [ + 110, + 107, + 499, + 354 + ], + "score": 0.973, + "type": "image", + "image_path": "8397d58044f6d9c8a19dcbb03ac376664021ffaa63d7eac01c5e0d5c288a1df7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 107, + 499, + 189.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 189.33333333333331, + 499, + 271.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 271.66666666666663, + 499, + 353.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 365, + 505, + 388 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "Figure 6: The schematic structure of training and inference for the NAT. The “distilled data” contains", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 376, + 464, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 464, + 390 + ], + "score": 1.0, + "content": "target sentences decoded by the autoregressive model and ground-truth source sentences.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 141, + 429, + 449, + 626 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 141, + 429, + 449, + 626 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 141, + 429, + 449, + 626 + ], + "spans": [ + { + "bbox": [ + 141, + 429, + 449, + 626 + ], + "score": 0.97, + "type": "image", + "image_path": "65e35fd7ac2e233ef8cdae03751ebce2756e57bc1378b23f1a59c83c62823be8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 141, + 429, + 449, + 494.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 141, + 494.6666666666667, + 449, + 560.3333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 141, + 560.3333333333334, + 449, + 626.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 641, + 505, + 708 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 505, + 654 + ], + "score": 1.0, + "content": "Figure 7: The translation latency, computed as the time to decode a single sentence without mini-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 653, + 504, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 653, + 504, + 664 + ], + "score": 1.0, + "content": "batching, for each sentence in the IWSLT development set as a function of its length. The autore-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 663, + 505, + 676 + ], + "score": 1.0, + "content": "gressive model has latency linear in the decoding length, while the latency of the NAT is nearly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 675, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 675, + 505, + 687 + ], + "score": 1.0, + "content": "constant for typical lengths, even with NPD with sample size 10. When using NPD with sample", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 506, + 698 + ], + "score": 1.0, + "content": "size 100, the level of parallelism is enough to more than saturate the GPU, leading again to linear", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 696, + 147, + 709 + ], + "spans": [ + { + "bbox": [ + 105, + 696, + 147, + 709 + ], + "score": 1.0, + "content": "latencies.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + } + ], + "index": 8.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 302, + 441, + 497 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 302, + 441, + 497 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 302, + 441, + 497 + ], + "spans": [ + { + "bbox": [ + 151, + 302, + 441, + 497 + ], + "score": 0.973, + "type": "image", + "image_path": "19326d4ac73a662073b026f5ada8aa0be903a2801d0ffdb32756ff0c9f97cdcc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 151, + 302, + 441, + 367.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 151, + 367.0, + 441, + 432.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 151, + 432.0, + 441, + 497.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 512, + 504, + 536 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "Figure 8: Learning curves for training and fine-tuning of the NAT on IWSLT. BLEU scores are on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 523, + 191, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 191, + 536 + ], + "score": 1.0, + "content": "the development set.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 151, + 302, + 441, + 497 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 151, + 302, + 441, + 497 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 151, + 302, + 441, + 497 + ], + "spans": [ + { + "bbox": [ + 151, + 302, + 441, + 497 + ], + "score": 0.973, + "type": "image", + "image_path": "19326d4ac73a662073b026f5ada8aa0be903a2801d0ffdb32756ff0c9f97cdcc.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 151, + 302, + 441, + 367.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 151, + 367.0, + 441, + 432.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 151, + 432.0, + 441, + 497.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 512, + 504, + 536 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 506, + 525 + ], + "score": 1.0, + "content": "Figure 8: Learning curves for training and fine-tuning of the NAT on IWSLT. BLEU scores are on", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 523, + 191, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 191, + 536 + ], + "score": 1.0, + "content": "the development set.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/B1l8BtlCb/B1l8BtlCb_model.json b/parse/train/B1l8BtlCb/B1l8BtlCb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6f5893e1d48b7a50423d478eee87d6a3b97d9926 --- /dev/null +++ b/parse/train/B1l8BtlCb/B1l8BtlCb_model.json @@ -0,0 +1,15538 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 686, + 1302, + 686, + 1302, + 1051, + 398, + 1051 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1188, + 1403, + 1188, + 1403, + 1401, + 299, + 1401 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1556, + 1403, + 1556, + 1403, + 1740, + 299, + 1740 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1417, + 1402, + 1417, + 1402, + 1541, + 299, + 1541 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 299, + 223, + 999, + 223, + 999, + 324, + 299, + 324 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 315, + 412, + 1078, + 412, + 1078, + 534, + 315, + 534 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 302, + 1916, + 1402, + 1916, + 1402, + 1981, + 302, + 1981 + ], + "score": 0.935 + }, + { + "category_id": 2, + "poly": [ + 330, + 2005, + 1195, + 2005, + 1195, + 2034, + 330, + 2034 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 302, + 1116, + 573, + 1116, + 573, + 1151, + 302, + 1151 + ], + "score": 0.895 + }, + { + "category_id": 0, + "poly": [ + 301, + 1786, + 557, + 1786, + 557, + 1823, + 301, + 1823 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 773, + 615, + 927, + 615, + 927, + 648, + 773, + 648 + ], + "score": 0.879 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.879 + }, + { + "category_id": 1, + "poly": [ + 314, + 374, + 1290, + 374, + 1290, + 409, + 314, + 409 + ], + "score": 0.854 + }, + { + "category_id": 0, + "poly": [ + 298, + 1857, + 994, + 1857, + 994, + 1890, + 298, + 1890 + ], + "score": 0.835 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.73 + }, + { + "category_id": 13, + "poly": [ + 690, + 1948, + 884, + 1948, + 884, + 1981, + 690, + 1981 + ], + "score": 0.93, + "latex": "Y = \\{ y _ { 1 } , . . . , y _ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 570, + 1917, + 779, + 1917, + 779, + 1951, + 570, + 1951 + ], + "score": 0.92, + "latex": "X = \\{ x _ { 1 } , . . . , x _ { T ^ { \\prime } } \\}" + }, + { + "category_id": 13, + "poly": [ + 392, + 375, + 444, + 375, + 444, + 408, + 392, + 408 + ], + "score": 0.82, + "latex": "\\mathbf { G u } ^ { \\dagger }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 220.0, + 795.0, + 220.0, + 795.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 276.0, + 1001.0, + 276.0, + 1001.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1999.0, + 1195.0, + 1999.0, + 1195.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1113.0, + 579.0, + 1113.0, + 579.0, + 1160.0, + 294.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1782.0, + 563.0, + 1782.0, + 563.0, + 1832.0, + 291.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 611.0, + 934.0, + 611.0, + 934.0, + 654.0, + 768.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1857.0, + 998.0, + 1857.0, + 998.0, + 1892.0, + 295.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 685.0, + 1306.0, + 685.0, + 1306.0, + 723.0, + 393.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 716.0, + 1307.0, + 716.0, + 1307.0, + 755.0, + 392.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 747.0, + 1307.0, + 747.0, + 1307.0, + 783.0, + 391.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 778.0, + 1306.0, + 778.0, + 1306.0, + 813.0, + 393.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 807.0, + 1307.0, + 807.0, + 1307.0, + 845.0, + 392.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 840.0, + 1306.0, + 840.0, + 1306.0, + 873.0, + 393.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 871.0, + 1304.0, + 871.0, + 1304.0, + 902.0, + 394.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 899.0, + 1308.0, + 899.0, + 1308.0, + 937.0, + 392.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 931.0, + 1306.0, + 931.0, + 1306.0, + 966.0, + 393.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 962.0, + 1305.0, + 962.0, + 1305.0, + 994.0, + 394.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 989.0, + 1307.0, + 989.0, + 1307.0, + 1027.0, + 392.0, + 1027.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1021.0, + 524.0, + 1021.0, + 524.0, + 1053.0, + 394.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1187.0, + 1402.0, + 1187.0, + 1402.0, + 1220.0, + 295.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1220.0, + 1404.0, + 1220.0, + 1404.0, + 1250.0, + 297.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1248.0, + 1405.0, + 1248.0, + 1405.0, + 1282.0, + 295.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1281.0, + 1405.0, + 1281.0, + 1405.0, + 1313.0, + 293.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1309.0, + 1407.0, + 1309.0, + 1407.0, + 1342.0, + 294.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1340.0, + 1407.0, + 1340.0, + 1407.0, + 1378.0, + 293.0, + 1378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1373.0, + 557.0, + 1373.0, + 557.0, + 1402.0, + 295.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1557.0, + 1404.0, + 1557.0, + 1404.0, + 1588.0, + 297.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1621.0, + 293.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1618.0, + 1405.0, + 1618.0, + 1405.0, + 1651.0, + 293.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1644.0, + 1404.0, + 1644.0, + 1404.0, + 1684.0, + 293.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1677.0, + 1403.0, + 1677.0, + 1403.0, + 1709.0, + 295.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1708.0, + 682.0, + 1708.0, + 682.0, + 1744.0, + 295.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1414.0, + 1404.0, + 1414.0, + 1404.0, + 1451.0, + 293.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1447.0, + 1404.0, + 1447.0, + 1404.0, + 1480.0, + 295.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1477.0, + 1407.0, + 1477.0, + 1407.0, + 1513.0, + 293.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1507.0, + 1333.0, + 1507.0, + 1333.0, + 1547.0, + 292.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 405.0, + 556.0, + 405.0, + 556.0, + 441.0, + 311.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 440.0, + 1080.0, + 440.0, + 1080.0, + 474.0, + 314.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 469.0, + 667.0, + 469.0, + 667.0, + 508.0, + 312.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 504.0, + 746.0, + 504.0, + 746.0, + 536.0, + 313.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1914.0, + 569.0, + 1914.0, + 569.0, + 1954.0, + 297.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 1914.0, + 1406.0, + 1914.0, + 1406.0, + 1954.0, + 780.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 689.0, + 1943.0, + 689.0, + 1985.0, + 295.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1943.0, + 1407.0, + 1943.0, + 1407.0, + 1985.0, + 885.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 373.0, + 391.0, + 373.0, + 391.0, + 414.0, + 308.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 373.0, + 1297.0, + 373.0, + 1297.0, + 414.0, + 445.0, + 414.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 776, + 1404, + 776, + 1404, + 994, + 297, + 994 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1410, + 990, + 1410, + 990, + 1656, + 298, + 1656 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1007, + 1404, + 1007, + 1404, + 1163, + 298, + 1163 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1241, + 1402, + 1241, + 1402, + 1396, + 297, + 1396 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1674, + 990, + 1674, + 990, + 1829, + 297, + 1829 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 519, + 1402, + 519, + 1402, + 614, + 298, + 614 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 377, + 1404, + 377, + 1404, + 502, + 298, + 502 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 1941, + 1402, + 1941, + 1402, + 2034, + 298, + 2034 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 1010, + 1767, + 1404, + 1767, + 1404, + 1923, + 1010, + 1923 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 1018, + 1419, + 1404, + 1419, + 1404, + 1733, + 1018, + 1733 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 604, + 276, + 1093, + 276, + 1093, + 367, + 604, + 367 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 332, + 1842, + 918, + 1842, + 918, + 1931, + 332, + 1931 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 523, + 627, + 1174, + 627, + 1174, + 717, + 523, + 717 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 298, + 230, + 617, + 230, + 617, + 262, + 298, + 262 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 300, + 726, + 1189, + 726, + 1189, + 760, + 300, + 760 + ], + "score": 0.92 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 816, + 74, + 816, + 106, + 298, + 106 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 302, + 1198, + 788, + 1198, + 788, + 1229, + 302, + 1229 + ], + "score": 0.908 + }, + { + "category_id": 9, + "poly": [ + 1366, + 656, + 1400, + 656, + 1400, + 687, + 1366, + 687 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 305, + 1400, + 305, + 1400, + 336, + 1366, + 336 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 953, + 1871, + 988, + 1871, + 988, + 1902, + 953, + 1902 + ], + "score": 0.84 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.738 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.173 + }, + { + "category_id": 14, + "poly": [ + 521, + 624, + 1177, + 624, + 1177, + 719, + 521, + 719 + ], + "score": 0.95, + "latex": "\\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { A R } } ( Y | X ; \\theta ) = \\sum _ { t = 1 } ^ { T + 1 } \\log p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) ." + }, + { + "category_id": 14, + "poly": [ + 605, + 272, + 1092, + 272, + 1092, + 368, + 605, + 368 + ], + "score": 0.95, + "latex": "p _ { \\mathcal { A } \\mathcal { R } } ( \\boldsymbol { Y } | \\boldsymbol { X } ; \\theta ) = \\prod _ { t = 1 } ^ { T + 1 } p ( y _ { t } | y _ { 0 : t - 1 } , x _ { 1 : T ^ { \\prime } } ; \\theta ) ," + }, + { + "category_id": 14, + "poly": [ + 333, + 1838, + 919, + 1838, + 919, + 1933, + 333, + 1933 + ], + "score": 0.93, + "latex": "p _ { \\mathcal N A } ( Y | X ; \\theta ) = p _ { L } ( T | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 : T ^ { \\prime } } ; \\theta ) ." + }, + { + "category_id": 13, + "poly": [ + 802, + 382, + 864, + 382, + 864, + 411, + 802, + 411 + ], + "score": 0.89, + "latex": "y _ { T + 1 }" + }, + { + "category_id": 13, + "poly": [ + 676, + 378, + 741, + 378, + 741, + 412, + 676, + 412 + ], + "score": 0.89, + "latex": "\\left. \\mathrm { b o s } \\right." + }, + { + "category_id": 13, + "poly": [ + 931, + 378, + 993, + 378, + 993, + 412, + 931, + 412 + ], + "score": 0.89, + "latex": "\\langle \\cos \\rangle" + }, + { + "category_id": 13, + "poly": [ + 347, + 1801, + 380, + 1801, + 380, + 1829, + 347, + 1829 + ], + "score": 0.87, + "latex": "p _ { L }" + }, + { + "category_id": 13, + "poly": [ + 579, + 383, + 609, + 383, + 609, + 411, + 579, + 411 + ], + "score": 0.84, + "latex": "y _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1768, + 398, + 1768, + 398, + 1794, + 374, + 1794 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1042, + 1799, + 1079, + 1799, + 1079, + 1828, + 1042, + 1828 + ], + "score": 0.26, + "latex": "^ { 6 6 } \\mathrm { X }" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1766.0, + 1405.0, + 1766.0, + 1405.0, + 1801.0, + 1008.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1797.0, + 1041.0, + 1797.0, + 1041.0, + 1832.0, + 1008.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1797.0, + 1405.0, + 1797.0, + 1405.0, + 1832.0, + 1080.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 1830.0, + 1405.0, + 1830.0, + 1405.0, + 1861.0, + 1008.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1859.0, + 1405.0, + 1859.0, + 1405.0, + 1892.0, + 1010.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 1891.0, + 1294.0, + 1891.0, + 1294.0, + 1923.0, + 1009.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1417.0, + 1192.0, + 1417.0, + 1192.0, + 1449.0, + 1078.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1422.0, + 1300.0, + 1422.0, + 1300.0, + 1445.0, + 1279.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1428.0, + 1368.0, + 1428.0, + 1368.0, + 1438.0, + 1359.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1434.0, + 1188.0, + 1434.0, + 1188.0, + 1483.0, + 1073.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1279.0, + 1443.0, + 1299.0, + 1443.0, + 1299.0, + 1465.0, + 1279.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1493.0, + 1175.0, + 1493.0, + 1175.0, + 1526.0, + 1086.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1497.0, + 1372.0, + 1497.0, + 1372.0, + 1523.0, + 1277.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1537.0, + 1183.0, + 1537.0, + 1183.0, + 1587.0, + 1071.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1576.0, + 1179.0, + 1576.0, + 1179.0, + 1602.0, + 1068.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1629.0, + 1159.0, + 1629.0, + 1159.0, + 1656.0, + 1098.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1629.0, + 1355.0, + 1629.0, + 1355.0, + 1656.0, + 1294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1674.0, + 1183.0, + 1674.0, + 1183.0, + 1714.0, + 1079.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1277.0, + 1678.0, + 1332.0, + 1678.0, + 1332.0, + 1708.0, + 1277.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1336.0, + 1680.0, + 1371.0, + 1680.0, + 1371.0, + 1703.0, + 1336.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1710.0, + 1106.0, + 1710.0, + 1106.0, + 1730.0, + 1084.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1710.0, + 1143.0, + 1710.0, + 1143.0, + 1730.0, + 1118.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1709.0, + 1178.0, + 1709.0, + 1178.0, + 1731.0, + 1156.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 1710.0, + 1305.0, + 1710.0, + 1305.0, + 1730.0, + 1280.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 1709.0, + 1343.0, + 1709.0, + 1343.0, + 1731.0, + 1310.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1710.0, + 1373.0, + 1710.0, + 1373.0, + 1730.0, + 1352.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1197.0, + 790.0, + 1197.0, + 790.0, + 1233.0, + 296.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 863.0, + 2085.0, + 863.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 776.0, + 1404.0, + 776.0, + 1404.0, + 814.0, + 295.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 808.0, + 1405.0, + 808.0, + 1405.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 842.0, + 1402.0, + 842.0, + 1402.0, + 873.0, + 296.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 870.0, + 1404.0, + 870.0, + 1404.0, + 904.0, + 292.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 1406.0, + 899.0, + 1406.0, + 937.0, + 292.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 927.0, + 1406.0, + 927.0, + 1406.0, + 968.0, + 291.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 962.0, + 449.0, + 962.0, + 449.0, + 994.0, + 295.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1408.0, + 990.0, + 1408.0, + 990.0, + 1445.0, + 295.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1441.0, + 991.0, + 1441.0, + 991.0, + 1475.0, + 296.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1472.0, + 992.0, + 1472.0, + 992.0, + 1503.0, + 294.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1502.0, + 993.0, + 1502.0, + 993.0, + 1537.0, + 295.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1534.0, + 990.0, + 1534.0, + 990.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1562.0, + 990.0, + 1562.0, + 990.0, + 1596.0, + 295.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1592.0, + 991.0, + 1592.0, + 991.0, + 1625.0, + 296.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1624.0, + 854.0, + 1624.0, + 854.0, + 1655.0, + 296.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1009.0, + 1404.0, + 1009.0, + 1404.0, + 1042.0, + 296.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1034.0, + 1404.0, + 1034.0, + 1404.0, + 1074.0, + 293.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1070.0, + 1404.0, + 1070.0, + 1404.0, + 1104.0, + 296.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1097.0, + 1404.0, + 1097.0, + 1404.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1114.0, + 1130.0, + 1114.0, + 1164.0, + 294.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1241.0, + 1403.0, + 1241.0, + 1403.0, + 1275.0, + 296.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1268.0, + 1406.0, + 1268.0, + 1406.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1303.0, + 1406.0, + 1303.0, + 1406.0, + 1337.0, + 293.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1334.0, + 1403.0, + 1334.0, + 1403.0, + 1368.0, + 296.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1361.0, + 1364.0, + 1361.0, + 1364.0, + 1398.0, + 293.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1674.0, + 992.0, + 1674.0, + 992.0, + 1708.0, + 298.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1705.0, + 990.0, + 1705.0, + 990.0, + 1740.0, + 295.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1733.0, + 993.0, + 1733.0, + 993.0, + 1772.0, + 293.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1764.0, + 373.0, + 1764.0, + 373.0, + 1800.0, + 294.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1764.0, + 991.0, + 1764.0, + 991.0, + 1800.0, + 399.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1797.0, + 346.0, + 1797.0, + 346.0, + 1829.0, + 296.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1797.0, + 542.0, + 1797.0, + 542.0, + 1829.0, + 381.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 520.0, + 1404.0, + 520.0, + 1404.0, + 554.0, + 296.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 548.0, + 1406.0, + 548.0, + 1406.0, + 590.0, + 292.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 577.0, + 613.0, + 577.0, + 613.0, + 620.0, + 292.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 375.0, + 578.0, + 375.0, + 578.0, + 417.0, + 292.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 375.0, + 675.0, + 375.0, + 675.0, + 417.0, + 610.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 375.0, + 801.0, + 375.0, + 801.0, + 417.0, + 742.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 375.0, + 930.0, + 375.0, + 930.0, + 417.0, + 865.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 375.0, + 1406.0, + 375.0, + 1406.0, + 417.0, + 994.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 408.0, + 1405.0, + 408.0, + 1405.0, + 444.0, + 294.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 441.0, + 1404.0, + 441.0, + 1404.0, + 473.0, + 294.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 469.0, + 1239.0, + 469.0, + 1239.0, + 504.0, + 293.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1975.0, + 296.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 615.0, + 2003.0, + 615.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 620.0, + 227.0, + 620.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 721.0, + 1193.0, + 721.0, + 1193.0, + 767.0, + 293.0, + 767.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1195, + 1403, + 1195, + 1403, + 1379, + 298, + 1379 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 323, + 222, + 1356, + 222, + 1356, + 807, + 323, + 807 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1026, + 1403, + 1026, + 1403, + 1180, + 298, + 1180 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1650, + 1404, + 1650, + 1404, + 1774, + 299, + 1774 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1402, + 1911, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1395, + 1403, + 1395, + 1403, + 1488, + 299, + 1488 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 297, + 835, + 1400, + 835, + 1400, + 929, + 297, + 929 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 300, + 1841, + 570, + 1841, + 570, + 1873, + 300, + 1873 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 300, + 1563, + 1089, + 1563, + 1089, + 1599, + 300, + 1599 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 302, + 957, + 756, + 957, + 756, + 989, + 302, + 989 + ], + "score": 0.798 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2112, + 842, + 2112 + ], + "score": 0.577 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.445 + }, + { + "category_id": 1, + "poly": [ + 302, + 957, + 756, + 957, + 756, + 989, + 302, + 989 + ], + "score": 0.097 + }, + { + "category_id": 13, + "poly": [ + 885, + 1058, + 946, + 1058, + 946, + 1092, + 885, + 1092 + ], + "score": 0.92, + "latex": "p ( y _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 1365, + 1060, + 1393, + 1060, + 1393, + 1086, + 1365, + 1086 + ], + "score": 0.79, + "latex": "X" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 227.0, + 1179.0, + 227.0, + 1179.0, + 258.0, + 884.0, + 258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 259.0, + 915.0, + 259.0, + 915.0, + 280.0, + 896.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 258.0, + 988.0, + 258.0, + 988.0, + 287.0, + 966.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 258.0, + 1059.0, + 258.0, + 1059.0, + 290.0, + 1038.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 255.0, + 1133.0, + 255.0, + 1133.0, + 290.0, + 1111.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1183.0, + 259.0, + 1202.0, + 259.0, + 1202.0, + 284.0, + 1183.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 261.0, + 1366.0, + 261.0, + 1366.0, + 298.0, + 1254.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 300.0, + 1013.0, + 300.0, + 1013.0, + 332.0, + 871.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 300.0, + 1227.0, + 300.0, + 1227.0, + 332.0, + 1014.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1259.0, + 285.0, + 1354.0, + 285.0, + 1354.0, + 320.0, + 1259.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 316.0, + 470.0, + 316.0, + 470.0, + 336.0, + 452.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 316.0, + 541.0, + 316.0, + 541.0, + 336.0, + 523.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 314.0, + 615.0, + 314.0, + 615.0, + 341.0, + 593.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 316.0, + 685.0, + 316.0, + 685.0, + 337.0, + 666.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 318.0, + 752.0, + 318.0, + 752.0, + 334.0, + 743.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 345.0, + 401.0, + 345.0, + 401.0, + 381.0, + 323.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 454.0, + 344.0, + 465.0, + 344.0, + 465.0, + 365.0, + 454.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 340.0, + 543.0, + 340.0, + 543.0, + 369.0, + 522.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 337.0, + 614.0, + 337.0, + 614.0, + 369.0, + 593.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 340.0, + 684.0, + 340.0, + 684.0, + 363.0, + 665.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 347.0, + 752.0, + 347.0, + 752.0, + 359.0, + 743.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 342.0, + 914.0, + 342.0, + 914.0, + 358.0, + 897.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 342.0, + 985.0, + 342.0, + 985.0, + 358.0, + 968.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 340.0, + 1058.0, + 340.0, + 1058.0, + 360.0, + 1038.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 342.0, + 1202.0, + 342.0, + 1202.0, + 358.0, + 1185.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 369.0, + 415.0, + 369.0, + 415.0, + 402.0, + 324.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 382.0, + 639.0, + 382.0, + 639.0, + 416.0, + 425.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 384.0, + 783.0, + 384.0, + 783.0, + 416.0, + 640.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 380.0, + 934.0, + 380.0, + 934.0, + 415.0, + 877.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 380.0, + 1005.0, + 380.0, + 1005.0, + 415.0, + 949.0, + 415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 380.0, + 1077.0, + 380.0, + 1077.0, + 413.0, + 1019.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 380.0, + 1149.0, + 380.0, + 1149.0, + 413.0, + 1092.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 380.0, + 1221.0, + 380.0, + 1221.0, + 413.0, + 1163.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 423.0, + 472.0, + 423.0, + 472.0, + 448.0, + 448.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 521.0, + 423.0, + 543.0, + 423.0, + 543.0, + 449.0, + 521.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 422.0, + 615.0, + 422.0, + 615.0, + 448.0, + 592.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 423.0, + 688.0, + 423.0, + 688.0, + 447.0, + 664.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 423.0, + 759.0, + 423.0, + 759.0, + 447.0, + 734.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 424.0, + 916.0, + 424.0, + 916.0, + 443.0, + 896.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 424.0, + 987.0, + 424.0, + 987.0, + 443.0, + 967.0, + 443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 426.0, + 1057.0, + 426.0, + 1057.0, + 442.0, + 1040.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 426.0, + 1127.0, + 426.0, + 1127.0, + 442.0, + 1114.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 424.0, + 1203.0, + 424.0, + 1203.0, + 444.0, + 1181.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 422.0, + 1351.0, + 422.0, + 1351.0, + 457.0, + 1261.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 457.0, + 1164.0, + 457.0, + 1164.0, + 487.0, + 934.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 447.0, + 1327.0, + 447.0, + 1327.0, + 480.0, + 1261.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 493.0, + 410.0, + 493.0, + 410.0, + 528.0, + 323.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 517.0, + 389.0, + 517.0, + 389.0, + 549.0, + 324.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 531.0, + 489.0, + 531.0, + 489.0, + 565.0, + 428.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 530.0, + 561.0, + 530.0, + 561.0, + 564.0, + 503.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 529.0, + 633.0, + 529.0, + 633.0, + 567.0, + 574.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 531.0, + 703.0, + 531.0, + 703.0, + 566.0, + 646.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 531.0, + 773.0, + 531.0, + 773.0, + 564.0, + 717.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 534.0, + 1186.0, + 534.0, + 1186.0, + 563.0, + 915.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 544.0, + 1243.0, + 544.0, + 1243.0, + 555.0, + 1234.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 557.0, + 814.0, + 557.0, + 814.0, + 563.0, + 808.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 564.0, + 815.0, + 564.0, + 815.0, + 574.0, + 806.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 580.0, + 464.0, + 580.0, + 464.0, + 592.0, + 455.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 579.0, + 606.0, + 579.0, + 606.0, + 588.0, + 598.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 569.0, + 818.0, + 569.0, + 818.0, + 625.0, + 803.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 583.0, + 1197.0, + 583.0, + 1197.0, + 593.0, + 1186.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 572.0, + 1337.0, + 572.0, + 1337.0, + 617.0, + 1261.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 493.0, + 616.0, + 714.0, + 616.0, + 714.0, + 645.0, + 493.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 808.0, + 624.0, + 814.0, + 624.0, + 814.0, + 630.0, + 808.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 614.0, + 1162.0, + 614.0, + 1162.0, + 648.0, + 937.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 661.0, + 464.0, + 661.0, + 464.0, + 669.0, + 455.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 661.0, + 535.0, + 661.0, + 535.0, + 669.0, + 526.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 661.0, + 606.0, + 661.0, + 606.0, + 669.0, + 596.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 660.0, + 913.0, + 660.0, + 913.0, + 674.0, + 897.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 668.0, + 423.0, + 668.0, + 423.0, + 702.0, + 373.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 667.0, + 1297.0, + 667.0, + 1297.0, + 701.0, + 1248.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 726.0, + 417.0, + 726.0, + 417.0, + 736.0, + 408.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 713.0, + 488.0, + 713.0, + 488.0, + 746.0, + 434.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 713.0, + 560.0, + 713.0, + 560.0, + 746.0, + 504.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 713.0, + 631.0, + 713.0, + 631.0, + 746.0, + 576.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 713.0, + 703.0, + 713.0, + 703.0, + 746.0, + 647.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 713.0, + 773.0, + 713.0, + 773.0, + 746.0, + 719.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 713.0, + 933.0, + 713.0, + 933.0, + 746.0, + 878.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 713.0, + 1005.0, + 713.0, + 1005.0, + 746.0, + 949.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1021.0, + 713.0, + 1076.0, + 713.0, + 1076.0, + 746.0, + 1021.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 713.0, + 1148.0, + 713.0, + 1148.0, + 746.0, + 1093.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 713.0, + 1220.0, + 713.0, + 1220.0, + 746.0, + 1164.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 723.0, + 1250.0, + 723.0, + 1250.0, + 740.0, + 1229.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 725.0, + 1300.0, + 725.0, + 1300.0, + 735.0, + 1286.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 732.0, + 853.0, + 732.0, + 853.0, + 766.0, + 801.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 775.0, + 641.0, + 775.0, + 641.0, + 809.0, + 438.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 777.0, + 688.0, + 777.0, + 688.0, + 807.0, + 660.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 793.0, + 752.0, + 793.0, + 752.0, + 803.0, + 742.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 775.0, + 1156.0, + 775.0, + 1156.0, + 808.0, + 884.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 792.0, + 1196.0, + 792.0, + 1196.0, + 803.0, + 1186.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.75, + 341.0, + 1121.75, + 341.0, + 1121.75, + 357.5, + 1117.75, + 357.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.75, + 723.5, + 362.75, + 723.5, + 362.75, + 735.5, + 351.75, + 735.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 835.0, + 1402.0, + 835.0, + 1402.0, + 871.0, + 295.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 867.0, + 1405.0, + 867.0, + 1405.0, + 901.0, + 295.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 899.0, + 1276.0, + 899.0, + 1276.0, + 929.0, + 297.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1838.0, + 576.0, + 1838.0, + 576.0, + 1878.0, + 293.0, + 1878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1559.0, + 1092.0, + 1559.0, + 1092.0, + 1605.0, + 291.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 956.0, + 759.0, + 956.0, + 759.0, + 993.0, + 295.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1194.0, + 1405.0, + 1194.0, + 1405.0, + 1231.0, + 293.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1221.0, + 1407.0, + 1221.0, + 1407.0, + 1264.0, + 292.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1259.0, + 1405.0, + 1259.0, + 1405.0, + 1291.0, + 296.0, + 1291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1287.0, + 1405.0, + 1287.0, + 1405.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1319.0, + 1405.0, + 1319.0, + 1405.0, + 1350.0, + 296.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 805.0, + 1348.0, + 805.0, + 1382.0, + 293.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1026.0, + 1405.0, + 1026.0, + 1405.0, + 1063.0, + 294.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1059.0, + 884.0, + 1059.0, + 884.0, + 1092.0, + 296.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1059.0, + 1364.0, + 1059.0, + 1364.0, + 1092.0, + 947.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1059.0, + 1402.0, + 1059.0, + 1402.0, + 1092.0, + 1394.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1088.0, + 1403.0, + 1088.0, + 1403.0, + 1122.0, + 297.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1118.0, + 1405.0, + 1118.0, + 1405.0, + 1153.0, + 293.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1148.0, + 1207.0, + 1148.0, + 1207.0, + 1186.0, + 292.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1651.0, + 1404.0, + 1651.0, + 1404.0, + 1684.0, + 295.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1678.0, + 1404.0, + 1678.0, + 1404.0, + 1718.0, + 292.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 815.0, + 1741.0, + 815.0, + 1779.0, + 294.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1402.0, + 1910.0, + 1402.0, + 1945.0, + 294.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1945.0, + 1403.0, + 1945.0, + 1403.0, + 1974.0, + 297.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1398.0, + 2002.0, + 1398.0, + 2036.0, + 293.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1394.0, + 1403.0, + 1394.0, + 1403.0, + 1432.0, + 295.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1461.0, + 295.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1454.0, + 1089.0, + 1454.0, + 1089.0, + 1493.0, + 294.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 956.0, + 759.0, + 956.0, + 759.0, + 993.0, + 295.0, + 993.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1279, + 1402, + 1279, + 1402, + 1405, + 299, + 1405 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1090, + 1403, + 1090, + 1403, + 1184, + 298, + 1184 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 919, + 1403, + 919, + 1403, + 1073, + 298, + 1073 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 368, + 1404, + 368, + 1404, + 583, + 297, + 583 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 301, + 1496, + 1403, + 1496, + 1403, + 1590, + 301, + 1590 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 301, + 1604, + 1400, + 1604, + 1400, + 1697, + 301, + 1697 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 296, + 286, + 1400, + 286, + 1400, + 349, + 296, + 349 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 581, + 1194, + 1119, + 1194, + 1119, + 1270, + 581, + 1270 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 297, + 1720, + 1406, + 1720, + 1406, + 1956, + 297, + 1956 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 297, + 1976, + 1404, + 1976, + 1404, + 2034, + 297, + 2034 + ], + "score": 0.925 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 572, + 229, + 572, + 261, + 300, + 261 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 298, + 605, + 1405, + 605, + 1405, + 893, + 298, + 893 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1217, + 1399, + 1217, + 1399, + 1246, + 1366, + 1246 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.784 + }, + { + "category_id": 0, + "poly": [ + 299, + 1438, + 1170, + 1438, + 1170, + 1472, + 299, + 1472 + ], + "score": 0.718 + }, + { + "category_id": 1, + "poly": [ + 299, + 1438, + 1170, + 1438, + 1170, + 1472, + 299, + 1472 + ], + "score": 0.213 + }, + { + "category_id": 14, + "poly": [ + 580, + 1192, + 1117, + 1192, + 1117, + 1270, + 580, + 1270 + ], + "score": 0.93, + "latex": "{ \\mathrm { A t t e n t i o n } } ( Q , K , V ) = { \\mathrm { s o f t m a x } } \\left( { \\frac { Q K ^ { T } } { \\sqrt { d _ { \\mathrm { m o d e l } } } } } \\right) \\cdot V ," + }, + { + "category_id": 13, + "poly": [ + 1191, + 1893, + 1294, + 1893, + 1294, + 1927, + 1191, + 1927 + ], + "score": 0.93, + "latex": "p ( y | x , z )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1282, + 440, + 1282, + 440, + 1312, + 372, + 1312 + ], + "score": 0.92, + "latex": "d _ { \\mathrm { m o d e l } }" + }, + { + "category_id": 13, + "poly": [ + 770, + 1973, + 1055, + 1973, + 1055, + 2008, + 770, + 2008 + ], + "score": 0.9, + "latex": "p ( j , k ) = \\sin { ( j / 1 0 0 0 0 ^ { k / d } ) }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1973, + 1397, + 1973, + 1397, + 2009, + 1217, + 2009 + ], + "score": 0.89, + "latex": "\\cos { \\left( j / 1 0 0 0 0 ^ { k / d } \\right) }" + }, + { + "category_id": 13, + "poly": [ + 486, + 2007, + 501, + 2007, + 501, + 2034, + 486, + 2034 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 759, + 2008, + 775, + 2008, + 775, + 2030, + 759, + 2030 + ], + "score": 0.8, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1183, + 606, + 1274, + 606, + 1274, + 640, + 1183, + 640 + ], + "score": 0.78, + "latex": "( T ^ { \\prime } t / T )" + }, + { + "category_id": 13, + "poly": [ + 913, + 610, + 927, + 610, + 927, + 634, + 913, + 634 + ], + "score": 0.78, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1176, + 1504, + 1196, + 1504, + 1196, + 1525, + 1176, + 1525 + ], + "score": 0.75, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1000, + 1534, + 1019, + 1534, + 1019, + 1556, + 1000, + 1556 + ], + "score": 0.74, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 385, + 2007, + 402, + 2007, + 402, + 2031, + 385, + 2031 + ], + "score": 0.74, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 417, + 1799, + 437, + 1799, + 437, + 1820, + 417, + 1820 + ], + "score": 0.74, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 591, + 1984, + 609, + 1984, + 609, + 2007, + 591, + 2007 + ], + "score": 0.72, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 443, + 1565, + 461, + 1565, + 461, + 1586, + 443, + 1586 + ], + "score": 0.71, + "latex": "z" + }, + { + "category_id": 13, + "poly": [ + 1159, + 1979, + 1176, + 1979, + 1176, + 2003, + 1159, + 2003 + ], + "score": 0.67, + "latex": "k" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1969.0, + 590.0, + 1969.0, + 590.0, + 2011.0, + 330.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1969.0, + 769.0, + 1969.0, + 769.0, + 2011.0, + 610.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1969.0, + 1158.0, + 1969.0, + 1158.0, + 2011.0, + 1056.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1969.0, + 1216.0, + 1969.0, + 1216.0, + 2011.0, + 1177.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1969.0, + 1404.0, + 1969.0, + 1404.0, + 2011.0, + 1398.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 384.0, + 2003.0, + 384.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 2003.0, + 485.0, + 2003.0, + 485.0, + 2036.0, + 403.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 2003.0, + 758.0, + 2003.0, + 758.0, + 2036.0, + 502.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 2003.0, + 990.0, + 2003.0, + 990.0, + 2036.0, + 776.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 223.0, + 578.0, + 223.0, + 578.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1172.0, + 1438.0, + 1172.0, + 1474.0, + 295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1280.0, + 371.0, + 1280.0, + 371.0, + 1316.0, + 295.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1280.0, + 1407.0, + 1280.0, + 1407.0, + 1316.0, + 441.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1310.0, + 1404.0, + 1310.0, + 1404.0, + 1346.0, + 294.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1342.0, + 1406.0, + 1342.0, + 1406.0, + 1376.0, + 293.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1369.0, + 1353.0, + 1369.0, + 1353.0, + 1410.0, + 292.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1090.0, + 1405.0, + 1090.0, + 1405.0, + 1124.0, + 294.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1124.0, + 1404.0, + 1124.0, + 1404.0, + 1154.0, + 296.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1152.0, + 783.0, + 1152.0, + 783.0, + 1186.0, + 296.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 918.0, + 1404.0, + 918.0, + 1404.0, + 954.0, + 294.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 948.0, + 1404.0, + 948.0, + 1404.0, + 983.0, + 294.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 981.0, + 1404.0, + 981.0, + 1404.0, + 1015.0, + 296.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1046.0, + 294.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1041.0, + 1245.0, + 1041.0, + 1245.0, + 1074.0, + 297.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 367.0, + 1405.0, + 367.0, + 1405.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 399.0, + 1406.0, + 399.0, + 1406.0, + 435.0, + 294.0, + 435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 431.0, + 1406.0, + 431.0, + 1406.0, + 463.0, + 293.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 461.0, + 1404.0, + 461.0, + 1404.0, + 495.0, + 295.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 489.0, + 1406.0, + 489.0, + 1406.0, + 527.0, + 294.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 520.0, + 1404.0, + 520.0, + 1404.0, + 554.0, + 295.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 556.0, + 586.0, + 556.0, + 586.0, + 584.0, + 295.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1497.0, + 1175.0, + 1497.0, + 1175.0, + 1531.0, + 297.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1497.0, + 1404.0, + 1497.0, + 1404.0, + 1531.0, + 1197.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1528.0, + 999.0, + 1528.0, + 999.0, + 1562.0, + 296.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 1528.0, + 1405.0, + 1528.0, + 1405.0, + 1562.0, + 1020.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1559.0, + 442.0, + 1559.0, + 442.0, + 1593.0, + 296.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 1559.0, + 981.0, + 1559.0, + 981.0, + 1593.0, + 462.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1603.0, + 1404.0, + 1603.0, + 1404.0, + 1637.0, + 297.0, + 1637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 1404.0, + 1635.0, + 1404.0, + 1669.0, + 293.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1666.0, + 468.0, + 1666.0, + 468.0, + 1698.0, + 296.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 283.0, + 1402.0, + 283.0, + 1402.0, + 323.0, + 292.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 315.0, + 557.0, + 315.0, + 557.0, + 353.0, + 294.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1719.0, + 1404.0, + 1719.0, + 1404.0, + 1758.0, + 294.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1753.0, + 817.0, + 1753.0, + 817.0, + 1786.0, + 321.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1791.0, + 416.0, + 1791.0, + 416.0, + 1828.0, + 297.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1791.0, + 1404.0, + 1791.0, + 1404.0, + 1828.0, + 438.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1822.0, + 1404.0, + 1822.0, + 1404.0, + 1859.0, + 320.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1851.0, + 1124.0, + 1851.0, + 1124.0, + 1891.0, + 320.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1893.0, + 1190.0, + 1893.0, + 1190.0, + 1930.0, + 295.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1893.0, + 1404.0, + 1893.0, + 1404.0, + 1930.0, + 1295.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1921.0, + 1287.0, + 1921.0, + 1287.0, + 1960.0, + 320.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 605.0, + 912.0, + 605.0, + 912.0, + 642.0, + 296.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 605.0, + 1182.0, + 605.0, + 1182.0, + 642.0, + 928.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 605.0, + 1406.0, + 605.0, + 1406.0, + 642.0, + 1275.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 637.0, + 1403.0, + 637.0, + 1403.0, + 671.0, + 320.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 669.0, + 1287.0, + 669.0, + 1287.0, + 703.0, + 321.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 708.0, + 1405.0, + 708.0, + 1405.0, + 742.0, + 300.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 739.0, + 1405.0, + 739.0, + 1405.0, + 773.0, + 321.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 765.0, + 1406.0, + 765.0, + 1406.0, + 805.0, + 320.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 796.0, + 1405.0, + 796.0, + 1405.0, + 836.0, + 321.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 830.0, + 1405.0, + 830.0, + 1405.0, + 864.0, + 323.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 862.0, + 1105.0, + 862.0, + 1105.0, + 896.0, + 322.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1172.0, + 1438.0, + 1172.0, + 1474.0, + 295.0, + 1474.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1596, + 1404, + 1596, + 1404, + 1751, + 297, + 1751 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 385, + 298, + 385 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 723, + 1403, + 723, + 1403, + 849, + 297, + 849 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1377, + 1404, + 1377, + 1404, + 1501, + 297, + 1501 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 398, + 1402, + 398, + 1402, + 493, + 299, + 493 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1848, + 1402, + 1848, + 1402, + 1942, + 298, + 1942 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 293, + 1764, + 1402, + 1764, + 1402, + 1828, + 293, + 1828 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 297, + 959, + 1405, + 959, + 1405, + 1349, + 297, + 1349 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 869, + 1400, + 869, + 1400, + 932, + 298, + 932 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 338, + 509, + 1318, + 509, + 1318, + 616, + 338, + 616 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 474, + 1958, + 1225, + 1958, + 1225, + 2019, + 474, + 2019 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 292, + 635, + 1404, + 635, + 1404, + 706, + 292, + 706 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 298, + 1538, + 1048, + 1538, + 1048, + 1571, + 298, + 1571 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1968, + 1400, + 1968, + 1400, + 1997, + 1366, + 1997 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1366, + 548, + 1400, + 548, + 1400, + 578, + 1366, + 578 + ], + "score": 0.864 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.757 + }, + { + "category_id": 14, + "poly": [ + 339, + 506, + 1324, + 506, + 1324, + 617, + 339, + 617 + ], + "score": 0.95, + "latex": "p _ { \\mathcal N \\mathcal M } ( Y | X ; \\theta ) = \\sum _ { f _ { 1 } , . . . , f _ { T ^ { \\prime } } \\in \\mathcal F } \\left( \\prod _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot \\prod _ { t = 1 } ^ { T } p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)" + }, + { + "category_id": 13, + "poly": [ + 919, + 672, + 982, + 672, + 982, + 706, + 919, + 706 + ], + "score": 0.94, + "latex": "x \\{ f \\}" + }, + { + "category_id": 13, + "poly": [ + 371, + 633, + 864, + 633, + 864, + 677, + 371, + 677 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathcal { F } = \\{ f _ { 1 } , . . . , f _ { T ^ { \\prime } } | \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } f _ { t ^ { \\prime } } = T , f _ { t ^ { \\prime } } \\in \\mathbb { Z } ^ { * } \\} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 741, + 1688, + 1001, + 1688, + 1001, + 1722, + 741, + 1722 + ], + "score": 0.92, + "latex": "Y \\overset { \\cdot } { = } G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 1026, + 725, + 1178, + 725, + 1178, + 759, + 1026, + 759 + ], + "score": 0.91, + "latex": "p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } \\big )" + }, + { + "category_id": 14, + "poly": [ + 477, + 1956, + 1223, + 1956, + 1223, + 2021, + 477, + 2021 + ], + "score": 0.89, + "latex": "\\hat { Y } _ { \\mathrm { a r g m a x } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e \\ } \\hat { f } _ { t ^ { \\prime } } = \\underset { f } { \\operatorname { a r g m a x } } p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 1092, + 760, + 1178, + 760, + 1178, + 785, + 1092, + 785 + ], + "score": 0.86, + "latex": "L = 5 0" + }, + { + "category_id": 13, + "poly": [ + 1308, + 674, + 1327, + 674, + 1327, + 705, + 1308, + 705 + ], + "score": 0.85, + "latex": "f" + }, + { + "category_id": 13, + "poly": [ + 1260, + 464, + 1284, + 464, + 1284, + 488, + 1260, + 488 + ], + "score": 0.84, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 677, + 462, + 705, + 462, + 705, + 488, + 677, + 488 + ], + "score": 0.82, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 844, + 675, + 869, + 675, + 869, + 700, + 844, + 700 + ], + "score": 0.81, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 1184, + 678, + 1204, + 678, + 1204, + 700, + 1184, + 700 + ], + "score": 0.76, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 476, + 1959, + 789, + 1959, + 789, + 2001, + 476, + 2001 + ], + "score": 0.53, + "latex": "\\hat { Y } _ { \\mathrm { a r g m a x } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1537.0, + 1052.0, + 1537.0, + 1052.0, + 1575.0, + 293.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1594.0, + 1405.0, + 1594.0, + 1405.0, + 1634.0, + 294.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1627.0, + 1402.0, + 1627.0, + 1402.0, + 1664.0, + 294.0, + 1664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1658.0, + 1402.0, + 1658.0, + 1402.0, + 1695.0, + 292.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1688.0, + 740.0, + 1688.0, + 740.0, + 1725.0, + 294.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1688.0, + 1405.0, + 1688.0, + 1405.0, + 1725.0, + 1002.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1719.0, + 943.0, + 1719.0, + 943.0, + 1754.0, + 292.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 325.0, + 296.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 319.0, + 1404.0, + 319.0, + 1404.0, + 357.0, + 293.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 352.0, + 691.0, + 352.0, + 691.0, + 385.0, + 296.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 722.0, + 1025.0, + 722.0, + 1025.0, + 761.0, + 291.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 722.0, + 1403.0, + 722.0, + 1403.0, + 761.0, + 1179.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 754.0, + 1091.0, + 754.0, + 1091.0, + 792.0, + 292.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 754.0, + 1405.0, + 754.0, + 1405.0, + 792.0, + 1179.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 786.0, + 1404.0, + 786.0, + 1404.0, + 824.0, + 292.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 817.0, + 1229.0, + 817.0, + 1229.0, + 852.0, + 294.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1376.0, + 1406.0, + 1376.0, + 1406.0, + 1410.0, + 295.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1409.0, + 1405.0, + 1409.0, + 1405.0, + 1441.0, + 294.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1439.0, + 1406.0, + 1439.0, + 1406.0, + 1475.0, + 294.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 834.0, + 1467.0, + 834.0, + 1506.0, + 294.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 396.0, + 1406.0, + 396.0, + 1406.0, + 436.0, + 293.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 428.0, + 1407.0, + 428.0, + 1407.0, + 466.0, + 294.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 461.0, + 676.0, + 461.0, + 676.0, + 497.0, + 293.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 461.0, + 1259.0, + 461.0, + 1259.0, + 497.0, + 706.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 461.0, + 1322.0, + 461.0, + 1322.0, + 497.0, + 1285.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1848.0, + 1406.0, + 1848.0, + 1406.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1877.0, + 1403.0, + 1877.0, + 1403.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1910.0, + 529.0, + 1910.0, + 529.0, + 1943.0, + 293.0, + 1943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1761.0, + 1404.0, + 1761.0, + 1404.0, + 1802.0, + 293.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1797.0, + 1142.0, + 1797.0, + 1142.0, + 1829.0, + 296.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 957.0, + 1403.0, + 957.0, + 1403.0, + 995.0, + 294.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 991.0, + 1047.0, + 991.0, + 1047.0, + 1026.0, + 320.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1030.0, + 1403.0, + 1030.0, + 1403.0, + 1069.0, + 296.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1099.0, + 321.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1094.0, + 1405.0, + 1094.0, + 1405.0, + 1130.0, + 323.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1126.0, + 1406.0, + 1126.0, + 1406.0, + 1158.0, + 324.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1154.0, + 1405.0, + 1154.0, + 1405.0, + 1190.0, + 321.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1183.0, + 1191.0, + 1183.0, + 1191.0, + 1218.0, + 320.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1225.0, + 1405.0, + 1225.0, + 1405.0, + 1262.0, + 311.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1294.0, + 321.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1287.0, + 1405.0, + 1287.0, + 1405.0, + 1324.0, + 323.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1317.0, + 974.0, + 1317.0, + 974.0, + 1352.0, + 323.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 869.0, + 1405.0, + 869.0, + 1405.0, + 904.0, + 296.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 899.0, + 878.0, + 899.0, + 878.0, + 935.0, + 296.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 618.0, + 370.0, + 618.0, + 370.0, + 711.0, + 285.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 618.0, + 918.0, + 618.0, + 918.0, + 711.0, + 870.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 983.0, + 618.0, + 1183.0, + 618.0, + 1183.0, + 711.0, + 983.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 618.0, + 1307.0, + 618.0, + 1307.0, + 711.0, + 1205.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 618.0, + 1415.0, + 618.0, + 1415.0, + 711.0, + 1328.0, + 711.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1771, + 1403, + 1771, + 1403, + 1958, + 298, + 1958 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 450, + 1401, + 450, + 1401, + 574, + 298, + 574 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 745, + 1403, + 745, + 1403, + 869, + 298, + 869 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1413, + 1404, + 1413, + 1404, + 1568, + 298, + 1568 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1582, + 1404, + 1582, + 1404, + 1676, + 298, + 1676 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 298, + 986, + 1404, + 986, + 1404, + 1081, + 298, + 1081 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 293, + 668, + 1400, + 668, + 1400, + 732, + 293, + 732 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 292, + 228, + 1402, + 228, + 1402, + 294, + 292, + 294 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 322, + 1118, + 1342, + 1118, + 1342, + 1382, + 322, + 1382 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 516, + 591, + 1182, + 591, + 1182, + 652, + 516, + 652 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1400, + 1971, + 1400, + 2035, + 297, + 2035 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 412, + 306, + 1279, + 306, + 1279, + 412, + 412, + 412 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 299, + 916, + 498, + 916, + 498, + 951, + 299, + 951 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 816, + 74, + 816, + 105, + 298, + 105 + ], + "score": 0.914 + }, + { + "category_id": 9, + "poly": [ + 1366, + 595, + 1400, + 595, + 1400, + 625, + 1366, + 625 + ], + "score": 0.874 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1237, + 1401, + 1237, + 1401, + 1266, + 1365, + 1266 + ], + "score": 0.869 + }, + { + "category_id": 9, + "poly": [ + 1365, + 343, + 1401, + 343, + 1401, + 374, + 1365, + 374 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2112, + 840, + 2112 + ], + "score": 0.792 + }, + { + "category_id": 0, + "poly": [ + 297, + 1713, + 930, + 1713, + 930, + 1747, + 297, + 1747 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 297, + 1713, + 930, + 1713, + 930, + 1747, + 297, + 1747 + ], + "score": 0.121 + }, + { + "category_id": 14, + "poly": [ + 321, + 1118, + 1347, + 1118, + 1347, + 1384, + 321, + 1384 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } & { \\mathcal { L } _ { \\mathrm { M L } } = \\log p _ { \\mathcal { N A } } ( Y | X ; \\theta ) = \\log \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } } \\in \\mathcal { F } } p _ { F } ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) \\cdot p ( y _ { 1 : T } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) } \\\\ & { \\geq \\displaystyle \\sum _ { f _ { 1 : T ^ { \\prime } \\sim \\mathcal { I } } } \\left( \\underbrace { T } _ { \\mathrm { t - 1 } } \\log p ( y _ { t } | x _ { 1 } \\{ f _ { 1 } \\} , . . . , x _ { T ^ { \\prime } } \\{ f _ { T ^ { \\prime } } \\} ; \\theta ) \\right)} _ { \\mathrm { T r a n s l a t i o n ~ L o s s } } + \\underbrace { \\sum _ { t ^ { \\prime } = 1 } ^ { T ^ { \\prime } } \\log p _ { F } ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta ) } _ { \\mathrm { F e r t i l i t y ~ L o s s } } + \\mathcal { H } ( q ) \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 412, + 301, + 1286, + 301, + 1286, + 415, + 412, + 415 + ], + "score": 0.92, + "latex": "\\hat { Y } _ { \\mathrm { a v e r a g e } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta ) , \\mathrm { w h e r e } \\hat { f } _ { t ^ { \\prime } } = \\mathrm { R o u n d } \\left( \\sum _ { f _ { t ^ { \\prime } } = 1 } ^ { L } p _ { F } \\big ( f _ { t ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } ; \\theta \\big ) f _ { t ^ { \\prime } } \\right)" + }, + { + "category_id": 14, + "poly": [ + 517, + 587, + 1183, + 587, + 1183, + 652, + 517, + 652 + ], + "score": 0.92, + "latex": "\\hat { Y } _ { \\mathrm { N P D } } = G ( x _ { 1 : T ^ { \\prime } } , \\underset { f _ { t ^ { \\prime } } \\sim p _ { F } } { \\mathrm { a r g m a x } } p _ { \\mathcal { A R } } ( G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta ) | X ; \\theta ) ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 551, + 1017, + 798, + 1017, + 798, + 1051, + 551, + 1051 + ], + "score": 0.92, + "latex": "p ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1048, + 516, + 1048, + 516, + 1083, + 297, + 1083 + ], + "score": 0.92, + "latex": "q ( f _ { 1 : T ^ { \\prime } } | x _ { 1 : T ^ { \\prime } } , y _ { 1 : T } )" + }, + { + "category_id": 13, + "poly": [ + 1102, + 988, + 1159, + 988, + 1159, + 1020, + 1102, + 1020 + ], + "score": 0.89, + "latex": "f _ { 1 : T ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 932, + 1647, + 968, + 1647, + 968, + 1676, + 932, + 1676 + ], + "score": 0.85, + "latex": "p _ { F }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1648, + 391, + 1648, + 391, + 1676, + 373, + 1676 + ], + "score": 0.79, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 558, + 1540, + 575, + 1540, + 575, + 1568, + 558, + 1568 + ], + "score": 0.78, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 678, + 1420, + 696, + 1420, + 696, + 1447, + 678, + 1447 + ], + "score": 0.75, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 414, + 336, + 728, + 336, + 728, + 377, + 414, + 377 + ], + "score": 0.29, + "latex": "\\hat { Y } _ { \\mathrm { a v e r a g e } } = G ( x _ { 1 : T ^ { \\prime } } , \\hat { f } _ { 1 : T ^ { \\prime } } ; \\theta )" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 912.0, + 502.0, + 912.0, + 502.0, + 958.0, + 291.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 818.0, + 72.0, + 818.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2116.0, + 840.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1714.0, + 932.0, + 1714.0, + 932.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1839.0, + 293.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1404.0, + 1834.0, + 1404.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 1404.0, + 1865.0, + 1404.0, + 1897.0, + 294.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1894.0, + 1405.0, + 1894.0, + 1405.0, + 1930.0, + 293.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1926.0, + 946.0, + 1926.0, + 946.0, + 1958.0, + 298.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 488.0, + 293.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 479.0, + 1406.0, + 479.0, + 1406.0, + 518.0, + 293.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 511.0, + 1403.0, + 511.0, + 1403.0, + 548.0, + 293.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 543.0, + 603.0, + 543.0, + 603.0, + 575.0, + 294.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 746.0, + 1403.0, + 746.0, + 1403.0, + 779.0, + 294.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 778.0, + 1405.0, + 778.0, + 1405.0, + 810.0, + 294.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 808.0, + 1407.0, + 808.0, + 1407.0, + 844.0, + 293.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 839.0, + 768.0, + 839.0, + 768.0, + 871.0, + 294.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1413.0, + 677.0, + 1413.0, + 677.0, + 1451.0, + 296.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 697.0, + 1413.0, + 1404.0, + 1413.0, + 1404.0, + 1451.0, + 697.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1441.0, + 1404.0, + 1441.0, + 1404.0, + 1480.0, + 295.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1473.0, + 1404.0, + 1473.0, + 1404.0, + 1510.0, + 294.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1502.0, + 1404.0, + 1502.0, + 1404.0, + 1542.0, + 293.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1537.0, + 557.0, + 1537.0, + 557.0, + 1570.0, + 296.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1537.0, + 759.0, + 1537.0, + 759.0, + 1570.0, + 576.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1578.0, + 1404.0, + 1578.0, + 1404.0, + 1619.0, + 293.0, + 1619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1612.0, + 1404.0, + 1612.0, + 1404.0, + 1648.0, + 295.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1640.0, + 372.0, + 1640.0, + 372.0, + 1681.0, + 292.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1640.0, + 931.0, + 1640.0, + 931.0, + 1681.0, + 392.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1640.0, + 980.0, + 1640.0, + 980.0, + 1681.0, + 969.0, + 1681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 985.0, + 1101.0, + 985.0, + 1101.0, + 1023.0, + 294.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 985.0, + 1405.0, + 985.0, + 1405.0, + 1023.0, + 1160.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 550.0, + 1016.0, + 550.0, + 1054.0, + 293.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1016.0, + 1406.0, + 1016.0, + 1406.0, + 1054.0, + 799.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1047.0, + 1405.0, + 1047.0, + 1405.0, + 1082.0, + 517.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 667.0, + 1403.0, + 667.0, + 1403.0, + 704.0, + 293.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 699.0, + 1403.0, + 699.0, + 1403.0, + 735.0, + 295.0, + 735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 1403.0, + 227.0, + 1403.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 494.0, + 261.0, + 494.0, + 295.0, + 295.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1404.0, + 2004.0, + 1404.0, + 2034.0, + 295.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1714.0, + 932.0, + 1714.0, + 932.0, + 1750.0, + 295.0, + 1750.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1424, + 1405, + 1424, + 1405, + 1671, + 297, + 1671 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 446, + 1403, + 446, + 1403, + 600, + 297, + 600 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 353, + 298, + 353 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 901, + 1404, + 901, + 1404, + 994, + 297, + 994 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1796, + 1404, + 1796, + 1404, + 1920, + 299, + 1920 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1688, + 1400, + 1688, + 1400, + 1783, + 298, + 1783 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 298, + 615, + 1402, + 615, + 1402, + 708, + 298, + 708 + ], + "score": 0.963 + }, + { + "category_id": 2, + "poly": [ + 331, + 1943, + 802, + 1943, + 802, + 2035, + 331, + 2035 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 296, + 823, + 1406, + 823, + 1406, + 887, + 296, + 887 + ], + "score": 0.95 + }, + { + "category_id": 8, + "poly": [ + 398, + 720, + 1300, + 720, + 1300, + 812, + 398, + 812 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 299, + 1382, + 681, + 1382, + 681, + 1414, + 299, + 1414 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 299, + 1313, + 558, + 1313, + 558, + 1349, + 299, + 1349 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 299, + 389, + 532, + 389, + 532, + 421, + 299, + 421 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 299, + 1176, + 1401, + 1176, + 1401, + 1271, + 299, + 1271 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1352, + 749, + 1399, + 749, + 1399, + 780, + 1352, + 780 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1354, + 1067, + 1400, + 1067, + 1400, + 1098, + 1354, + 1098 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.713 + }, + { + "category_id": 3, + "poly": [ + 309, + 1006, + 1341, + 1006, + 1341, + 1165, + 309, + 1165 + ], + "score": 0.552 + }, + { + "category_id": 8, + "poly": [ + 309, + 1006, + 1341, + 1006, + 1341, + 1165, + 309, + 1165 + ], + "score": 0.527 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.132 + }, + { + "category_id": 4, + "poly": [ + 299, + 1176, + 1401, + 1176, + 1401, + 1271, + 299, + 1271 + ], + "score": 0.129 + }, + { + "category_id": 14, + "poly": [ + 308, + 1002, + 1344, + 1002, + 1344, + 1167, + 308, + 1167 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { \\mathrm { F T } } = \\lambda \\left( \\underbrace { \\mathbb { E } } _ { \\int _ { \\mathrm { I : T : } } r \\sim p _ { F } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) - \\mathcal { L } _ { \\mathrm { R K L } } \\left( \\bar { f } _ { 1 : T ^ { \\prime } } \\right) \\right) + \\underbrace { \\mathbb { E } } _ { f _ { 1 : T ^ { \\prime } \\sim q } } \\left( \\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } \\right) \\right) _ { \\mathcal { L } _ { \\mathrm { R P } } } \\right) + ( 1 - \\lambda ) \\mathcal { L } _ { \\mathrm { K D } } ," + }, + { + "category_id": 14, + "poly": [ + 399, + 716, + 1299, + 716, + 1299, + 815, + 399, + 815 + ], + "score": 0.93, + "latex": "\\mathcal { L } _ { \\mathrm { R K L } } \\left( f _ { 1 : T ^ { \\prime } } ; \\theta \\right) = \\sum _ { t = 1 } ^ { T } \\sum _ { y _ { t } } \\left[ \\log p _ { \\mathcal { A R } } \\left( y _ { t } | \\hat { y } _ { 1 : t - 1 } , x _ { 1 : T ^ { \\prime } } \\right) \\cdot p _ { \\mathcal { N A } } \\left( y _ { t } | x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta \\right) \\right] ," + }, + { + "category_id": 13, + "poly": [ + 371, + 824, + 654, + 824, + 654, + 858, + 371, + 858 + ], + "score": 0.92, + "latex": "\\hat { y } _ { 1 : T } = G ( x _ { 1 : T ^ { \\prime } } , f _ { 1 : T ^ { \\prime } } ; \\theta )" + }, + { + "category_id": 13, + "poly": [ + 376, + 1175, + 434, + 1175, + 434, + 1210, + 376, + 1210 + ], + "score": 0.91, + "latex": "{ \\bar { f } } _ { 1 : T ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 454, + 1209, + 502, + 1209, + 502, + 1239, + 454, + 1239 + ], + "score": 0.9, + "latex": "{ \\mathcal { L } } _ { \\mathrm { R L } }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 1208, + 1342, + 1208, + 1342, + 1240, + 1295, + 1240 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { \\mathrm { B P } }" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1940.0, + 539.0, + 1940.0, + 539.0, + 1974.0, + 331.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1970.0, + 803.0, + 1970.0, + 803.0, + 2006.0, + 331.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 2000.0, + 804.0, + 2000.0, + 804.0, + 2038.0, + 331.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1380.0, + 684.0, + 1380.0, + 684.0, + 1417.0, + 294.0, + 1417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1309.0, + 562.0, + 1309.0, + 562.0, + 1354.0, + 291.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 386.0, + 537.0, + 386.0, + 537.0, + 426.0, + 293.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.75, + 1133.0, + 1056.75, + 1133.0, + 1056.75, + 1167.5, + 1012.75, + 1167.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1175.0, + 375.0, + 1175.0, + 375.0, + 1213.0, + 295.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1213.0, + 435.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 453.0, + 1204.0, + 453.0, + 1243.0, + 294.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1204.0, + 1294.0, + 1204.0, + 1294.0, + 1243.0, + 503.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1204.0, + 1405.0, + 1204.0, + 1405.0, + 1243.0, + 1343.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 822.0, + 1239.0, + 822.0, + 1273.0, + 295.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1454.0, + 1405.0, + 1454.0, + 1405.0, + 1487.0, + 295.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1485.0, + 1405.0, + 1485.0, + 1405.0, + 1520.0, + 294.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1516.0, + 1406.0, + 1516.0, + 1406.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1584.0, + 294.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1576.0, + 1402.0, + 1576.0, + 1402.0, + 1610.0, + 295.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1607.0, + 1406.0, + 1607.0, + 1406.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1637.0, + 620.0, + 1637.0, + 620.0, + 1674.0, + 295.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 448.0, + 1405.0, + 448.0, + 1405.0, + 481.0, + 296.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 474.0, + 1405.0, + 474.0, + 1405.0, + 514.0, + 292.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 507.0, + 1406.0, + 507.0, + 1406.0, + 545.0, + 292.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 535.0, + 1404.0, + 535.0, + 1404.0, + 577.0, + 292.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 571.0, + 537.0, + 571.0, + 537.0, + 602.0, + 293.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 231.0, + 1403.0, + 231.0, + 1403.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 262.0, + 1405.0, + 262.0, + 1405.0, + 294.0, + 293.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 292.0, + 1405.0, + 292.0, + 1405.0, + 325.0, + 294.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1241.0, + 321.0, + 1241.0, + 356.0, + 292.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 902.0, + 1404.0, + 902.0, + 1404.0, + 935.0, + 295.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 933.0, + 1406.0, + 933.0, + 1406.0, + 967.0, + 295.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 964.0, + 972.0, + 964.0, + 972.0, + 995.0, + 296.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1796.0, + 1405.0, + 1796.0, + 1405.0, + 1832.0, + 294.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1829.0, + 1404.0, + 1829.0, + 1404.0, + 1862.0, + 295.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1859.0, + 1404.0, + 1859.0, + 1404.0, + 1892.0, + 295.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 1077.0, + 1888.0, + 1077.0, + 1923.0, + 293.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1686.0, + 1405.0, + 1686.0, + 1405.0, + 1727.0, + 293.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1718.0, + 1406.0, + 1718.0, + 1406.0, + 1756.0, + 294.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1748.0, + 1130.0, + 1748.0, + 1130.0, + 1788.0, + 292.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 614.0, + 1403.0, + 614.0, + 1403.0, + 652.0, + 294.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 647.0, + 1404.0, + 647.0, + 1404.0, + 681.0, + 296.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 675.0, + 713.0, + 675.0, + 713.0, + 711.0, + 294.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 823.0, + 370.0, + 823.0, + 370.0, + 859.0, + 296.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 823.0, + 1404.0, + 823.0, + 1404.0, + 859.0, + 655.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 854.0, + 948.0, + 854.0, + 948.0, + 888.0, + 296.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1175.0, + 375.0, + 1175.0, + 375.0, + 1213.0, + 295.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1175.0, + 1405.0, + 1175.0, + 1405.0, + 1213.0, + 435.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1204.0, + 453.0, + 1204.0, + 453.0, + 1243.0, + 294.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1204.0, + 1294.0, + 1204.0, + 1294.0, + 1243.0, + 503.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1204.0, + 1405.0, + 1204.0, + 1405.0, + 1243.0, + 1343.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 822.0, + 1239.0, + 822.0, + 1273.0, + 295.0, + 1273.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 295, + 224, + 1411, + 224, + 1411, + 495, + 295, + 495 + ], + "score": 0.984, + "html": "
ModelsWMT14WMT16IWSLT16
En→DeDe→EnEn→RoRo→EnEn→DeLatency / Speedup
NAT17.3520.6226.2227.8325.2039 ms15.6×
NAT (+FT)17.6921.4727.2929.0626.5239 ms15.6×
NAT (+FT + NPD s = 10)18.6622.4129.0230.7627.4479 ms7.68×
NAT (+FT + NPD s = 100)19.1723.2029.7931.4428.16257 ms2.36×
Autoregressive (b = 1)22.7126.3931.3531.0328.89408 ms1.49×
Autoregressive (b = 4)23.4527.0231.9131.7629.70607ms1.00×
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1096, + 878, + 1096, + 878, + 1279, + 298, + 1279 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 676, + 880, + 676, + 880, + 890, + 297, + 890 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 917, + 879, + 917, + 879, + 1069, + 297, + 1069 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1817, + 1403, + 1817, + 1403, + 1941, + 299, + 1941 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1679, + 1403, + 1679, + 1403, + 1801, + 300, + 1801 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1306, + 1403, + 1306, + 1403, + 1459, + 299, + 1459 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 912, + 663, + 1354, + 663, + 1354, + 957, + 912, + 957 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 902, + 989, + 1404, + 989, + 1404, + 1204, + 902, + 1204 + ], + "score": 0.969 + }, + { + "category_id": 2, + "poly": [ + 330, + 1974, + 769, + 1974, + 769, + 2034, + 330, + 2034 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 297, + 518, + 1405, + 518, + 1405, + 641, + 297, + 641 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 300, + 1485, + 1396, + 1485, + 1396, + 1518, + 300, + 1518 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 296, + 1543, + 1244, + 1543, + 1244, + 1575, + 296, + 1575 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 300, + 1618, + 474, + 1618, + 474, + 1650, + 300, + 1650 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.794 + }, + { + "category_id": 13, + "poly": [ + 926, + 1399, + 1092, + 1399, + 1092, + 1432, + 926, + 1432 + ], + "score": 0.88, + "latex": "t _ { \\mathrm { w a r m u p } } = 7 4 6 )" + }, + { + "category_id": 13, + "poly": [ + 646, + 1431, + 751, + 1431, + 751, + 1457, + 646, + 1457 + ], + "score": 0.87, + "latex": "\\lambda = 0 . 2 5" + }, + { + "category_id": 13, + "poly": [ + 475, + 429, + 541, + 429, + 541, + 455, + 475, + 455 + ], + "score": 0.66, + "latex": "( b = 1 )" + }, + { + "category_id": 13, + "poly": [ + 1166, + 1543, + 1233, + 1543, + 1233, + 1574, + 1166, + 1574 + ], + "score": 0.62, + "latex": "\\mathrm { N A T ^ { 6 } }" + }, + { + "category_id": 13, + "poly": [ + 372, + 333, + 429, + 333, + 429, + 358, + 372, + 358 + ], + "score": 0.55, + "latex": "( + \\mathrm { F T } )" + }, + { + "category_id": 13, + "poly": [ + 754, + 1400, + 867, + 1400, + 867, + 1430, + 754, + 1430 + ], + "score": 0.54, + "latex": "n _ { \\mathrm { h e a d } } = 2" + }, + { + "category_id": 13, + "poly": [ + 474, + 458, + 544, + 458, + 544, + 484, + 474, + 484 + ], + "score": 0.5, + "latex": "( b = 4 )" + }, + { + "category_id": 13, + "poly": [ + 301, + 1399, + 859, + 1399, + 859, + 1430, + 301, + 1430 + ], + "score": 0.45, + "latex": "( d _ { \\mathrm { m o d e l } } = 2 8 7 , d _ { \\mathrm { h i d d e n } } = 5 0 7 , n _ { \\mathrm { l a y e r } } = 5 , n _ { \\mathrm { h e a d } } = \\mathrm { \\Omega } ^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 305, + 1399, + 457, + 1399, + 457, + 1430, + 305, + 1430 + ], + "score": 0.42, + "latex": "\\dot { d } _ { \\mathrm { m o d e l } } = 2 8 7" + }, + { + "category_id": 13, + "poly": [ + 973, + 1879, + 1038, + 1879, + 1038, + 1908, + 973, + 1908 + ], + "score": 0.39, + "latex": "8 0 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 629, + 1400, + 745, + 1400, + 745, + 1430, + 629, + 1430 + ], + "score": 0.38, + "latex": "n _ { \\mathrm { l a y e r } } = 5" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 676.0, + 967.0, + 676.0, + 967.0, + 699.0, + 928.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 675.0, + 1338.0, + 675.0, + 1338.0, + 683.0, + 1327.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 713.0, + 970.0, + 713.0, + 970.0, + 737.0, + 927.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 751.0, + 967.0, + 751.0, + 967.0, + 831.0, + 912.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 755.0, + 1310.0, + 755.0, + 1310.0, + 775.0, + 1249.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 769.0, + 1336.0, + 769.0, + 1336.0, + 796.0, + 1247.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 791.0, + 965.0, + 791.0, + 965.0, + 813.0, + 928.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 786.0, + 1333.0, + 786.0, + 1333.0, + 812.0, + 1247.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 805.0, + 1346.0, + 805.0, + 1346.0, + 827.0, + 1248.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 828.0, + 968.0, + 828.0, + 968.0, + 853.0, + 927.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 867.0, + 967.0, + 867.0, + 967.0, + 891.0, + 928.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 900.0, + 990.0, + 900.0, + 990.0, + 910.0, + 980.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 914.0, + 1000.0, + 914.0, + 1000.0, + 939.0, + 969.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 917.0, + 1085.0, + 917.0, + 1085.0, + 938.0, + 1054.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 916.0, + 1263.0, + 916.0, + 1263.0, + 939.0, + 1231.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 933.0, + 1199.0, + 933.0, + 1199.0, + 959.0, + 1119.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.75, + 914.5, + 1178.75, + 914.5, + 1178.75, + 941.5, + 1138.75, + 941.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1314.25, + 916.5, + 1351.25, + 916.5, + 1351.25, + 939.5, + 1314.25, + 939.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 987.0, + 1405.0, + 987.0, + 1405.0, + 1024.0, + 897.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1024.0, + 1404.0, + 1024.0, + 1404.0, + 1054.0, + 902.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1050.0, + 1405.0, + 1050.0, + 1405.0, + 1086.0, + 899.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1082.0, + 1405.0, + 1082.0, + 1405.0, + 1118.0, + 899.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1113.0, + 1404.0, + 1113.0, + 1404.0, + 1147.0, + 900.0, + 1147.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1144.0, + 1404.0, + 1144.0, + 1404.0, + 1177.0, + 899.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1173.0, + 1111.0, + 1173.0, + 1111.0, + 1205.0, + 900.0, + 1205.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1971.0, + 678.0, + 1971.0, + 678.0, + 2005.0, + 331.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 2002.0, + 772.0, + 2002.0, + 772.0, + 2036.0, + 332.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1613.0, + 478.0, + 1613.0, + 478.0, + 1656.0, + 293.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1096.0, + 747.0, + 1096.0, + 747.0, + 1131.0, + 296.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 1095.0, + 882.0, + 1095.0, + 882.0, + 1129.0, + 794.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1128.0, + 882.0, + 1128.0, + 882.0, + 1161.0, + 295.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1158.0, + 880.0, + 1158.0, + 880.0, + 1191.0, + 295.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1185.0, + 884.0, + 1185.0, + 884.0, + 1222.0, + 293.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1217.0, + 882.0, + 1217.0, + 882.0, + 1252.0, + 295.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1249.0, + 726.0, + 1249.0, + 726.0, + 1281.0, + 297.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 676.0, + 883.0, + 676.0, + 883.0, + 707.0, + 295.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 705.0, + 880.0, + 705.0, + 880.0, + 738.0, + 295.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 737.0, + 881.0, + 737.0, + 881.0, + 770.0, + 295.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 882.0, + 766.0, + 882.0, + 802.0, + 294.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 796.0, + 882.0, + 796.0, + 882.0, + 832.0, + 293.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 829.0, + 883.0, + 829.0, + 883.0, + 862.0, + 295.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 861.0, + 580.0, + 861.0, + 580.0, + 892.0, + 296.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 916.0, + 880.0, + 916.0, + 880.0, + 947.0, + 297.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 880.0, + 948.0, + 880.0, + 978.0, + 295.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 977.0, + 881.0, + 977.0, + 881.0, + 1011.0, + 295.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1010.0, + 881.0, + 1010.0, + 881.0, + 1040.0, + 296.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1040.0, + 647.0, + 1040.0, + 647.0, + 1071.0, + 295.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1851.0, + 295.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1404.0, + 1850.0, + 1404.0, + 1881.0, + 294.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 972.0, + 1880.0, + 972.0, + 1913.0, + 294.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1880.0, + 1403.0, + 1880.0, + 1403.0, + 1913.0, + 1039.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1907.0, + 1296.0, + 1907.0, + 1296.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1680.0, + 1403.0, + 1680.0, + 1403.0, + 1712.0, + 296.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1711.0, + 1403.0, + 1711.0, + 1403.0, + 1743.0, + 295.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1741.0, + 1407.0, + 1741.0, + 1407.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1771.0, + 1344.0, + 1771.0, + 1344.0, + 1804.0, + 296.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1305.0, + 1405.0, + 1305.0, + 1405.0, + 1342.0, + 295.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1335.0, + 1405.0, + 1335.0, + 1405.0, + 1373.0, + 293.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1364.0, + 1405.0, + 1364.0, + 1405.0, + 1404.0, + 293.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1394.0, + 300.0, + 1394.0, + 300.0, + 1438.0, + 293.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1394.0, + 925.0, + 1394.0, + 925.0, + 1438.0, + 868.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1394.0, + 1407.0, + 1394.0, + 1407.0, + 1438.0, + 1093.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 645.0, + 1426.0, + 645.0, + 1463.0, + 294.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1426.0, + 762.0, + 1426.0, + 762.0, + 1463.0, + 752.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 517.0, + 1405.0, + 517.0, + 1405.0, + 552.0, + 294.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 549.0, + 1403.0, + 549.0, + 1403.0, + 582.0, + 295.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 580.0, + 1407.0, + 580.0, + 1407.0, + 614.0, + 292.0, + 614.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 609.0, + 1337.0, + 609.0, + 1337.0, + 641.0, + 292.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 1399.0, + 1481.0, + 1399.0, + 1522.0, + 294.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1538.0, + 1165.0, + 1538.0, + 1165.0, + 1582.0, + 291.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1538.0, + 1246.0, + 1538.0, + 1246.0, + 1582.0, + 1234.0, + 1582.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1404, + 1942, + 1404, + 2035, + 300, + 2035 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 1133, + 1404, + 1133, + 1404, + 1258, + 297, + 1258 + ], + "score": 0.969 + }, + { + "category_id": 5, + "poly": [ + 341, + 503, + 1358, + 503, + 1358, + 954, + 341, + 954 + ], + "score": 0.963, + "html": "
Distillation b=1 b=4Decoder Inputs +uniform+fertility+PosAttFine-tuning +LKD +LBPBLEU +CRLBLEU (T)
~2
√16.51
√18.87
√√√20.72
√√21.12
<√24.0243.91
√25.2045.41
√√22.44 厂
兴 √√××
√√××
√√√√25.7646.11
√√√√26.5247.38
" + }, + { + "category_id": 1, + "poly": [ + 298, + 287, + 1404, + 287, + 1404, + 473, + 298, + 473 + ], + "score": 0.962 + }, + { + "category_id": 0, + "poly": [ + 299, + 230, + 581, + 230, + 581, + 261, + 299, + 261 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.794 + }, + { + "category_id": 5, + "poly": [ + 314, + 1316, + 1370, + 1316, + 1370, + 1764, + 314, + 1764 + ], + "score": 0.602, + "html": "
Source:politicians try to pick wordsand use words to shape realityand control reality,butin fact,reality changes words
far more than words can ever change reality.
Target: AR:Politiker versuchen Worte zu benutzen,um die Realitat zu formen und die Realitat zu kontrolieren,aber
tatsächlich verandert die Realitat Worte viel mehr,als Worte die Realitat jemals verändern konnten.
Politikerversuchen Worter zu wahlen und Worter zur Realitat zu gestalten und Realitätzu steuern,aber in
NAT:Wirklichkeit verändert sich die Realitat viel mehrals Worte,die die Realität verändern konnen.
Politikerversuchen,Worter wahlen und zu verwenden,um Realitatzu formen und Realitatzu formen,aber
NAT+NPD:tatsächlich ändert Realitat Realitat viel mehrals Wortedie Realität Realitat verändern.
Politikerversuchen,Worter wahlenund zu verwenden,um Realitat Realitat formenund die Realitatzu formen， aber tatsächlich ändert die Realität Worte viel mehrals Worte jemals die Realität verändern konnen.
Source:Isee wheelchairs bought and sold like used cars.
Target:ich erlebe,dass Rollstuhle gekauft und verkauft werden wie Gebrauchtwagen
AR: NAT:ich sehe Rollstuhlen,die wie Autos verkauft und verkauft werden.
NAT+NPD:ich sehe,dass Stuhle Stuhle und verkauftwie Autos verkauft.
ich sehe Rollühle kauften und verkaufte wie Autos.
" + }, + { + "category_id": 6, + "poly": [ + 297, + 1814, + 1404, + 1814, + 1404, + 1909, + 297, + 1909 + ], + "score": 0.521 + }, + { + "category_id": 1, + "poly": [ + 296, + 977, + 1404, + 977, + 1404, + 1101, + 296, + 1101 + ], + "score": 0.476 + }, + { + "category_id": 3, + "poly": [ + 314, + 1316, + 1370, + 1316, + 1370, + 1764, + 314, + 1764 + ], + "score": 0.377 + }, + { + "category_id": 4, + "poly": [ + 296, + 977, + 1404, + 977, + 1404, + 1101, + 296, + 1101 + ], + "score": 0.168 + }, + { + "category_id": 13, + "poly": [ + 1141, + 1135, + 1189, + 1135, + 1189, + 1165, + 1141, + 1165 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { \\mathrm { B P } }" + }, + { + "category_id": 13, + "poly": [ + 1269, + 1010, + 1296, + 1010, + 1296, + 1037, + 1269, + 1037 + ], + "score": 0.74, + "latex": "\\times" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 586.0, + 227.0, + 586.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1814.0, + 1402.0, + 1814.0, + 1402.0, + 1852.0, + 295.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1846.0, + 1402.0, + 1846.0, + 1402.0, + 1880.0, + 295.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1872.0, + 745.0, + 1872.0, + 745.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1318.0, + 428.0, + 1318.0, + 428.0, + 1347.0, + 357.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1320.0, + 1355.0, + 1320.0, + 1355.0, + 1345.0, + 439.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1343.0, + 810.0, + 1343.0, + 810.0, + 1372.0, + 438.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1374.0, + 429.0, + 1374.0, + 429.0, + 1407.0, + 361.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1375.0, + 1298.0, + 1375.0, + 1298.0, + 1403.0, + 437.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1401.0, + 1260.0, + 1401.0, + 1260.0, + 1429.0, + 437.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1432.0, + 429.0, + 1432.0, + 429.0, + 1460.0, + 389.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1432.0, + 1306.0, + 1432.0, + 1306.0, + 1461.0, + 437.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1458.0, + 1213.0, + 1458.0, + 1213.0, + 1486.0, + 437.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1490.0, + 430.0, + 1490.0, + 430.0, + 1519.0, + 377.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1490.0, + 1325.0, + 1490.0, + 1325.0, + 1518.0, + 437.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1519.0, + 1139.0, + 1519.0, + 1139.0, + 1542.0, + 440.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 1551.0, + 428.0, + 1551.0, + 428.0, + 1574.0, + 333.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1549.0, + 1356.0, + 1549.0, + 1356.0, + 1577.0, + 437.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1573.0, + 1257.0, + 1573.0, + 1257.0, + 1602.0, + 437.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1607.0, + 428.0, + 1607.0, + 428.0, + 1635.0, + 357.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1610.0, + 843.0, + 1610.0, + 843.0, + 1633.0, + 440.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 1634.0, + 431.0, + 1634.0, + 431.0, + 1670.0, + 359.0, + 1670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1638.0, + 1076.0, + 1638.0, + 1076.0, + 1665.0, + 439.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1668.0, + 429.0, + 1668.0, + 429.0, + 1695.0, + 389.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1668.0, + 983.0, + 1668.0, + 983.0, + 1694.0, + 439.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1698.0, + 429.0, + 1698.0, + 429.0, + 1723.0, + 378.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1699.0, + 947.0, + 1699.0, + 947.0, + 1722.0, + 440.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1726.0, + 429.0, + 1726.0, + 429.0, + 1754.0, + 331.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1726.0, + 863.0, + 1726.0, + 863.0, + 1754.0, + 437.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 975.0, + 1405.0, + 975.0, + 1405.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1008.0, + 1268.0, + 1008.0, + 1268.0, + 1042.0, + 291.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1008.0, + 1406.0, + 1008.0, + 1406.0, + 1042.0, + 1297.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1405.0, + 1039.0, + 1405.0, + 1075.0, + 294.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1068.0, + 1254.0, + 1068.0, + 1254.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1405.0, + 1974.0, + 1405.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1133.0, + 1140.0, + 1133.0, + 1140.0, + 1169.0, + 294.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1133.0, + 1405.0, + 1133.0, + 1405.0, + 1169.0, + 1190.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1161.0, + 1405.0, + 1161.0, + 1405.0, + 1203.0, + 292.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1194.0, + 1402.0, + 1194.0, + 1402.0, + 1229.0, + 294.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1224.0, + 622.0, + 1224.0, + 622.0, + 1264.0, + 292.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 286.0, + 1405.0, + 286.0, + 1405.0, + 323.0, + 295.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 351.0, + 1404.0, + 351.0, + 1404.0, + 383.0, + 296.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 374.0, + 1405.0, + 374.0, + 1405.0, + 420.0, + 292.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 405.0, + 1406.0, + 405.0, + 1406.0, + 450.0, + 292.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 441.0, + 881.0, + 441.0, + 881.0, + 476.0, + 293.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 975.0, + 1405.0, + 975.0, + 1405.0, + 1013.0, + 294.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1008.0, + 1268.0, + 1008.0, + 1268.0, + 1042.0, + 291.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1008.0, + 1406.0, + 1008.0, + 1406.0, + 1042.0, + 1297.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1039.0, + 1405.0, + 1039.0, + 1405.0, + 1075.0, + 294.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1068.0, + 1254.0, + 1068.0, + 1254.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1140, + 1404, + 1140, + 1404, + 1263, + 299, + 1263 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 822, + 1404, + 822, + 1404, + 944, + 298, + 944 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 355, + 293, + 839, + 293, + 839, + 563, + 355, + 563 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 887, + 293, + 1345, + 293, + 1345, + 564, + 887, + 564 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 298, + 960, + 1400, + 960, + 1400, + 1022, + 298, + 1022 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 300, + 1364, + 1400, + 1364, + 1400, + 1426, + 300, + 1426 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 300, + 1068, + 544, + 1068, + 544, + 1105, + 300, + 1105 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 815, + 76, + 815, + 104, + 300, + 104 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 299, + 1312, + 488, + 1312, + 488, + 1345, + 299, + 1345 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 837, + 2088, + 865, + 2088, + 865, + 2112, + 837, + 2112 + ], + "score": 0.837 + }, + { + "category_id": 1, + "poly": [ + 296, + 609, + 1406, + 609, + 1406, + 763, + 296, + 763 + ], + "score": 0.805 + }, + { + "category_id": 1, + "poly": [ + 361, + 238, + 729, + 238, + 729, + 265, + 361, + 265 + ], + "score": 0.75 + }, + { + "category_id": 1, + "poly": [ + 292, + 2002, + 1301, + 2002, + 1301, + 2034, + 292, + 2034 + ], + "score": 0.741 + }, + { + "category_id": 1, + "poly": [ + 299, + 1567, + 1399, + 1567, + 1399, + 1631, + 299, + 1631 + ], + "score": 0.707 + }, + { + "category_id": 1, + "poly": [ + 300, + 1653, + 1401, + 1653, + 1401, + 1717, + 300, + 1717 + ], + "score": 0.672 + }, + { + "category_id": 1, + "poly": [ + 297, + 1449, + 1406, + 1449, + 1406, + 1543, + 297, + 1543 + ], + "score": 0.67 + }, + { + "category_id": 1, + "poly": [ + 295, + 1740, + 1401, + 1740, + 1401, + 1806, + 295, + 1806 + ], + "score": 0.614 + }, + { + "category_id": 1, + "poly": [ + 299, + 1915, + 1406, + 1915, + 1406, + 1978, + 299, + 1978 + ], + "score": 0.576 + }, + { + "category_id": 1, + "poly": [ + 293, + 1828, + 1401, + 1828, + 1401, + 1892, + 293, + 1892 + ], + "score": 0.548 + }, + { + "category_id": 4, + "poly": [ + 296, + 609, + 1406, + 609, + 1406, + 763, + 296, + 763 + ], + "score": 0.324 + }, + { + "category_id": 0, + "poly": [ + 361, + 238, + 729, + 238, + 729, + 265, + 361, + 265 + ], + "score": 0.108 + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 549.0, + 1065.0, + 549.0, + 1113.0, + 292.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1312.0, + 490.0, + 1312.0, + 490.0, + 1349.0, + 297.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 611.0, + 1403.0, + 611.0, + 1403.0, + 644.0, + 295.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 637.0, + 1406.0, + 637.0, + 1406.0, + 677.0, + 293.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 667.0, + 1404.0, + 667.0, + 1404.0, + 707.0, + 293.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1408.0, + 699.0, + 1408.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 730.0, + 1072.0, + 730.0, + 1072.0, + 769.0, + 293.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 236.0, + 731.0, + 236.0, + 731.0, + 266.0, + 357.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1139.0, + 1404.0, + 1139.0, + 1404.0, + 1175.0, + 294.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1171.0, + 1405.0, + 1171.0, + 1405.0, + 1204.0, + 295.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1202.0, + 1407.0, + 1202.0, + 1407.0, + 1234.0, + 293.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1232.0, + 1219.0, + 1232.0, + 1219.0, + 1268.0, + 294.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 819.0, + 1405.0, + 819.0, + 1405.0, + 859.0, + 293.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 852.0, + 1406.0, + 852.0, + 1406.0, + 888.0, + 293.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 881.0, + 1405.0, + 881.0, + 1405.0, + 919.0, + 293.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 914.0, + 614.0, + 914.0, + 614.0, + 947.0, + 294.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 294.0, + 641.0, + 294.0, + 641.0, + 321.0, + 357.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 328.0, + 669.0, + 328.0, + 669.0, + 355.0, + 357.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 362.0, + 704.0, + 362.0, + 704.0, + 389.0, + 357.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 399.0, + 699.0, + 399.0, + 699.0, + 423.0, + 359.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 432.0, + 758.0, + 432.0, + 758.0, + 457.0, + 356.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 467.0, + 785.0, + 467.0, + 785.0, + 493.0, + 354.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 500.0, + 814.0, + 500.0, + 814.0, + 527.0, + 356.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 536.0, + 842.0, + 536.0, + 842.0, + 563.0, + 356.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 295.0, + 1199.0, + 295.0, + 1199.0, + 319.0, + 889.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 330.0, + 1205.0, + 330.0, + 1205.0, + 354.0, + 888.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 362.0, + 1184.0, + 362.0, + 1184.0, + 390.0, + 888.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 399.0, + 1227.0, + 399.0, + 1227.0, + 423.0, + 888.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 432.0, + 1226.0, + 432.0, + 1226.0, + 456.0, + 888.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 468.0, + 1323.0, + 468.0, + 1323.0, + 492.0, + 890.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 501.0, + 1336.0, + 501.0, + 1336.0, + 528.0, + 891.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 536.0, + 1348.0, + 536.0, + 1348.0, + 563.0, + 891.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 957.0, + 1402.0, + 957.0, + 1402.0, + 995.0, + 296.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 990.0, + 1117.0, + 990.0, + 1117.0, + 1026.0, + 293.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1360.0, + 1404.0, + 1360.0, + 1404.0, + 1401.0, + 294.0, + 1401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1394.0, + 844.0, + 1394.0, + 844.0, + 1427.0, + 321.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 611.0, + 1403.0, + 611.0, + 1403.0, + 644.0, + 295.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 637.0, + 1406.0, + 637.0, + 1406.0, + 677.0, + 293.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 667.0, + 1404.0, + 667.0, + 1404.0, + 707.0, + 293.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 699.0, + 1408.0, + 699.0, + 1408.0, + 736.0, + 294.0, + 736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 730.0, + 1072.0, + 730.0, + 1072.0, + 769.0, + 293.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 236.0, + 731.0, + 236.0, + 731.0, + 266.0, + 357.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1999.0, + 1304.0, + 1999.0, + 1304.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1566.0, + 1400.0, + 1566.0, + 1400.0, + 1602.0, + 296.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1599.0, + 775.0, + 1599.0, + 775.0, + 1632.0, + 321.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1653.0, + 1406.0, + 1653.0, + 1406.0, + 1689.0, + 296.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1685.0, + 692.0, + 1685.0, + 692.0, + 1717.0, + 324.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1451.0, + 1406.0, + 1451.0, + 1406.0, + 1485.0, + 294.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1481.0, + 1404.0, + 1481.0, + 1404.0, + 1515.0, + 321.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1511.0, + 397.0, + 1511.0, + 397.0, + 1545.0, + 323.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1741.0, + 1403.0, + 1741.0, + 1403.0, + 1777.0, + 295.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1773.0, + 1129.0, + 1773.0, + 1129.0, + 1805.0, + 322.0, + 1805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1403.0, + 1912.0, + 1403.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1945.0, + 1404.0, + 1945.0, + 1404.0, + 1980.0, + 323.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1828.0, + 1404.0, + 1828.0, + 1404.0, + 1864.0, + 295.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1860.0, + 1011.0, + 1860.0, + 1011.0, + 1892.0, + 324.0, + 1892.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 105, + 300, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2112, + 836, + 2112 + ], + "score": 0.834 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1402, + 229, + 1402, + 293, + 294, + 293 + ], + "score": 0.819 + }, + { + "category_id": 1, + "poly": [ + 295, + 311, + 1400, + 311, + 1400, + 377, + 295, + 377 + ], + "score": 0.688 + }, + { + "category_id": 1, + "poly": [ + 291, + 840, + 1402, + 840, + 1402, + 905, + 291, + 905 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 296, + 757, + 1398, + 757, + 1398, + 823, + 296, + 823 + ], + "score": 0.665 + }, + { + "category_id": 1, + "poly": [ + 294, + 674, + 1401, + 674, + 1401, + 738, + 294, + 738 + ], + "score": 0.651 + }, + { + "category_id": 1, + "poly": [ + 293, + 508, + 1401, + 508, + 1401, + 573, + 293, + 573 + ], + "score": 0.611 + }, + { + "category_id": 1, + "poly": [ + 293, + 592, + 1401, + 592, + 1401, + 656, + 293, + 656 + ], + "score": 0.609 + }, + { + "category_id": 1, + "poly": [ + 299, + 394, + 1400, + 394, + 1400, + 489, + 299, + 489 + ], + "score": 0.594 + }, + { + "category_id": 1, + "poly": [ + 300, + 923, + 1406, + 923, + 1406, + 1108, + 300, + 1108 + ], + "score": 0.455 + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 1405.0, + 226.0, + 1405.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 261.0, + 610.0, + 261.0, + 610.0, + 292.0, + 322.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1404.0, + 310.0, + 1404.0, + 350.0, + 294.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 343.0, + 908.0, + 343.0, + 908.0, + 375.0, + 324.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 842.0, + 1404.0, + 842.0, + 1404.0, + 878.0, + 294.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 873.0, + 908.0, + 873.0, + 908.0, + 905.0, + 323.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 760.0, + 1402.0, + 760.0, + 1402.0, + 793.0, + 297.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 789.0, + 1392.0, + 789.0, + 1392.0, + 825.0, + 324.0, + 825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 675.0, + 1402.0, + 675.0, + 1402.0, + 711.0, + 296.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 703.0, + 497.0, + 703.0, + 497.0, + 737.0, + 321.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 511.0, + 1402.0, + 511.0, + 1402.0, + 543.0, + 297.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 539.0, + 1026.0, + 539.0, + 1026.0, + 575.0, + 324.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 591.0, + 1402.0, + 591.0, + 1402.0, + 627.0, + 295.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 624.0, + 949.0, + 624.0, + 949.0, + 657.0, + 322.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 393.0, + 1401.0, + 393.0, + 1401.0, + 432.0, + 293.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 465.0, + 322.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 458.0, + 545.0, + 458.0, + 545.0, + 489.0, + 323.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 921.0, + 1406.0, + 921.0, + 1406.0, + 962.0, + 294.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 954.0, + 1404.0, + 954.0, + 1404.0, + 990.0, + 323.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 984.0, + 1405.0, + 984.0, + 1405.0, + 1021.0, + 320.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1014.0, + 1405.0, + 1014.0, + 1405.0, + 1052.0, + 322.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1042.0, + 1407.0, + 1042.0, + 1407.0, + 1083.0, + 319.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1072.0, + 396.0, + 1072.0, + 396.0, + 1111.0, + 320.0, + 1111.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 306, + 299, + 1388, + 299, + 1388, + 984, + 306, + 984 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 393, + 1193, + 1249, + 1193, + 1249, + 1741, + 393, + 1741 + ], + "score": 0.97 + }, + { + "category_id": 4, + "poly": [ + 295, + 1783, + 1405, + 1783, + 1405, + 1967, + 295, + 1967 + ], + "score": 0.958 + }, + { + "category_id": 4, + "poly": [ + 296, + 1015, + 1405, + 1015, + 1405, + 1080, + 296, + 1080 + ], + "score": 0.939 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.854 + }, + { + "category_id": 2, + "poly": [ + 302, + 228, + 754, + 228, + 754, + 261, + 302, + 261 + ], + "score": 0.648 + }, + { + "category_id": 0, + "poly": [ + 302, + 228, + 754, + 228, + 754, + 261, + 302, + 261 + ], + "score": 0.169 + }, + { + "category_id": 13, + "poly": [ + 698, + 1312, + 777, + 1312, + 777, + 1339, + 698, + 1339 + ], + "score": 0.35, + "latex": "\\scriptstyle \\left( 5 = 1 0 0 \\right)" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 336.0, + 772.0, + 336.0, + 772.0, + 379.0, + 567.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 336.0, + 1032.0, + 336.0, + 1032.0, + 363.0, + 893.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 357.0, + 430.0, + 357.0, + 430.0, + 392.0, + 314.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 355.0, + 1238.0, + 355.0, + 1238.0, + 395.0, + 1153.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 371.0, + 756.0, + 371.0, + 756.0, + 411.0, + 584.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 397.0, + 1202.0, + 397.0, + 1202.0, + 418.0, + 1182.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 446.0, + 670.0, + 446.0, + 670.0, + 481.0, + 517.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 493.0, + 881.0, + 493.0, + 881.0, + 520.0, + 790.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 530.0, + 459.0, + 530.0, + 459.0, + 554.0, + 389.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 525.0, + 744.0, + 525.0, + 744.0, + 560.0, + 601.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 550.0, + 463.0, + 550.0, + 463.0, + 573.0, + 371.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 460.0, + 569.0, + 476.0, + 569.0, + 476.0, + 623.0, + 460.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 480.0, + 584.0, + 574.0, + 584.0, + 574.0, + 612.0, + 480.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 581.0, + 854.0, + 581.0, + 854.0, + 615.0, + 752.0, + 615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 618.0, + 483.0, + 618.0, + 483.0, + 641.0, + 455.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 626.0, + 564.0, + 626.0, + 564.0, + 634.0, + 554.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 682.0, + 513.0, + 682.0, + 513.0, + 730.0, + 419.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 699.0, + 778.0, + 699.0, + 778.0, + 720.0, + 756.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 682.0, + 961.0, + 682.0, + 961.0, + 730.0, + 866.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1239.0, + 680.0, + 1390.0, + 680.0, + 1390.0, + 745.0, + 1239.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 740.0, + 1166.0, + 740.0, + 1166.0, + 772.0, + 1117.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 850.0, + 630.0, + 850.0, + 630.0, + 880.0, + 483.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 851.0, + 875.0, + 851.0, + 875.0, + 878.0, + 730.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 941.0, + 755.0, + 941.0, + 755.0, + 973.0, + 590.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 940.0, + 1053.0, + 940.0, + 1053.0, + 974.0, + 937.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 941.0, + 1223.0, + 941.0, + 1223.0, + 975.0, + 1160.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.75, + 363.5, + 1141.75, + 363.5, + 1141.75, + 383.5, + 1114.75, + 383.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1204.0, + 498.0, + 1204.0, + 498.0, + 1241.0, + 422.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 571.0, + 1211.0, + 749.0, + 1211.0, + 749.0, + 1248.0, + 571.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1268.0, + 496.0, + 1268.0, + 496.0, + 1302.0, + 423.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1245.0, + 774.0, + 1245.0, + 774.0, + 1309.0, + 573.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1284.0, + 975.0, + 1284.0, + 975.0, + 1294.0, + 966.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 1267.0, + 1119.0, + 1267.0, + 1119.0, + 1274.0, + 1111.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1292.0, + 1216.0, + 1292.0, + 1216.0, + 1304.0, + 1203.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1308.0, + 697.0, + 1308.0, + 697.0, + 1343.0, + 573.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1308.0, + 788.0, + 1308.0, + 788.0, + 1343.0, + 778.0, + 1343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1310.0, + 430.0, + 1310.0, + 430.0, + 1560.0, + 392.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1327.0, + 497.0, + 1327.0, + 497.0, + 1365.0, + 421.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1389.0, + 497.0, + 1389.0, + 497.0, + 1428.0, + 421.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 1437.0, + 1180.0, + 1437.0, + 1180.0, + 1457.0, + 1167.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1453.0, + 496.0, + 1453.0, + 496.0, + 1489.0, + 436.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1458.0, + 1215.0, + 1458.0, + 1215.0, + 1472.0, + 1203.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1468.0, + 1123.0, + 1468.0, + 1123.0, + 1488.0, + 1102.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 1515.0, + 496.0, + 1515.0, + 496.0, + 1550.0, + 437.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 1555.0, + 518.0, + 1555.0, + 518.0, + 1563.0, + 508.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1577.0, + 505.0, + 1577.0, + 505.0, + 1613.0, + 436.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 465.0, + 1641.0, + 493.0, + 1641.0, + 493.0, + 1673.0, + 465.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1643.0, + 1214.0, + 1643.0, + 1214.0, + 1651.0, + 1205.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1682.0, + 511.0, + 1682.0, + 511.0, + 1709.0, + 488.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1677.0, + 666.0, + 1677.0, + 666.0, + 1712.0, + 624.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 1679.0, + 809.0, + 1679.0, + 809.0, + 1711.0, + 770.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1679.0, + 953.0, + 1679.0, + 953.0, + 1711.0, + 915.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1677.0, + 1100.0, + 1677.0, + 1100.0, + 1713.0, + 1059.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1679.0, + 1251.0, + 1679.0, + 1251.0, + 1711.0, + 1197.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1709.0, + 1144.0, + 1709.0, + 1144.0, + 1743.0, + 578.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1782.0, + 1405.0, + 1782.0, + 1405.0, + 1818.0, + 294.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1814.0, + 1402.0, + 1814.0, + 1402.0, + 1846.0, + 295.0, + 1846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1843.0, + 1405.0, + 1843.0, + 1405.0, + 1880.0, + 292.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1875.0, + 1405.0, + 1875.0, + 1405.0, + 1911.0, + 294.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1904.0, + 1406.0, + 1904.0, + 1406.0, + 1941.0, + 292.0, + 1941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1934.0, + 409.0, + 1934.0, + 409.0, + 1970.0, + 292.0, + 1970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1014.0, + 1403.0, + 1014.0, + 1403.0, + 1050.0, + 295.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1045.0, + 1289.0, + 1045.0, + 1289.0, + 1084.0, + 293.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 223.0, + 757.0, + 223.0, + 757.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 223.0, + 757.0, + 223.0, + 757.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 420, + 840, + 1227, + 840, + 1227, + 1382, + 420, + 1382 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 292, + 1424, + 1401, + 1424, + 1401, + 1489, + 292, + 1489 + ], + "score": 0.936 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.844 + }, + { + "category_id": 15, + "poly": [ + 448.0, + 872.0, + 499.0, + 872.0, + 499.0, + 909.0, + 448.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 932.0, + 505.0, + 932.0, + 505.0, + 973.0, + 448.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 984.0, + 510.0, + 984.0, + 510.0, + 1164.0, + 439.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 998.0, + 457.0, + 998.0, + 457.0, + 1153.0, + 418.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1177.0, + 503.0, + 1177.0, + 503.0, + 1215.0, + 449.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1169.0, + 1093.0, + 1169.0, + 1093.0, + 1201.0, + 1033.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 1197.0, + 1186.0, + 1197.0, + 1186.0, + 1240.0, + 1033.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1238.0, + 501.0, + 1238.0, + 501.0, + 1277.0, + 449.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1232.0, + 1207.0, + 1232.0, + 1207.0, + 1267.0, + 1034.0, + 1267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1262.0, + 1210.0, + 1262.0, + 1210.0, + 1303.0, + 1032.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1323.0, + 544.0, + 1323.0, + 544.0, + 1351.0, + 522.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1319.0, + 656.0, + 1319.0, + 656.0, + 1354.0, + 596.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1318.0, + 847.0, + 1318.0, + 847.0, + 1354.0, + 672.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1319.0, + 928.0, + 1319.0, + 928.0, + 1354.0, + 858.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1319.0, + 1019.0, + 1319.0, + 1019.0, + 1354.0, + 961.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1319.0, + 1111.0, + 1319.0, + 1111.0, + 1354.0, + 1053.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1319.0, + 1202.0, + 1319.0, + 1202.0, + 1354.0, + 1141.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 1351.0, + 989.0, + 1351.0, + 989.0, + 1386.0, + 734.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1424.0, + 1406.0, + 1424.0, + 1406.0, + 1460.0, + 295.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1454.0, + 533.0, + 1454.0, + 533.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 818.0, + 73.0, + 818.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/BkxfshNYwB/BkxfshNYwB.md b/parse/train/BkxfshNYwB/BkxfshNYwB.md new file mode 100644 index 0000000000000000000000000000000000000000..f2616f1b00415dfa29a8f4d1e73ac88d6c249ba7 --- /dev/null +++ b/parse/train/BkxfshNYwB/BkxfshNYwB.md @@ -0,0 +1,333 @@ +# MINCUT POOLING IN GRAPH NEURAL NETWORKS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +The advance of node pooling operations in Graph Neural Networks (GNNs) has lagged behind the feverish design of new message-passing techniques, and pooling remains an important and challenging endeavor for the design of deep architectures. In this paper, we propose a pooling operation for GNNs that leverages a differentiable unsupervised loss based on the minCUT optimization objective. For each node, our method learns a soft cluster assignment vector that depends on the node features, the target inference task (e.g., a graph classification loss), and, thanks to the minCUT objective, also on the connectivity structure of the graph. Graph pooling is obtained by applying the matrix of assignment vectors to the adjacency matrix and the node features. The proposed method can also be used as a stand-alone module to cluster vertexes in annotated graphs and solve unsupervised problems. We validate the effectiveness of the proposed pooling method on downstream tasks, including supervised graph classification and a set of unsupervised tasks, which reveal the limitations of existing GNN pooling approaches. + +# 1 INTRODUCTION + +A fundamental component in deep convolutional neural networks is the pooling operation, which replaces the output of convolutions with local summaries of nearby points and is usually implemented by maximum or average operations (Lee et al., 2016). State-of-the-art architectures alternate convolutions, which extrapolate local patterns irrespective of the specific location on the input signal, and pooling, which lets the ensuing convolutions capture aggregated patterns. Pooling allows to learn abstract representations in deeper layers of the network by discarding information that is superfluous for the task, and keeps model complexity under control by limiting the growth of intermediate features. + +Graph Neural Networks (GNNs) extend the convolution operation from regular domains, such as images or time series, to data with arbitrary topologies and unordered structures described by graphs (Battaglia et al., 2018). The development of pooling strategies for GNNs, however, has lagged behind the design of newer and more effective message-passing (MP) operations (Gilmer et al., 2017), such as graph convolutions, mainly due to the difficulty of defining an aggregated version of the original graph that supports the pooled signal. + +A na¨ıve pooling strategy in GNNs is to average all nodes features (Li et al., 2016), but it has limited flexibility since it does not extract local summaries of the graph structure, and no further MP operations can be applied afterwards. An alternative approach consists in pre-computing coarsened versions of the original graph and then fit the data to these deterministic structures (Bruna et al., 2013). While this aggregation accounts for the connectivity of the graph, it ignores task-specific objectives as well as the node features. + +In this paper, we propose a differentiable pooling operation implemented as a neural network layer, which can be seamlessly combined with other MP layers (see Fig. 1). The parameters in the pooling layer are learned by combining the task-specific loss with an unsupervised regularization term, which optimizes a continuous relaxation of the normalized minCUT objective. The minCUT identifies dense graph components, where the nodes features become locally homogeneous after the message-passing. By gradually aggregating these components, the GNN learns to distil global properties from the graph. The proposed minCUT pooling operator (minCUTpool) yields partitions that 1) cluster together nodes which have similar features and are strongly connected on the graph, and 2) take into account the objective of the downstream task. + +![](images/2f1bdb5271287007bfb404fabc54a4bd48a828a415e925e71e65dd38e08b6950.jpg) +Figure 1: A deep GNN architecture where message-passing is followed by minCUT pooling. + +# 2 BACKGROUND + +# 2.1 MINCUT AND SPECTRAL CLUSTERING + +Given a graph $G = \{ \nu , \mathcal { E } \}$ , $| \nu | = N$ , and the associated adjacency matrix $\mathbf { A } \in \mathbb { R } ^ { N \times N }$ , the $K$ -way normalized minCUT (simply referred to as minCUT) aims at partitioning $\nu$ in $K$ disjoint subsets by removing the minimum volume of edges. The problem is equivalent to maximizing + +$$ +\frac { 1 } { K } \sum _ { k = 1 } ^ { K } \frac { \operatorname* { l i n k s } ( \mathcal { V } _ { k } ) } { \deg \mathrm { r e e } ( \mathcal { V } _ { k } ) } = \frac { 1 } { K } \sum _ { k = 1 } ^ { K } \frac { \sum _ { i , j \in \mathcal { V } _ { k } } \mathcal { E } _ { i , j } } { \sum _ { i \in \mathcal { V } _ { k } , j \in \mathcal { V } \backslash \mathcal { V } _ { k } } \mathcal { E } _ { i , j } } , +$$ + +where the numerator counts the edge volume within each cluster, and the denominator counts the edges between the nodes in a cluster and the rest of the graph (Shi & Malik, 2000). Let $\mathbf { C } \in \mathbb { R } ^ { N \times K }$ be a cluster assignment matrix, so that $\mathbf { C } _ { i , j } = 1$ if node $i$ belongs to cluster $j$ , and 0 otherwise. The minCUT problem can be expressed as + +$$ +\mathrm { m a x i m i z e } \frac { 1 } { K } \sum _ { k = 1 } ^ { K } \frac { { \bf C } _ { k } ^ { T } { \bf A } { \bf C } _ { k } } { { \bf C } _ { k } ^ { T } { \bf D } { \bf C } _ { k } } , \mathrm { s . t . } { \bf C } \in \{ 0 , 1 \} ^ { N \times K } , { \bf C } { \bf 1 } _ { K } = { \bf 1 } _ { N } , +$$ + +where $\mathbf { D } = \mathrm { d i a g } ( \mathbf { A } \mathbf { 1 } _ { N } )$ is the degree matrix (Dhillon et al., 2004). Since problem (2) is NP-hard, it is usually recast in a relaxed formulation that can be solved in polynomial time and guarantees a near-optimal solution (Yu & Shi, 2003): + +$$ +\arg \operatorname* { m a x } _ { \mathbf { Q } \in \mathbb { R } ^ { N \times K } } \frac { 1 } { K } \sum _ { k = 1 } ^ { K } \mathbf { Q } _ { k } ^ { T } \mathbf { A } \mathbf { Q } _ { k } , \quad \mathrm { s . t . } \mathbf { Q } = \mathbf { C } ( \mathbf { C } ^ { T } \mathbf { D } \mathbf { C } ) ^ { - \frac { 1 } { 2 } } , \ \mathbf { Q } ^ { T } \mathbf { Q } = \mathbf { I } _ { K } . +$$ + +While the optimization problem (3) is still non-convex, there exists an optimal solution $\mathbf { Q } ^ { * } = \mathbf { U } _ { K } \mathbf { O }$ , where ${ \bf U } _ { K } \dot { } \in \mathbb { R } ^ { N \times K }$ contains the eigenvectors of A corresponding to the $K$ largest eigenvalues, and $\mathbf { O } \in \mathbb { R } ^ { K \times K }$ is an orthogonal transformation (Ikebe et al., 1987). + +Since the elements of $\mathbf { Q } ^ { * }$ are real values rather than binary cluster indicators, the spectral clustering (SC) approach can be used to find discrete cluster assignments. In SC, the rows of $\mathbf { Q } ^ { * }$ are treated as node representations embedded in the eigenspace of the Laplacian, and are clustered together with standard algorithms such as $k$ -means (Von Luxburg, 2007). One of the main limitations of SC lies in the computation of the spectrum of A, which has a memory complexity of $\mathcal { O } ( N ^ { 2 } )$ and a computational complexity of $\mathcal { O } ( \bar { N } ^ { 3 } )$ . This prevents its applicability to large datasets. + +To deal with such scalability issues, the constrained optimization in (3) can be solved by gradient descent algorithms that refine the solution by iterating operations whose individual complexity is $\mathcal { O } ( N ^ { 2 } )$ , or even $\mathcal { O } ( N )$ (Han & Filippone, 2017). Those algorithms search the solution on the manifold induced by the orthogonality constraint on the columns of $\mathbf { Q }$ , by performing gradient updates along the geodesics (Wen & Yin, 2013; Collins et al., 2014). Alternative approaches rely on the QR factorization to constrain the space of feasible solutions (Damle et al., 2016), and alleviate the cost $\mathcal { O } ( N ^ { 3 } )$ of the factorization by ensuring that orthogonality holds only on one minibatch at a time (Shaham et al., 2018). + +Other works based on neural networks include an autoencoder trained to map the $i$ th row of the Laplacian to the ith components of the first $K$ eigenvectors, to avoid the spectral decomposition (Tian et al., 2014). Yi et al. (2017) use a soft orthogonality constraint to learn spectral embeddings as a volumetric reparametrization of a precomputed Laplacian eigenbase. Shaham et al. (2018); Kampffmeyer et al. (2019) propose differentiable loss functions to partition generic data and process out-of-sample data at inference time. Nazi et al. (2019) generate balanced node partitions with a GNN, but adopt an optimization that does not encourage cluster assignments to be orthogonal. + +# 2.2 GRAPH NEURAL NETWORKS + +Many approaches have been proposed to process graphs with neural networks, including recurrent architectures (Scarselli et al., 2009; Li et al., 2016) or convolutional operations inspired by filters used in graph signal processing (Defferrard et al., 2016; Bianchi et al., 2019). Since our focus is on graph pooling, we base our GNN implementation on a simple MP operation, which combines the features of each node with its 1st-order neighbors. To account for the initial node features, it is possible to introduce self-loops by adding a (scaled) identity matrix to the diagonal of A (Kipf & Welling, 2017). Since our pooling will modify the structure of the adjacency matrix, we prefer a MP implementation that leaves the original A unaltered and accounts for the initial node features by means of skip connections. + +Let $\tilde { \mathbf { A } } = \mathbf { D } ^ { - \frac { 1 } { 2 } } \mathbf { A } \mathbf { D } ^ { - \frac { 1 } { 2 } } \in \mathbb { R } ^ { N \times N }$ be the symmetrically normalized adjacency matrix and $\mathbf { X } \in \mathbb { R } ^ { N \times F }$ the matrix containing the node features. The output of the MP layer is + +$$ +\mathbf { X } ^ { ( t + 1 ) } = M P ( \mathbf { X } ^ { ( t ) } , \tilde { \mathbf { A } } ) = \operatorname { R e L U } ( \tilde { \mathbf { A } } \mathbf { X } ^ { ( t ) } \mathbf { W } _ { m } + \mathbf { X } ^ { ( t ) } \mathbf { W } _ { s } ) , +$$ + +where $\boldsymbol { \Theta } _ { M P } = \{ \mathbf { W } _ { m } , \mathbf { W } _ { s } \}$ are the trainable weights relative to the mixing and skip component of the layer, respectively. + +# 3 PROPOSED METHOD + +The minCUT pooling strategy computes a cluster assignment matrix $\mathbf { S } \in \mathbb { R } ^ { N \times K }$ by means of a multi-layer perceptron, which maps each node feature $\mathbf { x } _ { i }$ into the ith row of S: + +$$ +\begin{array} { r } { \mathbf { S } = s o f t m a x ( \mathbf { R e L U } ( \mathbf { X } \mathbf { W } _ { 1 } ) \mathbf { W } _ { 2 } ) , } \end{array} +$$ + +where $\boldsymbol { \Theta } _ { P o o l } = \{ \mathbf { W } _ { 1 } \in \mathbb { R } ^ { F \times H } , \mathbf { W } _ { 2 } \in \mathbb { R } ^ { H \times K } \}$ are trainable parameters. The softmax function guarantees that $s _ { i , j } \in [ 0 , 1 ]$ and enforces the constraints $\mathbf { S 1 } _ { K } = \mathbf { 1 } _ { N }$ inherited from the optimization problem in (2). The parameters $\Theta _ { M P }$ and $\Theta _ { P o o l }$ are jointly optimized by minimizing the usual task-specific loss, as well as an unsupervised loss $\mathcal { L } _ { u }$ , which is composed of two terms + +$$ +\mathcal { L } _ { u } = \mathcal { L } _ { c } + \mathcal { L } _ { o } = \underbrace { - \frac { T r ( \mathbf { S } ^ { T } \tilde { \mathbf { A } } \mathbf { S } ) } { T r ( \mathbf { S } ^ { T } \tilde { \mathbf { D } } \mathbf { S } ) } } _ { \mathcal { L } _ { c } } + \underbrace { \left\| \frac { \mathbf { S } ^ { T } \mathbf { S } } { \| \mathbf { S } ^ { T } \mathbf { S } \| _ { F } } - \frac { \mathbf { I } _ { K } } { \sqrt { K } } \right\| _ { F } } _ { \mathcal { L } _ { o } } , +$$ + +where $\| \cdot \| _ { F }$ indicates the Frobenius norm. + +The cut loss term, $\mathcal { L } _ { c }$ , evaluates the minCUT given by the cluster assignment S, and is bounded by $- 1 \leq \mathcal { L } _ { c } \leq 0$ . Minimizing $\mathcal { L } _ { c }$ encourages strongly connected nodes to be clustered together, since the inner product $\left. \mathbf { s } _ { i } , \mathbf { s } _ { j } \right.$ increases when $\tilde { a } _ { i , j }$ is large. $\mathcal { L } _ { c }$ has a single maximum, reached when the numerator $\begin{array} { r } { T r ( \mathbf { S } ^ { T } \tilde { \mathbf { A } } \mathbf { S } ) = \frac { 1 } { K } \sum _ { k = 1 } ^ { K } \mathbf { S } _ { k } ^ { T } \tilde { \mathbf { A } } \mathbf { S } _ { k } = 0 } \end{array}$ . This occurs if, for each pair of connected nodes (i.e., $\tilde { a } _ { i , j } > 0 \rangle$ ), the cluster assignments are orthogonal (i.e., $\langle \mathbf { s } _ { i } , \mathbf { s } _ { j } \rangle = 0 ,$ ). $\mathcal { L } _ { c }$ reaches its minimum, $- 1$ , when $T r ( \mathbf { S } ^ { T } \tilde { \mathbf { A } } \mathbf { S } ) = T r ( \mathbf { S } ^ { T } \tilde { \mathbf { D } } \mathbf { S } )$ . This occurs when in a graph with $K$ disconnected components the cluster assignments are equal for all the nodes in the same component and orthogonal to the cluster assignments of nodes in different components. However, $\mathcal { L } _ { c }$ is a non-convex function and its minimization can lead to local minima or degenerate solutions. For example, given a connected graph, a trivial optimal solution is the one that assigns all nodes to the same cluster. As a consequence of the continuous relaxation, another degenerate minimum occurs when the cluster assignments are all uniform, that is, all nodes are equally assigned to all clusters. This problem is exacerbated by prior message-passing operations, which make the node features more uniform. + +The orthogonality loss term, $\mathcal { L } _ { o }$ , penalizes the degenerate minima of $\mathcal { L } _ { c }$ by encouraging the cluster assignments to be orthogonal and the clusters to be of similar size. Since the two matrices in $\mathcal { L } _ { o }$ have unitary norm it is easy to see that $0 \leq \mathcal { L } _ { o } \leq 2$ . Therefore, $\mathcal { L } _ { o }$ does not dominate over $\mathcal { L } _ { c }$ and the two terms can be safely summed directly (see Fig. 4 for an example). ${ \mathbf { I } } _ { K }$ can be interpreted as a (rescaled) clustering matrix $\mathbf { I } _ { K } = \hat { \mathbf { S } } ^ { T } \hat { \mathbf { S } }$ , where $\hat { \bf S }$ assigns exactly $N / K$ points to each cluster. The value of the Frobenius norm between clustering matrices is not dominated by the performance on the largest clusters (Law et al., 2017) and, thus, can be used to optimize intra-cluster variance. + +Contrarily to SC methods that search for feasible solutions only within the space of orthogonal matrices, $\mathcal { L } _ { o }$ only introduces a soft constraint that could be violated during the learning procedure. Since $\mathcal { L } _ { c }$ is non-convex, the violation compromises the theoretical guarantee of convergence to the optimum of (3). However, we note that: + +1. the cluster assignments S are well initialized: after the MP operation, the features of the connected vertices become similar and, since the MLP is a smooth function (Nelles, 2013), it yields similar cluster assignments for those vertices; +2. in the GNN architecture, the minCUT objective is a regularization term and, therefore, a solution which is sub-optimal for (3) could instead be adequate for the specific objective of the downstream task; +3. optimizing the task-specific loss helps the GNN to avoid the degenerate minima of $\mathcal { L } _ { c }$ . + +# 3.1 COARSENING + +The coarsened version of the adjacency matrix and the graph signal are computed as + +$$ +{ \bf A } ^ { p o o l } = { \bf S } ^ { T } { \tilde { \bf A } } { \bf S } ; ~ { \bf X } ^ { p o o l } = { \bf S } ^ { T } { \bf X } , +$$ + +where the entry xpooi,j in $\mathbf { X } ^ { p o o l } ~ \in ~ \mathbb { R } ^ { K \times F }$ is the weighted average value of feature $j$ among the elements in cluster $i$ . $\mathbf { A } ^ { p o o l } ~ \in ~ \mathbb { R } ^ { K \times K }$ is a symmetric matrix, whose entries $a _ { i , i } ^ { p o o l }$ are the total number of edges between the nodes in the cluster i, while apooi,j is the number of edges between cluster $i$ and $j$ . Since $\mathbf { A } ^ { p o o l }$ corresponds to the numerator of $\mathcal { L } _ { c }$ in (7), the trace maximization yields clusters with many internal connections and weakly connected to each other. Hence, $\mathbf { A } ^ { p o o l }$ will be a diagonal-dominant matrix, which describes a graph with self-loops much stronger than any other connection. Because self-loops hamper the propagation across adjacent nodes in the MP operations following the pooling layer, we compute the new adjacency matrix $\tilde { \mathbf { A } } ^ { p o o l }$ by zeroing the diagonal and by applying the degree normalization + +$$ +\hat { \bf A } = { \bf A } ^ { p o o l } - { \bf I } _ { K } d i a g ( { \bf A } ^ { p o o l } ) ; \quad \tilde { \bf A } ^ { p o o l } = \hat { \bf D } ^ { - \frac { 1 } { 2 } } \hat { \bf A } \hat { \bf D } ^ { - \frac { 1 } { 2 } } . +$$ + +where $d i a g ( \cdot )$ returns the matrix diagonal. + +# 3.2 DISCUSSION AND RELATIONSHIP WITH SPECTRAL CLUSTERING + +The proposed method is straightforward to implement: the cluster assignments, the loss, graph coarsening, and feature pooling are all computed with standard linear algebra operations. + +There are several differences between minCUTpool and classic SC methods. SC partitions the graph based on the Laplacian, but does not account for the node features. Instead, the cluster assignments $\mathbf { s } _ { i }$ found by minCUTpool depend on $\mathbf { x } _ { i }$ , which works well if connected nodes have similar features. This is a reasonable assumption in GNNs since, even in disassortative graphs (i.e., networks where dissimilar nodes are likely to be connected (Newman, 2003)), the features tend to become similar due to the MP operations. + +Another difference is that SC handles a single graph and is not conceived for tasks with multiple graphs to be partitioned independently. Instead, thanks to the independence of the model parameters from the number of nodes $N$ and from the graph spectrum, minCUTpool can generalize to outof-sample data. This feature is fundamental in problems such as graph classification, where each sample is a graph with a different structure, and allows to train the model on small graphs and process larger ones at inference time. Finally, minCUTpool directly uses the soft cluster assignments rather than performing $k$ -means afterwards. + +# 4 RELATED WORK ON POOLING IN GNNS + +Trainable pooling methods. Similarly to our method, these approaches learn how to generate coarsened version of the graph through differentiable functions, which take as input the nodes features $\mathbf { X }$ and are parametrized by weights optimized on the task at hand. + +Diffpool (Ying et al., 2018) is a pooling module that includes two parallel MP layers: one to compute the new node features $\mathbf { X } ^ { ( t + 1 ) }$ and another to generate the cluster assignments S. Diffpool implements an unsupervised loss that consists of two terms. First, the link prediction term $\| \mathbf { A } - \mathbf { S } \mathbf { \dot { S } } ^ { T } \| _ { F }$ minimizes the Frobenius norm of the difference between the adjacency and the Gram matrix of the cluster assignments, encouraging nearby nodes to be clustered together. The second term $\begin{array} { r } { \frac { 1 } { N } \sum _ { i = 1 } ^ { N } H ( \mathbf { S } _ { i } ) } \end{array}$ minimizes the entropy of the cluster assignments to make them alike to one-hot vectors. Like minCUTpool, Diffpool clusters the vertices of annotated graphs, but yields completely different partitions, since it computes differently the clustering assignments, the coarsened adjacency matrix and, most importantly, the unsupervised loss. In Diffpool, such a loss shows pathological behaviors that are discussed later in the experiments. + +The approach dubbed Top- $K$ pooling (Hongyang Gao, 2019; Lee et al., 2019), learns a projection vector that is applied to each node feature to obtain a score. The nodes with the $K$ highest scores are retained, the others are dropped. Since the top- $K$ selection is not differentiable, the scores are also used as a gate/attention for the node features, letting the projection vector to be trained with backpropagation. Top- $K$ is memory efficient as it avoids generating cluster assignments. To prevent A from becoming disconnected after nodes removal, Top- $K$ drops the rows and the columns from ${ \bf A } ^ { 2 }$ and uses it as the new adjacency matrix. However, computing ${ \bf A } ^ { 2 }$ costs $\mathcal { O } ( N ^ { 2 } )$ and it is inefficient to implement with sparse operations. + +Topological pooling methods. These methods pre-compute a pyramid of coarsened graphs, only taking into account the topology (A), but not the node features $\mathbf { \Pi } ( \mathbf { X } )$ . During training, the node features are pooled with standard procedures and are fit into these deterministic graph structures. These methods are less flexible, but provide a stronger bias that can prevent degenerate solutions (e.g., coarsened graphs collapsing in a single node). + +The approach proposed by Bruna et al. (2013), which has been adopted also in other GNN architectures (Defferrard et al., 2016; Fey et al., 2018), exploits GRACLUS (Dhillon et al., 2004), a hierarchical algorithm based on SC. At each pooling level $l$ , GRACLUS indetifies the pairs of maximally similar nodes $i _ { l }$ and $j _ { l }$ to be clustered together into a new vertex $k _ { ( l + 1 ) }$ . At inference phase, max-pooling is used to determine which node in the pair is kept. Fake vertices are added so that the number of nodes can be halved each time, but this injects noisy information in the graph. + +Node decimation is a method originally proposed in graph signal processing literature (Shuman et al., 2016), which as been adapted also for GNNs (Simonovsky & Komodakis, 2017). The nodes are partitioned in two sets, according to the signs of the Laplacian eigenvector associated to the largest eigenvalue. One of the two sets is dropped, reducing the number of nodes each time approximately by half. Kron reduction is used to compute a pyramid of coarsened Laplacians from the remaining nodes. + +A procedure proposed in Gama et al. (2018) diffuses a signal from designated nodes on the graph and stores the observed sequence of diffused components. The resulting stream of information is interpreted as a time signal, where standard CNN pooling is applied. We also mention a pooling operation for coarsening binary unweighted graphs by aggregating maximal cliques (Luzhnica et al., 2019). Nodes assigned to the same clique are summarized by max or average pooling and become a new node in the coarsened graph. + +# 5 EXPERIMENTS + +We consider both supervised and unsupervised tasks, and compare minCUTpool with other GNN pooling strategies. The Appendix provides further details on the experiments and a schematic depiction of the architectures used in each task. In addition, the Appendix reports two additional experiments: i) graph reconstruction by means of an Auto Encoder with bottleneck, implemented with pooling and un-pooling layers, ii) an architecture with pooling for graph regression. + +# 5.1 CLUSTERING THE GRAPH NODES + +To study the effectiveness of the proposed loss, we perform different node clustering tasks with a simple GNN composed of a single MP layer followed by a pooling layer. The GNN is trained by minimizing $\mathcal { L } _ { u }$ only, so that its effect is evaluated without the “interference” of a supervised loss. + +Clustering on synthetic networks We consider two simple graphs: the first is a network with 6 communities and the second is a regular grid. The adjacency matrix A is binary and the features X are the 2-D node coordinates. Fig. 2 depicts the node partitions generated by SC (a, d), Diffpool (b, e), and minCUTpool (c, f). Cluster indexes for Diffpool and minCUTpool are obtained by taking the argmax of S row-wise. Compared to SC, Diffpool and minCUTpool leverage the information contained in X. minCUTpool generates very accurate and balanced partitions, demonstrating that the cluster assignment matrix S is well formed. On the other hand, Diffpool assigns some nodes to the wrong community in the first example, and produces an imbalanced partition of the grid. + +![](images/f54a1747a2f3fb20dd6df85e294051b41f7c8f617ba05ec5ba321652a3688701.jpg) +Figure 2: Node clustering on a community network $K { = } 6 )$ and on a grid graph $( K { = } 5 )$ + +Image segmentation Given an image, we build a Region Adjacency Graph (Tremeau & Colan- ´ toni, 2000) using as nodes the regions generated by an oversegmentation procedure (Felzenszwalb & Huttenlocher, 2004). The SC technique used in this example is the recursive normalized cut (Shi & Malik, 2000), which recursively clusters the nodes until convergence. For Diffpool and minCUTpool, we include node features consisting of the average and total color in each oversegmented region. We set the number of desired clusters to $K = 4$ . The results in Fig. 3 show that minCUTpool yields a more precise segmentation. On the other hand, SC and Diffpool aggregate wrong regions and, in addition, SC finds too many segments. + +Clustering on citation networks We cluster the nodes of three popular citation networks: Cora, Citeseer, and Pubmed. The nodes are documents represented by sparse bag-of-words feature vectors stored in $\mathbf { X }$ and the binary undirected edges in $\mathbf { A }$ indicate citation links between documents. Each node $i$ is labeled with the document class $y _ { i }$ . Once the training is over, to test the quality of the partitions generated by each method we check the agreement between the cluster assignments and the true class labels. Tab. 1 reports the Completeness Score $\begin{array} { r } { \mathrm { C S } ( \tilde { \bf y } , { \bf y } ) = 1 - \frac { H ( \tilde { \bf y } | { \bf y } ) } { H ( \tilde { \bf y } ) } } \end{array}$ and Normalized Mutual Information $\begin{array} { r } { \mathbf { N M I } ( \tilde { \mathbf { y } } , \mathbf { y } ) = \frac { H ( \tilde { \mathbf { y } } ) - H ( \tilde { \mathbf { y } } | \mathbf { y } ) } { \sqrt { H ( \tilde { \mathbf { y } } ) - H ( \mathbf { y } ) } } } \end{array}$ where $H ( \cdot )$ is the entropy. + +The GNN architecture configured with minCUTpool achieves a higher NMI score than SC, which does not account for the node features $\mathbf { X }$ when generating the partitions. Our pooling operation outperforms also Diffpool, since the minimization of the unsupervised loss in Diffpool yields degenerate solutions. The pathological behavior is shown in Fig. 4, which depicts the evolution of the NMI scores as the unsupervised losses in Diffpool and minCUTpool are minimized in training. + +![](images/81a3f7ffc1776e74fbc56b414ab1c318578c5dc39e05ac6ea084422888017b44.jpg) +Figure 3: Image segmentation by clustering the nodes of the Region Adjacency Graph. + +![](images/4067d319802597da0f8a99d17f9a05ae75ef1d9974bd1b46567b315150d067c6.jpg) +Figure 4: Unsupervised losses and NMI of Diffpool and minCUTpool on Cora. + +Table 1: NMI and CS obtained by clustering the nodes on citation networks over 10 different runs. The number of clusters $K$ is equal to the number of node classes. + +
DatasetKSpectral clusteringDiffpoolminCUTpool
NMIcsNMICSNMICS
Cora70.025 ± 0.0140.126 ± 0.0420.315 ± 0.0050.309 ±0.0050.404 ± 0.0180.392 ± 0.018
Citeseer60.014 ± 0.0030.033 ±0.0000.139 ± 0.0160.153 ± 0.0200.287 ±0.0470.283 ± 0.046
Pubmed30.182 ± 0.0000.261 ± 0.0000.079 ±0.0010.085 ±0.0010.200 ± 0.0200.197 ± 0.019
+ +# 5.2 SUPERVISED GRAPH CLASSIFICATION + +In this task, the $i$ -th datum is a graph with $N _ { i }$ nodes represented by a pair $\{ \mathbf { A } _ { i } , \mathbf { X } _ { i } \}$ and must be associated to the correct label $\mathbf { y } _ { i }$ . We test the models on different graph classification datasets. For featureless graphs, we used the node degree information and the clustering coefficient as surrogate node features. We evaluate model performance with a 10-fold train/test split, using $1 0 \%$ of the training set in each fold as validation for early stopping. We adopt a fixed network architecture, MP(32)-poolMP(32)-pool-MP(32)-GlobalAvgPool-softmax, where MP is the message-passing operation in (4) + +with 32 hidden units. The pooling module is implemented either by Graclus, Decimation pooling, Top- $K$ , SAGPool (Lee et al., 2019), Diffpool, or the proposed minCUTpool. Each pooling method is configured to drop half of the nodes in a graph $K = N / 2$ in Top- $K$ , Diffpool, and minCUTpool). As baselines, we consider the popular Weisfeiler-Lehman (WL) graph kernel (Shervashidze et al., 2011), a network with only MP layers (Flat), and a fully connected network (Dense). + +Table 2: Graph classification accuracy. Significantly better results $( p < 0 . 0 5 )$ are in bold. + +
DatasetWLDenseFlatGraclusDecim.DiffpoolTop-KSAGpoolminCUT
Bench-easy92.629.3±0.398.5±0.397.5±0.597.9±0.598.6±0.482.4±8.984.2±2.399.0±0.0
Bench-hard60.029.4±0.367.6±2.869.0±1.572.6±0.969.9±1.942.7±15.237.7±14.573.8±1.9
Mutagenicity81.7±1.168.4±0.378.0±1.374.4±1.877.8±2.377.6±2.771.9±3.772.4±2.479.9±2.1
Proteins71.2±2.668.7±3.372.6±4.868.6±4.673.3±3.772.7±3.869.6±3.570.5±2.676.5±2.6
DD78.6±2.770.6±5.276.8±1.570.5±4.872.0±3.179.3±2.469.4±7.871.5±4.580.8±2.3
COLLAB74.8±1.379.3±1.682.1±1.877.1±2.179.1±1.581.8±1.479.3±1.879.2±2.083.4±1.7
Reddit-Binary68.2±1.748.5±2.680.3±2.679.2±0.484.3±2.486.8±2.174.7±4.573.9±5.191.4±1.5
+ +Tab. 2 reports the classification results, highlighting those that are significantly better ( $\dot { p }$ -value $< ~ 0 . 0 5$ w.r.t. the method with the highest mean accuracy). The comparison with Flat helps to understand if a pooling operation is useful or not. The results of Dense, instead, help to quantify how much additional information is brought by the graph structure, with respect to the node features alone. It can be seen that minCUTpool obtains always equal or better results with respect to every other GNN architecture. On the other hand, some pooling procedures do not always improve the performance compared to the Flat baseline, making them not advisable to use in some cases. The WL kernel generally performs worse than the GNNs, except for the Mutagenicity dataset. This is probably because Mutagenicity has smaller graphs than the other datasets, and the adopted GNN architecture is overparametrized for this task. Interestingly, in some dataset such as Proteins and COLLAB it is possible to obtain fairly good classification accuracy with the Dense architecture, meaning that the graph structure only adds limited information. + +![](images/e375310d148f5976514d2295a7d3c8830c48da3e0966c6040f4a07d9c8821927.jpg) +Figure 5: Average duration of one epoch using the same GNN with different pooling operations. Times were computed with an Nvidia GeForce GTX 1050, on the DD dataset with batch size of 1. + +Fig. 5 reports a comparison of the execution time per training epoch for each pooling algorithm. Graclus and Decimation are understandably the fastest methods, since the coarsened graphs are precomputed. Among the differentiable pooling methods, minCUTpool is faster than Diffpool, which uses a slower MP layer rather than a MLP to compute cluster assignments, and than Top- $K$ , which computes the square of A at every forward pass. + +# 6 CONCLUSIONS + +We proposed a pooling layer for GNNs that coarsens a graph by taking into account both the the connectivity structure and the node features. The layer optimizes a regularization term based on the minCUT objective, which is minimized in conjunction with the task-specific loss to produce node partitions that are optimal for the task at hand. + +We tested the effectiveness of our pooling strategy on unsupervised node clustering tasks, by optimizing only the unsupervised clustering loss, as well as supervised graph classification tasks on several popular benchmark datasets. Results show that minCUTpool performs significantly better than existing pooling strategies for GNNs. + +# REFERENCES + +Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018. + +Filippo Maria Bianchi, Daniele Grattarola, L Livi, and C Alippi. Graph neural networks with convolutional arma filters. arXiv preprint arXiv:1901.01343, 2019. + +Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. + +Maxwell D Collins, Ji Liu, Jia Xu, Lopamudra Mukherjee, and Vikas Singh. Spectral clustering with a convex regularizer on millions of images. In European Conference on Computer Vision, pp. 282–298. Springer, 2014. + +Anil Damle, Victor Minden, and Lexing Ying. Robust and efficient multi-way spectral clustering. arXiv preprint arXiv:1609.08251, 2016. + +Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks ¨ on graphs with fast localized spectral filtering. In Advances in Neural Information Processing Systems, pp. 3844–3852, 2016. + +Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Kernel k-means: spectral clustering and normalized cuts. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 551–556. ACM, 2004. + +Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59(2):167–181, 2004. + +Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Muller. Splinecnn: Fast geomet- ¨ ric deep learning with continuous b-spline kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 869–877, 2018. + +Fernando Gama, Antonio G Marques, Geert Leus, and Alejandro Ribeiro. Convolutional neural network architectures for signals supported on graphs. IEEE Transactions on Signal Processing, 67(4):1034–1049, 2018. + +Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1263–1272. JMLR. org, 2017. + +Yufei Han and Maurizio Filippone. Mini-batch spectral clustering. In 2017 International Joint Conference on Neural Networks (IJCNN), pp. 3888–3895. IEEE, 2017. + +Shuiwang Ji Hongyang Gao. Graph u-nets. In Proceedings of the 36th International conference on Machine learning (ICML), 2019. + +Yasuhiko Ikebe, Toshiyuki Inagaki, and Sadaaki Miyamoto. The monotonicity theorem, cauchy’s interlace theorem, and the courant-fischer theorem. The American Mathematical Monthly, 94(4): 352–354, 1987. + +Michael Kampffmeyer, Sigurd Løkse, Filippo M. Bianchi, Lorenzo Livi, Arnt-Børre Salberg, and Robert Jenssen. Deep divergence-based approach to clustering. Neural Networks, 113:91 – 101, 2019. + +Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. International Conference of Learning Representations (ICLR), 2017. + +Marc T Law, Raquel Urtasun, and Richard S Zemel. Deep spectral clustering learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1985–1994. JMLR. org, 2017. + +Chen-Yu Lee, Patrick W Gallagher, and Zhuowen Tu. Generalizing pooling functions in convolutional neural networks: Mixed, gated, and tree. In Artificial Intelligence and Statistics, pp. 464–472, 2016. + +Junhyun Lee, Inyeop Lee, and Jaewoo Kang. Self-attention graph pooling. In Proceedings of the 36th International conference on Machine learning (ICML-19), 2019. + +Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. International Conference of Learning Representations (ICLR), 2016. + +Enxhell Luzhnica, Ben Day, and Pietro Lio. Clique pooling for graph classification. International Conference of Learning Representations (ICLR) – Representation Learning on Graphs and Manifolds workshop, 2019. + +Azade Nazi, Will Hang, Anna Goldie, Sujith Ravi, and Azalia Mirhoseini. Gap: Generalizable approximate graph partitioning framework. arXiv preprint arXiv:1903.00614, 2019. + +Oliver Nelles. Nonlinear system identification: from classical approaches to neural networks and fuzzy models. Springer Science & Business Media, 2013. + +Mark EJ Newman. Mixing patterns in networks. Physical Review E, 67(2):026126, 2003. + +Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2009. + +Uri Shaham, Kelly Stanton, Henry Li, Boaz Nadler, Ronen Basri, and Yuval Kluger. Spectralnet: Spectral clustering using deep neural networks. arXiv preprint arXiv:1801.01587, 2018. + +Nino Shervashidze, Pascal Schweitzer, Erik Jan van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12(Sep):2539– 2561, 2011. + +Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. Departmental Papers (CIS), pp. 107, 2000. + +David I Shuman, Mohammad Javad Faraji, and Pierre Vandergheynst. A multiscale pyramid transform for graph signals. IEEE Transactions on Signal Processing, 64(8):2119–2134, 2016. + +Martin Simonovsky and Nikos Komodakis. Dynamic edgeconditioned filters in convolutional neural networks on graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. + +Fei Tian, Bin Gao, Qing Cui, Enhong Chen, and Tie-Yan Liu. Learning deep representations for graph clustering. In AAAI, pp. 1293–1299, 2014. + +Alain Tremeau and Philippe Colantoni. Regions adjacency graph applied to color image segmenta-´ tion. IEEE Transactions on image processing, 9(4):735–744, 2000. + +Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416, 2007. + +Zaiwen Wen and Wotao Yin. A feasible method for optimization with orthogonality constraints. Mathematical Programming, 142(1-2):397–434, 2013. + +Li Yi, Hao Su, Xingwen Guo, and Leonidas J Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2282–2290, 2017. + +Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hierarchical graph representation learning with differentiable pooling. In Advances in Neural Information Processing Systems, pp. 4800–4810, 2018. + +Yu and Shi. Multiclass spectral clustering. In Proceedings Ninth IEEE International Conference on Computer Vision, pp. 313–319 vol.1, Oct 2003. + +# APPENDIX + +A ADDITIONAL EXPERIMENTS + +# A.1 GNN AUTOENCODER + +To compare the amount of information retained by the pooling layers in the coarsened graphs, we train an autoencoder (AE) to reconstruct a input graph signal $\mathbf { X }$ from its pooled version. The AE architecture is MP(32)-MP(32)-pool-unpool-MP(32)-MP(32)-MP, and is trained by minimizing the mean squared error between the original and the reconstructed graph signal, $\lVert \bf { X } - \bf { X } ^ { \mathrm { { r e c } } } \rVert ^ { 2 }$ . All the pooling operations are configured to retain $2 5 \%$ of the original nodes. + +In Diffpool and minCUTpool, the unpool step is simply implemented by transposing the original pooling operations + +$$ +{ \bf X } ^ { \mathrm { r e c } } = { \bf S } { \bf X } ^ { \mathrm { p o o l } } ; ~ { \bf A } ^ { \mathrm { r e c } } = { \bf S } { \bf A } ^ { \mathrm { p o o l } } { \bf S } ^ { T } . +$$ + +Top- $K$ does not generate a cluster assignment matrix, but returns a binary mask $\mathbf { m } = \{ 0 , 1 \} ^ { N }$ that indicates the nodes to drop (0) or to retain (1). Therefore, an upsamplig matrix $\mathbf { U }$ is built by dropping the columns of the identity matrix ${ \mathbf { I } } _ { N }$ that correspond to a 0 in $\mathbf { m }$ , $\bar { \mathbf { U } } = [ \mathbf { I } _ { N } ] _ { : , \mathbf { m } = = 1 }$ . The unpooling operation is performed by replacing S with $\mathbf { U }$ in (9), and the resulting upscaled graph is a version of the original graph with zeroes in correspondence of the dropped nodes. + +![](images/441bd5ff28bf0dbad772497856fe6d0256f85ad6ae1cfa337d7be79956b30a09.jpg) +Figure 6: AE reconstruction of a ring graph + +![](images/4b22636cf4edd79e814a6d27c3e71aeabc0f8293196e377e997385878fcd0602.jpg) +Figure 7: AE reconstruction of a grid graph + +Fig. 6 and 7 report the original graph signal $\mathbf { X }$ (the node features are the 2-D coordinates of the nodes) and the reconstruction $\mathbf { X } ^ { \mathrm { r e c } }$ obtained by using the different pooling methods, for a ring graph and a regular grid graph. The reconstruction produced by Diffpool is worse for the ring graph, but is almost perfect for the grid graph, while minCUTpool yields good results in both cases. On the other hand, Top- $K$ clearly fails in generating a coarsened representation that maintains enough information from the original graph. + +This experiment highlights a major issue in Top- $K$ pooling, which retains the nodes associated to the highest $K$ values of a score vector s, computed by projecting the node features onto a trainable vector p: $\mathbf { s } = \mathbf { X } \mathbf { p } $ . Nodes that are connected on the graph usually share similar features, and their similarity further increases after the MP operations, which combine the features of neighboring nodes. Retaining the nodes associated to the top $K$ scores in s corresponds to keeping those nodes that are alike and highly connected, as it can be seen in Fig. 6-7. Therefore, Top- $K$ discards entire portions of the graphs, which might contain important information. This explains why Top- $K$ fails to recover the original graph signal when used as bottleneck for the AE, and yields the worse performance among all GNN methods in the graph classification task. + +# A.2 GRAPH REGRESSION OF MOLECULAR PROPERTIES ON QM9 + +The QM9 chemical database is a collection of ${ \approx } 1 3 5 \mathrm { k }$ small organic molecules, associated to continuous labels describing several geometric, energetic, electronic, and thermodynamic properties1. Each molecule in the dataset is represented as a graph $\{ \mathbf { A } _ { i } , \mathbf { X } _ { i } \}$ , where atoms are associated to nodes, and edges represent chemical bonds. The atomic number of each atom (one-hot encoded; C, N, F, O) is taken as node feature and the type of bond (one-hot encoded; single, double, triple, aromatic) can be used as edge attribute. In this experiment, we ignore the edge attributes in order to use all pooling algorithms without modifications. + +The purpose of this experiment is to compare the trainable pooling methods also on a graph regression task, but it must be intended as a proof of concept. In fact, the graphs in this dataset are extremely small (the average number of nodes is 8) and, therefore, a pooling operation is arguably not necessary. We consider a GNN with architecture MP(32)-pool-MP(32)-GlobalAvgPool-Dense, where pool is implemented by Top- $K$ , Diffpool, or minCUTpool. The network is trained to predict a given chemical property from the input molecular graphs. Performance is evaluated with a 10-fold cross-validation, using $1 \dot { 0 } \%$ of the training set for validation in each split. The GNNs are trained for 50 epochs, using Adam with learning rate 5e-4, batch size 32, and ReLU activations. We use the mean squared error (MSE) as supervised loss. + +The MSE obtained on the prediction of each property for different pooling methods is reported in Tab. 3. As expected, the flat baseline with no pooling operation (MP(32)-MP(32)-GlobalAvgPoolDense) yields a lower error in most cases. Contrarily to the graph classification and the AE task, Top- $K$ achieves better results than Diffpool in average. Once again, minCUTpool significantly outperforms the other methods on each regression task and, in one case, also the flat baseline. + +
PropertyTop-KDiffpoolminCUTpoolFlat baseline
mu0.600±0.0850.651±0.0260.538±0.0120.559±0.007
alpha0.197±0.0870.114±0.0010.078±0.0070.065±0.006
homo0.698±0.1020.712±0.0150.526±0.0210.435±0.013
lumo0.601±0.0500.646±0.0130.540±0.0050.515±0.007
gap0.630±0.0440.698±0.0040.584±0.0070.552±0.008
r20.452±0.0870.440±0.0240.261±0.0060.204±0.006
zpve0.402±0.0320.410±0.0040.328±0.0050.284±0.005
uO_atom0.308±0.0550.245±0.0060.193±0.0020.163±0.001
cv0.291±0.1180.337±0.0180.148±0.0040.127±0.002
+ +Table 3: MSE on the graph regression task. The best results with a statistical significance of $p < 0 . 0 5$ are highlighted: the best overall are in bold, the best among pooling methods are underlined. + +# B EXPERIMENTAL DETAILS + +For the WL kernel, we used the implementation provided in the GraKeL library2. The pooling strategy based on Graclus, is taken from the ChebyNets repository3. + +# B.1 CLUSTERING ON CITATION NETWORKS + +Diffpool and minCUTpool are configured with 16 hidden neurons with linear activations in the MLP and MP layer, respectively used to compute the cluster assignment matrix S. The MP layer used to compute the propagated node features $\mathbf { X } ^ { ( 1 ) }$ uses an ELU activation in both architectures. The learning rate for Adam is 5e-4, and the models are trained for 10000 iterations. The details of the citation networks dataset are reported in Tab. 4. + +Table 4: Details of the citation networks datasets + +
DatasetNodesEdgesNode featuresNode classes
Cora2708542914337
Citeseer3327922837036
Pubmed19717886515003
+ +# B.2 GRAPH CLASSIFICATION + +We train the GNN architectures with Adam, an $\mathrm { L } _ { 2 }$ penalty loss with weight 1e-4, and 16 hidden units $( H )$ both in the MLP of minCUTpool and in the internal MP of Diffpool. Mutagenicity, Proteins, DD, COLLAB, and Reddit- ${ \it 2 k }$ are datasets representing real-world graphs and are taken from the repository of benchmark datasets for graph kernels4. Bench-easy and Bench-hard5 are datasets where the node features $\mathbf { X }$ and the adjacency matrix A are completely uninformative if considered alone. Hence, algorithms that account only for the node features or the graph structure will fail to classify the graphs. Since Bench-easy and Bench-hard come with a train/validation/test split, the 10-fold split is not necessary to evaluate the performance. The statistics of all the datasets are reported in Tab. 5. + +Table 5: Summary of statistics of the graph classification datasets + +
Datasetsamplesclassesavg. nodesavg. edgesnode attr.node labels
Bench-easy18003147.82922.66yes
Bench-hard18003148.32572.32yes
Mutagenicity4337230.3230.77yes
Proteins1113239.0672.821no
DD11782284.32715.661yes
COLLAB5000374.492457.78no
Reddit-2K20002429.63497.751no
+ +# C ARCHITECTURES SCHEMATA + +Fig. 8 reports the schematic representation of the minCUTpool layer; Fig. 9 the GNN architecture used in the clustering and segmentation tasks; Fig. 10 the GNN architecture used in the graph classification task; Fig. 12 the GNN architecture used in the graph regression task; Fig. 11 the graph autoencoder used in the graph signal reconstruction task. + +![](images/d9ade9085ab54361a1aa6626d8499f13441a1f331f8a266df6a2b0894a69a6db.jpg) +Figure 8: Schema of the minCUTpool layer. + +![](images/d710cb546df4f56d39a534e37d809d53060d1d9bc2b6ee1c45fe2367b58e3ef9.jpg) +Figure 9: Architecture for clustering/segmentation. + +![](images/da255a17d1270f108c2393a35c6f5725d6c431af5f91c472b5369ec2db9913d6.jpg) +Figure 10: Architecture for graph classification. + +![](images/969e11240083f473509da7280d39c4b88009ae72ad1c88f632f5a4207bc0742d.jpg) +Figure 11: Architecture for the autoencoder. + +![](images/031a97e555c2460b138d4ed99d26909b098af1b3775c29da264c3560104df77d.jpg) +Figure 12: Architecture for graph regression. \ No newline at end of file diff --git a/parse/train/BkxfshNYwB/BkxfshNYwB_content_list.json b/parse/train/BkxfshNYwB/BkxfshNYwB_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97e8338457febdd266a10831ba53749247029939 --- /dev/null +++ b/parse/train/BkxfshNYwB/BkxfshNYwB_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "MINCUT POOLING IN GRAPH NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 98, + 782, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 145, + 398, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 210, + 544, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advance of node pooling operations in Graph Neural Networks (GNNs) has lagged behind the feverish design of new message-passing techniques, and pooling remains an important and challenging endeavor for the design of deep architectures. In this paper, we propose a pooling operation for GNNs that leverages a differentiable unsupervised loss based on the minCUT optimization objective. For each node, our method learns a soft cluster assignment vector that depends on the node features, the target inference task (e.g., a graph classification loss), and, thanks to the minCUT objective, also on the connectivity structure of the graph. Graph pooling is obtained by applying the matrix of assignment vectors to the adjacency matrix and the node features. The proposed method can also be used as a stand-alone module to cluster vertexes in annotated graphs and solve unsupervised problems. We validate the effectiveness of the proposed pooling method on downstream tasks, including supervised graph classification and a set of unsupervised tasks, which reveal the limitations of existing GNN pooling approaches. ", + "bbox": [ + 233, + 242, + 764, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 467, + 334, + 482 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A fundamental component in deep convolutional neural networks is the pooling operation, which replaces the output of convolutions with local summaries of nearby points and is usually implemented by maximum or average operations (Lee et al., 2016). State-of-the-art architectures alternate convolutions, which extrapolate local patterns irrespective of the specific location on the input signal, and pooling, which lets the ensuing convolutions capture aggregated patterns. Pooling allows to learn abstract representations in deeper layers of the network by discarding information that is superfluous for the task, and keeps model complexity under control by limiting the growth of intermediate features. ", + "bbox": [ + 174, + 498, + 825, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graph Neural Networks (GNNs) extend the convolution operation from regular domains, such as images or time series, to data with arbitrary topologies and unordered structures described by graphs (Battaglia et al., 2018). The development of pooling strategies for GNNs, however, has lagged behind the design of newer and more effective message-passing (MP) operations (Gilmer et al., 2017), such as graph convolutions, mainly due to the difficulty of defining an aggregated version of the original graph that supports the pooled signal. ", + "bbox": [ + 174, + 617, + 825, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A na¨ıve pooling strategy in GNNs is to average all nodes features (Li et al., 2016), but it has limited flexibility since it does not extract local summaries of the graph structure, and no further MP operations can be applied afterwards. An alternative approach consists in pre-computing coarsened versions of the original graph and then fit the data to these deterministic structures (Bruna et al., 2013). While this aggregation accounts for the connectivity of the graph, it ignores task-specific objectives as well as the node features. ", + "bbox": [ + 174, + 708, + 823, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a differentiable pooling operation implemented as a neural network layer, which can be seamlessly combined with other MP layers (see Fig. 1). The parameters in the pooling layer are learned by combining the task-specific loss with an unsupervised regularization term, which optimizes a continuous relaxation of the normalized minCUT objective. The minCUT identifies dense graph components, where the nodes features become locally homogeneous after the message-passing. By gradually aggregating these components, the GNN learns to distil global properties from the graph. The proposed minCUT pooling operator (minCUTpool) yields partitions that 1) cluster together nodes which have similar features and are strongly connected on the graph, and 2) take into account the objective of the downstream task. ", + "bbox": [ + 174, + 799, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f1bdb5271287007bfb404fabc54a4bd48a828a415e925e71e65dd38e08b6950.jpg", + "image_caption": [ + "Figure 1: A deep GNN architecture where message-passing is followed by minCUT pooling. " + ], + "image_footnote": [], + "bbox": [ + 308, + 99, + 687, + 251 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 174, + 305, + 326, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 MINCUT AND SPECTRAL CLUSTERING ", + "text_level": 1, + "bbox": [ + 174, + 337, + 478, + 352 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a graph $G = \\{ \\nu , \\mathcal { E } \\}$ , $| \\nu | = N$ , and the associated adjacency matrix $\\mathbf { A } \\in \\mathbb { R } ^ { N \\times N }$ , the $K$ -way normalized minCUT (simply referred to as minCUT) aims at partitioning $\\nu$ in $K$ disjoint subsets by removing the minimum volume of edges. The problem is equivalent to maximizing ", + "bbox": [ + 174, + 362, + 825, + 406 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/bed538315cdc4b11b03e72a1f8508838e1bc29db0005bb3d50c0b887200c4dfa.jpg", + "text": "$$\n\\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\operatorname* { l i n k s } ( \\mathcal { V } _ { k } ) } { \\deg \\mathrm { r e e } ( \\mathcal { V } _ { k } ) } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\sum _ { i , j \\in \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } { \\sum _ { i \\in \\mathcal { V } _ { k } , j \\in \\mathcal { V } \\backslash \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } ,\n$$", + "text_format": "latex", + "bbox": [ + 333, + 411, + 665, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the numerator counts the edge volume within each cluster, and the denominator counts the edges between the nodes in a cluster and the rest of the graph (Shi & Malik, 2000). Let $\\mathbf { C } \\in \\mathbb { R } ^ { N \\times K }$ be a cluster assignment matrix, so that $\\mathbf { C } _ { i , j } = 1$ if node $i$ belongs to cluster $j$ , and 0 otherwise. The minCUT problem can be expressed as ", + "bbox": [ + 174, + 460, + 825, + 517 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/1bc8a238d93fae41b9a6b3f91c0e932f6d25110d313c304a11065a033efcef83.jpg", + "text": "$$\n\\mathrm { m a x i m i z e } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { { \\bf C } _ { k } ^ { T } { \\bf A } { \\bf C } _ { k } } { { \\bf C } _ { k } ^ { T } { \\bf D } { \\bf C } _ { k } } , \\mathrm { s . t . } { \\bf C } \\in \\{ 0 , 1 \\} ^ { N \\times K } , { \\bf C } { \\bf 1 } _ { K } = { \\bf 1 } _ { N } ,\n$$", + "text_format": "latex", + "bbox": [ + 274, + 523, + 722, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\mathbf { D } = \\mathrm { d i a g } ( \\mathbf { A } \\mathbf { 1 } _ { N } )$ is the degree matrix (Dhillon et al., 2004). Since problem (2) is NP-hard, it is usually recast in a relaxed formulation that can be solved in polynomial time and guarantees a near-optimal solution (Yu & Shi, 2003): ", + "bbox": [ + 174, + 574, + 826, + 617 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/8b4a96c5fa8a262b25a42766484b99a8cfede951176181e475f73ecda54a5854.jpg", + "text": "$$\n\\arg \\operatorname* { m a x } _ { \\mathbf { Q } \\in \\mathbb { R } ^ { N \\times K } } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { Q } _ { k } ^ { T } \\mathbf { A } \\mathbf { Q } _ { k } , \\quad \\mathrm { s . t . } \\mathbf { Q } = \\mathbf { C } ( \\mathbf { C } ^ { T } \\mathbf { D } \\mathbf { C } ) ^ { - \\frac { 1 } { 2 } } , \\ \\mathbf { Q } ^ { T } \\mathbf { Q } = \\mathbf { I } _ { K } .\n$$", + "text_format": "latex", + "bbox": [ + 269, + 623, + 728, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the optimization problem (3) is still non-convex, there exists an optimal solution $\\mathbf { Q } ^ { * } = \\mathbf { U } _ { K } \\mathbf { O }$ , where ${ \\bf U } _ { K } \\dot { } \\in \\mathbb { R } ^ { N \\times K }$ contains the eigenvectors of A corresponding to the $K$ largest eigenvalues, and $\\mathbf { O } \\in \\mathbb { R } ^ { K \\times K }$ is an orthogonal transformation (Ikebe et al., 1987). ", + "bbox": [ + 174, + 672, + 826, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the elements of $\\mathbf { Q } ^ { * }$ are real values rather than binary cluster indicators, the spectral clustering (SC) approach can be used to find discrete cluster assignments. In SC, the rows of $\\mathbf { Q } ^ { * }$ are treated as node representations embedded in the eigenspace of the Laplacian, and are clustered together with standard algorithms such as $k$ -means (Von Luxburg, 2007). One of the main limitations of SC lies in the computation of the spectrum of A, which has a memory complexity of $\\mathcal { O } ( N ^ { 2 } )$ and a computational complexity of $\\mathcal { O } ( \\bar { N } ^ { 3 } )$ . This prevents its applicability to large datasets. ", + "bbox": [ + 173, + 722, + 825, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To deal with such scalability issues, the constrained optimization in (3) can be solved by gradient descent algorithms that refine the solution by iterating operations whose individual complexity is $\\mathcal { O } ( N ^ { 2 } )$ , or even $\\mathcal { O } ( N )$ (Han & Filippone, 2017). Those algorithms search the solution on the manifold induced by the orthogonality constraint on the columns of $\\mathbf { Q }$ , by performing gradient updates along the geodesics (Wen & Yin, 2013; Collins et al., 2014). Alternative approaches rely on the QR factorization to constrain the space of feasible solutions (Damle et al., 2016), and alleviate the cost $\\mathcal { O } ( N ^ { 3 } )$ of the factorization by ensuring that orthogonality holds only on one minibatch at a time (Shaham et al., 2018). ", + "bbox": [ + 173, + 811, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other works based on neural networks include an autoencoder trained to map the $i$ th row of the Laplacian to the ith components of the first $K$ eigenvectors, to avoid the spectral decomposition (Tian et al., 2014). Yi et al. (2017) use a soft orthogonality constraint to learn spectral embeddings as a volumetric reparametrization of a precomputed Laplacian eigenbase. Shaham et al. (2018); Kampffmeyer et al. (2019) propose differentiable loss functions to partition generic data and process out-of-sample data at inference time. Nazi et al. (2019) generate balanced node partitions with a GNN, but adopt an optimization that does not encourage cluster assignments to be orthogonal. ", + "bbox": [ + 173, + 103, + 825, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 GRAPH NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 219, + 415, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many approaches have been proposed to process graphs with neural networks, including recurrent architectures (Scarselli et al., 2009; Li et al., 2016) or convolutional operations inspired by filters used in graph signal processing (Defferrard et al., 2016; Bianchi et al., 2019). Since our focus is on graph pooling, we base our GNN implementation on a simple MP operation, which combines the features of each node with its 1st-order neighbors. To account for the initial node features, it is possible to introduce self-loops by adding a (scaled) identity matrix to the diagonal of A (Kipf & Welling, 2017). Since our pooling will modify the structure of the adjacency matrix, we prefer a MP implementation that leaves the original A unaltered and accounts for the initial node features by means of skip connections. ", + "bbox": [ + 173, + 244, + 825, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\tilde { \\mathbf { A } } = \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\mathbf { A } \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\in \\mathbb { R } ^ { N \\times N }$ be the symmetrically normalized adjacency matrix and $\\mathbf { X } \\in \\mathbb { R } ^ { N \\times F }$ the matrix containing the node features. The output of the MP layer is ", + "bbox": [ + 171, + 378, + 823, + 409 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/34c777764a42e47800439e8acb679c0fc313d00d93b57909084dfaca360baee8.jpg", + "text": "$$\n\\mathbf { X } ^ { ( t + 1 ) } = M P ( \\mathbf { X } ^ { ( t ) } , \\tilde { \\mathbf { A } } ) = \\operatorname { R e L U } ( \\tilde { \\mathbf { A } } \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { m } + \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { s } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 299, + 416, + 696, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\boldsymbol { \\Theta } _ { M P } = \\{ \\mathbf { W } _ { m } , \\mathbf { W } _ { s } \\}$ are the trainable weights relative to the mixing and skip component of the layer, respectively. ", + "bbox": [ + 173, + 443, + 823, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 PROPOSED METHOD ", + "text_level": 1, + "bbox": [ + 176, + 493, + 372, + 510 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The minCUT pooling strategy computes a cluster assignment matrix $\\mathbf { S } \\in \\mathbb { R } ^ { N \\times K }$ by means of a multi-layer perceptron, which maps each node feature $\\mathbf { x } _ { i }$ into the ith row of S: ", + "bbox": [ + 171, + 525, + 825, + 554 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/eb9b8566cc8d18dce23ac90d2d6d28532fcd23484c72c59fadfa52de34d9d2e9.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbf { S } = s o f t m a x ( \\mathbf { R e L U } ( \\mathbf { X } \\mathbf { W } _ { 1 } ) \\mathbf { W } _ { 2 } ) , } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 379, + 563, + 616, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\boldsymbol { \\Theta } _ { P o o l } = \\{ \\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { F \\times H } , \\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { H \\times K } \\}$ are trainable parameters. The softmax function guarantees that $s _ { i , j } \\in [ 0 , 1 ]$ and enforces the constraints $\\mathbf { S 1 } _ { K } = \\mathbf { 1 } _ { N }$ inherited from the optimization problem in (2). The parameters $\\Theta _ { M P }$ and $\\Theta _ { P o o l }$ are jointly optimized by minimizing the usual task-specific loss, as well as an unsupervised loss $\\mathcal { L } _ { u }$ , which is composed of two terms ", + "bbox": [ + 174, + 585, + 825, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/093c5d20482e4733720500e265f3b4b47793157e111ce4ec091ed8c0b6c46652.jpg", + "text": "$$\n\\mathcal { L } _ { u } = \\mathcal { L } _ { c } + \\mathcal { L } _ { o } = \\underbrace { - \\frac { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } ) } } _ { \\mathcal { L } _ { c } } + \\underbrace { \\left\\| \\frac { \\mathbf { S } ^ { T } \\mathbf { S } } { \\| \\mathbf { S } ^ { T } \\mathbf { S } \\| _ { F } } - \\frac { \\mathbf { I } _ { K } } { \\sqrt { K } } \\right\\| _ { F } } _ { \\mathcal { L } _ { o } } ,\n$$", + "text_format": "latex", + "bbox": [ + 305, + 651, + 689, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _ { F }$ indicates the Frobenius norm. ", + "bbox": [ + 174, + 715, + 455, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The cut loss term, $\\mathcal { L } _ { c }$ , evaluates the minCUT given by the cluster assignment S, and is bounded by $- 1 \\leq \\mathcal { L } _ { c } \\leq 0$ . Minimizing $\\mathcal { L } _ { c }$ encourages strongly connected nodes to be clustered together, since the inner product $\\left. \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\right.$ increases when $\\tilde { a } _ { i , j }$ is large. $\\mathcal { L } _ { c }$ has a single maximum, reached when the numerator $\\begin{array} { r } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { S } _ { k } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } _ { k } = 0 } \\end{array}$ . This occurs if, for each pair of connected nodes (i.e., $\\tilde { a } _ { i , j } > 0 \\rangle$ ), the cluster assignments are orthogonal (i.e., $\\langle \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\rangle = 0 ,$ ). $\\mathcal { L } _ { c }$ reaches its minimum, $- 1$ , when $T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } )$ . This occurs when in a graph with $K$ disconnected components the cluster assignments are equal for all the nodes in the same component and orthogonal to the cluster assignments of nodes in different components. However, $\\mathcal { L } _ { c }$ is a non-convex function and its minimization can lead to local minima or degenerate solutions. For example, given a connected graph, a trivial optimal solution is the one that assigns all nodes to the same cluster. As a consequence of the continuous relaxation, another degenerate minimum occurs when the cluster assignments are all uniform, that is, all nodes are equally assigned to all clusters. This problem is exacerbated by prior message-passing operations, which make the node features more uniform. ", + "bbox": [ + 173, + 736, + 825, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The orthogonality loss term, $\\mathcal { L } _ { o }$ , penalizes the degenerate minima of $\\mathcal { L } _ { c }$ by encouraging the cluster assignments to be orthogonal and the clusters to be of similar size. Since the two matrices in $\\mathcal { L } _ { o }$ have unitary norm it is easy to see that $0 \\leq \\mathcal { L } _ { o } \\leq 2$ . Therefore, $\\mathcal { L } _ { o }$ does not dominate over $\\mathcal { L } _ { c }$ and the two terms can be safely summed directly (see Fig. 4 for an example). ${ \\mathbf { I } } _ { K }$ can be interpreted as a (rescaled) clustering matrix $\\mathbf { I } _ { K } = \\hat { \\mathbf { S } } ^ { T } \\hat { \\mathbf { S } }$ , where $\\hat { \\bf S }$ assigns exactly $N / K$ points to each cluster. The value of the Frobenius norm between clustering matrices is not dominated by the performance on the largest clusters (Law et al., 2017) and, thus, can be used to optimize intra-cluster variance. ", + "bbox": [ + 173, + 103, + 825, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Contrarily to SC methods that search for feasible solutions only within the space of orthogonal matrices, $\\mathcal { L } _ { o }$ only introduces a soft constraint that could be violated during the learning procedure. Since $\\mathcal { L } _ { c }$ is non-convex, the violation compromises the theoretical guarantee of convergence to the optimum of (3). However, we note that: ", + "bbox": [ + 176, + 209, + 825, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1. the cluster assignments S are well initialized: after the MP operation, the features of the connected vertices become similar and, since the MLP is a smooth function (Nelles, 2013), it yields similar cluster assignments for those vertices; \n2. in the GNN architecture, the minCUT objective is a regularization term and, therefore, a solution which is sub-optimal for (3) could instead be adequate for the specific objective of the downstream task; \n3. optimizing the task-specific loss helps the GNN to avoid the degenerate minima of $\\mathcal { L } _ { c }$ . ", + "bbox": [ + 210, + 276, + 825, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 COARSENING ", + "text_level": 1, + "bbox": [ + 174, + 401, + 310, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The coarsened version of the adjacency matrix and the graph signal are computed as ", + "bbox": [ + 174, + 426, + 725, + 443 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/c14e7c61c40a2f766857a297640952b7ad643866d8ac0d1f6858958d0c1eb67a.jpg", + "text": "$$\n{ \\bf A } ^ { p o o l } = { \\bf S } ^ { T } { \\tilde { \\bf A } } { \\bf S } ; ~ { \\bf X } ^ { p o o l } = { \\bf S } ^ { T } { \\bf X } ,\n$$", + "text_format": "latex", + "bbox": [ + 382, + 446, + 612, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the entry xpooi,j in $\\mathbf { X } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times F }$ is the weighted average value of feature $j$ among the elements in cluster $i$ . $\\mathbf { A } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times K }$ is a symmetric matrix, whose entries $a _ { i , i } ^ { p o o l }$ are the total number of edges between the nodes in the cluster i, while apooi,j is the number of edges between cluster $i$ and $j$ . Since $\\mathbf { A } ^ { p o o l }$ corresponds to the numerator of $\\mathcal { L } _ { c }$ in (7), the trace maximization yields clusters with many internal connections and weakly connected to each other. Hence, $\\mathbf { A } ^ { p o o l }$ will be a diagonal-dominant matrix, which describes a graph with self-loops much stronger than any other connection. Because self-loops hamper the propagation across adjacent nodes in the MP operations following the pooling layer, we compute the new adjacency matrix $\\tilde { \\mathbf { A } } ^ { p o o l }$ by zeroing the diagonal and by applying the degree normalization ", + "bbox": [ + 173, + 470, + 825, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/cdc85379486e51b65f69e07f035aec1a89b20127fdc77e297c858fa7fe2513dd.jpg", + "text": "$$\n\\hat { \\bf A } = { \\bf A } ^ { p o o l } - { \\bf I } _ { K } d i a g ( { \\bf A } ^ { p o o l } ) ; \\quad \\tilde { \\bf A } ^ { p o o l } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\hat { \\bf A } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } .\n$$", + "text_format": "latex", + "bbox": [ + 312, + 617, + 684, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d i a g ( \\cdot )$ returns the matrix diagonal. ", + "bbox": [ + 176, + 643, + 452, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 DISCUSSION AND RELATIONSHIP WITH SPECTRAL CLUSTERING ", + "text_level": 1, + "bbox": [ + 176, + 675, + 653, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed method is straightforward to implement: the cluster assignments, the loss, graph coarsening, and feature pooling are all computed with standard linear algebra operations. ", + "bbox": [ + 174, + 700, + 825, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are several differences between minCUTpool and classic SC methods. SC partitions the graph based on the Laplacian, but does not account for the node features. Instead, the cluster assignments $\\mathbf { s } _ { i }$ found by minCUTpool depend on $\\mathbf { x } _ { i }$ , which works well if connected nodes have similar features. This is a reasonable assumption in GNNs since, even in disassortative graphs (i.e., networks where dissimilar nodes are likely to be connected (Newman, 2003)), the features tend to become similar due to the MP operations. ", + "bbox": [ + 174, + 734, + 825, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Another difference is that SC handles a single graph and is not conceived for tasks with multiple graphs to be partitioned independently. Instead, thanks to the independence of the model parameters from the number of nodes $N$ and from the graph spectrum, minCUTpool can generalize to outof-sample data. This feature is fundamental in problems such as graph classification, where each sample is a graph with a different structure, and allows to train the model on small graphs and process larger ones at inference time. Finally, minCUTpool directly uses the soft cluster assignments rather than performing $k$ -means afterwards. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 RELATED WORK ON POOLING IN GNNS ", + "text_level": 1, + "bbox": [ + 174, + 102, + 534, + 117 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trainable pooling methods. Similarly to our method, these approaches learn how to generate coarsened version of the graph through differentiable functions, which take as input the nodes features $\\mathbf { X }$ and are parametrized by weights optimized on the task at hand. ", + "bbox": [ + 174, + 133, + 823, + 175 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffpool (Ying et al., 2018) is a pooling module that includes two parallel MP layers: one to compute the new node features $\\mathbf { X } ^ { ( t + 1 ) }$ and another to generate the cluster assignments S. Diffpool implements an unsupervised loss that consists of two terms. First, the link prediction term $\\| \\mathbf { A } - \\mathbf { S } \\mathbf { \\dot { S } } ^ { T } \\| _ { F }$ minimizes the Frobenius norm of the difference between the adjacency and the Gram matrix of the cluster assignments, encouraging nearby nodes to be clustered together. The second term $\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } H ( \\mathbf { S } _ { i } ) } \\end{array}$ minimizes the entropy of the cluster assignments to make them alike to one-hot vectors. Like minCUTpool, Diffpool clusters the vertices of annotated graphs, but yields completely different partitions, since it computes differently the clustering assignments, the coarsened adjacency matrix and, most importantly, the unsupervised loss. In Diffpool, such a loss shows pathological behaviors that are discussed later in the experiments. ", + "bbox": [ + 174, + 183, + 825, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The approach dubbed Top- $K$ pooling (Hongyang Gao, 2019; Lee et al., 2019), learns a projection vector that is applied to each node feature to obtain a score. The nodes with the $K$ highest scores are retained, the others are dropped. Since the top- $K$ selection is not differentiable, the scores are also used as a gate/attention for the node features, letting the projection vector to be trained with backpropagation. Top- $K$ is memory efficient as it avoids generating cluster assignments. To prevent A from becoming disconnected after nodes removal, Top- $K$ drops the rows and the columns from ${ \\bf A } ^ { 2 }$ and uses it as the new adjacency matrix. However, computing ${ \\bf A } ^ { 2 }$ costs $\\mathcal { O } ( N ^ { 2 } )$ and it is inefficient to implement with sparse operations. ", + "bbox": [ + 174, + 332, + 825, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Topological pooling methods. These methods pre-compute a pyramid of coarsened graphs, only taking into account the topology (A), but not the node features $\\mathbf { \\Pi } ( \\mathbf { X } )$ . During training, the node features are pooled with standard procedures and are fit into these deterministic graph structures. These methods are less flexible, but provide a stronger bias that can prevent degenerate solutions (e.g., coarsened graphs collapsing in a single node). ", + "bbox": [ + 174, + 460, + 823, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The approach proposed by Bruna et al. (2013), which has been adopted also in other GNN architectures (Defferrard et al., 2016; Fey et al., 2018), exploits GRACLUS (Dhillon et al., 2004), a hierarchical algorithm based on SC. At each pooling level $l$ , GRACLUS indetifies the pairs of maximally similar nodes $i _ { l }$ and $j _ { l }$ to be clustered together into a new vertex $k _ { ( l + 1 ) }$ . At inference phase, max-pooling is used to determine which node in the pair is kept. Fake vertices are added so that the number of nodes can be halved each time, but this injects noisy information in the graph. ", + "bbox": [ + 174, + 536, + 825, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Node decimation is a method originally proposed in graph signal processing literature (Shuman et al., 2016), which as been adapted also for GNNs (Simonovsky & Komodakis, 2017). The nodes are partitioned in two sets, according to the signs of the Laplacian eigenvector associated to the largest eigenvalue. One of the two sets is dropped, reducing the number of nodes each time approximately by half. Kron reduction is used to compute a pyramid of coarsened Laplacians from the remaining nodes. ", + "bbox": [ + 174, + 627, + 823, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A procedure proposed in Gama et al. (2018) diffuses a signal from designated nodes on the graph and stores the observed sequence of diffused components. The resulting stream of information is interpreted as a time signal, where standard CNN pooling is applied. We also mention a pooling operation for coarsening binary unweighted graphs by aggregating maximal cliques (Luzhnica et al., 2019). Nodes assigned to the same clique are summarized by max or average pooling and become a new node in the coarsened graph. ", + "bbox": [ + 174, + 717, + 825, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 821, + 326, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider both supervised and unsupervised tasks, and compare minCUTpool with other GNN pooling strategies. The Appendix provides further details on the experiments and a schematic depiction of the architectures used in each task. In addition, the Appendix reports two additional experiments: i) graph reconstruction by means of an Auto Encoder with bottleneck, implemented with pooling and un-pooling layers, ii) an architecture with pooling for graph regression. ", + "bbox": [ + 174, + 853, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1 CLUSTERING THE GRAPH NODES ", + "text_level": 1, + "bbox": [ + 176, + 104, + 442, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To study the effectiveness of the proposed loss, we perform different node clustering tasks with a simple GNN composed of a single MP layer followed by a pooling layer. The GNN is trained by minimizing $\\mathcal { L } _ { u }$ only, so that its effect is evaluated without the “interference” of a supervised loss. ", + "bbox": [ + 176, + 128, + 823, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clustering on synthetic networks We consider two simple graphs: the first is a network with 6 communities and the second is a regular grid. The adjacency matrix A is binary and the features X are the 2-D node coordinates. Fig. 2 depicts the node partitions generated by SC (a, d), Diffpool (b, e), and minCUTpool (c, f). Cluster indexes for Diffpool and minCUTpool are obtained by taking the argmax of S row-wise. Compared to SC, Diffpool and minCUTpool leverage the information contained in X. minCUTpool generates very accurate and balanced partitions, demonstrating that the cluster assignment matrix S is well formed. On the other hand, Diffpool assigns some nodes to the wrong community in the first example, and produces an imbalanced partition of the grid. ", + "bbox": [ + 173, + 186, + 825, + 299 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f54a1747a2f3fb20dd6df85e294051b41f7c8f617ba05ec5ba321652a3688701.jpg", + "image_caption": [ + "Figure 2: Node clustering on a community network $K { = } 6 )$ and on a grid graph $( K { = } 5 )$ " + ], + "image_footnote": [], + "bbox": [ + 259, + 316, + 738, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Image segmentation Given an image, we build a Region Adjacency Graph (Tremeau & Colan- ´ toni, 2000) using as nodes the regions generated by an oversegmentation procedure (Felzenszwalb & Huttenlocher, 2004). The SC technique used in this example is the recursive normalized cut (Shi & Malik, 2000), which recursively clusters the nodes until convergence. For Diffpool and minCUTpool, we include node features consisting of the average and total color in each oversegmented region. We set the number of desired clusters to $K = 4$ . The results in Fig. 3 show that minCUTpool yields a more precise segmentation. On the other hand, SC and Diffpool aggregate wrong regions and, in addition, SC finds too many segments. ", + "bbox": [ + 173, + 647, + 825, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clustering on citation networks We cluster the nodes of three popular citation networks: Cora, Citeseer, and Pubmed. The nodes are documents represented by sparse bag-of-words feature vectors stored in $\\mathbf { X }$ and the binary undirected edges in $\\mathbf { A }$ indicate citation links between documents. Each node $i$ is labeled with the document class $y _ { i }$ . Once the training is over, to test the quality of the partitions generated by each method we check the agreement between the cluster assignments and the true class labels. Tab. 1 reports the Completeness Score $\\begin{array} { r } { \\mathrm { C S } ( \\tilde { \\bf y } , { \\bf y } ) = 1 - \\frac { H ( \\tilde { \\bf y } | { \\bf y } ) } { H ( \\tilde { \\bf y } ) } } \\end{array}$ and Normalized Mutual Information $\\begin{array} { r } { \\mathbf { N M I } ( \\tilde { \\mathbf { y } } , \\mathbf { y } ) = \\frac { H ( \\tilde { \\mathbf { y } } ) - H ( \\tilde { \\mathbf { y } } | \\mathbf { y } ) } { \\sqrt { H ( \\tilde { \\mathbf { y } } ) - H ( \\mathbf { y } ) } } } \\end{array}$ where $H ( \\cdot )$ is the entropy. ", + "bbox": [ + 173, + 775, + 825, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The GNN architecture configured with minCUTpool achieves a higher NMI score than SC, which does not account for the node features $\\mathbf { X }$ when generating the partitions. Our pooling operation outperforms also Diffpool, since the minimization of the unsupervised loss in Diffpool yields degenerate solutions. The pathological behavior is shown in Fig. 4, which depicts the evolution of the NMI scores as the unsupervised losses in Diffpool and minCUTpool are minimized in training. ", + "bbox": [ + 173, + 895, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81a3f7ffc1776e74fbc56b414ab1c318578c5dc39e05ac6ea084422888017b44.jpg", + "image_caption": [ + "Figure 3: Image segmentation by clustering the nodes of the Region Adjacency Graph. " + ], + "image_footnote": [], + "bbox": [ + 199, + 108, + 807, + 337 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4067d319802597da0f8a99d17f9a05ae75ef1d9974bd1b46567b315150d067c6.jpg", + "image_caption": [ + "Figure 4: Unsupervised losses and NMI of Diffpool and minCUTpool on Cora. " + ], + "image_footnote": [], + "bbox": [ + 183, + 390, + 812, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 604, + 825, + 647 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4158255af3330a03321c8d25f82fd169f90fca4a4f87f98e57d954019d96471f.jpg", + "table_caption": [ + "Table 1: NMI and CS obtained by clustering the nodes on citation networks over 10 different runs. The number of clusters $K$ is equal to the number of node classes. " + ], + "table_footnote": [], + "table_body": "
DatasetKSpectral clusteringDiffpoolminCUTpool
NMIcsNMICSNMICS
Cora70.025 ± 0.0140.126 ± 0.0420.315 ± 0.0050.309 ±0.0050.404 ± 0.0180.392 ± 0.018
Citeseer60.014 ± 0.0030.033 ±0.0000.139 ± 0.0160.153 ± 0.0200.287 ±0.0470.283 ± 0.046
Pubmed30.182 ± 0.0000.261 ± 0.0000.079 ±0.0010.085 ±0.0010.200 ± 0.0200.197 ± 0.019
", + "bbox": [ + 187, + 703, + 810, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 SUPERVISED GRAPH CLASSIFICATION ", + "text_level": 1, + "bbox": [ + 176, + 814, + 473, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this task, the $i$ -th datum is a graph with $N _ { i }$ nodes represented by a pair $\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}$ and must be associated to the correct label $\\mathbf { y } _ { i }$ . We test the models on different graph classification datasets. For featureless graphs, we used the node degree information and the clustering coefficient as surrogate node features. We evaluate model performance with a 10-fold train/test split, using $1 0 \\%$ of the training set in each fold as validation for early stopping. We adopt a fixed network architecture, MP(32)-poolMP(32)-pool-MP(32)-GlobalAvgPool-softmax, where MP is the message-passing operation in (4) ", + "bbox": [ + 173, + 839, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with 32 hidden units. The pooling module is implemented either by Graclus, Decimation pooling, Top- $K$ , SAGPool (Lee et al., 2019), Diffpool, or the proposed minCUTpool. Each pooling method is configured to drop half of the nodes in a graph $K = N / 2$ in Top- $K$ , Diffpool, and minCUTpool). As baselines, we consider the popular Weisfeiler-Lehman (WL) graph kernel (Shervashidze et al., 2011), a network with only MP layers (Flat), and a fully connected network (Dense). ", + "bbox": [ + 174, + 103, + 825, + 174 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f42ea0a0a3ed9e12ba04e59f7afa9d42c08e5d9a16acdb713d983a07264519c7.jpg", + "table_caption": [ + "Table 2: Graph classification accuracy. Significantly better results $( p < 0 . 0 5 )$ are in bold. " + ], + "table_footnote": [], + "table_body": "
DatasetWLDenseFlatGraclusDecim.DiffpoolTop-KSAGpoolminCUT
Bench-easy92.629.3±0.398.5±0.397.5±0.597.9±0.598.6±0.482.4±8.984.2±2.399.0±0.0
Bench-hard60.029.4±0.367.6±2.869.0±1.572.6±0.969.9±1.942.7±15.237.7±14.573.8±1.9
Mutagenicity81.7±1.168.4±0.378.0±1.374.4±1.877.8±2.377.6±2.771.9±3.772.4±2.479.9±2.1
Proteins71.2±2.668.7±3.372.6±4.868.6±4.673.3±3.772.7±3.869.6±3.570.5±2.676.5±2.6
DD78.6±2.770.6±5.276.8±1.570.5±4.872.0±3.179.3±2.469.4±7.871.5±4.580.8±2.3
COLLAB74.8±1.379.3±1.682.1±1.877.1±2.179.1±1.581.8±1.479.3±1.879.2±2.083.4±1.7
Reddit-Binary68.2±1.748.5±2.680.3±2.679.2±0.484.3±2.486.8±2.174.7±4.573.9±5.191.4±1.5
", + "bbox": [ + 174, + 214, + 828, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tab. 2 reports the classification results, highlighting those that are significantly better ( $\\dot { p }$ -value $< ~ 0 . 0 5$ w.r.t. the method with the highest mean accuracy). The comparison with Flat helps to understand if a pooling operation is useful or not. The results of Dense, instead, help to quantify how much additional information is brought by the graph structure, with respect to the node features alone. It can be seen that minCUTpool obtains always equal or better results with respect to every other GNN architecture. On the other hand, some pooling procedures do not always improve the performance compared to the Flat baseline, making them not advisable to use in some cases. The WL kernel generally performs worse than the GNNs, except for the Mutagenicity dataset. This is probably because Mutagenicity has smaller graphs than the other datasets, and the adopted GNN architecture is overparametrized for this task. Interestingly, in some dataset such as Proteins and COLLAB it is possible to obtain fairly good classification accuracy with the Dense architecture, meaning that the graph structure only adds limited information. ", + "bbox": [ + 173, + 349, + 825, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e375310d148f5976514d2295a7d3c8830c48da3e0966c6040f4a07d9c8821927.jpg", + "image_caption": [ + "Figure 5: Average duration of one epoch using the same GNN with different pooling operations. Times were computed with an Nvidia GeForce GTX 1050, on the DD dataset with batch size of 1. " + ], + "image_footnote": [], + "bbox": [ + 178, + 535, + 459, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 reports a comparison of the execution time per training epoch for each pooling algorithm. Graclus and Decimation are understandably the fastest methods, since the coarsened graphs are precomputed. Among the differentiable pooling methods, minCUTpool is faster than Diffpool, which uses a slower MP layer rather than a MLP to compute cluster assignments, and than Top- $K$ , which computes the square of A at every forward pass. ", + "bbox": [ + 173, + 646, + 825, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 736, + 328, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a pooling layer for GNNs that coarsens a graph by taking into account both the the connectivity structure and the node features. The layer optimizes a regularization term based on the minCUT objective, which is minimized in conjunction with the task-specific loss to produce node partitions that are optimal for the task at hand. ", + "bbox": [ + 174, + 768, + 825, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We tested the effectiveness of our pooling strategy on unsupervised node clustering tasks, by optimizing only the unsupervised clustering loss, as well as supervised graph classification tasks on several popular benchmark datasets. Results show that minCUTpool performs significantly better than existing pooling strategies for GNNs. ", + "bbox": [ + 174, + 830, + 823, + 887 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 285, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi, Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261, 2018. ", + "bbox": [ + 173, + 126, + 825, + 181 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Filippo Maria Bianchi, Daniele Grattarola, L Livi, and C Alippi. Graph neural networks with convolutional arma filters. arXiv preprint arXiv:1901.01343, 2019. ", + "bbox": [ + 171, + 193, + 823, + 222 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013. ", + "bbox": [ + 171, + 229, + 823, + 260 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Maxwell D Collins, Ji Liu, Jia Xu, Lopamudra Mukherjee, and Vikas Singh. Spectral clustering with a convex regularizer on millions of images. In European Conference on Computer Vision, pp. 282–298. Springer, 2014. ", + "bbox": [ + 173, + 268, + 825, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Anil Damle, Victor Minden, and Lexing Ying. Robust and efficient multi-way spectral clustering. arXiv preprint arXiv:1609.08251, 2016. ", + "bbox": [ + 171, + 320, + 821, + 351 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks ¨ on graphs with fast localized spectral filtering. In Advances in Neural Information Processing Systems, pp. 3844–3852, 2016. ", + "bbox": [ + 173, + 359, + 825, + 404 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Kernel k-means: spectral clustering and normalized cuts. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 551–556. ACM, 2004. ", + "bbox": [ + 169, + 412, + 823, + 455 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59(2):167–181, 2004. ", + "bbox": [ + 168, + 464, + 823, + 494 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Muller. Splinecnn: Fast geomet- ¨ ric deep learning with continuous b-spline kernels. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 869–877, 2018. ", + "bbox": [ + 176, + 503, + 823, + 547 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fernando Gama, Antonio G Marques, Geert Leus, and Alejandro Ribeiro. Convolutional neural network architectures for signals supported on graphs. IEEE Transactions on Signal Processing, 67(4):1034–1049, 2018. ", + "bbox": [ + 174, + 555, + 823, + 598 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1263–1272. JMLR. org, 2017. ", + "bbox": [ + 173, + 608, + 825, + 651 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yufei Han and Maurizio Filippone. Mini-batch spectral clustering. In 2017 International Joint Conference on Neural Networks (IJCNN), pp. 3888–3895. IEEE, 2017. ", + "bbox": [ + 174, + 660, + 825, + 690 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Shuiwang Ji Hongyang Gao. Graph u-nets. In Proceedings of the 36th International conference on Machine learning (ICML), 2019. ", + "bbox": [ + 174, + 699, + 823, + 728 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Yasuhiko Ikebe, Toshiyuki Inagaki, and Sadaaki Miyamoto. The monotonicity theorem, cauchy’s interlace theorem, and the courant-fischer theorem. The American Mathematical Monthly, 94(4): 352–354, 1987. ", + "bbox": [ + 173, + 737, + 825, + 780 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michael Kampffmeyer, Sigurd Løkse, Filippo M. Bianchi, Lorenzo Livi, Arnt-Børre Salberg, and Robert Jenssen. Deep divergence-based approach to clustering. Neural Networks, 113:91 – 101, 2019. ", + "bbox": [ + 176, + 790, + 825, + 833 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. International Conference of Learning Representations (ICLR), 2017. ", + "bbox": [ + 171, + 843, + 823, + 872 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Marc T Law, Raquel Urtasun, and Richard S Zemel. Deep spectral clustering learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1985–1994. JMLR. org, 2017. ", + "bbox": [ + 174, + 881, + 823, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Chen-Yu Lee, Patrick W Gallagher, and Zhuowen Tu. Generalizing pooling functions in convolutional neural networks: Mixed, gated, and tree. In Artificial Intelligence and Statistics, pp. 464–472, 2016. ", + "bbox": [ + 173, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Junhyun Lee, Inyeop Lee, and Jaewoo Kang. Self-attention graph pooling. In Proceedings of the 36th International conference on Machine learning (ICML-19), 2019. ", + "bbox": [ + 171, + 155, + 823, + 184 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural networks. International Conference of Learning Representations (ICLR), 2016. ", + "bbox": [ + 173, + 191, + 823, + 222 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Enxhell Luzhnica, Ben Day, and Pietro Lio. Clique pooling for graph classification. International Conference of Learning Representations (ICLR) – Representation Learning on Graphs and Manifolds workshop, 2019. ", + "bbox": [ + 174, + 229, + 823, + 273 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Azade Nazi, Will Hang, Anna Goldie, Sujith Ravi, and Azalia Mirhoseini. Gap: Generalizable approximate graph partitioning framework. arXiv preprint arXiv:1903.00614, 2019. ", + "bbox": [ + 173, + 281, + 823, + 311 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Oliver Nelles. Nonlinear system identification: from classical approaches to neural networks and fuzzy models. Springer Science & Business Media, 2013. ", + "bbox": [ + 174, + 319, + 825, + 348 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Mark EJ Newman. Mixing patterns in networks. Physical Review E, 67(2):026126, 2003. ", + "bbox": [ + 171, + 357, + 763, + 373 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2009. ", + "bbox": [ + 174, + 381, + 820, + 410 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Uri Shaham, Kelly Stanton, Henry Li, Boaz Nadler, Ronen Basri, and Yuval Kluger. Spectralnet: Spectral clustering using deep neural networks. arXiv preprint arXiv:1801.01587, 2018. ", + "bbox": [ + 176, + 417, + 821, + 448 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nino Shervashidze, Pascal Schweitzer, Erik Jan van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12(Sep):2539– 2561, 2011. ", + "bbox": [ + 174, + 455, + 825, + 500 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. Departmental Papers (CIS), pp. 107, 2000. ", + "bbox": [ + 173, + 507, + 821, + 537 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "David I Shuman, Mohammad Javad Faraji, and Pierre Vandergheynst. A multiscale pyramid transform for graph signals. IEEE Transactions on Signal Processing, 64(8):2119–2134, 2016. ", + "bbox": [ + 173, + 545, + 820, + 575 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Martin Simonovsky and Nikos Komodakis. Dynamic edgeconditioned filters in convolutional neural networks on graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017. ", + "bbox": [ + 173, + 583, + 825, + 626 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Fei Tian, Bin Gao, Qing Cui, Enhong Chen, and Tie-Yan Liu. Learning deep representations for graph clustering. In AAAI, pp. 1293–1299, 2014. ", + "bbox": [ + 173, + 635, + 823, + 665 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alain Tremeau and Philippe Colantoni. Regions adjacency graph applied to color image segmenta-´ tion. IEEE Transactions on image processing, 9(4):735–744, 2000. ", + "bbox": [ + 173, + 672, + 823, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416, 2007. ", + "bbox": [ + 173, + 710, + 823, + 739 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zaiwen Wen and Wotao Yin. A feasible method for optimization with orthogonality constraints. Mathematical Programming, 142(1-2):397–434, 2013. ", + "bbox": [ + 173, + 747, + 823, + 777 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Li Yi, Hao Su, Xingwen Guo, and Leonidas J Guibas. Syncspeccnn: Synchronized spectral cnn for 3d shape segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 2282–2290, 2017. ", + "bbox": [ + 173, + 785, + 825, + 829 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hierarchical graph representation learning with differentiable pooling. In Advances in Neural Information Processing Systems, pp. 4800–4810, 2018. ", + "bbox": [ + 173, + 837, + 823, + 881 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yu and Shi. Multiclass spectral clustering. In Proceedings Ninth IEEE International Conference on Computer Vision, pp. 313–319 vol.1, Oct 2003. ", + "bbox": [ + 174, + 888, + 823, + 917 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 263, + 118 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A ADDITIONAL EXPERIMENTS ", + "bbox": [ + 178, + 137, + 442, + 154 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A.1 GNN AUTOENCODER ", + "text_level": 1, + "bbox": [ + 176, + 169, + 367, + 184 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "To compare the amount of information retained by the pooling layers in the coarsened graphs, we train an autoencoder (AE) to reconstruct a input graph signal $\\mathbf { X }$ from its pooled version. The AE architecture is MP(32)-MP(32)-pool-unpool-MP(32)-MP(32)-MP, and is trained by minimizing the mean squared error between the original and the reconstructed graph signal, $\\lVert \\bf { X } - \\bf { X } ^ { \\mathrm { { r e c } } } \\rVert ^ { 2 }$ . All the pooling operations are configured to retain $2 5 \\%$ of the original nodes. ", + "bbox": [ + 174, + 195, + 825, + 267 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Diffpool and minCUTpool, the unpool step is simply implemented by transposing the original pooling operations ", + "bbox": [ + 173, + 272, + 823, + 301 + ], + "page_idx": 10 + }, + { + "type": "equation", + "img_path": "images/8bbce8dd36526a6342dc4574ef8b5dab6566229c68e0a0326d22de1f3f93de0b.jpg", + "text": "$$\n{ \\bf X } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf X } ^ { \\mathrm { p o o l } } ; ~ { \\bf A } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf A } ^ { \\mathrm { p o o l } } { \\bf S } ^ { T } .\n$$", + "text_format": "latex", + "bbox": [ + 379, + 301, + 619, + 319 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Top- $K$ does not generate a cluster assignment matrix, but returns a binary mask $\\mathbf { m } = \\{ 0 , 1 \\} ^ { N }$ that indicates the nodes to drop (0) or to retain (1). Therefore, an upsamplig matrix $\\mathbf { U }$ is built by dropping the columns of the identity matrix ${ \\mathbf { I } } _ { N }$ that correspond to a 0 in $\\mathbf { m }$ , $\\bar { \\mathbf { U } } = [ \\mathbf { I } _ { N } ] _ { : , \\mathbf { m } = = 1 }$ . The unpooling operation is performed by replacing S with $\\mathbf { U }$ in (9), and the resulting upscaled graph is a version of the original graph with zeroes in correspondence of the dropped nodes. ", + "bbox": [ + 174, + 333, + 825, + 404 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/441bd5ff28bf0dbad772497856fe6d0256f85ad6ae1cfa337d7be79956b30a09.jpg", + "image_caption": [ + "Figure 6: AE reconstruction of a ring graph " + ], + "image_footnote": [], + "bbox": [ + 200, + 425, + 795, + 555 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/4b22636cf4edd79e814a6d27c3e71aeabc0f8293196e377e997385878fcd0602.jpg", + "image_caption": [ + "Figure 7: AE reconstruction of a grid graph " + ], + "image_footnote": [], + "bbox": [ + 202, + 614, + 795, + 744 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Fig. 6 and 7 report the original graph signal $\\mathbf { X }$ (the node features are the 2-D coordinates of the nodes) and the reconstruction $\\mathbf { X } ^ { \\mathrm { r e c } }$ obtained by using the different pooling methods, for a ring graph and a regular grid graph. The reconstruction produced by Diffpool is worse for the ring graph, but is almost perfect for the grid graph, while minCUTpool yields good results in both cases. On the other hand, Top- $K$ clearly fails in generating a coarsened representation that maintains enough information from the original graph. ", + "bbox": [ + 173, + 790, + 825, + 875 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This experiment highlights a major issue in Top- $K$ pooling, which retains the nodes associated to the highest $K$ values of a score vector s, computed by projecting the node features onto a trainable vector p: $\\mathbf { s } = \\mathbf { X } \\mathbf { p } $ . Nodes that are connected on the graph usually share similar features, and their similarity further increases after the MP operations, which combine the features of neighboring nodes. Retaining the nodes associated to the top $K$ scores in s corresponds to keeping those nodes that are alike and highly connected, as it can be seen in Fig. 6-7. Therefore, Top- $K$ discards entire portions of the graphs, which might contain important information. This explains why Top- $K$ fails to recover the original graph signal when used as bottleneck for the AE, and yields the worse performance among all GNN methods in the graph classification task. ", + "bbox": [ + 176, + 881, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.2 GRAPH REGRESSION OF MOLECULAR PROPERTIES ON QM9", + "text_level": 1, + "bbox": [ + 176, + 212, + 629, + 226 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The QM9 chemical database is a collection of ${ \\approx } 1 3 5 \\mathrm { k }$ small organic molecules, associated to continuous labels describing several geometric, energetic, electronic, and thermodynamic properties1. Each molecule in the dataset is represented as a graph $\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}$ , where atoms are associated to nodes, and edges represent chemical bonds. The atomic number of each atom (one-hot encoded; C, N, F, O) is taken as node feature and the type of bond (one-hot encoded; single, double, triple, aromatic) can be used as edge attribute. In this experiment, we ignore the edge attributes in order to use all pooling algorithms without modifications. ", + "bbox": [ + 174, + 239, + 825, + 337 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The purpose of this experiment is to compare the trainable pooling methods also on a graph regression task, but it must be intended as a proof of concept. In fact, the graphs in this dataset are extremely small (the average number of nodes is 8) and, therefore, a pooling operation is arguably not necessary. We consider a GNN with architecture MP(32)-pool-MP(32)-GlobalAvgPool-Dense, where pool is implemented by Top- $K$ , Diffpool, or minCUTpool. The network is trained to predict a given chemical property from the input molecular graphs. Performance is evaluated with a 10-fold cross-validation, using $1 \\dot { 0 } \\%$ of the training set for validation in each split. The GNNs are trained for 50 epochs, using Adam with learning rate 5e-4, batch size 32, and ReLU activations. We use the mean squared error (MSE) as supervised loss. ", + "bbox": [ + 173, + 344, + 825, + 469 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The MSE obtained on the prediction of each property for different pooling methods is reported in Tab. 3. As expected, the flat baseline with no pooling operation (MP(32)-MP(32)-GlobalAvgPoolDense) yields a lower error in most cases. Contrarily to the graph classification and the AE task, Top- $K$ achieves better results than Diffpool in average. Once again, minCUTpool significantly outperforms the other methods on each regression task and, in one case, also the flat baseline. ", + "bbox": [ + 174, + 477, + 823, + 546 + ], + "page_idx": 11 + }, + { + "type": "table", + "img_path": "images/abf2e28f0c4594a1f795c78d14e892f9e76511066dff22a973092bd24eabf709.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PropertyTop-KDiffpoolminCUTpoolFlat baseline
mu0.600±0.0850.651±0.0260.538±0.0120.559±0.007
alpha0.197±0.0870.114±0.0010.078±0.0070.065±0.006
homo0.698±0.1020.712±0.0150.526±0.0210.435±0.013
lumo0.601±0.0500.646±0.0130.540±0.0050.515±0.007
gap0.630±0.0440.698±0.0040.584±0.0070.552±0.008
r20.452±0.0870.440±0.0240.261±0.0060.204±0.006
zpve0.402±0.0320.410±0.0040.328±0.0050.284±0.005
uO_atom0.308±0.0550.245±0.0060.193±0.0020.163±0.001
cv0.291±0.1180.337±0.0180.148±0.0040.127±0.002
", + "bbox": [ + 259, + 568, + 736, + 712 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Table 3: MSE on the graph regression task. The best results with a statistical significance of $p < 0 . 0 5$ are highlighted: the best overall are in bold, the best among pooling methods are underlined. ", + "bbox": [ + 169, + 726, + 825, + 755 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B EXPERIMENTAL DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 796, + 416, + 811 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "For the WL kernel, we used the implementation provided in the GraKeL library2. The pooling strategy based on Graclus, is taken from the ChebyNets repository3. ", + "bbox": [ + 171, + 830, + 823, + 859 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "B.1 CLUSTERING ON CITATION NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 485, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Diffpool and minCUTpool are configured with 16 hidden neurons with linear activations in the MLP and MP layer, respectively used to compute the cluster assignment matrix S. The MP layer used to compute the propagated node features $\\mathbf { X } ^ { ( 1 ) }$ uses an ELU activation in both architectures. The learning rate for Adam is 5e-4, and the models are trained for 10000 iterations. The details of the citation networks dataset are reported in Tab. 4. ", + "bbox": [ + 174, + 128, + 825, + 200 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/83643cb5b85febe056dfd089e5192e8b209287c5ac7b40081e179d6e516f65fb.jpg", + "table_caption": [ + "Table 4: Details of the citation networks datasets " + ], + "table_footnote": [], + "table_body": "
DatasetNodesEdgesNode featuresNode classes
Cora2708542914337
Citeseer3327922837036
Pubmed19717886515003
", + "bbox": [ + 312, + 239, + 684, + 308 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "B.2 GRAPH CLASSIFICATION ", + "text_level": 1, + "bbox": [ + 174, + 334, + 387, + 349 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We train the GNN architectures with Adam, an $\\mathrm { L } _ { 2 }$ penalty loss with weight 1e-4, and 16 hidden units $( H )$ both in the MLP of minCUTpool and in the internal MP of Diffpool. Mutagenicity, Proteins, DD, COLLAB, and Reddit- ${ \\it 2 k }$ are datasets representing real-world graphs and are taken from the repository of benchmark datasets for graph kernels4. Bench-easy and Bench-hard5 are datasets where the node features $\\mathbf { X }$ and the adjacency matrix A are completely uninformative if considered alone. Hence, algorithms that account only for the node features or the graph structure will fail to classify the graphs. Since Bench-easy and Bench-hard come with a train/validation/test split, the 10-fold split is not necessary to evaluate the performance. The statistics of all the datasets are reported in Tab. 5. ", + "bbox": [ + 173, + 359, + 826, + 486 + ], + "page_idx": 12 + }, + { + "type": "table", + "img_path": "images/8139375cff6db7f44f94bbb4867d33a66608ffa28a0ec65a6dcfdb33bd0f0efa.jpg", + "table_caption": [ + "Table 5: Summary of statistics of the graph classification datasets " + ], + "table_footnote": [], + "table_body": "
Datasetsamplesclassesavg. nodesavg. edgesnode attr.node labels
Bench-easy18003147.82922.66yes
Bench-hard18003148.32572.32yes
Mutagenicity4337230.3230.77yes
Proteins1113239.0672.821no
DD11782284.32715.661yes
COLLAB5000374.492457.78no
Reddit-2K20002429.63497.751no
", + "bbox": [ + 225, + 527, + 772, + 645 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "C ARCHITECTURES SCHEMATA ", + "text_level": 1, + "bbox": [ + 176, + 676, + 446, + 693 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Fig. 8 reports the schematic representation of the minCUTpool layer; Fig. 9 the GNN architecture used in the clustering and segmentation tasks; Fig. 10 the GNN architecture used in the graph classification task; Fig. 12 the GNN architecture used in the graph regression task; Fig. 11 the graph autoencoder used in the graph signal reconstruction task. ", + "bbox": [ + 173, + 707, + 825, + 763 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/d9ade9085ab54361a1aa6626d8499f13441a1f331f8a266df6a2b0894a69a6db.jpg", + "image_caption": [ + "Figure 8: Schema of the minCUTpool layer. " + ], + "image_footnote": [], + "bbox": [ + 299, + 103, + 699, + 236 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/d710cb546df4f56d39a534e37d809d53060d1d9bc2b6ee1c45fe2367b58e3ef9.jpg", + "image_caption": [ + "Figure 9: Architecture for clustering/segmentation. " + ], + "image_footnote": [], + "bbox": [ + 400, + 304, + 591, + 361 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/da255a17d1270f108c2393a35c6f5725d6c431af5f91c472b5369ec2db9913d6.jpg", + "image_caption": [ + "Figure 10: Architecture for graph classification. " + ], + "image_footnote": [], + "bbox": [ + 230, + 426, + 766, + 482 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/969e11240083f473509da7280d39c4b88009ae72ad1c88f632f5a4207bc0742d.jpg", + "image_caption": [ + "Figure 11: Architecture for the autoencoder. " + ], + "image_footnote": [], + "bbox": [ + 222, + 541, + 774, + 599 + ], + "page_idx": 13 + }, + { + "type": "image", + "img_path": "images/031a97e555c2460b138d4ed99d26909b098af1b3775c29da264c3560104df77d.jpg", + "image_caption": [ + "Figure 12: Architecture for graph regression. " + ], + "image_footnote": [], + "bbox": [ + 294, + 659, + 700, + 715 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/BkxfshNYwB/BkxfshNYwB_middle.json b/parse/train/BkxfshNYwB/BkxfshNYwB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..c79b973da3b8c1cba41a35f3cf98becacde9bfda --- /dev/null +++ b/parse/train/BkxfshNYwB/BkxfshNYwB_middle.json @@ -0,0 +1,36646 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 479, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 481, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 481, + 99 + ], + "score": 1.0, + "content": "MINCUT POOLING IN GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 179 + ], + "lines": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 192, + 468, + 346 + ], + "lines": [ + { + "bbox": [ + 141, + 192, + 469, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 192, + 469, + 205 + ], + "score": 1.0, + "content": "The advance of node pooling operations in Graph Neural Networks (GNNs) has", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 203, + 470, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 470, + 217 + ], + "score": 1.0, + "content": "lagged behind the feverish design of new message-passing techniques, and pool-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 214, + 470, + 227 + ], + "spans": [ + { + "bbox": [ + 141, + 214, + 470, + 227 + ], + "score": 1.0, + "content": "ing remains an important and challenging endeavor for the design of deep archi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "score": 1.0, + "content": "tectures. In this paper, we propose a pooling operation for GNNs that leverages", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "a differentiable unsupervised loss based on the minCUT optimization objective.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 247, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 247, + 469, + 259 + ], + "score": 1.0, + "content": "For each node, our method learns a soft cluster assignment vector that depends on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 258, + 469, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 469, + 271 + ], + "score": 1.0, + "content": "the node features, the target inference task (e.g., a graph classification loss), and,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "score": 1.0, + "content": "thanks to the minCUT objective, also on the connectivity structure of the graph.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "score": 1.0, + "content": "Graph pooling is obtained by applying the matrix of assignment vectors to the ad-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 292, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 292, + 470, + 303 + ], + "score": 1.0, + "content": "jacency matrix and the node features. The proposed method can also be used as a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 303, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 470, + 314 + ], + "score": 1.0, + "content": "stand-alone module to cluster vertexes in annotated graphs and solve unsupervised", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "score": 1.0, + "content": "problems. We validate the effectiveness of the proposed pooling method on down-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 325, + 469, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 469, + 336 + ], + "score": 1.0, + "content": "stream tasks, including supervised graph classification and a set of unsupervised", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 334, + 432, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 432, + 348 + ], + "score": 1.0, + "content": "tasks, which reveal the limitations of existing GNN pooling approaches.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 370, + 205, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 208, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 208, + 385 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "A fundamental component in deep convolutional neural networks is the pooling operation, which re-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "places the output of convolutions with local summaries of nearby points and is usually implemented", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "by maximum or average operations (Lee et al., 2016). State-of-the-art architectures alternate convo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "lutions, which extrapolate local patterns irrespective of the specific location on the input signal, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "pooling, which lets the ensuing convolutions capture aggregated patterns. Pooling allows to learn", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "abstract representations in deeper layers of the network by discarding information that is superflu-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "ous for the task, and keeps model complexity under control by limiting the growth of intermediate", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 472, + 144, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 144, + 484 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Graph Neural Networks (GNNs) extend the convolution operation from regular domains, such as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "images or time series, to data with arbitrary topologies and unordered structures described by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "graphs (Battaglia et al., 2018). The development of pooling strategies for GNNs, however, has", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "lagged behind the design of newer and more effective message-passing (MP) operations (Gilmer", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "et al., 2017), such as graph convolutions, mainly due to the difficulty of defining an aggregated", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 349, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 349, + 557 + ], + "score": 1.0, + "content": "version of the original graph that supports the pooled signal.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "A na¨ıve pooling strategy in GNNs is to average all nodes features (Li et al., 2016), but it has lim-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "ited flexibility since it does not extract local summaries of the graph structure, and no further MP", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "operations can be applied afterwards. An alternative approach consists in pre-computing coarsened", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "versions of the original graph and then fit the data to these deterministic structures (Bruna et al.,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "2013). While this aggregation accounts for the connectivity of the graph, it ignores task-specific", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 616, + 263, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 263, + 627 + ], + "score": 1.0, + "content": "objectives as well as the node features.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "In this paper, we propose a differentiable pooling operation implemented as a neural network layer,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "which can be seamlessly combined with other MP layers (see Fig. 1). The parameters in the pool-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "ing layer are learned by combining the task-specific loss with an unsupervised regularization term,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "which optimizes a continuous relaxation of the normalized minCUT objective. The minCUT iden-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "tifies dense graph components, where the nodes features become locally homogeneous after the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 686, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 505, + 702 + ], + "score": 1.0, + "content": "message-passing. By gradually aggregating these components, the GNN learns to distil global prop-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "erties from the graph. The proposed minCUT pooling operator (minCUTpool) yields partitions that", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "1) cluster together nodes which have similar features and are strongly connected on the graph, and", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 338, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 338, + 732 + ], + "score": 1.0, + "content": "2) take into account the objective of the downstream task.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 479, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 77, + 481, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 77, + 481, + 99 + ], + "score": 1.0, + "content": "MINCUT POOLING IN GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 112, + 115, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "spans": [ + { + "bbox": [ + 113, + 115, + 201, + 127 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "spans": [ + { + "bbox": [ + 111, + 126, + 245, + 138 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 111, + 115, + 245, + 138 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 167, + 333, + 179 + ], + "lines": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "spans": [ + { + "bbox": [ + 277, + 166, + 335, + 180 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 192, + 468, + 346 + ], + "lines": [ + { + "bbox": [ + 141, + 192, + 469, + 205 + ], + "spans": [ + { + "bbox": [ + 141, + 192, + 469, + 205 + ], + "score": 1.0, + "content": "The advance of node pooling operations in Graph Neural Networks (GNNs) has", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 203, + 470, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 470, + 217 + ], + "score": 1.0, + "content": "lagged behind the feverish design of new message-passing techniques, and pool-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 214, + 470, + 227 + ], + "spans": [ + { + "bbox": [ + 141, + 214, + 470, + 227 + ], + "score": 1.0, + "content": "ing remains an important and challenging endeavor for the design of deep archi-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 470, + 239 + ], + "score": 1.0, + "content": "tectures. In this paper, we propose a pooling operation for GNNs that leverages", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 469, + 249 + ], + "score": 1.0, + "content": "a differentiable unsupervised loss based on the minCUT optimization objective.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 247, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 247, + 469, + 259 + ], + "score": 1.0, + "content": "For each node, our method learns a soft cluster assignment vector that depends on", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 258, + 469, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 258, + 469, + 271 + ], + "score": 1.0, + "content": "the node features, the target inference task (e.g., a graph classification loss), and,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "score": 1.0, + "content": "thanks to the minCUT objective, also on the connectivity structure of the graph.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 293 + ], + "score": 1.0, + "content": "Graph pooling is obtained by applying the matrix of assignment vectors to the ad-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 292, + 470, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 292, + 470, + 303 + ], + "score": 1.0, + "content": "jacency matrix and the node features. The proposed method can also be used as a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 303, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 303, + 470, + 314 + ], + "score": 1.0, + "content": "stand-alone module to cluster vertexes in annotated graphs and solve unsupervised", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 470, + 326 + ], + "score": 1.0, + "content": "problems. We validate the effectiveness of the proposed pooling method on down-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 325, + 469, + 336 + ], + "spans": [ + { + "bbox": [ + 141, + 325, + 469, + 336 + ], + "score": 1.0, + "content": "stream tasks, including supervised graph classification and a set of unsupervised", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 334, + 432, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 432, + 348 + ], + "score": 1.0, + "content": "tasks, which reveal the limitations of existing GNN pooling approaches.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 10.5, + "bbox_fs": [ + 141, + 192, + 470, + 348 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 370, + 205, + 382 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 208, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 208, + 385 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 395, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 505, + 408 + ], + "score": 1.0, + "content": "A fundamental component in deep convolutional neural networks is the pooling operation, which re-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 505, + 419 + ], + "score": 1.0, + "content": "places the output of convolutions with local summaries of nearby points and is usually implemented", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 505, + 430 + ], + "score": 1.0, + "content": "by maximum or average operations (Lee et al., 2016). State-of-the-art architectures alternate convo-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "lutions, which extrapolate local patterns irrespective of the specific location on the input signal, and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 506, + 453 + ], + "score": 1.0, + "content": "pooling, which lets the ensuing convolutions capture aggregated patterns. Pooling allows to learn", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 505, + 464 + ], + "score": 1.0, + "content": "abstract representations in deeper layers of the network by discarding information that is superflu-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "ous for the task, and keeps model complexity under control by limiting the growth of intermediate", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 472, + 144, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 144, + 484 + ], + "score": 1.0, + "content": "features.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 395, + 506, + 484 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "Graph Neural Networks (GNNs) extend the convolution operation from regular domains, such as", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "images or time series, to data with arbitrary topologies and unordered structures described by", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 524 + ], + "score": 1.0, + "content": "graphs (Battaglia et al., 2018). The development of pooling strategies for GNNs, however, has", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 505, + 535 + ], + "score": 1.0, + "content": "lagged behind the design of newer and more effective message-passing (MP) operations (Gilmer", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "score": 1.0, + "content": "et al., 2017), such as graph convolutions, mainly due to the difficulty of defining an aggregated", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 545, + 349, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 349, + 557 + ], + "score": 1.0, + "content": "version of the original graph that supports the pooled signal.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 489, + 506, + 557 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 561, + 504, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "A na¨ıve pooling strategy in GNNs is to average all nodes features (Li et al., 2016), but it has lim-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "ited flexibility since it does not extract local summaries of the graph structure, and no further MP", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "operations can be applied afterwards. An alternative approach consists in pre-computing coarsened", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "versions of the original graph and then fit the data to these deterministic structures (Bruna et al.,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "2013). While this aggregation accounts for the connectivity of the graph, it ignores task-specific", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 616, + 263, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 263, + 627 + ], + "score": 1.0, + "content": "objectives as well as the node features.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 561, + 506, + 627 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "In this paper, we propose a differentiable pooling operation implemented as a neural network layer,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "which can be seamlessly combined with other MP layers (see Fig. 1). The parameters in the pool-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "ing layer are learned by combining the task-specific loss with an unsupervised regularization term,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "which optimizes a continuous relaxation of the normalized minCUT objective. The minCUT iden-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "tifies dense graph components, where the nodes features become locally homogeneous after the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 686, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 505, + 702 + ], + "score": 1.0, + "content": "message-passing. By gradually aggregating these components, the GNN learns to distil global prop-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "erties from the graph. The proposed minCUT pooling operator (minCUTpool) yields partitions that", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "1) cluster together nodes which have similar features and are strongly connected on the graph, and", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 338, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 338, + 732 + ], + "score": 1.0, + "content": "2) take into account the objective of the downstream task.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 43, + "bbox_fs": [ + 104, + 632, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 79, + 421, + 199 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 189, + 79, + 421, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 189, + 79, + 421, + 199 + ], + "spans": [ + { + "bbox": [ + 189, + 79, + 421, + 199 + ], + "score": 0.969, + "type": "image", + "image_path": "2f1bdb5271287007bfb404fabc54a4bd48a828a415e925e71e65dd38e08b6950.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 79, + 421, + 94.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 189, + 94.0, + 421, + 109.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 189, + 109.0, + 421, + 124.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 189, + 124.0, + 421, + 139.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 189, + 139.0, + 421, + 154.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 189, + 154.0, + 421, + 169.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 189, + 169.0, + 421, + 184.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 189, + 184.0, + 421, + 199.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 119, + 211, + 491, + 224 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 209, + 493, + 227 + ], + "spans": [ + { + "bbox": [ + 117, + 209, + 493, + 227 + ], + "score": 1.0, + "content": "Figure 1: A deep GNN architecture where message-passing is followed by minCUT pooling.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + }, + { + "type": "title", + "bbox": [ + 107, + 242, + 200, + 255 + ], + "lines": [ + { + "bbox": [ + 104, + 241, + 201, + 258 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 201, + 258 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 107, + 267, + 293, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 294, + 280 + ], + "score": 1.0, + "content": "2.1 MINCUT AND SPECTRAL CLUSTERING", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 287, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 104, + 285, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 165, + 302 + ], + "score": 1.0, + "content": "Given a graph", + "type": "text" + }, + { + "bbox": [ + 165, + 288, + 215, + 300 + ], + "score": 0.79, + "content": "G = \\{ \\nu , \\mathcal { E } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 285, + 219, + 302 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 219, + 288, + 255, + 300 + ], + "score": 0.73, + "content": "| \\nu | = N", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 285, + 405, + 302 + ], + "score": 1.0, + "content": ", and the associated adjacency matrix", + "type": "text" + }, + { + "bbox": [ + 406, + 287, + 455, + 298 + ], + "score": 0.92, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { N \\times N }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 285, + 474, + 302 + ], + "score": 1.0, + "content": ", the", + "type": "text" + }, + { + "bbox": [ + 474, + 288, + 484, + 298 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 285, + 506, + 302 + ], + "score": 1.0, + "content": "-way", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 396, + 312 + ], + "score": 1.0, + "content": "normalized minCUT (simply referred to as minCUT) aims at partitioning", + "type": "text" + }, + { + "bbox": [ + 397, + 299, + 405, + 309 + ], + "score": 0.78, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 298, + 416, + 312 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 417, + 299, + 427, + 309 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 298, + 505, + 312 + ], + "score": 1.0, + "content": "disjoint subsets by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 309, + 441, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 441, + 323 + ], + "score": 1.0, + "content": "removing the minimum volume of edges. The problem is equivalent to maximizing", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 326, + 407, + 361 + ], + "lines": [ + { + "bbox": [ + 204, + 326, + 407, + 361 + ], + "spans": [ + { + "bbox": [ + 204, + 326, + 407, + 361 + ], + "score": 0.94, + "content": "\\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\operatorname* { l i n k s } ( \\mathcal { V } _ { k } ) } { \\deg \\mathrm { r e e } ( \\mathcal { V } _ { k } ) } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\sum _ { i , j \\in \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } { \\sum _ { i \\in \\mathcal { V } _ { k } , j \\in \\mathcal { V } \\backslash \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } ,", + "type": "interline_equation", + "image_path": "bed538315cdc4b11b03e72a1f8508838e1bc29db0005bb3d50c0b887200c4dfa.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 326, + 407, + 343.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 204, + 343.5, + 407, + 361.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "where the numerator counts the edge volume within each cluster, and the denominator counts the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 376, + 503, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 454, + 389 + ], + "score": 1.0, + "content": "edges between the nodes in a cluster and the rest of the graph (Shi & Malik, 2000). Let", + "type": "text" + }, + { + "bbox": [ + 454, + 376, + 503, + 387 + ], + "score": 0.92, + "content": "\\mathbf { C } \\in \\mathbb { R } ^ { N \\times K }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 262, + 401 + ], + "score": 1.0, + "content": "be a cluster assignment matrix, so that", + "type": "text" + }, + { + "bbox": [ + 262, + 388, + 299, + 400 + ], + "score": 0.92, + "content": "\\mathbf { C } _ { i , j } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 387, + 330, + 401 + ], + "score": 1.0, + "content": "if node", + "type": "text" + }, + { + "bbox": [ + 330, + 389, + 335, + 398 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 387, + 408, + 401 + ], + "score": 1.0, + "content": "belongs to cluster", + "type": "text" + }, + { + "bbox": [ + 409, + 389, + 415, + 399 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 387, + 505, + 401 + ], + "score": 1.0, + "content": ", and 0 otherwise. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 258, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 258, + 411 + ], + "score": 1.0, + "content": "minCUT problem can be expressed as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 415, + 442, + 450 + ], + "lines": [ + { + "bbox": [ + 168, + 415, + 442, + 450 + ], + "spans": [ + { + "bbox": [ + 168, + 415, + 442, + 450 + ], + "score": 0.92, + "content": "\\mathrm { m a x i m i z e } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { { \\bf C } _ { k } ^ { T } { \\bf A } { \\bf C } _ { k } } { { \\bf C } _ { k } ^ { T } { \\bf D } { \\bf C } _ { k } } , \\mathrm { s . t . } { \\bf C } \\in \\{ 0 , 1 \\} ^ { N \\times K } , { \\bf C } { \\bf 1 } _ { K } = { \\bf 1 } _ { N } ,", + "type": "interline_equation", + "image_path": "1bc8a238d93fae41b9a6b3f91c0e932f6d25110d313c304a11065a033efcef83.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 168, + 415, + 442, + 426.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 168, + 426.6666666666667, + 442, + 438.33333333333337 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 168, + 438.33333333333337, + 442, + 450.00000000000006 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 506, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 133, + 468 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 455, + 205, + 467 + ], + "score": 0.9, + "content": "\\mathbf { D } = \\mathrm { d i a g } ( \\mathbf { A } \\mathbf { 1 } _ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "is the degree matrix (Dhillon et al., 2004). Since problem (2) is NP-hard,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "it is usually recast in a relaxed formulation that can be solved in polynomial time and guarantees a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 478, + 268, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 268, + 489 + ], + "score": 1.0, + "content": "near-optimal solution (Yu & Shi, 2003):", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 494, + 446, + 529 + ], + "lines": [ + { + "bbox": [ + 165, + 494, + 446, + 529 + ], + "spans": [ + { + "bbox": [ + 165, + 494, + 446, + 529 + ], + "score": 0.93, + "content": "\\arg \\operatorname* { m a x } _ { \\mathbf { Q } \\in \\mathbb { R } ^ { N \\times K } } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { Q } _ { k } ^ { T } \\mathbf { A } \\mathbf { Q } _ { k } , \\quad \\mathrm { s . t . } \\mathbf { Q } = \\mathbf { C } ( \\mathbf { C } ^ { T } \\mathbf { D } \\mathbf { C } ) ^ { - \\frac { 1 } { 2 } } , \\ \\mathbf { Q } ^ { T } \\mathbf { Q } = \\mathbf { I } _ { K } .", + "type": "interline_equation", + "image_path": "8b4a96c5fa8a262b25a42766484b99a8cfede951176181e475f73ecda54a5854.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 165, + 494, + 446, + 505.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 165, + 505.6666666666667, + 446, + 517.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 165, + 517.3333333333334, + 446, + 529.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 506, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 448, + 546 + ], + "score": 1.0, + "content": "While the optimization problem (3) is still non-convex, there exists an optimal solution", + "type": "text" + }, + { + "bbox": [ + 449, + 533, + 501, + 545 + ], + "score": 0.91, + "content": "\\mathbf { Q } ^ { * } = \\mathbf { U } _ { K } \\mathbf { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 532, + 505, + 546 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 541, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 134, + 559 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 543, + 195, + 555 + ], + "score": 0.91, + "content": "{ \\bf U } _ { K } \\dot { } \\in \\mathbb { R } ^ { N \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 541, + 411, + 559 + ], + "score": 1.0, + "content": "contains the eigenvectors of A corresponding to the", + "type": "text" + }, + { + "bbox": [ + 412, + 545, + 422, + 554 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 541, + 506, + 559 + ], + "score": 1.0, + "content": "largest eigenvalues,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 553, + 384, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 123, + 568 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 555, + 173, + 565 + ], + "score": 0.9, + "content": "\\mathbf { O } \\in \\mathbb { R } ^ { K \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 553, + 384, + 568 + ], + "score": 1.0, + "content": "is an orthogonal transformation (Ikebe et al., 1987).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 194, + 586 + ], + "score": 1.0, + "content": "Since the elements of", + "type": "text" + }, + { + "bbox": [ + 194, + 572, + 208, + 583 + ], + "score": 0.9, + "content": "\\mathbf { Q } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "are real values rather than binary cluster indicators, the spectral clustering", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 444, + 596 + ], + "score": 1.0, + "content": "(SC) approach can be used to find discrete cluster assignments. In SC, the rows of", + "type": "text" + }, + { + "bbox": [ + 444, + 583, + 459, + 594 + ], + "score": 0.89, + "content": "\\mathbf { Q } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "are treated", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "as node representations embedded in the eigenspace of the Laplacian, and are clustered together", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 240, + 617 + ], + "score": 1.0, + "content": "with standard algorithms such as", + "type": "text" + }, + { + "bbox": [ + 240, + 605, + 247, + 615 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "-means (Von Luxburg, 2007). One of the main limitations of SC", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 447, + 629 + ], + "score": 1.0, + "content": "lies in the computation of the spectrum of A, which has a memory complexity of", + "type": "text" + }, + { + "bbox": [ + 447, + 615, + 478, + 628 + ], + "score": 0.93, + "content": "\\mathcal { O } ( N ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "and a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 449, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 224, + 640 + ], + "score": 1.0, + "content": "computational complexity of", + "type": "text" + }, + { + "bbox": [ + 225, + 627, + 254, + 639 + ], + "score": 0.93, + "content": "\\mathcal { O } ( \\bar { N } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 626, + 449, + 640 + ], + "score": 1.0, + "content": ". This prevents its applicability to large datasets.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "To deal with such scalability issues, the constrained optimization in (3) can be solved by gradient", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "descent algorithms that refine the solution by iterating operations whose individual complexity is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 664, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 107, + 665, + 137, + 678 + ], + "score": 0.92, + "content": "\\mathcal { O } ( N ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 664, + 177, + 679 + ], + "score": 1.0, + "content": ", or even", + "type": "text" + }, + { + "bbox": [ + 178, + 666, + 204, + 678 + ], + "score": 0.92, + "content": "\\mathcal { O } ( N )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 664, + 506, + 679 + ], + "score": 1.0, + "content": "(Han & Filippone, 2017). Those algorithms search the solution on the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 392, + 690 + ], + "score": 1.0, + "content": "manifold induced by the orthogonality constraint on the columns of", + "type": "text" + }, + { + "bbox": [ + 392, + 677, + 402, + 688 + ], + "score": 0.68, + "content": "\\mathbf { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 676, + 506, + 690 + ], + "score": 1.0, + "content": ", by performing gradient", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "updates along the geodesics (Wen & Yin, 2013; Collins et al., 2014). Alternative approaches rely on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "the QR factorization to constrain the space of feasible solutions (Damle et al., 2016), and alleviate", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 708, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 140, + 723 + ], + "score": 1.0, + "content": "the cost", + "type": "text" + }, + { + "bbox": [ + 140, + 709, + 170, + 722 + ], + "score": 0.92, + "content": "\\mathcal { O } ( N ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 708, + 507, + 723 + ], + "score": 1.0, + "content": "of the factorization by ensuring that orthogonality holds only on one minibatch at a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "score": 1.0, + "content": "time (Shaham et al., 2018).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 189, + 79, + 421, + 199 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 189, + 79, + 421, + 199 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 189, + 79, + 421, + 199 + ], + "spans": [ + { + "bbox": [ + 189, + 79, + 421, + 199 + ], + "score": 0.969, + "type": "image", + "image_path": "2f1bdb5271287007bfb404fabc54a4bd48a828a415e925e71e65dd38e08b6950.jpg" + } + ] + } + ], + "index": 3.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 79, + 421, + 94.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 189, + 94.0, + 421, + 109.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 189, + 109.0, + 421, + 124.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 189, + 124.0, + 421, + 139.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 189, + 139.0, + 421, + 154.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 189, + 154.0, + 421, + 169.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 189, + 169.0, + 421, + 184.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 189, + 184.0, + 421, + 199.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 119, + 211, + 491, + 224 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 209, + 493, + 227 + ], + "spans": [ + { + "bbox": [ + 117, + 209, + 493, + 227 + ], + "score": 1.0, + "content": "Figure 1: A deep GNN architecture where message-passing is followed by minCUT pooling.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 5.75 + }, + { + "type": "title", + "bbox": [ + 107, + 242, + 200, + 255 + ], + "lines": [ + { + "bbox": [ + 104, + 241, + 201, + 258 + ], + "spans": [ + { + "bbox": [ + 104, + 241, + 201, + 258 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 107, + 267, + 293, + 279 + ], + "lines": [ + { + "bbox": [ + 106, + 267, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 106, + 267, + 294, + 280 + ], + "score": 1.0, + "content": "2.1 MINCUT AND SPECTRAL CLUSTERING", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 287, + 505, + 322 + ], + "lines": [ + { + "bbox": [ + 104, + 285, + 506, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 285, + 165, + 302 + ], + "score": 1.0, + "content": "Given a graph", + "type": "text" + }, + { + "bbox": [ + 165, + 288, + 215, + 300 + ], + "score": 0.79, + "content": "G = \\{ \\nu , \\mathcal { E } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 285, + 219, + 302 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 219, + 288, + 255, + 300 + ], + "score": 0.73, + "content": "| \\nu | = N", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 285, + 405, + 302 + ], + "score": 1.0, + "content": ", and the associated adjacency matrix", + "type": "text" + }, + { + "bbox": [ + 406, + 287, + 455, + 298 + ], + "score": 0.92, + "content": "\\mathbf { A } \\in \\mathbb { R } ^ { N \\times N }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 285, + 474, + 302 + ], + "score": 1.0, + "content": ", the", + "type": "text" + }, + { + "bbox": [ + 474, + 288, + 484, + 298 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 285, + 506, + 302 + ], + "score": 1.0, + "content": "-way", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 298, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 396, + 312 + ], + "score": 1.0, + "content": "normalized minCUT (simply referred to as minCUT) aims at partitioning", + "type": "text" + }, + { + "bbox": [ + 397, + 299, + 405, + 309 + ], + "score": 0.78, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 298, + 416, + 312 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 417, + 299, + 427, + 309 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 428, + 298, + 505, + 312 + ], + "score": 1.0, + "content": "disjoint subsets by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 309, + 441, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 441, + 323 + ], + "score": 1.0, + "content": "removing the minimum volume of edges. The problem is equivalent to maximizing", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 104, + 285, + 506, + 323 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 204, + 326, + 407, + 361 + ], + "lines": [ + { + "bbox": [ + 204, + 326, + 407, + 361 + ], + "spans": [ + { + "bbox": [ + 204, + 326, + 407, + 361 + ], + "score": 0.94, + "content": "\\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\operatorname* { l i n k s } ( \\mathcal { V } _ { k } ) } { \\deg \\mathrm { r e e } ( \\mathcal { V } _ { k } ) } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\sum _ { i , j \\in \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } { \\sum _ { i \\in \\mathcal { V } _ { k } , j \\in \\mathcal { V } \\backslash \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } ,", + "type": "interline_equation", + "image_path": "bed538315cdc4b11b03e72a1f8508838e1bc29db0005bb3d50c0b887200c4dfa.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 204, + 326, + 407, + 343.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 204, + 343.5, + 407, + 361.0 + ], + "spans": [], + "index": 15 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 505, + 377 + ], + "score": 1.0, + "content": "where the numerator counts the edge volume within each cluster, and the denominator counts the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 376, + 503, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 454, + 389 + ], + "score": 1.0, + "content": "edges between the nodes in a cluster and the rest of the graph (Shi & Malik, 2000). Let", + "type": "text" + }, + { + "bbox": [ + 454, + 376, + 503, + 387 + ], + "score": 0.92, + "content": "\\mathbf { C } \\in \\mathbb { R } ^ { N \\times K }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 262, + 401 + ], + "score": 1.0, + "content": "be a cluster assignment matrix, so that", + "type": "text" + }, + { + "bbox": [ + 262, + 388, + 299, + 400 + ], + "score": 0.92, + "content": "\\mathbf { C } _ { i , j } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 387, + 330, + 401 + ], + "score": 1.0, + "content": "if node", + "type": "text" + }, + { + "bbox": [ + 330, + 389, + 335, + 398 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 387, + 408, + 401 + ], + "score": 1.0, + "content": "belongs to cluster", + "type": "text" + }, + { + "bbox": [ + 409, + 389, + 415, + 399 + ], + "score": 0.8, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 387, + 505, + 401 + ], + "score": 1.0, + "content": ", and 0 otherwise. The", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 399, + 258, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 258, + 411 + ], + "score": 1.0, + "content": "minCUT problem can be expressed as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 365, + 505, + 411 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 415, + 442, + 450 + ], + "lines": [ + { + "bbox": [ + 168, + 415, + 442, + 450 + ], + "spans": [ + { + "bbox": [ + 168, + 415, + 442, + 450 + ], + "score": 0.92, + "content": "\\mathrm { m a x i m i z e } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { { \\bf C } _ { k } ^ { T } { \\bf A } { \\bf C } _ { k } } { { \\bf C } _ { k } ^ { T } { \\bf D } { \\bf C } _ { k } } , \\mathrm { s . t . } { \\bf C } \\in \\{ 0 , 1 \\} ^ { N \\times K } , { \\bf C } { \\bf 1 } _ { K } = { \\bf 1 } _ { N } ,", + "type": "interline_equation", + "image_path": "1bc8a238d93fae41b9a6b3f91c0e932f6d25110d313c304a11065a033efcef83.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 168, + 415, + 442, + 426.6666666666667 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 168, + 426.6666666666667, + 442, + 438.33333333333337 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 168, + 438.33333333333337, + 442, + 450.00000000000006 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 455, + 506, + 489 + ], + "lines": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 133, + 468 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 455, + 205, + 467 + ], + "score": 0.9, + "content": "\\mathbf { D } = \\mathrm { d i a g } ( \\mathbf { A } \\mathbf { 1 } _ { N } )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "is the degree matrix (Dhillon et al., 2004). Since problem (2) is NP-hard,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "it is usually recast in a relaxed formulation that can be solved in polynomial time and guarantees a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 478, + 268, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 268, + 489 + ], + "score": 1.0, + "content": "near-optimal solution (Yu & Shi, 2003):", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 455, + 506, + 489 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 165, + 494, + 446, + 529 + ], + "lines": [ + { + "bbox": [ + 165, + 494, + 446, + 529 + ], + "spans": [ + { + "bbox": [ + 165, + 494, + 446, + 529 + ], + "score": 0.93, + "content": "\\arg \\operatorname* { m a x } _ { \\mathbf { Q } \\in \\mathbb { R } ^ { N \\times K } } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { Q } _ { k } ^ { T } \\mathbf { A } \\mathbf { Q } _ { k } , \\quad \\mathrm { s . t . } \\mathbf { Q } = \\mathbf { C } ( \\mathbf { C } ^ { T } \\mathbf { D } \\mathbf { C } ) ^ { - \\frac { 1 } { 2 } } , \\ \\mathbf { Q } ^ { T } \\mathbf { Q } = \\mathbf { I } _ { K } .", + "type": "interline_equation", + "image_path": "8b4a96c5fa8a262b25a42766484b99a8cfede951176181e475f73ecda54a5854.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 165, + 494, + 446, + 505.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 165, + 505.6666666666667, + 446, + 517.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 165, + 517.3333333333334, + 446, + 529.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 506, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 532, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 448, + 546 + ], + "score": 1.0, + "content": "While the optimization problem (3) is still non-convex, there exists an optimal solution", + "type": "text" + }, + { + "bbox": [ + 449, + 533, + 501, + 545 + ], + "score": 0.91, + "content": "\\mathbf { Q } ^ { * } = \\mathbf { U } _ { K } \\mathbf { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 532, + 505, + 546 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 541, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 104, + 541, + 134, + 559 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 543, + 195, + 555 + ], + "score": 0.91, + "content": "{ \\bf U } _ { K } \\dot { } \\in \\mathbb { R } ^ { N \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 541, + 411, + 559 + ], + "score": 1.0, + "content": "contains the eigenvectors of A corresponding to the", + "type": "text" + }, + { + "bbox": [ + 412, + 545, + 422, + 554 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 541, + 506, + 559 + ], + "score": 1.0, + "content": "largest eigenvalues,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 553, + 384, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 123, + 568 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 555, + 173, + 565 + ], + "score": 0.9, + "content": "\\mathbf { O } \\in \\mathbb { R } ^ { K \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 553, + 384, + 568 + ], + "score": 1.0, + "content": "is an orthogonal transformation (Ikebe et al., 1987).", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 104, + 532, + 506, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 572, + 505, + 639 + ], + "lines": [ + { + "bbox": [ + 105, + 570, + 506, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 194, + 586 + ], + "score": 1.0, + "content": "Since the elements of", + "type": "text" + }, + { + "bbox": [ + 194, + 572, + 208, + 583 + ], + "score": 0.9, + "content": "\\mathbf { Q } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 570, + 506, + 586 + ], + "score": 1.0, + "content": "are real values rather than binary cluster indicators, the spectral clustering", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 583, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 444, + 596 + ], + "score": 1.0, + "content": "(SC) approach can be used to find discrete cluster assignments. In SC, the rows of", + "type": "text" + }, + { + "bbox": [ + 444, + 583, + 459, + 594 + ], + "score": 0.89, + "content": "\\mathbf { Q } ^ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 583, + 506, + 596 + ], + "score": 1.0, + "content": "are treated", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 506, + 606 + ], + "score": 1.0, + "content": "as node representations embedded in the eigenspace of the Laplacian, and are clustered together", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 240, + 617 + ], + "score": 1.0, + "content": "with standard algorithms such as", + "type": "text" + }, + { + "bbox": [ + 240, + 605, + 247, + 615 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 604, + 506, + 617 + ], + "score": 1.0, + "content": "-means (Von Luxburg, 2007). One of the main limitations of SC", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 615, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 447, + 629 + ], + "score": 1.0, + "content": "lies in the computation of the spectrum of A, which has a memory complexity of", + "type": "text" + }, + { + "bbox": [ + 447, + 615, + 478, + 628 + ], + "score": 0.93, + "content": "\\mathcal { O } ( N ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 615, + 506, + 629 + ], + "score": 1.0, + "content": "and a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 626, + 449, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 224, + 640 + ], + "score": 1.0, + "content": "computational complexity of", + "type": "text" + }, + { + "bbox": [ + 225, + 627, + 254, + 639 + ], + "score": 0.93, + "content": "\\mathcal { O } ( \\bar { N } ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 626, + 449, + 640 + ], + "score": 1.0, + "content": ". This prevents its applicability to large datasets.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 570, + 506, + 640 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 643, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 657 + ], + "score": 1.0, + "content": "To deal with such scalability issues, the constrained optimization in (3) can be solved by gradient", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "descent algorithms that refine the solution by iterating operations whose individual complexity is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 107, + 664, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 107, + 665, + 137, + 678 + ], + "score": 0.92, + "content": "\\mathcal { O } ( N ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 664, + 177, + 679 + ], + "score": 1.0, + "content": ", or even", + "type": "text" + }, + { + "bbox": [ + 178, + 666, + 204, + 678 + ], + "score": 0.92, + "content": "\\mathcal { O } ( N )", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 664, + 506, + 679 + ], + "score": 1.0, + "content": "(Han & Filippone, 2017). Those algorithms search the solution on the", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 676, + 392, + 690 + ], + "score": 1.0, + "content": "manifold induced by the orthogonality constraint on the columns of", + "type": "text" + }, + { + "bbox": [ + 392, + 677, + 402, + 688 + ], + "score": 0.68, + "content": "\\mathbf { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 676, + 506, + 690 + ], + "score": 1.0, + "content": ", by performing gradient", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 701 + ], + "score": 1.0, + "content": "updates along the geodesics (Wen & Yin, 2013; Collins et al., 2014). Alternative approaches rely on", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "the QR factorization to constrain the space of feasible solutions (Damle et al., 2016), and alleviate", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 708, + 507, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 140, + 723 + ], + "score": 1.0, + "content": "the cost", + "type": "text" + }, + { + "bbox": [ + 140, + 709, + 170, + 722 + ], + "score": 0.92, + "content": "\\mathcal { O } ( N ^ { 3 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 708, + 507, + 723 + ], + "score": 1.0, + "content": "of the factorization by ensuring that orthogonality holds only on one minibatch at a", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 217, + 733 + ], + "score": 1.0, + "content": "time (Shaham et al., 2018).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 41.5, + "bbox_fs": [ + 104, + 643, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 445, + 95 + ], + "score": 1.0, + "content": "Other works based on neural networks include an autoencoder trained to map the", + "type": "text" + }, + { + "bbox": [ + 445, + 84, + 450, + 92 + ], + "score": 0.31, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "th row of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 295, + 106 + ], + "score": 1.0, + "content": "Laplacian to the ith components of the first", + "type": "text" + }, + { + "bbox": [ + 295, + 94, + 307, + 104 + ], + "score": 0.75, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "eigenvectors, to avoid the spectral decomposi-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "tion (Tian et al., 2014). Yi et al. (2017) use a soft orthogonality constraint to learn spectral em-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "beddings as a volumetric reparametrization of a precomputed Laplacian eigenbase. Shaham et al.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "(2018); Kampffmeyer et al. (2019) propose differentiable loss functions to partition generic data and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "process out-of-sample data at inference time. Nazi et al. (2019) generate balanced node partitions", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "with a GNN, but adopt an optimization that does not encourage cluster assignments to be orthogonal.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 108, + 174, + 254, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 255, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 255, + 187 + ], + "score": 1.0, + "content": "2.2 GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "Many approaches have been proposed to process graphs with neural networks, including recurrent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 219 + ], + "score": 1.0, + "content": "architectures (Scarselli et al., 2009; Li et al., 2016) or convolutional operations inspired by filters", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 217, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 230 + ], + "score": 1.0, + "content": "used in graph signal processing (Defferrard et al., 2016; Bianchi et al., 2019). Since our focus is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 228, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 506, + 241 + ], + "score": 1.0, + "content": "on graph pooling, we base our GNN implementation on a simple MP operation, which combines", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "the features of each node with its 1st-order neighbors. To account for the initial node features, it is", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "possible to introduce self-loops by adding a (scaled) identity matrix to the diagonal of A (Kipf &", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 507, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 507, + 275 + ], + "score": 1.0, + "content": "Welling, 2017). Since our pooling will modify the structure of the adjacency matrix, we prefer a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 285 + ], + "score": 1.0, + "content": "MP implementation that leaves the original A unaltered and accounts for the initial node features", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 283, + 229, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 229, + 296 + ], + "score": 1.0, + "content": "by means of skip connections.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 105, + 300, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 104, + 298, + 504, + 316 + ], + "spans": [ + { + "bbox": [ + 104, + 298, + 122, + 316 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 299, + 236, + 312 + ], + "score": 0.92, + "content": "\\tilde { \\mathbf { A } } = \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\mathbf { A } \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\in \\mathbb { R } ^ { N \\times N }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 298, + 455, + 316 + ], + "score": 1.0, + "content": "be the symmetrically normalized adjacency matrix and", + "type": "text" + }, + { + "bbox": [ + 455, + 300, + 504, + 312 + ], + "score": 0.91, + "content": "\\mathbf { X } \\in \\mathbb { R } ^ { N \\times F }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 312, + 388, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 388, + 325 + ], + "score": 1.0, + "content": "the matrix containing the node features. The output of the MP layer is", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 330, + 426, + 346 + ], + "lines": [ + { + "bbox": [ + 183, + 330, + 426, + 346 + ], + "spans": [ + { + "bbox": [ + 183, + 330, + 426, + 346 + ], + "score": 0.9, + "content": "\\mathbf { X } ^ { ( t + 1 ) } = M P ( \\mathbf { X } ^ { ( t ) } , \\tilde { \\mathbf { A } } ) = \\operatorname { R e L U } ( \\tilde { \\mathbf { A } } \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { m } + \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { s } ) ,", + "type": "interline_equation", + "image_path": "34c777764a42e47800439e8acb679c0fc313d00d93b57909084dfaca360baee8.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 183, + 330, + 426, + 346 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 133, + 364 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 351, + 222, + 364 + ], + "score": 0.93, + "content": "\\boldsymbol { \\Theta } _ { M P } = \\{ \\mathbf { W } _ { m } , \\mathbf { W } _ { s } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 350, + 506, + 364 + ], + "score": 1.0, + "content": "are the trainable weights relative to the mixing and skip component of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 362, + 198, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 198, + 376 + ], + "score": 1.0, + "content": "the layer, respectively.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 108, + 391, + 228, + 404 + ], + "lines": [ + { + "bbox": [ + 104, + 390, + 230, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 390, + 230, + 406 + ], + "score": 1.0, + "content": "3 PROPOSED METHOD", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 416, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 507, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 390, + 430 + ], + "score": 1.0, + "content": "The minCUT pooling strategy computes a cluster assignment matrix", + "type": "text" + }, + { + "bbox": [ + 390, + 416, + 442, + 427 + ], + "score": 0.92, + "content": "\\mathbf { S } \\in \\mathbb { R } ^ { N \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 413, + 507, + 430 + ], + "score": 1.0, + "content": "by means of a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 428, + 423, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 324, + 440 + ], + "score": 1.0, + "content": "multi-layer perceptron, which maps each node feature", + "type": "text" + }, + { + "bbox": [ + 324, + 430, + 335, + 439 + ], + "score": 0.86, + "content": "\\mathbf { x } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 428, + 423, + 440 + ], + "score": 1.0, + "content": "into the ith row of S:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 446, + 377, + 459 + ], + "lines": [ + { + "bbox": [ + 232, + 446, + 377, + 459 + ], + "spans": [ + { + "bbox": [ + 232, + 446, + 377, + 459 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathbf { S } = s o f t m a x ( \\mathbf { R e L U } ( \\mathbf { X } \\mathbf { W } _ { 1 } ) \\mathbf { W } _ { 2 } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "eb9b8566cc8d18dce23ac90d2d6d28532fcd23484c72c59fadfa52de34d9d2e9.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 232, + 446, + 377, + 459 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 463, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 463, + 133, + 479 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 464, + 309, + 478 + ], + "score": 0.83, + "content": "\\boldsymbol { \\Theta } _ { P o o l } = \\{ \\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { F \\times H } , \\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { H \\times K } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 463, + 506, + 479 + ], + "score": 1.0, + "content": "are trainable parameters. The softmax function", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 168, + 490 + ], + "score": 1.0, + "content": "guarantees that", + "type": "text" + }, + { + "bbox": [ + 169, + 478, + 216, + 489 + ], + "score": 0.91, + "content": "s _ { i , j } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 477, + 329, + 490 + ], + "score": 1.0, + "content": "and enforces the constraints", + "type": "text" + }, + { + "bbox": [ + 330, + 477, + 377, + 488 + ], + "score": 0.92, + "content": "\\mathbf { S 1 } _ { K } = \\mathbf { 1 } _ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 477, + 506, + 490 + ], + "score": 1.0, + "content": "inherited from the optimization", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 240, + 500 + ], + "score": 1.0, + "content": "problem in (2). The parameters", + "type": "text" + }, + { + "bbox": [ + 240, + 488, + 266, + 499 + ], + "score": 0.92, + "content": "\\Theta _ { M P }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 487, + 285, + 500 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 488, + 313, + 499 + ], + "score": 0.91, + "content": "\\Theta _ { P o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "are jointly optimized by minimizing the usual", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 498, + 455, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 305, + 511 + ], + "score": 1.0, + "content": "task-specific loss, as well as an unsupervised loss", + "type": "text" + }, + { + "bbox": [ + 306, + 499, + 318, + 510 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 498, + 455, + 511 + ], + "score": 1.0, + "content": ", which is composed of two terms", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 516, + 422, + 561 + ], + "lines": [ + { + "bbox": [ + 187, + 516, + 422, + 561 + ], + "spans": [ + { + "bbox": [ + 187, + 516, + 422, + 561 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { u } = \\mathcal { L } _ { c } + \\mathcal { L } _ { o } = \\underbrace { - \\frac { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } ) } } _ { \\mathcal { L } _ { c } } + \\underbrace { \\left\\| \\frac { \\mathbf { S } ^ { T } \\mathbf { S } } { \\| \\mathbf { S } ^ { T } \\mathbf { S } \\| _ { F } } - \\frac { \\mathbf { I } _ { K } } { \\sqrt { K } } \\right\\| _ { F } } _ { \\mathcal { L } _ { o } } ,", + "type": "interline_equation", + "image_path": "093c5d20482e4733720500e265f3b4b47793157e111ce4ec091ed8c0b6c46652.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 187, + 516, + 422, + 531.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 187, + 531.0, + 422, + 546.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 187, + 546.0, + 422, + 561.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 279, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 280, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 133, + 580 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 568, + 158, + 579 + ], + "score": 0.92, + "content": "\\| \\cdot \\| _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 567, + 280, + 580 + ], + "score": 1.0, + "content": "indicates the Frobenius norm.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 583, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 181, + 597 + ], + "score": 1.0, + "content": "The cut loss term,", + "type": "text" + }, + { + "bbox": [ + 182, + 585, + 194, + 595 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 584, + 505, + 597 + ], + "score": 1.0, + "content": ", evaluates the minCUT given by the cluster assignment S, and is bounded by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 164, + 606 + ], + "score": 0.91, + "content": "- 1 \\leq \\mathcal { L } _ { c } \\leq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 595, + 218, + 609 + ], + "score": 1.0, + "content": ". Minimizing", + "type": "text" + }, + { + "bbox": [ + 218, + 596, + 231, + 606 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "encourages strongly connected nodes to be clustered together, since", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 178, + 619 + ], + "score": 1.0, + "content": "the inner product", + "type": "text" + }, + { + "bbox": [ + 178, + 606, + 207, + 618 + ], + "score": 0.92, + "content": "\\left. \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 606, + 272, + 619 + ], + "score": 1.0, + "content": "increases when", + "type": "text" + }, + { + "bbox": [ + 272, + 607, + 288, + 618 + ], + "score": 0.9, + "content": "\\tilde { a } _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 606, + 324, + 619 + ], + "score": 1.0, + "content": "is large.", + "type": "text" + }, + { + "bbox": [ + 324, + 606, + 336, + 617 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "has a single maximum, reached when the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 101, + 616, + 508, + 635 + ], + "spans": [ + { + "bbox": [ + 101, + 616, + 151, + 635 + ], + "score": 1.0, + "content": "numerator", + "type": "text" + }, + { + "bbox": [ + 151, + 618, + 304, + 633 + ], + "score": 0.92, + "content": "\\begin{array} { r } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { S } _ { k } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } _ { k } = 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 616, + 508, + 635 + ], + "score": 1.0, + "content": ". This occurs if, for each pair of connected nodes", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 127, + 644 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 127, + 631, + 163, + 643 + ], + "score": 0.9, + "content": "\\tilde { a } _ { i , j } > 0 \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 631, + 346, + 644 + ], + "score": 1.0, + "content": "), the cluster assignments are orthogonal (i.e.,", + "type": "text" + }, + { + "bbox": [ + 347, + 631, + 396, + 643 + ], + "score": 0.91, + "content": "\\langle \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\rangle = 0 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 631, + 402, + 644 + ], + "score": 1.0, + "content": ").", + "type": "text" + }, + { + "bbox": [ + 403, + 631, + 415, + 642 + ], + "score": 0.86, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "reaches its minimum,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 642, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 107, + 644, + 120, + 655 + ], + "score": 0.79, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 642, + 150, + 657 + ], + "score": 1.0, + "content": ", when", + "type": "text" + }, + { + "bbox": [ + 151, + 642, + 264, + 656 + ], + "score": 0.92, + "content": "T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } )", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 642, + 414, + 657 + ], + "score": 1.0, + "content": ". This occurs when in a graph with", + "type": "text" + }, + { + "bbox": [ + 414, + 644, + 425, + 654 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 642, + 505, + 657 + ], + "score": 1.0, + "content": "disconnected com-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "ponents the cluster assignments are equal for all the nodes in the same component and orthogonal", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 399, + 678 + ], + "score": 1.0, + "content": "to the cluster assignments of nodes in different components. However,", + "type": "text" + }, + { + "bbox": [ + 400, + 666, + 412, + 677 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "is a non-convex func-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "tion and its minimization can lead to local minima or degenerate solutions. For example, given a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "connected graph, a trivial optimal solution is the one that assigns all nodes to the same cluster. As", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "a consequence of the continuous relaxation, another degenerate minimum occurs when the cluster", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "assignments are all uniform, that is, all nodes are equally assigned to all clusters. This problem is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 488, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 488, + 733 + ], + "score": 1.0, + "content": "exacerbated by prior message-passing operations, which make the node features more uniform.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 40 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 160 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 445, + 95 + ], + "score": 1.0, + "content": "Other works based on neural networks include an autoencoder trained to map the", + "type": "text" + }, + { + "bbox": [ + 445, + 84, + 450, + 92 + ], + "score": 0.31, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "th row of the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 295, + 106 + ], + "score": 1.0, + "content": "Laplacian to the ith components of the first", + "type": "text" + }, + { + "bbox": [ + 295, + 94, + 307, + 104 + ], + "score": 0.75, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "eigenvectors, to avoid the spectral decomposi-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "tion (Tian et al., 2014). Yi et al. (2017) use a soft orthogonality constraint to learn spectral em-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "beddings as a volumetric reparametrization of a precomputed Laplacian eigenbase. Shaham et al.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "(2018); Kampffmeyer et al. (2019) propose differentiable loss functions to partition generic data and", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "process out-of-sample data at inference time. Nazi et al. (2019) generate balanced node partitions", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "with a GNN, but adopt an optimization that does not encourage cluster assignments to be orthogonal.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 82, + 505, + 162 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 174, + 254, + 185 + ], + "lines": [ + { + "bbox": [ + 105, + 173, + 255, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 255, + 187 + ], + "score": 1.0, + "content": "2.2 GRAPH NEURAL NETWORKS", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 505, + 294 + ], + "lines": [ + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 506, + 208 + ], + "score": 1.0, + "content": "Many approaches have been proposed to process graphs with neural networks, including recurrent", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 219 + ], + "score": 1.0, + "content": "architectures (Scarselli et al., 2009; Li et al., 2016) or convolutional operations inspired by filters", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 217, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 106, + 217, + 505, + 230 + ], + "score": 1.0, + "content": "used in graph signal processing (Defferrard et al., 2016; Bianchi et al., 2019). Since our focus is", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 228, + 506, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 506, + 241 + ], + "score": 1.0, + "content": "on graph pooling, we base our GNN implementation on a simple MP operation, which combines", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "the features of each node with its 1st-order neighbors. To account for the initial node features, it is", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 250, + 505, + 263 + ], + "score": 1.0, + "content": "possible to introduce self-loops by adding a (scaled) identity matrix to the diagonal of A (Kipf &", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 507, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 507, + 275 + ], + "score": 1.0, + "content": "Welling, 2017). Since our pooling will modify the structure of the adjacency matrix, we prefer a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 506, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 506, + 285 + ], + "score": 1.0, + "content": "MP implementation that leaves the original A unaltered and accounts for the initial node features", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 283, + 229, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 229, + 296 + ], + "score": 1.0, + "content": "by means of skip connections.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 195, + 507, + 296 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 300, + 504, + 324 + ], + "lines": [ + { + "bbox": [ + 104, + 298, + 504, + 316 + ], + "spans": [ + { + "bbox": [ + 104, + 298, + 122, + 316 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 299, + 236, + 312 + ], + "score": 0.92, + "content": "\\tilde { \\mathbf { A } } = \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\mathbf { A } \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\in \\mathbb { R } ^ { N \\times N }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 298, + 455, + 316 + ], + "score": 1.0, + "content": "be the symmetrically normalized adjacency matrix and", + "type": "text" + }, + { + "bbox": [ + 455, + 300, + 504, + 312 + ], + "score": 0.91, + "content": "\\mathbf { X } \\in \\mathbb { R } ^ { N \\times F }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 312, + 388, + 325 + ], + "spans": [ + { + "bbox": [ + 106, + 312, + 388, + 325 + ], + "score": 1.0, + "content": "the matrix containing the node features. The output of the MP layer is", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 298, + 504, + 325 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 330, + 426, + 346 + ], + "lines": [ + { + "bbox": [ + 183, + 330, + 426, + 346 + ], + "spans": [ + { + "bbox": [ + 183, + 330, + 426, + 346 + ], + "score": 0.9, + "content": "\\mathbf { X } ^ { ( t + 1 ) } = M P ( \\mathbf { X } ^ { ( t ) } , \\tilde { \\mathbf { A } } ) = \\operatorname { R e L U } ( \\tilde { \\mathbf { A } } \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { m } + \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { s } ) ,", + "type": "interline_equation", + "image_path": "34c777764a42e47800439e8acb679c0fc313d00d93b57909084dfaca360baee8.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 183, + 330, + 426, + 346 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 504, + 375 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 133, + 364 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 351, + 222, + 364 + ], + "score": 0.93, + "content": "\\boldsymbol { \\Theta } _ { M P } = \\{ \\mathbf { W } _ { m } , \\mathbf { W } _ { s } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 350, + 506, + 364 + ], + "score": 1.0, + "content": "are the trainable weights relative to the mixing and skip component of", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 362, + 198, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 198, + 376 + ], + "score": 1.0, + "content": "the layer, respectively.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 350, + 506, + 376 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 391, + 228, + 404 + ], + "lines": [ + { + "bbox": [ + 104, + 390, + 230, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 390, + 230, + 406 + ], + "score": 1.0, + "content": "3 PROPOSED METHOD", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 416, + 505, + 439 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 507, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 390, + 430 + ], + "score": 1.0, + "content": "The minCUT pooling strategy computes a cluster assignment matrix", + "type": "text" + }, + { + "bbox": [ + 390, + 416, + 442, + 427 + ], + "score": 0.92, + "content": "\\mathbf { S } \\in \\mathbb { R } ^ { N \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 413, + 507, + 430 + ], + "score": 1.0, + "content": "by means of a", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 428, + 423, + 440 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 324, + 440 + ], + "score": 1.0, + "content": "multi-layer perceptron, which maps each node feature", + "type": "text" + }, + { + "bbox": [ + 324, + 430, + 335, + 439 + ], + "score": 0.86, + "content": "\\mathbf { x } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 428, + 423, + 440 + ], + "score": 1.0, + "content": "into the ith row of S:", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 413, + 507, + 440 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 446, + 377, + 459 + ], + "lines": [ + { + "bbox": [ + 232, + 446, + 377, + 459 + ], + "spans": [ + { + "bbox": [ + 232, + 446, + 377, + 459 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathbf { S } = s o f t m a x ( \\mathbf { R e L U } ( \\mathbf { X } \\mathbf { W } _ { 1 } ) \\mathbf { W } _ { 2 } ) , } \\end{array}", + "type": "interline_equation", + "image_path": "eb9b8566cc8d18dce23ac90d2d6d28532fcd23484c72c59fadfa52de34d9d2e9.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 232, + 446, + 377, + 459 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 510 + ], + "lines": [ + { + "bbox": [ + 104, + 463, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 463, + 133, + 479 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 464, + 309, + 478 + ], + "score": 0.83, + "content": "\\boldsymbol { \\Theta } _ { P o o l } = \\{ \\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { F \\times H } , \\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { H \\times K } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 463, + 506, + 479 + ], + "score": 1.0, + "content": "are trainable parameters. The softmax function", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 477, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 168, + 490 + ], + "score": 1.0, + "content": "guarantees that", + "type": "text" + }, + { + "bbox": [ + 169, + 478, + 216, + 489 + ], + "score": 0.91, + "content": "s _ { i , j } \\in [ 0 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 477, + 329, + 490 + ], + "score": 1.0, + "content": "and enforces the constraints", + "type": "text" + }, + { + "bbox": [ + 330, + 477, + 377, + 488 + ], + "score": 0.92, + "content": "\\mathbf { S 1 } _ { K } = \\mathbf { 1 } _ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 477, + 506, + 490 + ], + "score": 1.0, + "content": "inherited from the optimization", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 487, + 505, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 240, + 500 + ], + "score": 1.0, + "content": "problem in (2). The parameters", + "type": "text" + }, + { + "bbox": [ + 240, + 488, + 266, + 499 + ], + "score": 0.92, + "content": "\\Theta _ { M P }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 487, + 285, + 500 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 286, + 488, + 313, + 499 + ], + "score": 0.91, + "content": "\\Theta _ { P o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 487, + 505, + 500 + ], + "score": 1.0, + "content": "are jointly optimized by minimizing the usual", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 498, + 455, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 498, + 305, + 511 + ], + "score": 1.0, + "content": "task-specific loss, as well as an unsupervised loss", + "type": "text" + }, + { + "bbox": [ + 306, + 499, + 318, + 510 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 319, + 498, + 455, + 511 + ], + "score": 1.0, + "content": ", which is composed of two terms", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 463, + 506, + 511 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 516, + 422, + 561 + ], + "lines": [ + { + "bbox": [ + 187, + 516, + 422, + 561 + ], + "spans": [ + { + "bbox": [ + 187, + 516, + 422, + 561 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { u } = \\mathcal { L } _ { c } + \\mathcal { L } _ { o } = \\underbrace { - \\frac { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } ) } } _ { \\mathcal { L } _ { c } } + \\underbrace { \\left\\| \\frac { \\mathbf { S } ^ { T } \\mathbf { S } } { \\| \\mathbf { S } ^ { T } \\mathbf { S } \\| _ { F } } - \\frac { \\mathbf { I } _ { K } } { \\sqrt { K } } \\right\\| _ { F } } _ { \\mathcal { L } _ { o } } ,", + "type": "interline_equation", + "image_path": "093c5d20482e4733720500e265f3b4b47793157e111ce4ec091ed8c0b6c46652.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 187, + 516, + 422, + 531.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 187, + 531.0, + 422, + 546.0 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 187, + 546.0, + 422, + 561.0 + ], + "spans": [], + "index": 32 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 279, + 579 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 280, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 133, + 580 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 568, + 158, + 579 + ], + "score": 0.92, + "content": "\\| \\cdot \\| _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 567, + 280, + 580 + ], + "score": 1.0, + "content": "indicates the Frobenius norm.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33, + "bbox_fs": [ + 106, + 567, + 280, + 580 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 583, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 181, + 597 + ], + "score": 1.0, + "content": "The cut loss term,", + "type": "text" + }, + { + "bbox": [ + 182, + 585, + 194, + 595 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 584, + 505, + 597 + ], + "score": 1.0, + "content": ", evaluates the minCUT given by the cluster assignment S, and is bounded by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 164, + 606 + ], + "score": 0.91, + "content": "- 1 \\leq \\mathcal { L } _ { c } \\leq 0", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 595, + 218, + 609 + ], + "score": 1.0, + "content": ". Minimizing", + "type": "text" + }, + { + "bbox": [ + 218, + 596, + 231, + 606 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "encourages strongly connected nodes to be clustered together, since", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 178, + 619 + ], + "score": 1.0, + "content": "the inner product", + "type": "text" + }, + { + "bbox": [ + 178, + 606, + 207, + 618 + ], + "score": 0.92, + "content": "\\left. \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\right.", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 606, + 272, + 619 + ], + "score": 1.0, + "content": "increases when", + "type": "text" + }, + { + "bbox": [ + 272, + 607, + 288, + 618 + ], + "score": 0.9, + "content": "\\tilde { a } _ { i , j }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 606, + 324, + 619 + ], + "score": 1.0, + "content": "is large.", + "type": "text" + }, + { + "bbox": [ + 324, + 606, + 336, + 617 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "has a single maximum, reached when the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 101, + 616, + 508, + 635 + ], + "spans": [ + { + "bbox": [ + 101, + 616, + 151, + 635 + ], + "score": 1.0, + "content": "numerator", + "type": "text" + }, + { + "bbox": [ + 151, + 618, + 304, + 633 + ], + "score": 0.92, + "content": "\\begin{array} { r } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { S } _ { k } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } _ { k } = 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 616, + 508, + 635 + ], + "score": 1.0, + "content": ". This occurs if, for each pair of connected nodes", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 631, + 505, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 127, + 644 + ], + "score": 1.0, + "content": "(i.e.,", + "type": "text" + }, + { + "bbox": [ + 127, + 631, + 163, + 643 + ], + "score": 0.9, + "content": "\\tilde { a } _ { i , j } > 0 \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 631, + 346, + 644 + ], + "score": 1.0, + "content": "), the cluster assignments are orthogonal (i.e.,", + "type": "text" + }, + { + "bbox": [ + 347, + 631, + 396, + 643 + ], + "score": 0.91, + "content": "\\langle \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\rangle = 0 ,", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 631, + 402, + 644 + ], + "score": 1.0, + "content": ").", + "type": "text" + }, + { + "bbox": [ + 403, + 631, + 415, + 642 + ], + "score": 0.86, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 631, + 505, + 644 + ], + "score": 1.0, + "content": "reaches its minimum,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 107, + 642, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 107, + 644, + 120, + 655 + ], + "score": 0.79, + "content": "- 1", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 642, + 150, + 657 + ], + "score": 1.0, + "content": ", when", + "type": "text" + }, + { + "bbox": [ + 151, + 642, + 264, + 656 + ], + "score": 0.92, + "content": "T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } )", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 642, + 414, + 657 + ], + "score": 1.0, + "content": ". This occurs when in a graph with", + "type": "text" + }, + { + "bbox": [ + 414, + 644, + 425, + 654 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 642, + 505, + 657 + ], + "score": 1.0, + "content": "disconnected com-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 104, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "ponents the cluster assignments are equal for all the nodes in the same component and orthogonal", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 399, + 678 + ], + "score": 1.0, + "content": "to the cluster assignments of nodes in different components. However,", + "type": "text" + }, + { + "bbox": [ + 400, + 666, + 412, + 677 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "is a non-convex func-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 506, + 689 + ], + "score": 1.0, + "content": "tion and its minimization can lead to local minima or degenerate solutions. For example, given a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 505, + 699 + ], + "score": 1.0, + "content": "connected graph, a trivial optimal solution is the one that assigns all nodes to the same cluster. As", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 104, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "a consequence of the continuous relaxation, another degenerate minimum occurs when the cluster", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "assignments are all uniform, that is, all nodes are equally assigned to all clusters. This problem is", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 720, + 488, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 488, + 733 + ], + "score": 1.0, + "content": "exacerbated by prior message-passing operations, which make the node features more uniform.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 40, + "bbox_fs": [ + 101, + 584, + 508, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 222, + 96 + ], + "score": 1.0, + "content": "The orthogonality loss term,", + "type": "text" + }, + { + "bbox": [ + 222, + 83, + 235, + 93 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 81, + 383, + 96 + ], + "score": 1.0, + "content": ", penalizes the degenerate minima of", + "type": "text" + }, + { + "bbox": [ + 383, + 83, + 396, + 93 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "by encouraging the cluster", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 504, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 491, + 107 + ], + "score": 1.0, + "content": "assignments to be orthogonal and the clusters to be of similar size. Since the two matrices in", + "type": "text" + }, + { + "bbox": [ + 491, + 94, + 504, + 105 + ], + "score": 0.87, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 264, + 118 + ], + "score": 1.0, + "content": "have unitary norm it is easy to see that", + "type": "text" + }, + { + "bbox": [ + 264, + 105, + 315, + 115 + ], + "score": 0.92, + "content": "0 \\leq \\mathcal { L } _ { o } \\leq 2", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 104, + 364, + 118 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 364, + 105, + 377, + 115 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 104, + 474, + 118 + ], + "score": 1.0, + "content": "does not dominate over", + "type": "text" + }, + { + "bbox": [ + 475, + 105, + 487, + 115 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 398, + 128 + ], + "score": 1.0, + "content": "the two terms can be safely summed directly (see Fig. 4 for an example).", + "type": "text" + }, + { + "bbox": [ + 399, + 115, + 412, + 127 + ], + "score": 0.87, + "content": "{ \\mathbf { I } } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 116, + 506, + 128 + ], + "score": 1.0, + "content": "can be interpreted as a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 219, + 141 + ], + "score": 1.0, + "content": "(rescaled) clustering matrix", + "type": "text" + }, + { + "bbox": [ + 219, + 126, + 266, + 139 + ], + "score": 0.93, + "content": "\\mathbf { I } _ { K } = \\hat { \\mathbf { S } } ^ { T } \\hat { \\mathbf { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 125, + 297, + 141 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 298, + 126, + 306, + 138 + ], + "score": 0.78, + "content": "\\hat { \\bf S }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 125, + 370, + 141 + ], + "score": 1.0, + "content": "assigns exactly", + "type": "text" + }, + { + "bbox": [ + 370, + 128, + 394, + 140 + ], + "score": 0.91, + "content": "N / K", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 125, + 506, + 141 + ], + "score": 1.0, + "content": "points to each cluster. The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 139, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 506, + 152 + ], + "score": 1.0, + "content": "value of the Frobenius norm between clustering matrices is not dominated by the performance on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 151, + 483, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 483, + 162 + ], + "score": 1.0, + "content": "the largest clusters (Law et al., 2017) and, thus, can be used to optimize intra-cluster variance.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 108, + 166, + 505, + 211 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 178 + ], + "score": 1.0, + "content": "Contrarily to SC methods that search for feasible solutions only within the space of orthogonal", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 145, + 190 + ], + "score": 1.0, + "content": "matrices,", + "type": "text" + }, + { + "bbox": [ + 146, + 178, + 158, + 189 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "only introduces a soft constraint that could be violated during the learning procedure.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 188, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 131, + 202 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 189, + 144, + 200 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 188, + 505, + 202 + ], + "score": 1.0, + "content": "is non-convex, the violation compromises the theoretical guarantee of convergence to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 200, + 267, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 267, + 212 + ], + "score": 1.0, + "content": "optimum of (3). However, we note that:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 129, + 219, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 505, + 231 + ], + "score": 1.0, + "content": "1. the cluster assignments S are well initialized: after the MP operation, the features of the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "connected vertices become similar and, since the MLP is a smooth function (Nelles, 2013),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 241, + 361, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 241, + 361, + 255 + ], + "score": 1.0, + "content": "it yields similar cluster assignments for those vertices;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 128, + 256, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 128, + 256, + 506, + 269 + ], + "score": 1.0, + "content": "2. in the GNN architecture, the minCUT objective is a regularization term and, therefore, a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "solution which is sub-optimal for (3) could instead be adequate for the specific objective of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 279, + 229, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 279, + 229, + 291 + ], + "score": 1.0, + "content": "the downstream task;", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 129, + 293, + 489, + 307 + ], + "spans": [ + { + "bbox": [ + 129, + 293, + 474, + 307 + ], + "score": 1.0, + "content": "3. optimizing the task-specific loss helps the GNN to avoid the degenerate minima of", + "type": "text" + }, + { + "bbox": [ + 474, + 294, + 486, + 304 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 293, + 489, + 307 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 107, + 318, + 190, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 190, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 190, + 330 + ], + "score": 1.0, + "content": "3.1 COARSENING", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 444, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 446, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 446, + 353 + ], + "score": 1.0, + "content": "The coarsened version of the adjacency matrix and the graph signal are computed as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 354, + 375, + 369 + ], + "lines": [ + { + "bbox": [ + 234, + 354, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 234, + 354, + 375, + 369 + ], + "score": 0.91, + "content": "{ \\bf A } ^ { p o o l } = { \\bf S } ^ { T } { \\tilde { \\bf A } } { \\bf S } ; ~ { \\bf X } ^ { p o o l } = { \\bf S } ^ { T } { \\bf X } ,", + "type": "interline_equation", + "image_path": "c14e7c61c40a2f766857a297640952b7ad643866d8ac0d1f6858958d0c1eb67a.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 234, + 354, + 375, + 369 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 194, + 390 + ], + "score": 1.0, + "content": "where the entry xpooi,j", + "type": "text" + }, + { + "bbox": [ + 197, + 371, + 210, + 390 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 210, + 374, + 279, + 386 + ], + "score": 0.92, + "content": "\\mathbf { X } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times F }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 371, + 451, + 390 + ], + "score": 1.0, + "content": "is the weighted average value of feature", + "type": "text" + }, + { + "bbox": [ + 451, + 376, + 457, + 387 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 371, + 506, + 390 + ], + "score": 1.0, + "content": "among the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 386, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 386, + 187, + 406 + ], + "score": 1.0, + "content": "elements in cluster", + "type": "text" + }, + { + "bbox": [ + 187, + 391, + 192, + 400 + ], + "score": 0.59, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 386, + 199, + 406 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 200, + 388, + 270, + 400 + ], + "score": 0.9, + "content": "\\mathbf { A } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 386, + 429, + 406 + ], + "score": 1.0, + "content": "is a symmetric matrix, whose entries", + "type": "text" + }, + { + "bbox": [ + 429, + 388, + 450, + 403 + ], + "score": 0.93, + "content": "a _ { i , i } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 390, + 505, + 401 + ], + "score": 1.0, + "content": "are the total", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 103, + 401, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 103, + 401, + 368, + 420 + ], + "score": 1.0, + "content": "number of edges between the nodes in the cluster i, while apooi,j", + "type": "text" + }, + { + "bbox": [ + 370, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "is the number of edges between", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 416, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 135, + 432 + ], + "score": 1.0, + "content": "cluster", + "type": "text" + }, + { + "bbox": [ + 135, + 419, + 140, + 428 + ], + "score": 0.67, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 416, + 157, + 432 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 158, + 418, + 164, + 430 + ], + "score": 0.79, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 416, + 192, + 432 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 192, + 417, + 217, + 428 + ], + "score": 0.88, + "content": "\\mathbf { A } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 416, + 346, + 432 + ], + "score": 1.0, + "content": "corresponds to the numerator of", + "type": "text" + }, + { + "bbox": [ + 347, + 419, + 358, + 429 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 416, + 506, + 432 + ], + "score": 1.0, + "content": "in (7), the trace maximization yields", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 448, + 441 + ], + "score": 1.0, + "content": "clusters with many internal connections and weakly connected to each other. Hence,", + "type": "text" + }, + { + "bbox": [ + 449, + 429, + 473, + 439 + ], + "score": 0.89, + "content": "\\mathbf { A } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "will be", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "a diagonal-dominant matrix, which describes a graph with self-loops much stronger than any other", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 464 + ], + "score": 1.0, + "content": "connection. Because self-loops hamper the propagation across adjacent nodes in the MP operations", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 380, + 477 + ], + "score": 1.0, + "content": "following the pooling layer, we compute the new adjacency matrix", + "type": "text" + }, + { + "bbox": [ + 380, + 462, + 405, + 474 + ], + "score": 0.9, + "content": "\\tilde { \\mathbf { A } } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "by zeroing the diagonal", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 474, + 275, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 275, + 487 + ], + "score": 1.0, + "content": "and by applying the degree normalization", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 489, + 419, + 506 + ], + "lines": [ + { + "bbox": [ + 191, + 489, + 419, + 506 + ], + "spans": [ + { + "bbox": [ + 191, + 489, + 419, + 506 + ], + "score": 0.9, + "content": "\\hat { \\bf A } = { \\bf A } ^ { p o o l } - { \\bf I } _ { K } d i a g ( { \\bf A } ^ { p o o l } ) ; \\quad \\tilde { \\bf A } ^ { p o o l } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\hat { \\bf A } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "cdc85379486e51b65f69e07f035aec1a89b20127fdc77e297c858fa7fe2513dd.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 191, + 489, + 419, + 506 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 510, + 277, + 522 + ], + "lines": [ + { + "bbox": [ + 107, + 509, + 277, + 522 + ], + "spans": [ + { + "bbox": [ + 107, + 509, + 133, + 522 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 510, + 164, + 522 + ], + "score": 0.86, + "content": "d i a g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 509, + 277, + 522 + ], + "score": 1.0, + "content": "returns the matrix diagonal.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "title", + "bbox": [ + 108, + 535, + 400, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 401, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 401, + 547 + ], + "score": 1.0, + "content": "3.2 DISCUSSION AND RELATIONSHIP WITH SPECTRAL CLUSTERING", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "The proposed method is straightforward to implement: the cluster assignments, the loss, graph", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 463, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 463, + 579 + ], + "score": 1.0, + "content": "coarsening, and feature pooling are all computed with standard linear algebra operations.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "score": 1.0, + "content": "There are several differences between minCUTpool and classic SC methods. SC partitions the graph", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "based on the Laplacian, but does not account for the node features. Instead, the cluster assignments", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 115, + 616 + ], + "score": 0.84, + "content": "\\mathbf { s } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 604, + 253, + 618 + ], + "score": 1.0, + "content": "found by minCUTpool depend on", + "type": "text" + }, + { + "bbox": [ + 253, + 606, + 264, + 616 + ], + "score": 0.85, + "content": "\\mathbf { x } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 604, + 506, + 618 + ], + "score": 1.0, + "content": ", which works well if connected nodes have similar features.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "This is a reasonable assumption in GNNs since, even in disassortative graphs (i.e., networks where", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "dissimilar nodes are likely to be connected (Newman, 2003)), the features tend to become similar", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 637, + 211, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 211, + 651 + ], + "score": 1.0, + "content": "due to the MP operations.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "Another difference is that SC handles a single graph and is not conceived for tasks with multiple", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "graphs to be partitioned independently. Instead, thanks to the independence of the model parameters", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 218, + 690 + ], + "score": 1.0, + "content": "from the number of nodes", + "type": "text" + }, + { + "bbox": [ + 218, + 677, + 228, + 687 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "and from the graph spectrum, minCUTpool can generalize to out-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "of-sample data. This feature is fundamental in problems such as graph classification, where each", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "sample is a graph with a different structure, and allows to train the model on small graphs and process", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "larger ones at inference time. Finally, minCUTpool directly uses the soft cluster assignments rather", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 256, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 173, + 732 + ], + "score": 1.0, + "content": "than performing", + "type": "text" + }, + { + "bbox": [ + 173, + 721, + 180, + 730 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 721, + 256, + 732 + ], + "score": 1.0, + "content": "-means afterwards.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 222, + 96 + ], + "score": 1.0, + "content": "The orthogonality loss term,", + "type": "text" + }, + { + "bbox": [ + 222, + 83, + 235, + 93 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 81, + 383, + 96 + ], + "score": 1.0, + "content": ", penalizes the degenerate minima of", + "type": "text" + }, + { + "bbox": [ + 383, + 83, + 396, + 93 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 81, + 506, + 96 + ], + "score": 1.0, + "content": "by encouraging the cluster", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 504, + 107 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 491, + 107 + ], + "score": 1.0, + "content": "assignments to be orthogonal and the clusters to be of similar size. Since the two matrices in", + "type": "text" + }, + { + "bbox": [ + 491, + 94, + 504, + 105 + ], + "score": 0.87, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 264, + 118 + ], + "score": 1.0, + "content": "have unitary norm it is easy to see that", + "type": "text" + }, + { + "bbox": [ + 264, + 105, + 315, + 115 + ], + "score": 0.92, + "content": "0 \\leq \\mathcal { L } _ { o } \\leq 2", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 104, + 364, + 118 + ], + "score": 1.0, + "content": ". Therefore,", + "type": "text" + }, + { + "bbox": [ + 364, + 105, + 377, + 115 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 104, + 474, + 118 + ], + "score": 1.0, + "content": "does not dominate over", + "type": "text" + }, + { + "bbox": [ + 475, + 105, + 487, + 115 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 398, + 128 + ], + "score": 1.0, + "content": "the two terms can be safely summed directly (see Fig. 4 for an example).", + "type": "text" + }, + { + "bbox": [ + 399, + 115, + 412, + 127 + ], + "score": 0.87, + "content": "{ \\mathbf { I } } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 116, + 506, + 128 + ], + "score": 1.0, + "content": "can be interpreted as a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 125, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 125, + 219, + 141 + ], + "score": 1.0, + "content": "(rescaled) clustering matrix", + "type": "text" + }, + { + "bbox": [ + 219, + 126, + 266, + 139 + ], + "score": 0.93, + "content": "\\mathbf { I } _ { K } = \\hat { \\mathbf { S } } ^ { T } \\hat { \\mathbf { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 125, + 297, + 141 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 298, + 126, + 306, + 138 + ], + "score": 0.78, + "content": "\\hat { \\bf S }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 125, + 370, + 141 + ], + "score": 1.0, + "content": "assigns exactly", + "type": "text" + }, + { + "bbox": [ + 370, + 128, + 394, + 140 + ], + "score": 0.91, + "content": "N / K", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 125, + 506, + 141 + ], + "score": 1.0, + "content": "points to each cluster. The", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 139, + 506, + 152 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 506, + 152 + ], + "score": 1.0, + "content": "value of the Frobenius norm between clustering matrices is not dominated by the performance on", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 151, + 483, + 162 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 483, + 162 + ], + "score": 1.0, + "content": "the largest clusters (Law et al., 2017) and, thus, can be used to optimize intra-cluster variance.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 81, + 506, + 162 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 166, + 505, + 211 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 505, + 178 + ], + "score": 1.0, + "content": "Contrarily to SC methods that search for feasible solutions only within the space of orthogonal", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 178, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 145, + 190 + ], + "score": 1.0, + "content": "matrices,", + "type": "text" + }, + { + "bbox": [ + 146, + 178, + 158, + 189 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 178, + 505, + 190 + ], + "score": 1.0, + "content": "only introduces a soft constraint that could be violated during the learning procedure.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 188, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 131, + 202 + ], + "score": 1.0, + "content": "Since", + "type": "text" + }, + { + "bbox": [ + 131, + 189, + 144, + 200 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 188, + 505, + 202 + ], + "score": 1.0, + "content": "is non-convex, the violation compromises the theoretical guarantee of convergence to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 200, + 267, + 212 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 267, + 212 + ], + "score": 1.0, + "content": "optimum of (3). However, we note that:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 166, + 505, + 212 + ] + }, + { + "type": "list", + "bbox": [ + 129, + 219, + 505, + 306 + ], + "lines": [ + { + "bbox": [ + 130, + 221, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 130, + 221, + 505, + 231 + ], + "score": 1.0, + "content": "1. the cluster assignments S are well initialized: after the MP operation, the features of the", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 230, + 506, + 244 + ], + "spans": [ + { + "bbox": [ + 141, + 230, + 506, + 244 + ], + "score": 1.0, + "content": "connected vertices become similar and, since the MLP is a smooth function (Nelles, 2013),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 241, + 361, + 255 + ], + "spans": [ + { + "bbox": [ + 141, + 241, + 361, + 255 + ], + "score": 1.0, + "content": "it yields similar cluster assignments for those vertices;", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 128, + 256, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 128, + 256, + 506, + 269 + ], + "score": 1.0, + "content": "2. in the GNN architecture, the minCUT objective is a regularization term and, therefore, a", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "solution which is sub-optimal for (3) could instead be adequate for the specific objective of", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 142, + 279, + 229, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 279, + 229, + 291 + ], + "score": 1.0, + "content": "the downstream task;", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 293, + 489, + 307 + ], + "spans": [ + { + "bbox": [ + 129, + 293, + 474, + 307 + ], + "score": 1.0, + "content": "3. optimizing the task-specific loss helps the GNN to avoid the degenerate minima of", + "type": "text" + }, + { + "bbox": [ + 474, + 294, + 486, + 304 + ], + "score": 0.88, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 293, + 489, + 307 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 14, + "bbox_fs": [ + 128, + 221, + 506, + 307 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 318, + 190, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 190, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 190, + 330 + ], + "score": 1.0, + "content": "3.1 COARSENING", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 444, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 446, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 446, + 353 + ], + "score": 1.0, + "content": "The coarsened version of the adjacency matrix and the graph signal are computed as", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 337, + 446, + 353 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 354, + 375, + 369 + ], + "lines": [ + { + "bbox": [ + 234, + 354, + 375, + 369 + ], + "spans": [ + { + "bbox": [ + 234, + 354, + 375, + 369 + ], + "score": 0.91, + "content": "{ \\bf A } ^ { p o o l } = { \\bf S } ^ { T } { \\tilde { \\bf A } } { \\bf S } ; ~ { \\bf X } ^ { p o o l } = { \\bf S } ^ { T } { \\bf X } ,", + "type": "interline_equation", + "image_path": "c14e7c61c40a2f766857a297640952b7ad643866d8ac0d1f6858958d0c1eb67a.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 234, + 354, + 375, + 369 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 373, + 505, + 486 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 373, + 194, + 390 + ], + "score": 1.0, + "content": "where the entry xpooi,j", + "type": "text" + }, + { + "bbox": [ + 197, + 371, + 210, + 390 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 210, + 374, + 279, + 386 + ], + "score": 0.92, + "content": "\\mathbf { X } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times F }", + "type": "inline_equation" + }, + { + "bbox": [ + 280, + 371, + 451, + 390 + ], + "score": 1.0, + "content": "is the weighted average value of feature", + "type": "text" + }, + { + "bbox": [ + 451, + 376, + 457, + 387 + ], + "score": 0.83, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 371, + 506, + 390 + ], + "score": 1.0, + "content": "among the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 386, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 104, + 386, + 187, + 406 + ], + "score": 1.0, + "content": "elements in cluster", + "type": "text" + }, + { + "bbox": [ + 187, + 391, + 192, + 400 + ], + "score": 0.59, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 386, + 199, + 406 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 200, + 388, + 270, + 400 + ], + "score": 0.9, + "content": "\\mathbf { A } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times K }", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 386, + 429, + 406 + ], + "score": 1.0, + "content": "is a symmetric matrix, whose entries", + "type": "text" + }, + { + "bbox": [ + 429, + 388, + 450, + 403 + ], + "score": 0.93, + "content": "a _ { i , i } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 390, + 505, + 401 + ], + "score": 1.0, + "content": "are the total", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 103, + 401, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 103, + 401, + 368, + 420 + ], + "score": 1.0, + "content": "number of edges between the nodes in the cluster i, while apooi,j", + "type": "text" + }, + { + "bbox": [ + 370, + 403, + 506, + 417 + ], + "score": 1.0, + "content": "is the number of edges between", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 416, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 135, + 432 + ], + "score": 1.0, + "content": "cluster", + "type": "text" + }, + { + "bbox": [ + 135, + 419, + 140, + 428 + ], + "score": 0.67, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 416, + 157, + 432 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 158, + 418, + 164, + 430 + ], + "score": 0.79, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 416, + 192, + 432 + ], + "score": 1.0, + "content": ". Since", + "type": "text" + }, + { + "bbox": [ + 192, + 417, + 217, + 428 + ], + "score": 0.88, + "content": "\\mathbf { A } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 416, + 346, + 432 + ], + "score": 1.0, + "content": "corresponds to the numerator of", + "type": "text" + }, + { + "bbox": [ + 347, + 419, + 358, + 429 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { c }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 416, + 506, + 432 + ], + "score": 1.0, + "content": "in (7), the trace maximization yields", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 448, + 441 + ], + "score": 1.0, + "content": "clusters with many internal connections and weakly connected to each other. Hence,", + "type": "text" + }, + { + "bbox": [ + 449, + 429, + 473, + 439 + ], + "score": 0.89, + "content": "\\mathbf { A } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 473, + 428, + 506, + 441 + ], + "score": 1.0, + "content": "will be", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 452 + ], + "score": 1.0, + "content": "a diagonal-dominant matrix, which describes a graph with self-loops much stronger than any other", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 451, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 464 + ], + "score": 1.0, + "content": "connection. Because self-loops hamper the propagation across adjacent nodes in the MP operations", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 380, + 477 + ], + "score": 1.0, + "content": "following the pooling layer, we compute the new adjacency matrix", + "type": "text" + }, + { + "bbox": [ + 380, + 462, + 405, + 474 + ], + "score": 0.9, + "content": "\\tilde { \\mathbf { A } } ^ { p o o l }", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "by zeroing the diagonal", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 474, + 275, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 275, + 487 + ], + "score": 1.0, + "content": "and by applying the degree normalization", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25, + "bbox_fs": [ + 103, + 371, + 506, + 487 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 489, + 419, + 506 + ], + "lines": [ + { + "bbox": [ + 191, + 489, + 419, + 506 + ], + "spans": [ + { + "bbox": [ + 191, + 489, + 419, + 506 + ], + "score": 0.9, + "content": "\\hat { \\bf A } = { \\bf A } ^ { p o o l } - { \\bf I } _ { K } d i a g ( { \\bf A } ^ { p o o l } ) ; \\quad \\tilde { \\bf A } ^ { p o o l } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\hat { \\bf A } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } .", + "type": "interline_equation", + "image_path": "cdc85379486e51b65f69e07f035aec1a89b20127fdc77e297c858fa7fe2513dd.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 191, + 489, + 419, + 506 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 510, + 277, + 522 + ], + "lines": [ + { + "bbox": [ + 107, + 509, + 277, + 522 + ], + "spans": [ + { + "bbox": [ + 107, + 509, + 133, + 522 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 510, + 164, + 522 + ], + "score": 0.86, + "content": "d i a g ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 509, + 277, + 522 + ], + "score": 1.0, + "content": "returns the matrix diagonal.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31, + "bbox_fs": [ + 107, + 509, + 277, + 522 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 535, + 400, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 534, + 401, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 401, + 547 + ], + "score": 1.0, + "content": "3.2 DISCUSSION AND RELATIONSHIP WITH SPECTRAL CLUSTERING", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 555, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 568 + ], + "score": 1.0, + "content": "The proposed method is straightforward to implement: the cluster assignments, the loss, graph", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 566, + 463, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 463, + 579 + ], + "score": 1.0, + "content": "coarsening, and feature pooling are all computed with standard linear algebra operations.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 554, + 505, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 582, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 596 + ], + "score": 1.0, + "content": "There are several differences between minCUTpool and classic SC methods. SC partitions the graph", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "based on the Laplacian, but does not account for the node features. Instead, the cluster assignments", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 115, + 616 + ], + "score": 0.84, + "content": "\\mathbf { s } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 604, + 253, + 618 + ], + "score": 1.0, + "content": "found by minCUTpool depend on", + "type": "text" + }, + { + "bbox": [ + 253, + 606, + 264, + 616 + ], + "score": 0.85, + "content": "\\mathbf { x } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 604, + 506, + 618 + ], + "score": 1.0, + "content": ", which works well if connected nodes have similar features.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 628 + ], + "score": 1.0, + "content": "This is a reasonable assumption in GNNs since, even in disassortative graphs (i.e., networks where", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "dissimilar nodes are likely to be connected (Newman, 2003)), the features tend to become similar", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 637, + 211, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 211, + 651 + ], + "score": 1.0, + "content": "due to the MP operations.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 581, + 506, + 651 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "Another difference is that SC handles a single graph and is not conceived for tasks with multiple", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 678 + ], + "score": 1.0, + "content": "graphs to be partitioned independently. Instead, thanks to the independence of the model parameters", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 218, + 690 + ], + "score": 1.0, + "content": "from the number of nodes", + "type": "text" + }, + { + "bbox": [ + 218, + 677, + 228, + 687 + ], + "score": 0.79, + "content": "N", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "and from the graph spectrum, minCUTpool can generalize to out-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "of-sample data. This feature is fundamental in problems such as graph classification, where each", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "sample is a graph with a different structure, and allows to train the model on small graphs and process", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "larger ones at inference time. Finally, minCUTpool directly uses the soft cluster assignments rather", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 721, + 256, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 173, + 732 + ], + "score": 1.0, + "content": "than performing", + "type": "text" + }, + { + "bbox": [ + 173, + 721, + 180, + 730 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 721, + 256, + 732 + ], + "score": 1.0, + "content": "-means afterwards.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 654, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 327, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 328, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 328, + 95 + ], + "score": 1.0, + "content": "4 RELATED WORK ON POOLING IN GNNS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 504, + 139 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "score": 1.0, + "content": "Trainable pooling methods. Similarly to our method, these approaches learn how to generate", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 505, + 129 + ], + "score": 1.0, + "content": "coarsened version of the graph through differentiable functions, which take as input the nodes fea-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 393, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 128, + 140 + ], + "score": 1.0, + "content": "tures", + "type": "text" + }, + { + "bbox": [ + 128, + 128, + 138, + 138 + ], + "score": 0.42, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 128, + 393, + 140 + ], + "score": 1.0, + "content": "and are parametrized by weights optimized on the task at hand.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Diffpool (Ying et al., 2018) is a pooling module that includes two parallel MP layers: one to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 154, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 104, + 154, + 237, + 171 + ], + "score": 1.0, + "content": "compute the new node features", + "type": "text" + }, + { + "bbox": [ + 238, + 156, + 267, + 168 + ], + "score": 0.91, + "content": "\\mathbf { X } ^ { ( t + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 154, + 506, + 171 + ], + "score": 1.0, + "content": "and another to generate the cluster assignments S. Diff-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 181 + ], + "score": 1.0, + "content": "pool implements an unsupervised loss that consists of two terms. First, the link prediction term", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 178, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 107, + 178, + 163, + 191 + ], + "score": 0.92, + "content": "\\| \\mathbf { A } - \\mathbf { S } \\mathbf { \\dot { S } } ^ { T } \\| _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 178, + 506, + 192 + ], + "score": 1.0, + "content": "minimizes the Frobenius norm of the difference between the adjacency and the Gram", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "matrix of the cluster assignments, encouraging nearby nodes to be clustered together. The sec-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 146, + 219 + ], + "score": 1.0, + "content": "ond term", + "type": "text" + }, + { + "bbox": [ + 147, + 201, + 210, + 216 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } H ( \\mathbf { S } _ { i } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 199, + 506, + 219 + ], + "score": 1.0, + "content": "minimizes the entropy of the cluster assignments to make them alike to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 227 + ], + "score": 1.0, + "content": "one-hot vectors. Like minCUTpool, Diffpool clusters the vertices of annotated graphs, but yields", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "completely different partitions, since it computes differently the clustering assignments, the coars-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "ened adjacency matrix and, most importantly, the unsupervised loss. In Diffpool, such a loss shows", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 247, + 370, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 370, + 260 + ], + "score": 1.0, + "content": "pathological behaviors that are discussed later in the experiments.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 215, + 276 + ], + "score": 1.0, + "content": "The approach dubbed Top-", + "type": "text" + }, + { + "bbox": [ + 215, + 264, + 225, + 274 + ], + "score": 0.69, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "pooling (Hongyang Gao, 2019; Lee et al., 2019), learns a projection", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 433, + 287 + ], + "score": 1.0, + "content": "vector that is applied to each node feature to obtain a score. The nodes with the", + "type": "text" + }, + { + "bbox": [ + 434, + 275, + 444, + 285 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "highest scores", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 322, + 298 + ], + "score": 1.0, + "content": "are retained, the others are dropped. Since the top-", + "type": "text" + }, + { + "bbox": [ + 322, + 286, + 333, + 296 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "selection is not differentiable, the scores", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "are also used as a gate/attention for the node features, letting the projection vector to be trained", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 219, + 320 + ], + "score": 1.0, + "content": "with backpropagation. Top-", + "type": "text" + }, + { + "bbox": [ + 220, + 308, + 230, + 317 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "is memory efficient as it avoids generating cluster assignments. To", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 367, + 331 + ], + "score": 1.0, + "content": "prevent A from becoming disconnected after nodes removal, Top-", + "type": "text" + }, + { + "bbox": [ + 367, + 319, + 378, + 329 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "drops the rows and the columns", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 329, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 128, + 343 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 329, + 143, + 340 + ], + "score": 0.81, + "content": "{ \\bf A } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 329, + 398, + 343 + ], + "score": 1.0, + "content": "and uses it as the new adjacency matrix. However, computing", + "type": "text" + }, + { + "bbox": [ + 399, + 329, + 413, + 340 + ], + "score": 0.84, + "content": "{ \\bf A } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 329, + 438, + 343 + ], + "score": 1.0, + "content": "costs", + "type": "text" + }, + { + "bbox": [ + 438, + 329, + 468, + 342 + ], + "score": 0.93, + "content": "\\mathcal { O } ( N ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 329, + 506, + 343 + ], + "score": 1.0, + "content": "and it is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 341, + 298, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 298, + 353 + ], + "score": 1.0, + "content": "inefficient to implement with sparse operations.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 504, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "score": 1.0, + "content": "Topological pooling methods. These methods pre-compute a pyramid of coarsened graphs, only", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 373, + 388 + ], + "score": 1.0, + "content": "taking into account the topology (A), but not the node features", + "type": "text" + }, + { + "bbox": [ + 373, + 376, + 389, + 387 + ], + "score": 0.53, + "content": "\\mathbf { \\Pi } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 375, + 506, + 388 + ], + "score": 1.0, + "content": ". During training, the node", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "score": 1.0, + "content": "features are pooled with standard procedures and are fit into these deterministic graph structures.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "These methods are less flexible, but provide a stronger bias that can prevent degenerate solutions", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 315, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 315, + 421 + ], + "score": 1.0, + "content": "(e.g., coarsened graphs collapsing in a single node).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "The approach proposed by Bruna et al. (2013), which has been adopted also in other GNN ar-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "chitectures (Defferrard et al., 2016; Fey et al., 2018), exploits GRACLUS (Dhillon et al., 2004), a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 339, + 460 + ], + "score": 1.0, + "content": "hierarchical algorithm based on SC. At each pooling level", + "type": "text" + }, + { + "bbox": [ + 339, + 448, + 344, + 457 + ], + "score": 0.39, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 447, + 505, + 460 + ], + "score": 1.0, + "content": ", GRACLUS indetifies the pairs of max-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 190, + 472 + ], + "score": 1.0, + "content": "imally similar nodes", + "type": "text" + }, + { + "bbox": [ + 191, + 459, + 199, + 469 + ], + "score": 0.85, + "content": "i _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 458, + 217, + 472 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 459, + 225, + 470 + ], + "score": 0.85, + "content": "j _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 458, + 394, + 472 + ], + "score": 1.0, + "content": "to be clustered together into a new vertex", + "type": "text" + }, + { + "bbox": [ + 394, + 459, + 420, + 471 + ], + "score": 0.91, + "content": "k _ { ( l + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 458, + 506, + 472 + ], + "score": 1.0, + "content": ". At inference phase,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "max-pooling is used to determine which node in the pair is kept. Fake vertices are added so that the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 480, + 462, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 462, + 493 + ], + "score": 1.0, + "content": "number of nodes can be halved each time, but this injects noisy information in the graph.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "score": 1.0, + "content": "Node decimation is a method originally proposed in graph signal processing literature (Shuman", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "et al., 2016), which as been adapted also for GNNs (Simonovsky & Komodakis, 2017). The nodes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "are partitioned in two sets, according to the signs of the Laplacian eigenvector associated to the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "largest eigenvalue. One of the two sets is dropped, reducing the number of nodes each time approx-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "score": 1.0, + "content": "imately by half. Kron reduction is used to compute a pyramid of coarsened Laplacians from the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 177, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 177, + 564 + ], + "score": 1.0, + "content": "remaining nodes.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "A procedure proposed in Gama et al. (2018) diffuses a signal from designated nodes on the graph", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 592 + ], + "score": 1.0, + "content": "and stores the observed sequence of diffused components. The resulting stream of information", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "score": 1.0, + "content": "is interpreted as a time signal, where standard CNN pooling is applied. We also mention a pooling", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "operation for coarsening binary unweighted graphs by aggregating maximal cliques (Luzhnica et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "score": 1.0, + "content": "2019). Nodes assigned to the same clique are summarized by max or average pooling and become a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 624, + 242, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 242, + 636 + ], + "score": 1.0, + "content": "new node in the coarsened graph.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5 + }, + { + "type": "title", + "bbox": [ + 108, + 651, + 200, + 664 + ], + "lines": [ + { + "bbox": [ + 104, + 650, + 202, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 650, + 202, + 666 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "We consider both supervised and unsupervised tasks, and compare minCUTpool with other GNN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "pooling strategies. The Appendix provides further details on the experiments and a schematic de-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "piction of the architectures used in each task. In addition, the Appendix reports two additional", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "experiments: i) graph reconstruction by means of an Auto Encoder with bottleneck, implemented", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 462, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 462, + 734 + ], + "score": 1.0, + "content": "with pooling and un-pooling layers, ii) an architecture with pooling for graph regression.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 327, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 328, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 328, + 95 + ], + "score": 1.0, + "content": "4 RELATED WORK ON POOLING IN GNNS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 106, + 504, + 139 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 506, + 119 + ], + "score": 1.0, + "content": "Trainable pooling methods. Similarly to our method, these approaches learn how to generate", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 118, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 118, + 505, + 129 + ], + "score": 1.0, + "content": "coarsened version of the graph through differentiable functions, which take as input the nodes fea-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 128, + 393, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 128, + 140 + ], + "score": 1.0, + "content": "tures", + "type": "text" + }, + { + "bbox": [ + 128, + 128, + 138, + 138 + ], + "score": 0.42, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 128, + 393, + 140 + ], + "score": 1.0, + "content": "and are parametrized by weights optimized on the task at hand.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 106, + 506, + 140 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 145, + 505, + 258 + ], + "lines": [ + { + "bbox": [ + 106, + 145, + 506, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 145, + 506, + 158 + ], + "score": 1.0, + "content": "Diffpool (Ying et al., 2018) is a pooling module that includes two parallel MP layers: one to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 154, + 506, + 171 + ], + "spans": [ + { + "bbox": [ + 104, + 154, + 237, + 171 + ], + "score": 1.0, + "content": "compute the new node features", + "type": "text" + }, + { + "bbox": [ + 238, + 156, + 267, + 168 + ], + "score": 0.91, + "content": "\\mathbf { X } ^ { ( t + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 154, + 506, + 171 + ], + "score": 1.0, + "content": "and another to generate the cluster assignments S. Diff-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 168, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 105, + 168, + 505, + 181 + ], + "score": 1.0, + "content": "pool implements an unsupervised loss that consists of two terms. First, the link prediction term", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 178, + 506, + 192 + ], + "spans": [ + { + "bbox": [ + 107, + 178, + 163, + 191 + ], + "score": 0.92, + "content": "\\| \\mathbf { A } - \\mathbf { S } \\mathbf { \\dot { S } } ^ { T } \\| _ { F }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 178, + 506, + 192 + ], + "score": 1.0, + "content": "minimizes the Frobenius norm of the difference between the adjacency and the Gram", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "matrix of the cluster assignments, encouraging nearby nodes to be clustered together. The sec-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 506, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 146, + 219 + ], + "score": 1.0, + "content": "ond term", + "type": "text" + }, + { + "bbox": [ + 147, + 201, + 210, + 216 + ], + "score": 0.91, + "content": "\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } H ( \\mathbf { S } _ { i } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 199, + 506, + 219 + ], + "score": 1.0, + "content": "minimizes the entropy of the cluster assignments to make them alike to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 505, + 227 + ], + "score": 1.0, + "content": "one-hot vectors. Like minCUTpool, Diffpool clusters the vertices of annotated graphs, but yields", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "completely different partitions, since it computes differently the clustering assignments, the coars-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 236, + 505, + 248 + ], + "score": 1.0, + "content": "ened adjacency matrix and, most importantly, the unsupervised loss. In Diffpool, such a loss shows", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 247, + 370, + 260 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 370, + 260 + ], + "score": 1.0, + "content": "pathological behaviors that are discussed later in the experiments.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5, + "bbox_fs": [ + 104, + 145, + 506, + 260 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 263, + 505, + 352 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 215, + 276 + ], + "score": 1.0, + "content": "The approach dubbed Top-", + "type": "text" + }, + { + "bbox": [ + 215, + 264, + 225, + 274 + ], + "score": 0.69, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 264, + 505, + 276 + ], + "score": 1.0, + "content": "pooling (Hongyang Gao, 2019; Lee et al., 2019), learns a projection", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 275, + 505, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 433, + 287 + ], + "score": 1.0, + "content": "vector that is applied to each node feature to obtain a score. The nodes with the", + "type": "text" + }, + { + "bbox": [ + 434, + 275, + 444, + 285 + ], + "score": 0.79, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 275, + 505, + 287 + ], + "score": 1.0, + "content": "highest scores", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 286, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 322, + 298 + ], + "score": 1.0, + "content": "are retained, the others are dropped. Since the top-", + "type": "text" + }, + { + "bbox": [ + 322, + 286, + 333, + 296 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 286, + 505, + 298 + ], + "score": 1.0, + "content": "selection is not differentiable, the scores", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "are also used as a gate/attention for the node features, letting the projection vector to be trained", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 219, + 320 + ], + "score": 1.0, + "content": "with backpropagation. Top-", + "type": "text" + }, + { + "bbox": [ + 220, + 308, + 230, + 317 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "is memory efficient as it avoids generating cluster assignments. To", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 367, + 331 + ], + "score": 1.0, + "content": "prevent A from becoming disconnected after nodes removal, Top-", + "type": "text" + }, + { + "bbox": [ + 367, + 319, + 378, + 329 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "drops the rows and the columns", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 329, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 128, + 343 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 329, + 143, + 340 + ], + "score": 0.81, + "content": "{ \\bf A } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 329, + 398, + 343 + ], + "score": 1.0, + "content": "and uses it as the new adjacency matrix. However, computing", + "type": "text" + }, + { + "bbox": [ + 399, + 329, + 413, + 340 + ], + "score": 0.84, + "content": "{ \\bf A } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 414, + 329, + 438, + 343 + ], + "score": 1.0, + "content": "costs", + "type": "text" + }, + { + "bbox": [ + 438, + 329, + 468, + 342 + ], + "score": 0.93, + "content": "\\mathcal { O } ( N ^ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 329, + 506, + 343 + ], + "score": 1.0, + "content": "and it is", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 341, + 298, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 298, + 353 + ], + "score": 1.0, + "content": "inefficient to implement with sparse operations.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 264, + 506, + 353 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 504, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "score": 1.0, + "content": "Topological pooling methods. These methods pre-compute a pyramid of coarsened graphs, only", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 373, + 388 + ], + "score": 1.0, + "content": "taking into account the topology (A), but not the node features", + "type": "text" + }, + { + "bbox": [ + 373, + 376, + 389, + 387 + ], + "score": 0.53, + "content": "\\mathbf { \\Pi } ( \\mathbf { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 375, + 506, + 388 + ], + "score": 1.0, + "content": ". During training, the node", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 400 + ], + "score": 1.0, + "content": "features are pooled with standard procedures and are fit into these deterministic graph structures.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "These methods are less flexible, but provide a stronger bias that can prevent degenerate solutions", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 315, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 315, + 421 + ], + "score": 1.0, + "content": "(e.g., coarsened graphs collapsing in a single node).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 363, + 506, + 421 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 425, + 505, + 491 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "The approach proposed by Bruna et al. (2013), which has been adopted also in other GNN ar-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "chitectures (Defferrard et al., 2016; Fey et al., 2018), exploits GRACLUS (Dhillon et al., 2004), a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 339, + 460 + ], + "score": 1.0, + "content": "hierarchical algorithm based on SC. At each pooling level", + "type": "text" + }, + { + "bbox": [ + 339, + 448, + 344, + 457 + ], + "score": 0.39, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 447, + 505, + 460 + ], + "score": 1.0, + "content": ", GRACLUS indetifies the pairs of max-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 458, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 190, + 472 + ], + "score": 1.0, + "content": "imally similar nodes", + "type": "text" + }, + { + "bbox": [ + 191, + 459, + 199, + 469 + ], + "score": 0.85, + "content": "i _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 458, + 217, + 472 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 217, + 459, + 225, + 470 + ], + "score": 0.85, + "content": "j _ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 458, + 394, + 472 + ], + "score": 1.0, + "content": "to be clustered together into a new vertex", + "type": "text" + }, + { + "bbox": [ + 394, + 459, + 420, + 471 + ], + "score": 0.91, + "content": "k _ { ( l + 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 458, + 506, + 472 + ], + "score": 1.0, + "content": ". At inference phase,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 505, + 481 + ], + "score": 1.0, + "content": "max-pooling is used to determine which node in the pair is kept. Fake vertices are added so that the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 480, + 462, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 462, + 493 + ], + "score": 1.0, + "content": "number of nodes can be halved each time, but this injects noisy information in the graph.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 425, + 506, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 497, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "score": 1.0, + "content": "Node decimation is a method originally proposed in graph signal processing literature (Shuman", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 506, + 520 + ], + "score": 1.0, + "content": "et al., 2016), which as been adapted also for GNNs (Simonovsky & Komodakis, 2017). The nodes", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 506, + 532 + ], + "score": 1.0, + "content": "are partitioned in two sets, according to the signs of the Laplacian eigenvector associated to the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 543 + ], + "score": 1.0, + "content": "largest eigenvalue. One of the two sets is dropped, reducing the number of nodes each time approx-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 506, + 554 + ], + "score": 1.0, + "content": "imately by half. Kron reduction is used to compute a pyramid of coarsened Laplacians from the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 552, + 177, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 177, + 564 + ], + "score": 1.0, + "content": "remaining nodes.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 496, + 506, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 582 + ], + "score": 1.0, + "content": "A procedure proposed in Gama et al. (2018) diffuses a signal from designated nodes on the graph", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 505, + 592 + ], + "score": 1.0, + "content": "and stores the observed sequence of diffused components. The resulting stream of information", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 604 + ], + "score": 1.0, + "content": "is interpreted as a time signal, where standard CNN pooling is applied. We also mention a pooling", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 615 + ], + "score": 1.0, + "content": "operation for coarsening binary unweighted graphs by aggregating maximal cliques (Luzhnica et al.,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "score": 1.0, + "content": "2019). Nodes assigned to the same clique are summarized by max or average pooling and become a", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 624, + 242, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 242, + 636 + ], + "score": 1.0, + "content": "new node in the coarsened graph.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 569, + 506, + 636 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 651, + 200, + 664 + ], + "lines": [ + { + "bbox": [ + 104, + 650, + 202, + 666 + ], + "spans": [ + { + "bbox": [ + 104, + 650, + 202, + 666 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 506, + 691 + ], + "score": 1.0, + "content": "We consider both supervised and unsupervised tasks, and compare minCUTpool with other GNN", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "pooling strategies. The Appendix provides further details on the experiments and a schematic de-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "piction of the architectures used in each task. In addition, the Appendix reports two additional", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "experiments: i) graph reconstruction by means of an Auto Encoder with bottleneck, implemented", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 462, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 462, + 734 + ], + "score": 1.0, + "content": "with pooling and un-pooling layers, ii) an architecture with pooling for graph regression.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 271, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 272, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 272, + 95 + ], + "score": 1.0, + "content": "5.1 CLUSTERING THE GRAPH NODES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "To study the effectiveness of the proposed loss, we perform different node clustering tasks with a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "score": 1.0, + "content": "simple GNN composed of a single MP layer followed by a pooling layer. The GNN is trained by", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 125, + 497, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 155, + 137 + ], + "score": 1.0, + "content": "minimizing", + "type": "text" + }, + { + "bbox": [ + 155, + 126, + 168, + 136 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 125, + 497, + 137 + ], + "score": 1.0, + "content": "only, so that its effect is evaluated without the “interference” of a supervised loss.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 148, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "score": 1.0, + "content": "Clustering on synthetic networks We consider two simple graphs: the first is a network with 6", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "communities and the second is a regular grid. The adjacency matrix A is binary and the features X", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "are the 2-D node coordinates. Fig. 2 depicts the node partitions generated by SC (a, d), Diffpool (b,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 195 + ], + "score": 1.0, + "content": "e), and minCUTpool (c, f). Cluster indexes for Diffpool and minCUTpool are obtained by taking", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "the argmax of S row-wise. Compared to SC, Diffpool and minCUTpool leverage the information", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 506, + 216 + ], + "score": 1.0, + "content": "contained in X. minCUTpool generates very accurate and balanced partitions, demonstrating that", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "the cluster assignment matrix S is well formed. On the other hand, Diffpool assigns some nodes to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 225, + 476, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 476, + 237 + ], + "score": 1.0, + "content": "the wrong community in the first example, and produces an imbalanced partition of the grid.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5 + }, + { + "type": "image", + "bbox": [ + 159, + 251, + 452, + 472 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 159, + 251, + 452, + 472 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 251, + 452, + 472 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 452, + 472 + ], + "score": 0.978, + "type": "image", + "image_path": "f54a1747a2f3fb20dd6df85e294051b41f7c8f617ba05ec5ba321652a3688701.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 159, + 251, + 452, + 264.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 159, + 264.0, + 452, + 277.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 159, + 277.0, + 452, + 290.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 159, + 290.0, + 452, + 303.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 159, + 303.0, + 452, + 316.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 159, + 316.0, + 452, + 329.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 159, + 329.0, + 452, + 342.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 159, + 342.0, + 452, + 355.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 159, + 355.0, + 452, + 368.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 159, + 368.0, + 452, + 381.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 159, + 381.0, + 452, + 394.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 159, + 394.0, + 452, + 407.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 159, + 407.0, + 452, + 420.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 159, + 420.0, + 452, + 433.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 159, + 433.0, + 452, + 446.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 159, + 446.0, + 452, + 459.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 159, + 459.0, + 452, + 472.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 485, + 478, + 497 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 475, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 342, + 499 + ], + "score": 1.0, + "content": "Figure 2: Node clustering on a community network", + "type": "text" + }, + { + "bbox": [ + 343, + 486, + 366, + 496 + ], + "score": 0.79, + "content": "K { = } 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 484, + 451, + 499 + ], + "score": 1.0, + "content": "and on a grid graph", + "type": "text" + }, + { + "bbox": [ + 451, + 486, + 475, + 496 + ], + "score": 0.79, + "content": "( K { = } 5 )", + "type": "inline_equation" + } + ], + "index": 29 + } + ], + "index": 29 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 504, + 527 + ], + "score": 1.0, + "content": "Image segmentation Given an image, we build a Region Adjacency Graph (Tremeau & Colan- ´", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "toni, 2000) using as nodes the regions generated by an oversegmentation procedure (Felzenszwalb", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "& Huttenlocher, 2004). The SC technique used in this example is the recursive normalized cut (Shi", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "& Malik, 2000), which recursively clusters the nodes until convergence. For Diffpool and min-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "CUTpool, we include node features consisting of the average and total color in each oversegmented", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 296, + 582 + ], + "score": 1.0, + "content": "region. We set the number of desired clusters to", + "type": "text" + }, + { + "bbox": [ + 296, + 569, + 325, + 579 + ], + "score": 0.9, + "content": "K = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 568, + 506, + 582 + ], + "score": 1.0, + "content": ". The results in Fig. 3 show that minCUTpool", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 592 + ], + "score": 1.0, + "content": "yields a more precise segmentation. On the other hand, SC and Diffpool aggregate wrong regions", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 591, + 291, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 291, + 603 + ], + "score": 1.0, + "content": "and, in addition, SC finds too many segments.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "Clustering on citation networks We cluster the nodes of three popular citation networks: Cora,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "Citeseer, and Pubmed. The nodes are documents represented by sparse bag-of-words feature vectors", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 144, + 648 + ], + "score": 1.0, + "content": "stored in", + "type": "text" + }, + { + "bbox": [ + 144, + 636, + 154, + 646 + ], + "score": 0.39, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 636, + 297, + 648 + ], + "score": 1.0, + "content": "and the binary undirected edges in", + "type": "text" + }, + { + "bbox": [ + 297, + 636, + 307, + 646 + ], + "score": 0.33, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "indicate citation links between documents. Each", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 647, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 129, + 659 + ], + "score": 1.0, + "content": "node", + "type": "text" + }, + { + "bbox": [ + 129, + 648, + 134, + 657 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 647, + 280, + 659 + ], + "score": 1.0, + "content": "is labeled with the document class", + "type": "text" + }, + { + "bbox": [ + 280, + 649, + 289, + 659 + ], + "score": 0.85, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 647, + 505, + 659 + ], + "score": 1.0, + "content": ". Once the training is over, to test the quality of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 658, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 669 + ], + "score": 1.0, + "content": "partitions generated by each method we check the agreement between the cluster assignments and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 668, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 343, + 687 + ], + "score": 1.0, + "content": "the true class labels. Tab. 1 reports the Completeness Score", + "type": "text" + }, + { + "bbox": [ + 344, + 669, + 437, + 686 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathrm { C S } ( \\tilde { \\bf y } , { \\bf y } ) = 1 - \\frac { H ( \\tilde { \\bf y } | { \\bf y } ) } { H ( \\tilde { \\bf y } ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 668, + 506, + 687 + ], + "score": 1.0, + "content": "and Normalized", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 684, + 411, + 705 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 187, + 702 + ], + "score": 1.0, + "content": "Mutual Information", + "type": "text" + }, + { + "bbox": [ + 187, + 684, + 299, + 705 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { N M I } ( \\tilde { \\mathbf { y } } , \\mathbf { y } ) = \\frac { H ( \\tilde { \\mathbf { y } } ) - H ( \\tilde { \\mathbf { y } } | \\mathbf { y } ) } { \\sqrt { H ( \\tilde { \\mathbf { y } } ) - H ( \\mathbf { y } ) } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 686, + 330, + 701 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 330, + 687, + 351, + 699 + ], + "score": 0.91, + "content": "H ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 686, + 411, + 701 + ], + "score": 1.0, + "content": "is the entropy.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "The GNN architecture configured with minCUTpool achieves a higher NMI score than SC, which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 267, + 733 + ], + "score": 1.0, + "content": "does not account for the node features", + "type": "text" + }, + { + "bbox": [ + 268, + 721, + 277, + 730 + ], + "score": 0.68, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "when generating the partitions. Our pooling operation", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 45.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 83, + 271, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 272, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 272, + 95 + ], + "score": 1.0, + "content": "5.1 CLUSTERING THE GRAPH NODES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 506, + 115 + ], + "score": 1.0, + "content": "To study the effectiveness of the proposed loss, we perform different node clustering tasks with a", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 113, + 505, + 127 + ], + "score": 1.0, + "content": "simple GNN composed of a single MP layer followed by a pooling layer. The GNN is trained by", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 125, + 497, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 155, + 137 + ], + "score": 1.0, + "content": "minimizing", + "type": "text" + }, + { + "bbox": [ + 155, + 126, + 168, + 136 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 125, + 497, + 137 + ], + "score": 1.0, + "content": "only, so that its effect is evaluated without the “interference” of a supervised loss.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 103, + 506, + 137 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 148, + 505, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "score": 1.0, + "content": "Clustering on synthetic networks We consider two simple graphs: the first is a network with 6", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 171 + ], + "score": 1.0, + "content": "communities and the second is a regular grid. The adjacency matrix A is binary and the features X", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "are the 2-D node coordinates. Fig. 2 depicts the node partitions generated by SC (a, d), Diffpool (b,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 195 + ], + "score": 1.0, + "content": "e), and minCUTpool (c, f). Cluster indexes for Diffpool and minCUTpool are obtained by taking", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 505, + 205 + ], + "score": 1.0, + "content": "the argmax of S row-wise. Compared to SC, Diffpool and minCUTpool leverage the information", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 204, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 506, + 216 + ], + "score": 1.0, + "content": "contained in X. minCUTpool generates very accurate and balanced partitions, demonstrating that", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "the cluster assignment matrix S is well formed. On the other hand, Diffpool assigns some nodes to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 225, + 476, + 237 + ], + "spans": [ + { + "bbox": [ + 106, + 225, + 476, + 237 + ], + "score": 1.0, + "content": "the wrong community in the first example, and produces an imbalanced partition of the grid.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 147, + 506, + 237 + ] + }, + { + "type": "image", + "bbox": [ + 159, + 251, + 452, + 472 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 159, + 251, + 452, + 472 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 251, + 452, + 472 + ], + "spans": [ + { + "bbox": [ + 159, + 251, + 452, + 472 + ], + "score": 0.978, + "type": "image", + "image_path": "f54a1747a2f3fb20dd6df85e294051b41f7c8f617ba05ec5ba321652a3688701.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 159, + 251, + 452, + 264.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 159, + 264.0, + 452, + 277.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 159, + 277.0, + 452, + 290.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 159, + 290.0, + 452, + 303.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 159, + 303.0, + 452, + 316.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 159, + 316.0, + 452, + 329.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 159, + 329.0, + 452, + 342.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 159, + 342.0, + 452, + 355.0 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 159, + 355.0, + 452, + 368.0 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 159, + 368.0, + 452, + 381.0 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 159, + 381.0, + 452, + 394.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 159, + 394.0, + 452, + 407.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 159, + 407.0, + 452, + 420.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 159, + 420.0, + 452, + 433.0 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 159, + 433.0, + 452, + 446.0 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 159, + 446.0, + 452, + 459.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 159, + 459.0, + 452, + 472.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 131, + 485, + 478, + 497 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 484, + 475, + 499 + ], + "spans": [ + { + "bbox": [ + 130, + 484, + 342, + 499 + ], + "score": 1.0, + "content": "Figure 2: Node clustering on a community network", + "type": "text" + }, + { + "bbox": [ + 343, + 486, + 366, + 496 + ], + "score": 0.79, + "content": "K { = } 6 )", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 484, + 451, + 499 + ], + "score": 1.0, + "content": "and on a grid graph", + "type": "text" + }, + { + "bbox": [ + 451, + 486, + 475, + 496 + ], + "score": 0.79, + "content": "( K { = } 5 )", + "type": "inline_equation" + } + ], + "index": 29 + } + ], + "index": 29 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 505, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 504, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 504, + 527 + ], + "score": 1.0, + "content": "Image segmentation Given an image, we build a Region Adjacency Graph (Tremeau & Colan- ´", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 537 + ], + "score": 1.0, + "content": "toni, 2000) using as nodes the regions generated by an oversegmentation procedure (Felzenszwalb", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "& Huttenlocher, 2004). The SC technique used in this example is the recursive normalized cut (Shi", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 559 + ], + "score": 1.0, + "content": "& Malik, 2000), which recursively clusters the nodes until convergence. For Diffpool and min-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 570 + ], + "score": 1.0, + "content": "CUTpool, we include node features consisting of the average and total color in each oversegmented", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 296, + 582 + ], + "score": 1.0, + "content": "region. We set the number of desired clusters to", + "type": "text" + }, + { + "bbox": [ + 296, + 569, + 325, + 579 + ], + "score": 0.9, + "content": "K = 4", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 568, + 506, + 582 + ], + "score": 1.0, + "content": ". The results in Fig. 3 show that minCUTpool", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 580, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 592 + ], + "score": 1.0, + "content": "yields a more precise segmentation. On the other hand, SC and Diffpool aggregate wrong regions", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 591, + 291, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 291, + 603 + ], + "score": 1.0, + "content": "and, in addition, SC finds too many segments.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 33.5, + "bbox_fs": [ + 104, + 514, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 614, + 505, + 705 + ], + "lines": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "Clustering on citation networks We cluster the nodes of three popular citation networks: Cora,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "Citeseer, and Pubmed. The nodes are documents represented by sparse bag-of-words feature vectors", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 636, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 144, + 648 + ], + "score": 1.0, + "content": "stored in", + "type": "text" + }, + { + "bbox": [ + 144, + 636, + 154, + 646 + ], + "score": 0.39, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 636, + 297, + 648 + ], + "score": 1.0, + "content": "and the binary undirected edges in", + "type": "text" + }, + { + "bbox": [ + 297, + 636, + 307, + 646 + ], + "score": 0.33, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 636, + 505, + 648 + ], + "score": 1.0, + "content": "indicate citation links between documents. Each", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 647, + 505, + 659 + ], + "spans": [ + { + "bbox": [ + 105, + 647, + 129, + 659 + ], + "score": 1.0, + "content": "node", + "type": "text" + }, + { + "bbox": [ + 129, + 648, + 134, + 657 + ], + "score": 0.74, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 647, + 280, + 659 + ], + "score": 1.0, + "content": "is labeled with the document class", + "type": "text" + }, + { + "bbox": [ + 280, + 649, + 289, + 659 + ], + "score": 0.85, + "content": "y _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 647, + 505, + 659 + ], + "score": 1.0, + "content": ". Once the training is over, to test the quality of the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 658, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 658, + 506, + 669 + ], + "score": 1.0, + "content": "partitions generated by each method we check the agreement between the cluster assignments and", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 668, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 343, + 687 + ], + "score": 1.0, + "content": "the true class labels. Tab. 1 reports the Completeness Score", + "type": "text" + }, + { + "bbox": [ + 344, + 669, + 437, + 686 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathrm { C S } ( \\tilde { \\bf y } , { \\bf y } ) = 1 - \\frac { H ( \\tilde { \\bf y } | { \\bf y } ) } { H ( \\tilde { \\bf y } ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 438, + 668, + 506, + 687 + ], + "score": 1.0, + "content": "and Normalized", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 684, + 411, + 705 + ], + "spans": [ + { + "bbox": [ + 104, + 686, + 187, + 702 + ], + "score": 1.0, + "content": "Mutual Information", + "type": "text" + }, + { + "bbox": [ + 187, + 684, + 299, + 705 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\mathbf { N M I } ( \\tilde { \\mathbf { y } } , \\mathbf { y } ) = \\frac { H ( \\tilde { \\mathbf { y } } ) - H ( \\tilde { \\mathbf { y } } | \\mathbf { y } ) } { \\sqrt { H ( \\tilde { \\mathbf { y } } ) - H ( \\mathbf { y } ) } } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 686, + 330, + 701 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 330, + 687, + 351, + 699 + ], + "score": 0.91, + "content": "H ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 686, + 411, + 701 + ], + "score": 1.0, + "content": "is the entropy.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 41, + "bbox_fs": [ + 104, + 613, + 506, + 705 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "The GNN architecture configured with minCUTpool achieves a higher NMI score than SC, which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 267, + 733 + ], + "score": 1.0, + "content": "does not account for the node features", + "type": "text" + }, + { + "bbox": [ + 268, + 721, + 277, + 730 + ], + "score": 0.68, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "when generating the partitions. Our pooling operation", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "outperforms also Diffpool, since the minimization of the unsupervised loss in Diffpool yields de-", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "generate solutions. The pathological behavior is shown in Fig. 4, which depicts the evolution of the", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 501, + 487, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 487, + 514 + ], + "score": 1.0, + "content": "NMI scores as the unsupervised losses in Diffpool and minCUTpool are minimized in training.", + "type": "text", + "cross_page": true + } + ], + "index": 10 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 86, + 494, + 267 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 86, + 494, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 86, + 494, + 267 + ], + "spans": [ + { + "bbox": [ + 122, + 86, + 494, + 267 + ], + "score": 0.975, + "type": "image", + "image_path": "81a3f7ffc1776e74fbc56b414ab1c318578c5dc39e05ac6ea084422888017b44.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 86, + 494, + 146.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 146.33333333333334, + 494, + 206.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 206.66666666666669, + 494, + 267.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 130, + 279, + 479, + 292 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 278, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 129, + 278, + 479, + 294 + ], + "score": 1.0, + "content": "Figure 3: Image segmentation by clustering the nodes of the Region Adjacency Graph.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 112, + 309, + 497, + 433 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 309, + 497, + 433 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 309, + 497, + 433 + ], + "spans": [ + { + "bbox": [ + 112, + 309, + 497, + 433 + ], + "score": 0.97, + "type": "image", + "image_path": "4067d319802597da0f8a99d17f9a05ae75ef1d9974bd1b46567b315150d067c6.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 112, + 309, + 497, + 350.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 112, + 350.3333333333333, + 497, + 391.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 391.66666666666663, + 497, + 432.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 145, + 446, + 464, + 458 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 444, + 466, + 461 + ], + "spans": [ + { + "bbox": [ + 145, + 444, + 466, + 461 + ], + "score": 1.0, + "content": "Figure 4: Unsupervised losses and NMI of Diffpool and minCUTpool on Cora.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 513 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 492 + ], + "score": 1.0, + "content": "outperforms also Diffpool, since the minimization of the unsupervised loss in Diffpool yields de-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 503 + ], + "score": 1.0, + "content": "generate solutions. The pathological behavior is shown in Fig. 4, which depicts the evolution of the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 501, + 487, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 487, + 514 + ], + "score": 1.0, + "content": "NMI scores as the unsupervised losses in Diffpool and minCUTpool are minimized in training.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 115, + 557, + 496, + 621 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 532, + 503, + 555 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Table 1: NMI and CS obtained by clustering the nodes on citation networks over 10 different runs.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 543, + 369, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 201, + 556 + ], + "score": 1.0, + "content": "The number of clusters", + "type": "text" + }, + { + "bbox": [ + 201, + 545, + 212, + 554 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 543, + 369, + 556 + ], + "score": 1.0, + "content": "is equal to the number of node classes.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "table_body", + "bbox": [ + 115, + 557, + 496, + 621 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 557, + 496, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 557, + 496, + 621 + ], + "score": 0.978, + "html": "
DatasetKSpectral clusteringDiffpoolminCUTpool
NMIcsNMICSNMICS
Cora70.025 ± 0.0140.126 ± 0.0420.315 ± 0.0050.309 ±0.0050.404 ± 0.0180.392 ± 0.018
Citeseer60.014 ± 0.0030.033 ±0.0000.139 ± 0.0160.153 ± 0.0200.287 ±0.0470.283 ± 0.046
Pubmed30.182 ± 0.0000.261 ± 0.0000.079 ±0.0010.085 ±0.0010.200 ± 0.0200.197 ± 0.019
", + "type": "table", + "image_path": "4158255af3330a03321c8d25f82fd169f90fca4a4f87f98e57d954019d96471f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 115, + 557, + 496, + 578.3333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 115, + 578.3333333333334, + 496, + 599.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 115, + 599.6666666666667, + 496, + 621.0000000000001 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 290, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 291, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 291, + 657 + ], + "score": 1.0, + "content": "5.2 SUPERVISED GRAPH CLASSIFICATION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 169, + 677 + ], + "score": 1.0, + "content": "In this task, the", + "type": "text" + }, + { + "bbox": [ + 169, + 667, + 174, + 676 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 666, + 275, + 677 + ], + "score": 1.0, + "content": "-th datum is a graph with", + "type": "text" + }, + { + "bbox": [ + 275, + 666, + 288, + 677 + ], + "score": 0.88, + "content": "N _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 666, + 400, + 677 + ], + "score": 1.0, + "content": "nodes represented by a pair", + "type": "text" + }, + { + "bbox": [ + 400, + 666, + 439, + 678 + ], + "score": 0.93, + "content": "\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "and must be as-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 216, + 689 + ], + "score": 1.0, + "content": "sociated to the correct label", + "type": "text" + }, + { + "bbox": [ + 216, + 678, + 227, + 688 + ], + "score": 0.84, + "content": "\\mathbf { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 677, + 504, + 689 + ], + "score": 1.0, + "content": ". We test the models on different graph classification datasets. For fea-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "score": 1.0, + "content": "tureless graphs, we used the node degree information and the clustering coefficient as surrogate node", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 413, + 712 + ], + "score": 1.0, + "content": "features. We evaluate model performance with a 10-fold train/test split, using", + "type": "text" + }, + { + "bbox": [ + 414, + 699, + 433, + 709 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "of the training set", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "in each fold as validation for early stopping. We adopt a fixed network architecture, MP(32)-pool-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "MP(32)-pool-MP(32)-GlobalAvgPool-softmax, where MP is the message-passing operation in (4)", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 86, + 494, + 267 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 86, + 494, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 86, + 494, + 267 + ], + "spans": [ + { + "bbox": [ + 122, + 86, + 494, + 267 + ], + "score": 0.975, + "type": "image", + "image_path": "81a3f7ffc1776e74fbc56b414ab1c318578c5dc39e05ac6ea084422888017b44.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 86, + 494, + 146.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 146.33333333333334, + 494, + 206.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 206.66666666666669, + 494, + 267.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 130, + 279, + 479, + 292 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 129, + 278, + 479, + 294 + ], + "spans": [ + { + "bbox": [ + 129, + 278, + 479, + 294 + ], + "score": 1.0, + "content": "Figure 3: Image segmentation by clustering the nodes of the Region Adjacency Graph.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 112, + 309, + 497, + 433 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 309, + 497, + 433 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 309, + 497, + 433 + ], + "spans": [ + { + "bbox": [ + 112, + 309, + 497, + 433 + ], + "score": 0.97, + "type": "image", + "image_path": "4067d319802597da0f8a99d17f9a05ae75ef1d9974bd1b46567b315150d067c6.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 112, + 309, + 497, + 350.3333333333333 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 112, + 350.3333333333333, + 497, + 391.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 112, + 391.66666666666663, + 497, + 432.99999999999994 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 145, + 446, + 464, + 458 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 145, + 444, + 466, + 461 + ], + "spans": [ + { + "bbox": [ + 145, + 444, + 466, + 461 + ], + "score": 1.0, + "content": "Figure 4: Unsupervised losses and NMI of Diffpool and minCUTpool on Cora.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 107, + 479, + 505, + 513 + ], + "lines": [], + "index": 9, + "bbox_fs": [ + 105, + 479, + 505, + 514 + ], + "lines_deleted": true + }, + { + "type": "table", + "bbox": [ + 115, + 557, + 496, + 621 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 532, + 503, + 555 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "Table 1: NMI and CS obtained by clustering the nodes on citation networks over 10 different runs.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 543, + 369, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 201, + 556 + ], + "score": 1.0, + "content": "The number of clusters", + "type": "text" + }, + { + "bbox": [ + 201, + 545, + 212, + 554 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 212, + 543, + 369, + 556 + ], + "score": 1.0, + "content": "is equal to the number of node classes.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "table_body", + "bbox": [ + 115, + 557, + 496, + 621 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 557, + 496, + 621 + ], + "spans": [ + { + "bbox": [ + 115, + 557, + 496, + 621 + ], + "score": 0.978, + "html": "
DatasetKSpectral clusteringDiffpoolminCUTpool
NMIcsNMICSNMICS
Cora70.025 ± 0.0140.126 ± 0.0420.315 ± 0.0050.309 ±0.0050.404 ± 0.0180.392 ± 0.018
Citeseer60.014 ± 0.0030.033 ±0.0000.139 ± 0.0160.153 ± 0.0200.287 ±0.0470.283 ± 0.046
Pubmed30.182 ± 0.0000.261 ± 0.0000.079 ±0.0010.085 ±0.0010.200 ± 0.0200.197 ± 0.019
", + "type": "table", + "image_path": "4158255af3330a03321c8d25f82fd169f90fca4a4f87f98e57d954019d96471f.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 115, + 557, + 496, + 578.3333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 115, + 578.3333333333334, + 496, + 599.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 115, + 599.6666666666667, + 496, + 621.0000000000001 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 12.75 + }, + { + "type": "title", + "bbox": [ + 108, + 645, + 290, + 656 + ], + "lines": [ + { + "bbox": [ + 106, + 644, + 291, + 657 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 291, + 657 + ], + "score": 1.0, + "content": "5.2 SUPERVISED GRAPH CLASSIFICATION", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 665, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 169, + 677 + ], + "score": 1.0, + "content": "In this task, the", + "type": "text" + }, + { + "bbox": [ + 169, + 667, + 174, + 676 + ], + "score": 0.68, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 666, + 275, + 677 + ], + "score": 1.0, + "content": "-th datum is a graph with", + "type": "text" + }, + { + "bbox": [ + 275, + 666, + 288, + 677 + ], + "score": 0.88, + "content": "N _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 666, + 400, + 677 + ], + "score": 1.0, + "content": "nodes represented by a pair", + "type": "text" + }, + { + "bbox": [ + 400, + 666, + 439, + 678 + ], + "score": 0.93, + "content": "\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 440, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "and must be as-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 216, + 689 + ], + "score": 1.0, + "content": "sociated to the correct label", + "type": "text" + }, + { + "bbox": [ + 216, + 678, + 227, + 688 + ], + "score": 0.84, + "content": "\\mathbf { y } _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 677, + 504, + 689 + ], + "score": 1.0, + "content": ". We test the models on different graph classification datasets. For fea-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 504, + 700 + ], + "score": 1.0, + "content": "tureless graphs, we used the node degree information and the clustering coefficient as surrogate node", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 413, + 712 + ], + "score": 1.0, + "content": "features. We evaluate model performance with a 10-fold train/test split, using", + "type": "text" + }, + { + "bbox": [ + 414, + 699, + 433, + 709 + ], + "score": 0.88, + "content": "1 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "of the training set", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "in each fold as validation for early stopping. We adopt a fixed network architecture, MP(32)-pool-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 734 + ], + "score": 1.0, + "content": "MP(32)-pool-MP(32)-GlobalAvgPool-softmax, where MP is the message-passing operation in (4)", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 666, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "with 32 hidden units. The pooling module is implemented either by Graclus, Decimation pooling,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 125, + 106 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + }, + { + "bbox": [ + 126, + 94, + 135, + 104 + ], + "score": 0.6, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 94, + 505, + 106 + ], + "score": 1.0, + "content": ", SAGPool (Lee et al., 2019), Diffpool, or the proposed minCUTpool. Each pooling method", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 504, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 302, + 117 + ], + "score": 1.0, + "content": "is configured to drop half of the nodes in a graph", + "type": "text" + }, + { + "bbox": [ + 303, + 105, + 345, + 116 + ], + "score": 0.9, + "content": "K = N / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 104, + 374, + 117 + ], + "score": 1.0, + "content": "in Top-", + "type": "text" + }, + { + "bbox": [ + 375, + 105, + 385, + 115 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 104, + 504, + 117 + ], + "score": 1.0, + "content": ", Diffpool, and minCUTpool).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "As baselines, we consider the popular Weisfeiler-Lehman (WL) graph kernel (Shervashidze et al.,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 448, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 448, + 140 + ], + "score": 1.0, + "content": "2011), a network with only MP layers (Flat), and a fully connected network (Dense).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 107, + 170, + 507, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 123, + 157, + 484, + 169 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 156, + 484, + 170 + ], + "spans": [ + { + "bbox": [ + 126, + 156, + 392, + 170 + ], + "score": 1.0, + "content": "Table 2: Graph classification accuracy. Significantly better results", + "type": "text" + }, + { + "bbox": [ + 393, + 158, + 435, + 169 + ], + "score": 0.86, + "content": "( p < 0 . 0 5 )", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 156, + 484, + 170 + ], + "score": 1.0, + "content": "are in bold.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 170, + 507, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 170, + 507, + 264 + ], + "spans": [ + { + "bbox": [ + 107, + 170, + 507, + 264 + ], + "score": 0.979, + "html": "
DatasetWLDenseFlatGraclusDecim.DiffpoolTop-KSAGpoolminCUT
Bench-easy92.629.3±0.398.5±0.397.5±0.597.9±0.598.6±0.482.4±8.984.2±2.399.0±0.0
Bench-hard60.029.4±0.367.6±2.869.0±1.572.6±0.969.9±1.942.7±15.237.7±14.573.8±1.9
Mutagenicity81.7±1.168.4±0.378.0±1.374.4±1.877.8±2.377.6±2.771.9±3.772.4±2.479.9±2.1
Proteins71.2±2.668.7±3.372.6±4.868.6±4.673.3±3.772.7±3.869.6±3.570.5±2.676.5±2.6
DD78.6±2.770.6±5.276.8±1.570.5±4.872.0±3.179.3±2.469.4±7.871.5±4.580.8±2.3
COLLAB74.8±1.379.3±1.682.1±1.877.1±2.179.1±1.581.8±1.479.3±1.879.2±2.083.4±1.7
Reddit-Binary68.2±1.748.5±2.680.3±2.679.2±0.484.3±2.486.8±2.174.7±4.573.9±5.191.4±1.5
", + "type": "table", + "image_path": "f42ea0a0a3ed9e12ba04e59f7afa9d42c08e5d9a16acdb713d983a07264519c7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 107, + 170, + 507, + 201.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 201.33333333333334, + 507, + 232.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 107, + 232.66666666666669, + 507, + 264.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 473, + 291 + ], + "score": 1.0, + "content": "Tab. 2 reports the classification results, highlighting those that are significantly better (", + "type": "text" + }, + { + "bbox": [ + 473, + 280, + 479, + 289 + ], + "score": 0.74, + "content": "\\dot { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "-value", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 139, + 300 + ], + "score": 0.82, + "content": "< ~ 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "w.r.t. the method with the highest mean accuracy). The comparison with Flat helps to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "understand if a pooling operation is useful or not. The results of Dense, instead, help to quantify", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "score": 1.0, + "content": "how much additional information is brought by the graph structure, with respect to the node features", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "score": 1.0, + "content": "alone. It can be seen that minCUTpool obtains always equal or better results with respect to every", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "other GNN architecture. On the other hand, some pooling procedures do not always improve the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "performance compared to the Flat baseline, making them not advisable to use in some cases. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "WL kernel generally performs worse than the GNNs, except for the Mutagenicity dataset. This is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "probably because Mutagenicity has smaller graphs than the other datasets, and the adopted GNN", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 377, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 506, + 389 + ], + "score": 1.0, + "content": "architecture is overparametrized for this task. Interestingly, in some dataset such as Proteins and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "score": 1.0, + "content": "COLLAB it is possible to obtain fairly good classification accuracy with the Dense architecture,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 399, + 361, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 361, + 410 + ], + "score": 1.0, + "content": "meaning that the graph structure only adds limited information.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 14.5 + }, + { + "type": "image", + "bbox": [ + 109, + 424, + 281, + 498 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 424, + 281, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 424, + 281, + 498 + ], + "spans": [ + { + "bbox": [ + 109, + 424, + 281, + 498 + ], + "score": 0.968, + "type": "image", + "image_path": "e375310d148f5976514d2295a7d3c8830c48da3e0966c6040f4a07d9c8821927.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 109, + 424, + 281, + 438.8 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 109, + 438.8, + 281, + 453.6 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 109, + 453.6, + 281, + 468.40000000000003 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 109, + 468.40000000000003, + 281, + 483.20000000000005 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 109, + 483.20000000000005, + 281, + 498.00000000000006 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 295, + 438, + 505, + 482 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 295, + 437, + 504, + 450 + ], + "spans": [ + { + "bbox": [ + 295, + 437, + 504, + 450 + ], + "score": 1.0, + "content": "Figure 5: Average duration of one epoch using the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 295, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "same GNN with different pooling operations. Times", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 460, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 295, + 460, + 505, + 471 + ], + "score": 1.0, + "content": "were computed with an Nvidia GeForce GTX 1050,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 295, + 470, + 453, + 482 + ], + "spans": [ + { + "bbox": [ + 295, + 470, + 453, + 482 + ], + "score": 1.0, + "content": "on the DD dataset with batch size of 1.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + } + ], + "index": 24.75 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Fig. 5 reports a comparison of the execution time per training epoch for each pooling algorithm.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 523, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 504, + 536 + ], + "score": 1.0, + "content": "Graclus and Decimation are understandably the fastest methods, since the coarsened graphs are pre-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "computed. Among the differentiable pooling methods, minCUTpool is faster than Diffpool, which", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 464, + 559 + ], + "score": 1.0, + "content": "uses a slower MP layer rather than a MLP to compute cluster assignments, and than Top-", + "type": "text" + }, + { + "bbox": [ + 464, + 546, + 474, + 555 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 544, + 506, + 559 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 556, + 302, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 302, + 569 + ], + "score": 1.0, + "content": "computes the square of A at every forward pass.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 108, + 583, + 201, + 596 + ], + "lines": [ + { + "bbox": [ + 104, + 581, + 203, + 599 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 203, + 599 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "We proposed a pooling layer for GNNs that coarsens a graph by taking into account both the the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 620, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 631 + ], + "score": 1.0, + "content": "connectivity structure and the node features. The layer optimizes a regularization term based on the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "score": 1.0, + "content": "minCUT objective, which is minimized in conjunction with the task-specific loss to produce node", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 642, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 293, + 653 + ], + "score": 1.0, + "content": "partitions that are optimal for the task at hand.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "score": 1.0, + "content": "We tested the effectiveness of our pooling strategy on unsupervised node clustering tasks, by op-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "timizing only the unsupervised clustering loss, as well as supervised graph classification tasks on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 682, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 505, + 692 + ], + "score": 1.0, + "content": "several popular benchmark datasets. Results show that minCUTpool performs significantly better", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 692, + 277, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 277, + 703 + ], + "score": 1.0, + "content": "than existing pooling strategies for GNNs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 138 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "score": 1.0, + "content": "with 32 hidden units. The pooling module is implemented either by Graclus, Decimation pooling,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 125, + 106 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + }, + { + "bbox": [ + 126, + 94, + 135, + 104 + ], + "score": 0.6, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 94, + 505, + 106 + ], + "score": 1.0, + "content": ", SAGPool (Lee et al., 2019), Diffpool, or the proposed minCUTpool. Each pooling method", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 504, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 302, + 117 + ], + "score": 1.0, + "content": "is configured to drop half of the nodes in a graph", + "type": "text" + }, + { + "bbox": [ + 303, + 105, + 345, + 116 + ], + "score": 0.9, + "content": "K = N / 2", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 104, + 374, + 117 + ], + "score": 1.0, + "content": "in Top-", + "type": "text" + }, + { + "bbox": [ + 375, + 105, + 385, + 115 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 104, + 504, + 117 + ], + "score": 1.0, + "content": ", Diffpool, and minCUTpool).", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 505, + 128 + ], + "score": 1.0, + "content": "As baselines, we consider the popular Weisfeiler-Lehman (WL) graph kernel (Shervashidze et al.,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 448, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 448, + 140 + ], + "score": 1.0, + "content": "2011), a network with only MP layers (Flat), and a fully connected network (Dense).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 81, + 505, + 140 + ] + }, + { + "type": "table", + "bbox": [ + 107, + 170, + 507, + 264 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 123, + 157, + 484, + 169 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 126, + 156, + 484, + 170 + ], + "spans": [ + { + "bbox": [ + 126, + 156, + 392, + 170 + ], + "score": 1.0, + "content": "Table 2: Graph classification accuracy. Significantly better results", + "type": "text" + }, + { + "bbox": [ + 393, + 158, + 435, + 169 + ], + "score": 0.86, + "content": "( p < 0 . 0 5 )", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 156, + 484, + 170 + ], + "score": 1.0, + "content": "are in bold.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 170, + 507, + 264 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 170, + 507, + 264 + ], + "spans": [ + { + "bbox": [ + 107, + 170, + 507, + 264 + ], + "score": 0.979, + "html": "
DatasetWLDenseFlatGraclusDecim.DiffpoolTop-KSAGpoolminCUT
Bench-easy92.629.3±0.398.5±0.397.5±0.597.9±0.598.6±0.482.4±8.984.2±2.399.0±0.0
Bench-hard60.029.4±0.367.6±2.869.0±1.572.6±0.969.9±1.942.7±15.237.7±14.573.8±1.9
Mutagenicity81.7±1.168.4±0.378.0±1.374.4±1.877.8±2.377.6±2.771.9±3.772.4±2.479.9±2.1
Proteins71.2±2.668.7±3.372.6±4.868.6±4.673.3±3.772.7±3.869.6±3.570.5±2.676.5±2.6
DD78.6±2.770.6±5.276.8±1.570.5±4.872.0±3.179.3±2.469.4±7.871.5±4.580.8±2.3
COLLAB74.8±1.379.3±1.682.1±1.877.1±2.179.1±1.581.8±1.479.3±1.879.2±2.083.4±1.7
Reddit-Binary68.2±1.748.5±2.680.3±2.679.2±0.484.3±2.486.8±2.174.7±4.573.9±5.191.4±1.5
", + "type": "table", + "image_path": "f42ea0a0a3ed9e12ba04e59f7afa9d42c08e5d9a16acdb713d983a07264519c7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 107, + 170, + 507, + 201.33333333333334 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 107, + 201.33333333333334, + 507, + 232.66666666666669 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 107, + 232.66666666666669, + 507, + 264.0 + ], + "spans": [], + "index": 8 + } + ] + } + ], + "index": 6.0 + }, + { + "type": "text", + "bbox": [ + 106, + 277, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 506, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 473, + 291 + ], + "score": 1.0, + "content": "Tab. 2 reports the classification results, highlighting those that are significantly better (", + "type": "text" + }, + { + "bbox": [ + 473, + 280, + 479, + 289 + ], + "score": 0.74, + "content": "\\dot { p }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 277, + 506, + 291 + ], + "score": 1.0, + "content": "-value", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 106, + 289, + 139, + 300 + ], + "score": 0.82, + "content": "< ~ 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "w.r.t. the method with the highest mean accuracy). The comparison with Flat helps to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "understand if a pooling operation is useful or not. The results of Dense, instead, help to quantify", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 324 + ], + "score": 1.0, + "content": "how much additional information is brought by the graph structure, with respect to the node features", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 335 + ], + "score": 1.0, + "content": "alone. It can be seen that minCUTpool obtains always equal or better results with respect to every", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 345 + ], + "score": 1.0, + "content": "other GNN architecture. On the other hand, some pooling procedures do not always improve the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 344, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 505, + 356 + ], + "score": 1.0, + "content": "performance compared to the Flat baseline, making them not advisable to use in some cases. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 506, + 367 + ], + "score": 1.0, + "content": "WL kernel generally performs worse than the GNNs, except for the Mutagenicity dataset. This is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 104, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "probably because Mutagenicity has smaller graphs than the other datasets, and the adopted GNN", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 377, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 506, + 389 + ], + "score": 1.0, + "content": "architecture is overparametrized for this task. Interestingly, in some dataset such as Proteins and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 400 + ], + "score": 1.0, + "content": "COLLAB it is possible to obtain fairly good classification accuracy with the Dense architecture,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 399, + 361, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 361, + 410 + ], + "score": 1.0, + "content": "meaning that the graph structure only adds limited information.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 277, + 506, + 410 + ] + }, + { + "type": "image", + "bbox": [ + 109, + 424, + 281, + 498 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 424, + 281, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 424, + 281, + 498 + ], + "spans": [ + { + "bbox": [ + 109, + 424, + 281, + 498 + ], + "score": 0.968, + "type": "image", + "image_path": "e375310d148f5976514d2295a7d3c8830c48da3e0966c6040f4a07d9c8821927.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 109, + 424, + 281, + 438.8 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 109, + 438.8, + 281, + 453.6 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 109, + 453.6, + 281, + 468.40000000000003 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 109, + 468.40000000000003, + 281, + 483.20000000000005 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 109, + 483.20000000000005, + 281, + 498.00000000000006 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 295, + 438, + 505, + 482 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 295, + 437, + 504, + 450 + ], + "spans": [ + { + "bbox": [ + 295, + 437, + 504, + 450 + ], + "score": 1.0, + "content": "Figure 5: Average duration of one epoch using the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 448, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 295, + 448, + 505, + 460 + ], + "score": 1.0, + "content": "same GNN with different pooling operations. Times", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 460, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 295, + 460, + 505, + 471 + ], + "score": 1.0, + "content": "were computed with an Nvidia GeForce GTX 1050,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 295, + 470, + 453, + 482 + ], + "spans": [ + { + "bbox": [ + 295, + 470, + 453, + 482 + ], + "score": 1.0, + "content": "on the DD dataset with batch size of 1.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + } + ], + "index": 24.75 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Fig. 5 reports a comparison of the execution time per training epoch for each pooling algorithm.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 523, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 504, + 536 + ], + "score": 1.0, + "content": "Graclus and Decimation are understandably the fastest methods, since the coarsened graphs are pre-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "computed. Among the differentiable pooling methods, minCUTpool is faster than Diffpool, which", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 544, + 506, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 464, + 559 + ], + "score": 1.0, + "content": "uses a slower MP layer rather than a MLP to compute cluster assignments, and than Top-", + "type": "text" + }, + { + "bbox": [ + 464, + 546, + 474, + 555 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 544, + 506, + 559 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 556, + 302, + 569 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 302, + 569 + ], + "score": 1.0, + "content": "computes the square of A at every forward pass.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 512, + 506, + 569 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 583, + 201, + 596 + ], + "lines": [ + { + "bbox": [ + 104, + 581, + 203, + 599 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 203, + 599 + ], + "score": 1.0, + "content": "6 CONCLUSIONS", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 609, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 505, + 621 + ], + "score": 1.0, + "content": "We proposed a pooling layer for GNNs that coarsens a graph by taking into account both the the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 620, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 620, + 505, + 631 + ], + "score": 1.0, + "content": "connectivity structure and the node features. The layer optimizes a regularization term based on the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 505, + 642 + ], + "score": 1.0, + "content": "minCUT objective, which is minimized in conjunction with the task-specific loss to produce node", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 642, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 293, + 653 + ], + "score": 1.0, + "content": "partitions that are optimal for the task at hand.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 609, + 505, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 658, + 504, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "score": 1.0, + "content": "We tested the effectiveness of our pooling strategy on unsupervised node clustering tasks, by op-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 505, + 682 + ], + "score": 1.0, + "content": "timizing only the unsupervised clustering loss, as well as supervised graph classification tasks on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 682, + 505, + 692 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 505, + 692 + ], + "score": 1.0, + "content": "several popular benchmark datasets. Results show that minCUTpool performs significantly better", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 692, + 277, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 277, + 703 + ], + "score": 1.0, + "content": "than existing pooling strategies for GNNs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 657, + 505, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 122, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 114, + 122, + 506, + 135 + ], + "score": 1.0, + "content": "Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 142, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 142, + 145 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 105, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Filippo Maria Bianchi, Daniele Grattarola, L Livi, and C Alippi. Graph neural networks with con-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 373, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 373, + 176 + ], + "score": 1.0, + "content": "volutional arma filters. arXiv preprint arXiv:1901.01343, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 105, + 182, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 401, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 401, + 207 + ], + "score": 1.0, + "content": "connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 227 + ], + "score": 1.0, + "content": "Maxwell D Collins, Ji Liu, Jia Xu, Lopamudra Mukherjee, and Vikas Singh. Spectral clustering", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "with a convex regularizer on millions of images. In European Conference on Computer Vision,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 236, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 236, + 248 + ], + "score": 1.0, + "content": "pp. 282–298. Springer, 2014.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 105, + 254, + 503, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 270 + ], + "score": 1.0, + "content": "Anil Damle, Victor Minden, and Lexing Ying. Robust and efficient multi-way spectral clustering.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 266, + 278, + 278 + ], + "spans": [ + { + "bbox": [ + 116, + 266, + 278, + 278 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1609.08251, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 285, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks ¨", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "on graphs with fast localized spectral filtering. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 308, + 243, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 308, + 243, + 320 + ], + "score": 1.0, + "content": "Systems, pp. 3844–3852, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 104, + 327, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Kernel k-means: spectral clustering and nor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 337, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 337, + 505, + 352 + ], + "score": 1.0, + "content": "malized cuts. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 349, + 335, + 362 + ], + "spans": [ + { + "bbox": [ + 116, + 349, + 335, + 362 + ], + "score": 1.0, + "content": "discovery and data mining, pp. 551–556. ACM, 2004.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 103, + 368, + 504, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 382 + ], + "score": 1.0, + "content": "Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. Inter-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 380, + 352, + 392 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 352, + 392 + ], + "score": 1.0, + "content": "national journal of computer vision, 59(2):167–181, 2004.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 108, + 399, + 504, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Muller. Splinecnn: Fast geomet- ¨", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "ric deep learning with continuous b-spline kernels. In Proceedings of the IEEE Conference on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 420, + 368, + 434 + ], + "spans": [ + { + "bbox": [ + 116, + 420, + 368, + 434 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition, pp. 869–877, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "Fernando Gama, Antonio G Marques, Geert Leus, and Alejandro Ribeiro. Convolutional neural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 451, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 506, + 466 + ], + "score": 1.0, + "content": "network architectures for signals supported on graphs. IEEE Transactions on Signal Processing,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 462, + 216, + 475 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 216, + 475 + ], + "score": 1.0, + "content": "67(4):1034–1049, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 493, + 504, + 506 + ], + "spans": [ + { + "bbox": [ + 116, + 493, + 504, + 506 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In Proceedings of the 34th International Conference on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 504, + 379, + 518 + ], + "spans": [ + { + "bbox": [ + 116, + 504, + 379, + 518 + ], + "score": 1.0, + "content": "Machine Learning-Volume 70, pp. 1263–1272. JMLR. org, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Yufei Han and Maurizio Filippone. Mini-batch spectral clustering. In 2017 International Joint", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 534, + 403, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 534, + 403, + 547 + ], + "score": 1.0, + "content": "Conference on Neural Networks (IJCNN), pp. 3888–3895. IEEE, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 504, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "Shuiwang Ji Hongyang Gao. Graph u-nets. In Proceedings of the 36th International conference on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 565, + 251, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 565, + 251, + 578 + ], + "score": 1.0, + "content": "Machine learning (ICML), 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 584, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "Yasuhiko Ikebe, Toshiyuki Inagaki, and Sadaaki Miyamoto. The monotonicity theorem, cauchy’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 115, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "interlace theorem, and the courant-fischer theorem. The American Mathematical Monthly, 94(4):", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 607, + 181, + 618 + ], + "spans": [ + { + "bbox": [ + 116, + 607, + 181, + 618 + ], + "score": 1.0, + "content": "352–354, 1987.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 108, + 626, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Michael Kampffmeyer, Sigurd Løkse, Filippo M. Bianchi, Lorenzo Livi, Arnt-Børre Salberg, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "Robert Jenssen. Deep divergence-based approach to clustering. Neural Networks, 113:91 – 101,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 646, + 143, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 646, + 143, + 661 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 679, + 424, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 679, + 424, + 691 + ], + "score": 1.0, + "content": "works. International Conference of Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 504, + 712 + ], + "score": 1.0, + "content": "Marc T Law, Raquel Urtasun, and Richard S Zemel. Deep spectral clustering learning. In Pro-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "ceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1985–1994.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 719, + 190, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 190, + 733 + ], + "score": 1.0, + "content": "JMLR. org, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 505, + 144 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Peter W Battaglia, Jessica B Hamrick, Victor Bapst, Alvaro Sanchez-Gonzalez, Vinicius Zambaldi,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Mateusz Malinowski, Andrea Tacchetti, David Raposo, Adam Santoro, Ryan Faulkner, et al.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 122, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 114, + 122, + 506, + 135 + ], + "score": 1.0, + "content": "Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 142, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 142, + 145 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 100, + 506, + 145 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 153, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Filippo Maria Bianchi, Daniele Grattarola, L Livi, and C Alippi. Graph neural networks with con-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 163, + 373, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 163, + 373, + 176 + ], + "score": 1.0, + "content": "volutional arma filters. arXiv preprint arXiv:1901.01343, 2019.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 153, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 182, + 504, + 206 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 505, + 196 + ], + "score": 1.0, + "content": "Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. Spectral networks and locally", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 194, + 401, + 207 + ], + "spans": [ + { + "bbox": [ + 115, + 194, + 401, + 207 + ], + "score": 1.0, + "content": "connected networks on graphs. arXiv preprint arXiv:1312.6203, 2013.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 182, + 505, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 213, + 505, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 212, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 212, + 505, + 227 + ], + "score": 1.0, + "content": "Maxwell D Collins, Ji Liu, Jia Xu, Lopamudra Mukherjee, and Vikas Singh. Spectral clustering", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 116, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "with a convex regularizer on millions of images. In European Conference on Computer Vision,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 236, + 248 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 236, + 248 + ], + "score": 1.0, + "content": "pp. 282–298. Springer, 2014.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 212, + 505, + 248 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 254, + 503, + 278 + ], + "lines": [ + { + "bbox": [ + 105, + 252, + 506, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 270 + ], + "score": 1.0, + "content": "Anil Damle, Victor Minden, and Lexing Ying. Robust and efficient multi-way spectral clustering.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 266, + 278, + 278 + ], + "spans": [ + { + "bbox": [ + 116, + 266, + 278, + 278 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1609.08251, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 252, + 506, + 278 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 285, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 505, + 298 + ], + "score": 1.0, + "content": "Michael Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks ¨", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "on graphs with fast localized spectral filtering. In Advances in Neural Information Processing", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 308, + 243, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 308, + 243, + 320 + ], + "score": 1.0, + "content": "Systems, pp. 3844–3852, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 285, + 506, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 327, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "Inderjit S Dhillon, Yuqiang Guan, and Brian Kulis. Kernel k-means: spectral clustering and nor-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 337, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 337, + 505, + 352 + ], + "score": 1.0, + "content": "malized cuts. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 349, + 335, + 362 + ], + "spans": [ + { + "bbox": [ + 116, + 349, + 335, + 362 + ], + "score": 1.0, + "content": "discovery and data mining, pp. 551–556. ACM, 2004.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 106, + 327, + 505, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 103, + 368, + 504, + 392 + ], + "lines": [ + { + "bbox": [ + 105, + 367, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 367, + 505, + 382 + ], + "score": 1.0, + "content": "Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. Inter-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 380, + 352, + 392 + ], + "spans": [ + { + "bbox": [ + 115, + 380, + 352, + 392 + ], + "score": 1.0, + "content": "national journal of computer vision, 59(2):167–181, 2004.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 367, + 505, + 392 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 399, + 504, + 434 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "Matthias Fey, Jan Eric Lenssen, Frank Weichert, and Heinrich Muller. Splinecnn: Fast geomet- ¨", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 115, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "ric deep learning with continuous b-spline kernels. In Proceedings of the IEEE Conference on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 420, + 368, + 434 + ], + "spans": [ + { + "bbox": [ + 116, + 420, + 368, + 434 + ], + "score": 1.0, + "content": "Computer Vision and Pattern Recognition, pp. 869–877, 2018.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 398, + 505, + 434 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 440, + 504, + 474 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 453 + ], + "score": 1.0, + "content": "Fernando Gama, Antonio G Marques, Geert Leus, and Alejandro Ribeiro. Convolutional neural", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 451, + 506, + 466 + ], + "spans": [ + { + "bbox": [ + 115, + 451, + 506, + 466 + ], + "score": 1.0, + "content": "network architectures for signals supported on graphs. IEEE Transactions on Signal Processing,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 462, + 216, + 475 + ], + "spans": [ + { + "bbox": [ + 116, + 462, + 216, + 475 + ], + "score": 1.0, + "content": "67(4):1034–1049, 2018.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 440, + 506, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "score": 1.0, + "content": "Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 493, + 504, + 506 + ], + "spans": [ + { + "bbox": [ + 116, + 493, + 504, + 506 + ], + "score": 1.0, + "content": "message passing for quantum chemistry. In Proceedings of the 34th International Conference on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 504, + 379, + 518 + ], + "spans": [ + { + "bbox": [ + 116, + 504, + 379, + 518 + ], + "score": 1.0, + "content": "Machine Learning-Volume 70, pp. 1263–1272. JMLR. org, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 482, + 505, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 536 + ], + "score": 1.0, + "content": "Yufei Han and Maurizio Filippone. Mini-batch spectral clustering. In 2017 International Joint", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 534, + 403, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 534, + 403, + 547 + ], + "score": 1.0, + "content": "Conference on Neural Networks (IJCNN), pp. 3888–3895. IEEE, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 523, + 505, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 504, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "Shuiwang Ji Hongyang Gao. Graph u-nets. In Proceedings of the 36th International conference on", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 565, + 251, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 565, + 251, + 578 + ], + "score": 1.0, + "content": "Machine learning (ICML), 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 554, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 584, + 505, + 618 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "Yasuhiko Ikebe, Toshiyuki Inagaki, and Sadaaki Miyamoto. The monotonicity theorem, cauchy’s", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 115, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "interlace theorem, and the courant-fischer theorem. The American Mathematical Monthly, 94(4):", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 607, + 181, + 618 + ], + "spans": [ + { + "bbox": [ + 116, + 607, + 181, + 618 + ], + "score": 1.0, + "content": "352–354, 1987.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 584, + 505, + 618 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 626, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 639 + ], + "score": 1.0, + "content": "Michael Kampffmeyer, Sigurd Løkse, Filippo M. Bianchi, Lorenzo Livi, Arnt-Børre Salberg, and", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 115, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "Robert Jenssen. Deep divergence-based approach to clustering. Neural Networks, 113:91 – 101,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 646, + 143, + 661 + ], + "spans": [ + { + "bbox": [ + 115, + 646, + 143, + 661 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 626, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 668, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 681 + ], + "score": 1.0, + "content": "Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 679, + 424, + 691 + ], + "spans": [ + { + "bbox": [ + 116, + 679, + 424, + 691 + ], + "score": 1.0, + "content": "works. International Conference of Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 106, + 668, + 505, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 697, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 697, + 504, + 712 + ], + "score": 1.0, + "content": "Marc T Law, Raquel Urtasun, and Richard S Zemel. Deep spectral clustering learning. In Pro-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 116, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "ceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1985–1994.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 719, + 190, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 719, + 190, + 733 + ], + "score": 1.0, + "content": "JMLR. org, 2017.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 697, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Chen-Yu Lee, Patrick W Gallagher, and Zhuowen Tu. Generalizing pooling functions in convo-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "lutional neural networks: Mixed, gated, and tree. In Artificial Intelligence and Statistics, pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 181, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 181, + 115 + ], + "score": 1.0, + "content": "464–472, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 105, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Junhyun Lee, Inyeop Lee, and Jaewoo Kang. Self-attention graph pooling. In Proceedings of the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 397, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 397, + 147 + ], + "score": 1.0, + "content": "36th International conference on Machine learning (ICML-19), 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 152, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 436, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 436, + 176 + ], + "score": 1.0, + "content": "networks. International Conference of Learning Representations (ICLR), 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "Enxhell Luzhnica, Ben Day, and Pietro Lio. Clique pooling for graph classification. International", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Conference of Learning Representations (ICLR) – Representation Learning on Graphs and Man-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 210, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 210, + 217 + ], + "score": 1.0, + "content": "ifolds workshop, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 504, + 247 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Azade Nazi, Will Hang, Anna Goldie, Sujith Ravi, and Azalia Mirhoseini. Gap: Generalizable", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 455, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 455, + 247 + ], + "score": 1.0, + "content": "approximate graph partitioning framework. arXiv preprint arXiv:1903.00614, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 505, + 276 + ], + "lines": [ + { + "bbox": [ + 106, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "Oliver Nelles. Nonlinear system identification: from classical approaches to neural networks and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 265, + 347, + 277 + ], + "spans": [ + { + "bbox": [ + 114, + 265, + 347, + 277 + ], + "score": 1.0, + "content": "fuzzy models. Springer Science & Business Media, 2013.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 105, + 283, + 467, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 466, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 466, + 296 + ], + "score": 1.0, + "content": "Mark EJ Newman. Mixing patterns in networks. Physical Review E, 67(2):026126, 2003.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 502, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "score": 1.0, + "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 314, + 495, + 326 + ], + "spans": [ + { + "bbox": [ + 117, + 314, + 495, + 326 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2009.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 108, + 331, + 503, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "Uri Shaham, Kelly Stanton, Henry Li, Boaz Nadler, Ronen Basri, and Yuval Kluger. Spectralnet:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 343, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 343, + 470, + 356 + ], + "score": 1.0, + "content": "Spectral clustering using deep neural networks. arXiv preprint arXiv:1801.01587, 2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 375 + ], + "score": 1.0, + "content": "Nino Shervashidze, Pascal Schweitzer, Erik Jan van Leeuwen, Kurt Mehlhorn, and Karsten M Borg-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "wardt. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12(Sep):2539–", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 383, + 167, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 167, + 396 + ], + "score": 1.0, + "content": "2561, 2011.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 503, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 417 + ], + "score": 1.0, + "content": "Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. Departmental Papers", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 414, + 203, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 203, + 426 + ], + "score": 1.0, + "content": "(CIS), pp. 107, 2000.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 502, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 504, + 445 + ], + "score": 1.0, + "content": "David I Shuman, Mohammad Javad Faraji, and Pierre Vandergheynst. A multiscale pyramid trans-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 444, + 477, + 456 + ], + "spans": [ + { + "bbox": [ + 115, + 444, + 477, + 456 + ], + "score": 1.0, + "content": "form for graph signals. IEEE Transactions on Signal Processing, 64(8):2119–2134, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Martin Simonovsky and Nikos Komodakis. Dynamic edgeconditioned filters in convolutional neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "networks on graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 484, + 195, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 484, + 195, + 497 + ], + "score": 1.0, + "content": "Recognition, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "score": 1.0, + "content": "Fei Tian, Bin Gao, Qing Cui, Enhong Chen, and Tie-Yan Liu. Learning deep representations for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 515, + 315, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 315, + 526 + ], + "score": 1.0, + "content": "graph clustering. In AAAI, pp. 1293–1299, 2014.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 533, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 504, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 504, + 546 + ], + "score": 1.0, + "content": "Alain Tremeau and Philippe Colantoni. Regions adjacency graph applied to color image segmenta-´", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 544, + 387, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 387, + 556 + ], + "score": 1.0, + "content": "tion. IEEE Transactions on image processing, 9(4):735–744, 2000.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 504, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 573, + 144, + 587 + ], + "spans": [ + { + "bbox": [ + 114, + 573, + 144, + 587 + ], + "score": 1.0, + "content": "2007.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 504, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 504, + 606 + ], + "score": 1.0, + "content": "Zaiwen Wen and Wotao Yin. A feasible method for optimization with orthogonality constraints.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 604, + 337, + 616 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 337, + 616 + ], + "score": 1.0, + "content": "Mathematical Programming, 142(1-2):397–434, 2013.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Li Yi, Hao Su, Xingwen Guo, and Leonidas J Guibas. Syncspeccnn: Synchronized spectral cnn for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 633, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 115, + 633, + 506, + 647 + ], + "score": 1.0, + "content": "3d shape segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 645, + 259, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 645, + 259, + 657 + ], + "score": 1.0, + "content": "Recognition, pp. 2282–2290, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 663, + 504, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 674, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 116, + 674, + 505, + 688 + ], + "score": 1.0, + "content": "erarchical graph representation learning with differentiable pooling. In Advances in Neural Infor-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 686, + 319, + 698 + ], + "spans": [ + { + "bbox": [ + 116, + 686, + 319, + 698 + ], + "score": 1.0, + "content": "mation Processing Systems, pp. 4800–4810, 2018.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 704, + 504, + 727 + ], + "lines": [ + { + "bbox": [ + 106, + 703, + 505, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 718 + ], + "score": 1.0, + "content": "Yu and Shi. Multiclass spectral clustering. In Proceedings Ninth IEEE International Conference on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 715, + 308, + 728 + ], + "spans": [ + { + "bbox": [ + 116, + 715, + 308, + 728 + ], + "score": 1.0, + "content": "Computer Vision, pp. 313–319 vol.1, Oct 2003.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "Chen-Yu Lee, Patrick W Gallagher, and Zhuowen Tu. Generalizing pooling functions in convo-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 92, + 505, + 108 + ], + "score": 1.0, + "content": "lutional neural networks: Mixed, gated, and tree. In Artificial Intelligence and Statistics, pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 105, + 181, + 115 + ], + "spans": [ + { + "bbox": [ + 115, + 105, + 181, + 115 + ], + "score": 1.0, + "content": "464–472, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 81, + 505, + 115 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 123, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Junhyun Lee, Inyeop Lee, and Jaewoo Kang. Self-attention graph pooling. In Proceedings of the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 397, + 147 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 397, + 147 + ], + "score": 1.0, + "content": "36th International conference on Machine learning (ICML-19), 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 106, + 123, + 505, + 147 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 152, + 504, + 176 + ], + "lines": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 153, + 505, + 165 + ], + "score": 1.0, + "content": "Yujia Li, Daniel Tarlow, Marc Brockschmidt, and Richard Zemel. Gated graph sequence neural", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 436, + 176 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 436, + 176 + ], + "score": 1.0, + "content": "networks. International Conference of Learning Representations (ICLR), 2016.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 153, + 505, + 176 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 182, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 506, + 196 + ], + "score": 1.0, + "content": "Enxhell Luzhnica, Ben Day, and Pietro Lio. Clique pooling for graph classification. International", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 193, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 193, + 505, + 207 + ], + "score": 1.0, + "content": "Conference of Learning Representations (ICLR) – Representation Learning on Graphs and Man-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 210, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 210, + 217 + ], + "score": 1.0, + "content": "ifolds workshop, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 182, + 506, + 217 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 223, + 504, + 247 + ], + "lines": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "Azade Nazi, Will Hang, Anna Goldie, Sujith Ravi, and Azalia Mirhoseini. Gap: Generalizable", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 235, + 455, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 235, + 455, + 247 + ], + "score": 1.0, + "content": "approximate graph partitioning framework. arXiv preprint arXiv:1903.00614, 2019.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 223, + 505, + 247 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 253, + 505, + 276 + ], + "lines": [ + { + "bbox": [ + 106, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "Oliver Nelles. Nonlinear system identification: from classical approaches to neural networks and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 265, + 347, + 277 + ], + "spans": [ + { + "bbox": [ + 114, + 265, + 347, + 277 + ], + "score": 1.0, + "content": "fuzzy models. Springer Science & Business Media, 2013.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 253, + 506, + 277 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 283, + 467, + 296 + ], + "lines": [ + { + "bbox": [ + 106, + 283, + 466, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 283, + 466, + 296 + ], + "score": 1.0, + "content": "Mark EJ Newman. Mixing patterns in networks. Physical Review E, 67(2):026126, 2003.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 106, + 283, + 466, + 296 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 302, + 502, + 325 + ], + "lines": [ + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 505, + 315 + ], + "score": 1.0, + "content": "Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 117, + 314, + 495, + 326 + ], + "spans": [ + { + "bbox": [ + 117, + 314, + 495, + 326 + ], + "score": 1.0, + "content": "The graph neural network model. IEEE Transactions on Neural Networks, 20(1):61–80, 2009.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 301, + 505, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 331, + 503, + 355 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "Uri Shaham, Kelly Stanton, Henry Li, Boaz Nadler, Ronen Basri, and Yuval Kluger. Spectralnet:", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 343, + 470, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 343, + 470, + 356 + ], + "score": 1.0, + "content": "Spectral clustering using deep neural networks. arXiv preprint arXiv:1801.01587, 2018.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 331, + 505, + 356 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 361, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 360, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 360, + 505, + 375 + ], + "score": 1.0, + "content": "Nino Shervashidze, Pascal Schweitzer, Erik Jan van Leeuwen, Kurt Mehlhorn, and Karsten M Borg-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 372, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 115, + 372, + 506, + 385 + ], + "score": 1.0, + "content": "wardt. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12(Sep):2539–", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 383, + 167, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 167, + 396 + ], + "score": 1.0, + "content": "2561, 2011.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 360, + 506, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 402, + 503, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 505, + 417 + ], + "score": 1.0, + "content": "Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. Departmental Papers", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 414, + 203, + 426 + ], + "spans": [ + { + "bbox": [ + 115, + 414, + 203, + 426 + ], + "score": 1.0, + "content": "(CIS), pp. 107, 2000.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 400, + 505, + 426 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 502, + 456 + ], + "lines": [ + { + "bbox": [ + 106, + 432, + 504, + 445 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 504, + 445 + ], + "score": 1.0, + "content": "David I Shuman, Mohammad Javad Faraji, and Pierre Vandergheynst. A multiscale pyramid trans-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 444, + 477, + 456 + ], + "spans": [ + { + "bbox": [ + 115, + 444, + 477, + 456 + ], + "score": 1.0, + "content": "form for graph signals. IEEE Transactions on Signal Processing, 64(8):2119–2134, 2016.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 106, + 432, + 504, + 456 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 462, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 475 + ], + "score": 1.0, + "content": "Martin Simonovsky and Nikos Komodakis. Dynamic edgeconditioned filters in convolutional neural", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 115, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "networks on graphs. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 484, + 195, + 497 + ], + "spans": [ + { + "bbox": [ + 116, + 484, + 195, + 497 + ], + "score": 1.0, + "content": "Recognition, 2017.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 462, + 505, + 497 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 503, + 504, + 527 + ], + "lines": [ + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 506, + 517 + ], + "score": 1.0, + "content": "Fei Tian, Bin Gao, Qing Cui, Enhong Chen, and Tie-Yan Liu. Learning deep representations for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 515, + 315, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 515, + 315, + 526 + ], + "score": 1.0, + "content": "graph clustering. In AAAI, pp. 1293–1299, 2014.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 502, + 506, + 526 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 533, + 504, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 533, + 504, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 504, + 546 + ], + "score": 1.0, + "content": "Alain Tremeau and Philippe Colantoni. Regions adjacency graph applied to color image segmenta-´", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 544, + 387, + 556 + ], + "spans": [ + { + "bbox": [ + 115, + 544, + 387, + 556 + ], + "score": 1.0, + "content": "tion. IEEE Transactions on image processing, 9(4):735–744, 2000.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 533, + 504, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 563, + 504, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 576 + ], + "score": 1.0, + "content": "Ulrike Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395–416,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 573, + 144, + 587 + ], + "spans": [ + { + "bbox": [ + 114, + 573, + 144, + 587 + ], + "score": 1.0, + "content": "2007.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 561, + 505, + 587 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 504, + 616 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 504, + 606 + ], + "score": 1.0, + "content": "Zaiwen Wen and Wotao Yin. A feasible method for optimization with orthogonality constraints.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 604, + 337, + 616 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 337, + 616 + ], + "score": 1.0, + "content": "Mathematical Programming, 142(1-2):397–434, 2013.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 592, + 504, + 616 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 622, + 505, + 657 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Li Yi, Hao Su, Xingwen Guo, and Leonidas J Guibas. Syncspeccnn: Synchronized spectral cnn for", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 633, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 115, + 633, + 506, + 647 + ], + "score": 1.0, + "content": "3d shape segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 645, + 259, + 657 + ], + "spans": [ + { + "bbox": [ + 115, + 645, + 259, + 657 + ], + "score": 1.0, + "content": "Recognition, pp. 2282–2290, 2017.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 622, + 506, + 657 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 663, + 504, + 698 + ], + "lines": [ + { + "bbox": [ + 106, + 663, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 677 + ], + "score": 1.0, + "content": "Zhitao Ying, Jiaxuan You, Christopher Morris, Xiang Ren, Will Hamilton, and Jure Leskovec. Hi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 674, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 116, + 674, + 505, + 688 + ], + "score": 1.0, + "content": "erarchical graph representation learning with differentiable pooling. In Advances in Neural Infor-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 686, + 319, + 698 + ], + "spans": [ + { + "bbox": [ + 116, + 686, + 319, + 698 + ], + "score": 1.0, + "content": "mation Processing Systems, pp. 4800–4810, 2018.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 663, + 505, + 698 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 704, + 504, + 727 + ], + "lines": [ + { + "bbox": [ + 106, + 703, + 505, + 718 + ], + "spans": [ + { + "bbox": [ + 106, + 703, + 505, + 718 + ], + "score": 1.0, + "content": "Yu and Shi. Multiclass spectral clustering. In Proceedings Ninth IEEE International Conference on", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 715, + 308, + 728 + ], + "spans": [ + { + "bbox": [ + 116, + 715, + 308, + 728 + ], + "score": 1.0, + "content": "Computer Vision, pp. 313–319 vol.1, Oct 2003.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 106, + 703, + 505, + 728 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 161, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 109, + 109, + 271, + 122 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 271, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 271, + 123 + ], + "score": 1.0, + "content": "A ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 134, + 225, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 134, + 227, + 147 + ], + "spans": [ + { + "bbox": [ + 106, + 134, + 227, + 147 + ], + "score": 1.0, + "content": "A.1 GNN AUTOENCODER", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 155, + 505, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "To compare the amount of information retained by the pooling layers in the coarsened graphs, we", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 357, + 179 + ], + "score": 1.0, + "content": "train an autoencoder (AE) to reconstruct a input graph signal", + "type": "text" + }, + { + "bbox": [ + 357, + 167, + 367, + 177 + ], + "score": 0.47, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "from its pooled version. The AE", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "architecture is MP(32)-MP(32)-pool-unpool-MP(32)-MP(32)-MP, and is trained by minimizing the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 188, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 415, + 201 + ], + "score": 1.0, + "content": "mean squared error between the original and the reconstructed graph signal,", + "type": "text" + }, + { + "bbox": [ + 416, + 188, + 469, + 201 + ], + "score": 0.93, + "content": "\\lVert \\bf { X } - \\bf { X } ^ { \\mathrm { { r e c } } } \\rVert ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 188, + 506, + 201 + ], + "score": 1.0, + "content": ". All the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 200, + 387, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 279, + 213 + ], + "score": 1.0, + "content": "pooling operations are configured to retain", + "type": "text" + }, + { + "bbox": [ + 279, + 200, + 298, + 210 + ], + "score": 0.88, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 200, + 387, + 213 + ], + "score": 1.0, + "content": "of the original nodes.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "In Diffpool and minCUTpool, the unpool step is simply implemented by transposing the original", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 228, + 184, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 184, + 240 + ], + "score": 1.0, + "content": "pooling operations", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 239, + 379, + 253 + ], + "lines": [ + { + "bbox": [ + 232, + 239, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 232, + 239, + 379, + 253 + ], + "score": 0.9, + "content": "{ \\bf X } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf X } ^ { \\mathrm { p o o l } } ; ~ { \\bf A } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf A } ^ { \\mathrm { p o o l } } { \\bf S } ^ { T } .", + "type": "interline_equation", + "image_path": "8bbce8dd36526a6342dc4574ef8b5dab6566229c68e0a0326d22de1f3f93de0b.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 232, + 239, + 379, + 253 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 264, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 504, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 125, + 278 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + }, + { + "bbox": [ + 126, + 265, + 136, + 275 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 262, + 443, + 278 + ], + "score": 1.0, + "content": "does not generate a cluster assignment matrix, but returns a binary mask", + "type": "text" + }, + { + "bbox": [ + 443, + 264, + 504, + 276 + ], + "score": 0.93, + "content": "\\mathbf { m } = \\{ 0 , 1 \\} ^ { N }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 450, + 289 + ], + "score": 1.0, + "content": "that indicates the nodes to drop (0) or to retain (1). Therefore, an upsamplig matrix", + "type": "text" + }, + { + "bbox": [ + 450, + 276, + 460, + 286 + ], + "score": 0.47, + "content": "\\mathbf { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "is built by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 285, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 283, + 301 + ], + "score": 1.0, + "content": "dropping the columns of the identity matrix", + "type": "text" + }, + { + "bbox": [ + 283, + 287, + 297, + 298 + ], + "score": 0.88, + "content": "{ \\mathbf { I } } _ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 285, + 397, + 301 + ], + "score": 1.0, + "content": "that correspond to a 0 in", + "type": "text" + }, + { + "bbox": [ + 397, + 288, + 408, + 297 + ], + "score": 0.47, + "content": "\\mathbf { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 285, + 412, + 301 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 412, + 286, + 483, + 299 + ], + "score": 0.9, + "content": "\\bar { \\mathbf { U } } = [ \\mathbf { I } _ { N } ] _ { : , \\mathbf { m } = = 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 285, + 506, + 301 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 325, + 311 + ], + "score": 1.0, + "content": "unpooling operation is performed by replacing S with", + "type": "text" + }, + { + "bbox": [ + 325, + 298, + 335, + 308 + ], + "score": 0.38, + "content": "\\mathbf { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "in (9), and the resulting upscaled graph is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 308, + 441, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 441, + 322 + ], + "score": 1.0, + "content": "a version of the original graph with zeroes in correspondence of the dropped nodes.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 123, + 337, + 487, + 440 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 337, + 487, + 440 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 337, + 487, + 440 + ], + "spans": [ + { + "bbox": [ + 123, + 337, + 487, + 440 + ], + "score": 0.971, + "type": "image", + "image_path": "441bd5ff28bf0dbad772497856fe6d0256f85ad6ae1cfa337d7be79956b30a09.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 123, + 337, + 487, + 371.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 123, + 371.3333333333333, + 487, + 405.66666666666663 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 123, + 405.66666666666663, + 487, + 439.99999999999994 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 218, + 452, + 393, + 464 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 216, + 450, + 394, + 466 + ], + "spans": [ + { + "bbox": [ + 216, + 450, + 394, + 466 + ], + "score": 1.0, + "content": "Figure 6: AE reconstruction of a ring graph", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 18.0 + }, + { + "type": "image", + "bbox": [ + 124, + 487, + 487, + 590 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 487, + 487, + 590 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 487, + 487, + 590 + ], + "spans": [ + { + "bbox": [ + 124, + 487, + 487, + 590 + ], + "score": 0.973, + "type": "image", + "image_path": "4b22636cf4edd79e814a6d27c3e71aeabc0f8293196e377e997385878fcd0602.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 124, + 487, + 487, + 521.3333333333334 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 124, + 521.3333333333334, + 487, + 555.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 124, + 555.6666666666667, + 487, + 590.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 218, + 603, + 393, + 615 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 216, + 600, + 394, + 617 + ], + "spans": [ + { + "bbox": [ + 216, + 600, + 394, + 617 + ], + "score": 1.0, + "content": "Figure 7: AE reconstruction of a grid graph", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + } + ], + "index": 22.0 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 290, + 640 + ], + "score": 1.0, + "content": "Fig. 6 and 7 report the original graph signal", + "type": "text" + }, + { + "bbox": [ + 290, + 627, + 301, + 637 + ], + "score": 0.27, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "(the node features are the 2-D coordinates of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 636, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 226, + 651 + ], + "score": 1.0, + "content": "nodes) and the reconstruction", + "type": "text" + }, + { + "bbox": [ + 226, + 638, + 245, + 648 + ], + "score": 0.88, + "content": "\\mathbf { X } ^ { \\mathrm { r e c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 636, + 505, + 651 + ], + "score": 1.0, + "content": "obtained by using the different pooling methods, for a ring graph", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "and a regular grid graph. The reconstruction produced by Diffpool is worse for the ring graph,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "but is almost perfect for the grid graph, while minCUTpool yields good results in both cases. On", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 188, + 684 + ], + "score": 1.0, + "content": "the other hand, Top-", + "type": "text" + }, + { + "bbox": [ + 188, + 671, + 199, + 681 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "clearly fails in generating a coarsened representation that maintains enough", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 681, + 253, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 253, + 694 + ], + "score": 1.0, + "content": "information from the original graph.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 304, + 712 + ], + "score": 1.0, + "content": "This experiment highlights a major issue in Top-", + "type": "text" + }, + { + "bbox": [ + 304, + 700, + 314, + 709 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "pooling, which retains the nodes associated to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 154, + 723 + ], + "score": 1.0, + "content": "the highest", + "type": "text" + }, + { + "bbox": [ + 154, + 710, + 165, + 720 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "values of a score vector s, computed by projecting the node features onto a train-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 165, + 732 + ], + "score": 1.0, + "content": "able vector p:", + "type": "text" + }, + { + "bbox": [ + 166, + 721, + 201, + 732 + ], + "score": 0.88, + "content": "\\mathbf { s } = \\mathbf { X } \\mathbf { p } ", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 721, + 505, + 732 + ], + "score": 1.0, + "content": ". Nodes that are connected on the graph usually share similar features, and", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 161, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 109, + 109, + 271, + 122 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 271, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 271, + 123 + ], + "score": 1.0, + "content": "A ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 110, + 271, + 123 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 134, + 225, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 134, + 227, + 147 + ], + "spans": [ + { + "bbox": [ + 106, + 134, + 227, + 147 + ], + "score": 1.0, + "content": "A.1 GNN AUTOENCODER", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 155, + 505, + 212 + ], + "lines": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 505, + 168 + ], + "score": 1.0, + "content": "To compare the amount of information retained by the pooling layers in the coarsened graphs, we", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 166, + 506, + 179 + ], + "spans": [ + { + "bbox": [ + 105, + 166, + 357, + 179 + ], + "score": 1.0, + "content": "train an autoencoder (AE) to reconstruct a input graph signal", + "type": "text" + }, + { + "bbox": [ + 357, + 167, + 367, + 177 + ], + "score": 0.47, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 166, + 506, + 179 + ], + "score": 1.0, + "content": "from its pooled version. The AE", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 191 + ], + "score": 1.0, + "content": "architecture is MP(32)-MP(32)-pool-unpool-MP(32)-MP(32)-MP, and is trained by minimizing the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 188, + 506, + 201 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 415, + 201 + ], + "score": 1.0, + "content": "mean squared error between the original and the reconstructed graph signal,", + "type": "text" + }, + { + "bbox": [ + 416, + 188, + 469, + 201 + ], + "score": 0.93, + "content": "\\lVert \\bf { X } - \\bf { X } ^ { \\mathrm { { r e c } } } \\rVert ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 188, + 506, + 201 + ], + "score": 1.0, + "content": ". All the", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 200, + 387, + 213 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 279, + 213 + ], + "score": 1.0, + "content": "pooling operations are configured to retain", + "type": "text" + }, + { + "bbox": [ + 279, + 200, + 298, + 210 + ], + "score": 0.88, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 200, + 387, + 213 + ], + "score": 1.0, + "content": "of the original nodes.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 156, + 506, + 213 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 216, + 504, + 239 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 230 + ], + "score": 1.0, + "content": "In Diffpool and minCUTpool, the unpool step is simply implemented by transposing the original", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 228, + 184, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 184, + 240 + ], + "score": 1.0, + "content": "pooling operations", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 216, + 505, + 240 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 232, + 239, + 379, + 253 + ], + "lines": [ + { + "bbox": [ + 232, + 239, + 379, + 253 + ], + "spans": [ + { + "bbox": [ + 232, + 239, + 379, + 253 + ], + "score": 0.9, + "content": "{ \\bf X } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf X } ^ { \\mathrm { p o o l } } ; ~ { \\bf A } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf A } ^ { \\mathrm { p o o l } } { \\bf S } ^ { T } .", + "type": "interline_equation", + "image_path": "8bbce8dd36526a6342dc4574ef8b5dab6566229c68e0a0326d22de1f3f93de0b.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 232, + 239, + 379, + 253 + ], + "spans": [], + "index": 10 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 264, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 504, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 125, + 278 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + }, + { + "bbox": [ + 126, + 265, + 136, + 275 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 262, + 443, + 278 + ], + "score": 1.0, + "content": "does not generate a cluster assignment matrix, but returns a binary mask", + "type": "text" + }, + { + "bbox": [ + 443, + 264, + 504, + 276 + ], + "score": 0.93, + "content": "\\mathbf { m } = \\{ 0 , 1 \\} ^ { N }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 450, + 289 + ], + "score": 1.0, + "content": "that indicates the nodes to drop (0) or to retain (1). Therefore, an upsamplig matrix", + "type": "text" + }, + { + "bbox": [ + 450, + 276, + 460, + 286 + ], + "score": 0.47, + "content": "\\mathbf { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "is built by", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 285, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 283, + 301 + ], + "score": 1.0, + "content": "dropping the columns of the identity matrix", + "type": "text" + }, + { + "bbox": [ + 283, + 287, + 297, + 298 + ], + "score": 0.88, + "content": "{ \\mathbf { I } } _ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 285, + 397, + 301 + ], + "score": 1.0, + "content": "that correspond to a 0 in", + "type": "text" + }, + { + "bbox": [ + 397, + 288, + 408, + 297 + ], + "score": 0.47, + "content": "\\mathbf { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 285, + 412, + 301 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 412, + 286, + 483, + 299 + ], + "score": 0.9, + "content": "\\bar { \\mathbf { U } } = [ \\mathbf { I } _ { N } ] _ { : , \\mathbf { m } = = 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 285, + 506, + 301 + ], + "score": 1.0, + "content": ". The", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 325, + 311 + ], + "score": 1.0, + "content": "unpooling operation is performed by replacing S with", + "type": "text" + }, + { + "bbox": [ + 325, + 298, + 335, + 308 + ], + "score": 0.38, + "content": "\\mathbf { U }", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 297, + 506, + 311 + ], + "score": 1.0, + "content": "in (9), and the resulting upscaled graph is", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 308, + 441, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 441, + 322 + ], + "score": 1.0, + "content": "a version of the original graph with zeroes in correspondence of the dropped nodes.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 262, + 506, + 322 + ] + }, + { + "type": "image", + "bbox": [ + 123, + 337, + 487, + 440 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 337, + 487, + 440 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 337, + 487, + 440 + ], + "spans": [ + { + "bbox": [ + 123, + 337, + 487, + 440 + ], + "score": 0.971, + "type": "image", + "image_path": "441bd5ff28bf0dbad772497856fe6d0256f85ad6ae1cfa337d7be79956b30a09.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 123, + 337, + 487, + 371.3333333333333 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 123, + 371.3333333333333, + 487, + 405.66666666666663 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 123, + 405.66666666666663, + 487, + 439.99999999999994 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 218, + 452, + 393, + 464 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 216, + 450, + 394, + 466 + ], + "spans": [ + { + "bbox": [ + 216, + 450, + 394, + 466 + ], + "score": 1.0, + "content": "Figure 6: AE reconstruction of a ring graph", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 18.0 + }, + { + "type": "image", + "bbox": [ + 124, + 487, + 487, + 590 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 487, + 487, + 590 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 487, + 487, + 590 + ], + "spans": [ + { + "bbox": [ + 124, + 487, + 487, + 590 + ], + "score": 0.973, + "type": "image", + "image_path": "4b22636cf4edd79e814a6d27c3e71aeabc0f8293196e377e997385878fcd0602.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 124, + 487, + 487, + 521.3333333333334 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 124, + 521.3333333333334, + 487, + 555.6666666666667 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 124, + 555.6666666666667, + 487, + 590.0000000000001 + ], + "spans": [], + "index": 22 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 218, + 603, + 393, + 615 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 216, + 600, + 394, + 617 + ], + "spans": [ + { + "bbox": [ + 216, + 600, + 394, + 617 + ], + "score": 1.0, + "content": "Figure 7: AE reconstruction of a grid graph", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + } + ], + "index": 22.0 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 290, + 640 + ], + "score": 1.0, + "content": "Fig. 6 and 7 report the original graph signal", + "type": "text" + }, + { + "bbox": [ + 290, + 627, + 301, + 637 + ], + "score": 0.27, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "(the node features are the 2-D coordinates of the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 636, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 104, + 636, + 226, + 651 + ], + "score": 1.0, + "content": "nodes) and the reconstruction", + "type": "text" + }, + { + "bbox": [ + 226, + 638, + 245, + 648 + ], + "score": 0.88, + "content": "\\mathbf { X } ^ { \\mathrm { r e c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 636, + 505, + 651 + ], + "score": 1.0, + "content": "obtained by using the different pooling methods, for a ring graph", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "and a regular grid graph. The reconstruction produced by Diffpool is worse for the ring graph,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "but is almost perfect for the grid graph, while minCUTpool yields good results in both cases. On", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 188, + 684 + ], + "score": 1.0, + "content": "the other hand, Top-", + "type": "text" + }, + { + "bbox": [ + 188, + 671, + 199, + 681 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "clearly fails in generating a coarsened representation that maintains enough", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 681, + 253, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 253, + 694 + ], + "score": 1.0, + "content": "information from the original graph.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 626, + 506, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 304, + 712 + ], + "score": 1.0, + "content": "This experiment highlights a major issue in Top-", + "type": "text" + }, + { + "bbox": [ + 304, + 700, + 314, + 709 + ], + "score": 0.82, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "pooling, which retains the nodes associated to", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 154, + 723 + ], + "score": 1.0, + "content": "the highest", + "type": "text" + }, + { + "bbox": [ + 154, + 710, + 165, + 720 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 165, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "values of a score vector s, computed by projecting the node features onto a train-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 721, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 165, + 732 + ], + "score": 1.0, + "content": "able vector p:", + "type": "text" + }, + { + "bbox": [ + 166, + 721, + 201, + 732 + ], + "score": 0.88, + "content": "\\mathbf { s } = \\mathbf { X } \\mathbf { p } ", + "type": "inline_equation" + }, + { + "bbox": [ + 202, + 721, + 505, + 732 + ], + "score": 1.0, + "content": ". Nodes that are connected on the graph usually share similar features, and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "their similarity further increases after the MP operations, which combine the features of neighbor-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 324, + 106 + ], + "score": 1.0, + "content": "ing nodes. Retaining the nodes associated to the top", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 325, + 94, + 335, + 104 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 336, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "scores in s corresponds to keeping those", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 458, + 117 + ], + "score": 1.0, + "content": "nodes that are alike and highly connected, as it can be seen in Fig. 6-7. Therefore, Top-", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 459, + 105, + 469, + 114 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 469, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "discards", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 504, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 493, + 127 + ], + "score": 1.0, + "content": "entire portions of the graphs, which might contain important information. This explains why Top-", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 494, + 115, + 504, + 125 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "fails to recover the original graph signal when used as bottleneck for the AE, and yields the worse", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 387, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 387, + 149 + ], + "score": 1.0, + "content": "performance among all GNN methods in the graph classification task.", + "type": "text", + "cross_page": true + } + ], + "index": 5 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "their similarity further increases after the MP operations, which combine the features of neighbor-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 324, + 106 + ], + "score": 1.0, + "content": "ing nodes. Retaining the nodes associated to the top", + "type": "text" + }, + { + "bbox": [ + 325, + 94, + 335, + 104 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "scores in s corresponds to keeping those", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 458, + 117 + ], + "score": 1.0, + "content": "nodes that are alike and highly connected, as it can be seen in Fig. 6-7. Therefore, Top-", + "type": "text" + }, + { + "bbox": [ + 459, + 105, + 469, + 114 + ], + "score": 0.77, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 469, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "discards", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 504, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 493, + 127 + ], + "score": 1.0, + "content": "entire portions of the graphs, which might contain important information. This explains why Top-", + "type": "text" + }, + { + "bbox": [ + 494, + 115, + 504, + 125 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "fails to recover the original graph signal when used as bottleneck for the AE, and yields the worse", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 387, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 387, + 149 + ], + "score": 1.0, + "content": "performance among all GNN methods in the graph classification task.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 108, + 168, + 385, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 386, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 386, + 180 + ], + "score": 1.0, + "content": "A.2 GRAPH REGRESSION OF MOLECULAR PROPERTIES ON QM9", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 295, + 203 + ], + "score": 1.0, + "content": "The QM9 chemical database is a collection of", + "type": "text" + }, + { + "bbox": [ + 296, + 191, + 325, + 201 + ], + "score": 0.88, + "content": "{ \\approx } 1 3 5 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "small organic molecules, associated to con-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "tinuous labels describing several geometric, energetic, electronic, and thermodynamic properties1.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 334, + 225 + ], + "score": 1.0, + "content": "Each molecule in the dataset is represented as a graph", + "type": "text" + }, + { + "bbox": [ + 335, + 213, + 374, + 225 + ], + "score": 0.93, + "content": "\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 213, + 505, + 225 + ], + "score": 1.0, + "content": ", where atoms are associated to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "nodes, and edges represent chemical bonds. The atomic number of each atom (one-hot encoded;", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "C, N, F, O) is taken as node feature and the type of bond (one-hot encoded; single, double, triple,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "aromatic) can be used as edge attribute. In this experiment, we ignore the edge attributes in order to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 305, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 305, + 269 + ], + "score": 1.0, + "content": "use all pooling algorithms without modifications.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "The purpose of this experiment is to compare the trainable pooling methods also on a graph re-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "gression task, but it must be intended as a proof of concept. In fact, the graphs in this dataset are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "extremely small (the average number of nodes is 8) and, therefore, a pooling operation is arguably", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "not necessary. We consider a GNN with architecture MP(32)-pool-MP(32)-GlobalAvgPool-Dense,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 249, + 330 + ], + "score": 1.0, + "content": "where pool is implemented by Top-", + "type": "text" + }, + { + "bbox": [ + 249, + 318, + 259, + 327 + ], + "score": 0.72, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 317, + 506, + 330 + ], + "score": 1.0, + "content": ", Diffpool, or minCUTpool. The network is trained to predict", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "a given chemical property from the input molecular graphs. Performance is evaluated with a 10-fold", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 199, + 351 + ], + "score": 1.0, + "content": "cross-validation, using", + "type": "text" + }, + { + "bbox": [ + 199, + 339, + 218, + 350 + ], + "score": 0.88, + "content": "1 \\dot { 0 } \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "of the training set for validation in each split. The GNNs are trained for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "50 epochs, using Adam with learning rate 5e-4, batch size 32, and ReLU activations. We use the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 361, + 291, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 291, + 374 + ], + "score": 1.0, + "content": "mean squared error (MSE) as supervised loss.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "The MSE obtained on the prediction of each property for different pooling methods is reported in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "Tab. 3. As expected, the flat baseline with no pooling operation (MP(32)-MP(32)-GlobalAvgPool-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 414 + ], + "score": 1.0, + "content": "Dense) yields a lower error in most cases. Contrarily to the graph classification and the AE task,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 410, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 125, + 424 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + }, + { + "bbox": [ + 126, + 411, + 136, + 421 + ], + "score": 0.74, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 410, + 505, + 424 + ], + "score": 1.0, + "content": "achieves better results than Diffpool in average. Once again, minCUTpool significantly", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 422, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 481, + 435 + ], + "score": 1.0, + "content": "outperforms the other methods on each regression task and, in one case, also the flat baseline.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 159, + 450, + 451, + 564 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 159, + 450, + 451, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 450, + 451, + 564 + ], + "spans": [ + { + "bbox": [ + 159, + 450, + 451, + 564 + ], + "score": 0.983, + "html": "
PropertyTop-KDiffpoolminCUTpoolFlat baseline
mu0.600±0.0850.651±0.0260.538±0.0120.559±0.007
alpha0.197±0.0870.114±0.0010.078±0.0070.065±0.006
homo0.698±0.1020.712±0.0150.526±0.0210.435±0.013
lumo0.601±0.0500.646±0.0130.540±0.0050.515±0.007
gap0.630±0.0440.698±0.0040.584±0.0070.552±0.008
r20.452±0.0870.440±0.0240.261±0.0060.204±0.006
zpve0.402±0.0320.410±0.0040.328±0.0050.284±0.005
uO_atom0.308±0.0550.245±0.0060.193±0.0020.163±0.001
cv0.291±0.1180.337±0.0180.148±0.0040.127±0.002
", + "type": "table", + "image_path": "abf2e28f0c4594a1f795c78d14e892f9e76511066dff22a973092bd24eabf709.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 159, + 450, + 451, + 488.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 159, + 488.0, + 451, + 526.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 159, + 526.0, + 451, + 564.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 104, + 575, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 466, + 588 + ], + "score": 1.0, + "content": "Table 3: MSE on the graph regression task. The best results with a statistical significance of", + "type": "text" + }, + { + "bbox": [ + 466, + 576, + 504, + 587 + ], + "score": 0.86, + "content": "p < 0 . 0 5", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 477, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 477, + 598 + ], + "score": 1.0, + "content": "are highlighted: the best overall are in bold, the best among pooling methods are underlined.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 631, + 255, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 257, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 257, + 645 + ], + "score": 1.0, + "content": "B EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 105, + 658, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 672 + ], + "score": 1.0, + "content": "For the WL kernel, we used the implementation provided in the GraKeL library2. The pooling", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 668, + 379, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 379, + 682 + ], + "score": 1.0, + "content": "strategy based on Graclus, is taken from the ChebyNets repository3.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 699, + 318, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 698, + 319, + 711 + ], + "spans": [ + { + "bbox": [ + 119, + 698, + 319, + 711 + ], + "score": 1.0, + "content": "1http://quantum-machine.org/datasets/", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 708, + 308, + 722 + ], + "spans": [ + { + "bbox": [ + 118, + 708, + 308, + 722 + ], + "score": 1.0, + "content": "2https://ysig.github.io/GraKeL/dev/", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 719, + 308, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 719, + 308, + 734 + ], + "score": 1.0, + "content": "3https://github.com/mdeff/cnn_graph", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [], + "index": 2.5, + "bbox_fs": [ + 105, + 83, + 505, + 149 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 168, + 385, + 179 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 386, + 180 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 386, + 180 + ], + "score": 1.0, + "content": "A.2 GRAPH REGRESSION OF MOLECULAR PROPERTIES ON QM9", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 190, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 190, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 295, + 203 + ], + "score": 1.0, + "content": "The QM9 chemical database is a collection of", + "type": "text" + }, + { + "bbox": [ + 296, + 191, + 325, + 201 + ], + "score": 0.88, + "content": "{ \\approx } 1 3 5 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 190, + 505, + 203 + ], + "score": 1.0, + "content": "small organic molecules, associated to con-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 214 + ], + "score": 1.0, + "content": "tinuous labels describing several geometric, energetic, electronic, and thermodynamic properties1.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 213, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 334, + 225 + ], + "score": 1.0, + "content": "Each molecule in the dataset is represented as a graph", + "type": "text" + }, + { + "bbox": [ + 335, + 213, + 374, + 225 + ], + "score": 0.93, + "content": "\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 213, + 505, + 225 + ], + "score": 1.0, + "content": ", where atoms are associated to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 105, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "nodes, and edges represent chemical bonds. The atomic number of each atom (one-hot encoded;", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 248 + ], + "score": 1.0, + "content": "C, N, F, O) is taken as node feature and the type of bond (one-hot encoded; single, double, triple,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 505, + 258 + ], + "score": 1.0, + "content": "aromatic) can be used as edge attribute. In this experiment, we ignore the edge attributes in order to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 257, + 305, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 257, + 305, + 269 + ], + "score": 1.0, + "content": "use all pooling algorithms without modifications.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 190, + 505, + 269 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 273, + 505, + 372 + ], + "lines": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "The purpose of this experiment is to compare the trainable pooling methods also on a graph re-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 284, + 506, + 297 + ], + "score": 1.0, + "content": "gression task, but it must be intended as a proof of concept. In fact, the graphs in this dataset are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "extremely small (the average number of nodes is 8) and, therefore, a pooling operation is arguably", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 319 + ], + "score": 1.0, + "content": "not necessary. We consider a GNN with architecture MP(32)-pool-MP(32)-GlobalAvgPool-Dense,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 317, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 249, + 330 + ], + "score": 1.0, + "content": "where pool is implemented by Top-", + "type": "text" + }, + { + "bbox": [ + 249, + 318, + 259, + 327 + ], + "score": 0.72, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 317, + 506, + 330 + ], + "score": 1.0, + "content": ", Diffpool, or minCUTpool. The network is trained to predict", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 340 + ], + "score": 1.0, + "content": "a given chemical property from the input molecular graphs. Performance is evaluated with a 10-fold", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 339, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 199, + 351 + ], + "score": 1.0, + "content": "cross-validation, using", + "type": "text" + }, + { + "bbox": [ + 199, + 339, + 218, + 350 + ], + "score": 0.88, + "content": "1 \\dot { 0 } \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 339, + 505, + 351 + ], + "score": 1.0, + "content": "of the training set for validation in each split. The GNNs are trained for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 505, + 363 + ], + "score": 1.0, + "content": "50 epochs, using Adam with learning rate 5e-4, batch size 32, and ReLU activations. We use the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 361, + 291, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 361, + 291, + 374 + ], + "score": 1.0, + "content": "mean squared error (MSE) as supervised loss.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 274, + 506, + 374 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "spans": [ + { + "bbox": [ + 105, + 377, + 506, + 391 + ], + "score": 1.0, + "content": "The MSE obtained on the prediction of each property for different pooling methods is reported in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 388, + 506, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 506, + 402 + ], + "score": 1.0, + "content": "Tab. 3. As expected, the flat baseline with no pooling operation (MP(32)-MP(32)-GlobalAvgPool-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 399, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 506, + 414 + ], + "score": 1.0, + "content": "Dense) yields a lower error in most cases. Contrarily to the graph classification and the AE task,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 410, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 125, + 424 + ], + "score": 1.0, + "content": "Top-", + "type": "text" + }, + { + "bbox": [ + 126, + 411, + 136, + 421 + ], + "score": 0.74, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 410, + 505, + 424 + ], + "score": 1.0, + "content": "achieves better results than Diffpool in average. Once again, minCUTpool significantly", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 422, + 481, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 422, + 481, + 435 + ], + "score": 1.0, + "content": "outperforms the other methods on each regression task and, in one case, also the flat baseline.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 377, + 506, + 435 + ] + }, + { + "type": "table", + "bbox": [ + 159, + 450, + 451, + 564 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 159, + 450, + 451, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 159, + 450, + 451, + 564 + ], + "spans": [ + { + "bbox": [ + 159, + 450, + 451, + 564 + ], + "score": 0.983, + "html": "
PropertyTop-KDiffpoolminCUTpoolFlat baseline
mu0.600±0.0850.651±0.0260.538±0.0120.559±0.007
alpha0.197±0.0870.114±0.0010.078±0.0070.065±0.006
homo0.698±0.1020.712±0.0150.526±0.0210.435±0.013
lumo0.601±0.0500.646±0.0130.540±0.0050.515±0.007
gap0.630±0.0440.698±0.0040.584±0.0070.552±0.008
r20.452±0.0870.440±0.0240.261±0.0060.204±0.006
zpve0.402±0.0320.410±0.0040.328±0.0050.284±0.005
uO_atom0.308±0.0550.245±0.0060.193±0.0020.163±0.001
cv0.291±0.1180.337±0.0180.148±0.0040.127±0.002
", + "type": "table", + "image_path": "abf2e28f0c4594a1f795c78d14e892f9e76511066dff22a973092bd24eabf709.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 159, + 450, + 451, + 488.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 159, + 488.0, + 451, + 526.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 159, + 526.0, + 451, + 564.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 104, + 575, + 505, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 575, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 575, + 466, + 588 + ], + "score": 1.0, + "content": "Table 3: MSE on the graph regression task. The best results with a statistical significance of", + "type": "text" + }, + { + "bbox": [ + 466, + 576, + 504, + 587 + ], + "score": 0.86, + "content": "p < 0 . 0 5", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 586, + 477, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 477, + 598 + ], + "score": 1.0, + "content": "are highlighted: the best overall are in bold, the best among pooling methods are underlined.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 575, + 504, + 598 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 631, + 255, + 643 + ], + "lines": [ + { + "bbox": [ + 105, + 630, + 257, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 257, + 645 + ], + "score": 1.0, + "content": "B EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 105, + 658, + 504, + 681 + ], + "lines": [ + { + "bbox": [ + 105, + 657, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 672 + ], + "score": 1.0, + "content": "For the WL kernel, we used the implementation provided in the GraKeL library2. The pooling", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 668, + 379, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 379, + 682 + ], + "score": 1.0, + "content": "strategy based on Graclus, is taken from the ChebyNets repository3.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 657, + 506, + 682 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 298, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 298, + 95 + ], + "score": 1.0, + "content": "B.1 CLUSTERING ON CITATION NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "Diffpool and minCUTpool are configured with 16 hidden neurons with linear activations in the MLP", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "and MP layer, respectively used to compute the cluster assignment matrix S. The MP layer used", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 275, + 139 + ], + "score": 1.0, + "content": "to compute the propagated node features", + "type": "text" + }, + { + "bbox": [ + 276, + 125, + 296, + 136 + ], + "score": 0.84, + "content": "\\mathbf { X } ^ { ( 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 124, + 505, + 139 + ], + "score": 1.0, + "content": "uses an ELU activation in both architectures. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "score": 1.0, + "content": "learning rate for Adam is 5e-4, and the models are trained for 10000 iterations. The details of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 148, + 298, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 298, + 160 + ], + "score": 1.0, + "content": "citation networks dataset are reported in Tab. 4.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 191, + 190, + 419, + 244 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 207, + 179, + 403, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 177, + 404, + 191 + ], + "spans": [ + { + "bbox": [ + 206, + 177, + 404, + 191 + ], + "score": 1.0, + "content": "Table 4: Details of the citation networks datasets", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 191, + 190, + 419, + 244 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 191, + 190, + 419, + 244 + ], + "spans": [ + { + "bbox": [ + 191, + 190, + 419, + 244 + ], + "score": 0.976, + "html": "
DatasetNodesEdgesNode featuresNode classes
Cora2708542914337
Citeseer3327922837036
Pubmed19717886515003
", + "type": "table", + "image_path": "83643cb5b85febe056dfd089e5192e8b209287c5ac7b40081e179d6e516f65fb.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 190, + 419, + 203.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 191, + 203.5, + 419, + 217.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 191, + 217.0, + 419, + 230.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 191, + 230.5, + 419, + 244.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 107, + 265, + 237, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 239, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 239, + 278 + ], + "score": 1.0, + "content": "B.2 GRAPH CLASSIFICATION", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 285, + 506, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 303, + 299 + ], + "score": 1.0, + "content": "We train the GNN architectures with Adam, an", + "type": "text" + }, + { + "bbox": [ + 303, + 287, + 315, + 298 + ], + "score": 0.82, + "content": "\\mathrm { L } _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "penalty loss with weight 1e-4, and 16 hidden", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 296, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 130, + 310 + ], + "score": 1.0, + "content": "units", + "type": "text" + }, + { + "bbox": [ + 131, + 298, + 147, + 308 + ], + "score": 0.59, + "content": "( H )", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 296, + 505, + 310 + ], + "score": 1.0, + "content": "both in the MLP of minCUTpool and in the internal MP of Diffpool. Mutagenicity,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 256, + 321 + ], + "score": 1.0, + "content": "Proteins, DD, COLLAB, and Reddit-", + "type": "text" + }, + { + "bbox": [ + 257, + 308, + 268, + 318 + ], + "score": 0.69, + "content": "{ \\it 2 k }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "are datasets representing real-world graphs and are taken", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "from the repository of benchmark datasets for graph kernels4. Bench-easy and Bench-hard5 are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 242, + 343 + ], + "score": 1.0, + "content": "datasets where the node features", + "type": "text" + }, + { + "bbox": [ + 243, + 330, + 253, + 340 + ], + "score": 0.3, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "and the adjacency matrix A are completely uninformative if", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "considered alone. Hence, algorithms that account only for the node features or the graph structure", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "will fail to classify the graphs. Since Bench-easy and Bench-hard come with a train/validation/test", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "split, the 10-fold split is not necessary to evaluate the performance. The statistics of all the datasets", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 374, + 197, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 197, + 385 + ], + "score": 1.0, + "content": "are reported in Tab. 5.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 138, + 418, + 473, + 511 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 173, + 405, + 438, + 416 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 174, + 404, + 437, + 417 + ], + "spans": [ + { + "bbox": [ + 174, + 404, + 437, + 417 + ], + "score": 1.0, + "content": "Table 5: Summary of statistics of the graph classification datasets", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "table_body", + "bbox": [ + 138, + 418, + 473, + 511 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 138, + 418, + 473, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 473, + 511 + ], + "score": 0.983, + "html": "
Datasetsamplesclassesavg. nodesavg. edgesnode attr.node labels
Bench-easy18003147.82922.66yes
Bench-hard18003148.32572.32yes
Mutagenicity4337230.3230.77yes
Proteins1113239.0672.821no
DD11782284.32715.661yes
COLLAB5000374.492457.78no
Reddit-2K20002429.63497.751no
", + "type": "table", + "image_path": "8139375cff6db7f44f94bbb4867d33a66608ffa28a0ec65a6dcfdb33bd0f0efa.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 138, + 418, + 473, + 449.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 138, + 449.0, + 473, + 480.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 138, + 480.0, + 473, + 511.0 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 22.0 + }, + { + "type": "title", + "bbox": [ + 108, + 536, + 273, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 275, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 275, + 551 + ], + "score": 1.0, + "content": "C ARCHITECTURES SCHEMATA", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "Fig. 8 reports the schematic representation of the minCUTpool layer; Fig. 9 the GNN architecture", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "used in the clustering and segmentation tasks; Fig. 10 the GNN architecture used in the graph clas-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "sification task; Fig. 12 the GNN architecture used in the graph regression task; Fig. 11 the graph", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 594, + 335, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 335, + 606 + ], + "score": 1.0, + "content": "autoencoder used in the graph signal reconstruction task.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 700, + 486, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 698, + 487, + 713 + ], + "spans": [ + { + "bbox": [ + 118, + 698, + 487, + 713 + ], + "score": 1.0, + "content": "4https://ls11-www.cs.tu-dortmund.de/staff/morris/graphkerneldatasets", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 709, + 433, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 433, + 725 + ], + "score": 1.0, + "content": "5https://github.com/FilippoMB/Benchmark_dataset_for_graph_", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 185, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 185, + 732 + ], + "score": 1.0, + "content": "classification", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 298, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 298, + 95 + ], + "score": 1.0, + "content": "B.1 CLUSTERING ON CITATION NETWORKS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 102, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "Diffpool and minCUTpool are configured with 16 hidden neurons with linear activations in the MLP", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "and MP layer, respectively used to compute the cluster assignment matrix S. The MP layer used", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 275, + 139 + ], + "score": 1.0, + "content": "to compute the propagated node features", + "type": "text" + }, + { + "bbox": [ + 276, + 125, + 296, + 136 + ], + "score": 0.84, + "content": "\\mathbf { X } ^ { ( 1 ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 124, + 505, + 139 + ], + "score": 1.0, + "content": "uses an ELU activation in both architectures. The", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 150 + ], + "score": 1.0, + "content": "learning rate for Adam is 5e-4, and the models are trained for 10000 iterations. The details of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 148, + 298, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 148, + 298, + 160 + ], + "score": 1.0, + "content": "citation networks dataset are reported in Tab. 4.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 102, + 505, + 160 + ] + }, + { + "type": "table", + "bbox": [ + 191, + 190, + 419, + 244 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 207, + 179, + 403, + 190 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 206, + 177, + 404, + 191 + ], + "spans": [ + { + "bbox": [ + 206, + 177, + 404, + 191 + ], + "score": 1.0, + "content": "Table 4: Details of the citation networks datasets", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "table_body", + "bbox": [ + 191, + 190, + 419, + 244 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 191, + 190, + 419, + 244 + ], + "spans": [ + { + "bbox": [ + 191, + 190, + 419, + 244 + ], + "score": 0.976, + "html": "
DatasetNodesEdgesNode featuresNode classes
Cora2708542914337
Citeseer3327922837036
Pubmed19717886515003
", + "type": "table", + "image_path": "83643cb5b85febe056dfd089e5192e8b209287c5ac7b40081e179d6e516f65fb.jpg" + } + ] + } + ], + "index": 8.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 190, + 419, + 203.5 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 191, + 203.5, + 419, + 217.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 191, + 217.0, + 419, + 230.5 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 191, + 230.5, + 419, + 244.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.25 + }, + { + "type": "title", + "bbox": [ + 107, + 265, + 237, + 277 + ], + "lines": [ + { + "bbox": [ + 105, + 265, + 239, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 265, + 239, + 278 + ], + "score": 1.0, + "content": "B.2 GRAPH CLASSIFICATION", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 285, + 506, + 385 + ], + "lines": [ + { + "bbox": [ + 106, + 286, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 303, + 299 + ], + "score": 1.0, + "content": "We train the GNN architectures with Adam, an", + "type": "text" + }, + { + "bbox": [ + 303, + 287, + 315, + 298 + ], + "score": 0.82, + "content": "\\mathrm { L } _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 286, + 506, + 299 + ], + "score": 1.0, + "content": "penalty loss with weight 1e-4, and 16 hidden", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 296, + 505, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 130, + 310 + ], + "score": 1.0, + "content": "units", + "type": "text" + }, + { + "bbox": [ + 131, + 298, + 147, + 308 + ], + "score": 0.59, + "content": "( H )", + "type": "inline_equation" + }, + { + "bbox": [ + 147, + 296, + 505, + 310 + ], + "score": 1.0, + "content": "both in the MLP of minCUTpool and in the internal MP of Diffpool. Mutagenicity,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 256, + 321 + ], + "score": 1.0, + "content": "Proteins, DD, COLLAB, and Reddit-", + "type": "text" + }, + { + "bbox": [ + 257, + 308, + 268, + 318 + ], + "score": 0.69, + "content": "{ \\it 2 k }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "are datasets representing real-world graphs and are taken", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 331 + ], + "score": 1.0, + "content": "from the repository of benchmark datasets for graph kernels4. Bench-easy and Bench-hard5 are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 330, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 106, + 330, + 242, + 343 + ], + "score": 1.0, + "content": "datasets where the node features", + "type": "text" + }, + { + "bbox": [ + 243, + 330, + 253, + 340 + ], + "score": 0.3, + "content": "\\mathbf { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 330, + 506, + 343 + ], + "score": 1.0, + "content": "and the adjacency matrix A are completely uninformative if", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "considered alone. Hence, algorithms that account only for the node features or the graph structure", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "will fail to classify the graphs. Since Bench-easy and Bench-hard come with a train/validation/test", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "split, the 10-fold split is not necessary to evaluate the performance. The statistics of all the datasets", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 374, + 197, + 385 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 197, + 385 + ], + "score": 1.0, + "content": "are reported in Tab. 5.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 286, + 506, + 385 + ] + }, + { + "type": "table", + "bbox": [ + 138, + 418, + 473, + 511 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 173, + 405, + 438, + 416 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 174, + 404, + 437, + 417 + ], + "spans": [ + { + "bbox": [ + 174, + 404, + 437, + 417 + ], + "score": 1.0, + "content": "Table 5: Summary of statistics of the graph classification datasets", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "table_body", + "bbox": [ + 138, + 418, + 473, + 511 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 138, + 418, + 473, + 511 + ], + "spans": [ + { + "bbox": [ + 138, + 418, + 473, + 511 + ], + "score": 0.983, + "html": "
Datasetsamplesclassesavg. nodesavg. edgesnode attr.node labels
Bench-easy18003147.82922.66yes
Bench-hard18003148.32572.32yes
Mutagenicity4337230.3230.77yes
Proteins1113239.0672.821no
DD11782284.32715.661yes
COLLAB5000374.492457.78no
Reddit-2K20002429.63497.751no
", + "type": "table", + "image_path": "8139375cff6db7f44f94bbb4867d33a66608ffa28a0ec65a6dcfdb33bd0f0efa.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 138, + 418, + 473, + 449.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 138, + 449.0, + 473, + 480.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 138, + 480.0, + 473, + 511.0 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "index": 22.0 + }, + { + "type": "title", + "bbox": [ + 108, + 536, + 273, + 549 + ], + "lines": [ + { + "bbox": [ + 105, + 535, + 275, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 275, + 551 + ], + "score": 1.0, + "content": "C ARCHITECTURES SCHEMATA", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 560, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 573 + ], + "score": 1.0, + "content": "Fig. 8 reports the schematic representation of the minCUTpool layer; Fig. 9 the GNN architecture", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 505, + 584 + ], + "score": 1.0, + "content": "used in the clustering and segmentation tasks; Fig. 10 the GNN architecture used in the graph clas-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "sification task; Fig. 12 the GNN architecture used in the graph regression task; Fig. 11 the graph", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 594, + 335, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 335, + 606 + ], + "score": 1.0, + "content": "autoencoder used in the graph signal reconstruction task.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 561, + 505, + 606 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 183, + 82, + 428, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 183, + 82, + 428, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 183, + 82, + 428, + 187 + ], + "spans": [ + { + "bbox": [ + 183, + 82, + 428, + 187 + ], + "score": 0.969, + "type": "image", + "image_path": "d9ade9085ab54361a1aa6626d8499f13441a1f331f8a266df6a2b0894a69a6db.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 183, + 82, + 428, + 117.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 183, + 117.0, + 428, + 152.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 183, + 152.0, + 428, + 187.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 216, + 200, + 394, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 199, + 395, + 216 + ], + "spans": [ + { + "bbox": [ + 215, + 199, + 395, + 216 + ], + "score": 1.0, + "content": "Figure 8: Schema of the minCUTpool layer.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 245, + 241, + 362, + 286 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 245, + 241, + 362, + 286 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 245, + 241, + 362, + 286 + ], + "spans": [ + { + "bbox": [ + 245, + 241, + 362, + 286 + ], + "score": 0.941, + "type": "image", + "image_path": "d710cb546df4f56d39a534e37d809d53060d1d9bc2b6ee1c45fe2367b58e3ef9.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 245, + 241, + 362, + 263.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 245, + 263.5, + 362, + 286.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 203, + 300, + 407, + 312 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 201, + 297, + 409, + 315 + ], + "spans": [ + { + "bbox": [ + 201, + 297, + 409, + 315 + ], + "score": 1.0, + "content": "Figure 9: Architecture for clustering/segmentation.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 5.25 + }, + { + "type": "image", + "bbox": [ + 141, + 338, + 469, + 382 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 141, + 338, + 469, + 382 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 141, + 338, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 469, + 382 + ], + "score": 0.967, + "type": "image", + "image_path": "da255a17d1270f108c2393a35c6f5725d6c431af5f91c472b5369ec2db9913d6.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 141, + 338, + 469, + 352.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 141, + 352.6666666666667, + 469, + 367.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 141, + 367.33333333333337, + 469, + 382.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 209, + 392, + 402, + 404 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 208, + 391, + 402, + 404 + ], + "spans": [ + { + "bbox": [ + 208, + 391, + 402, + 404 + ], + "score": 1.0, + "content": "Figure 10: Architecture for graph classification.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "image", + "bbox": [ + 136, + 429, + 474, + 475 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 429, + 474, + 475 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 136, + 429, + 474, + 475 + ], + "spans": [ + { + "bbox": [ + 136, + 429, + 474, + 475 + ], + "score": 0.963, + "type": "image", + "image_path": "969e11240083f473509da7280d39c4b88009ae72ad1c88f632f5a4207bc0742d.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 136, + 429, + 474, + 444.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 136, + 444.3333333333333, + 474, + 459.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 136, + 459.66666666666663, + 474, + 474.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 217, + 484, + 394, + 497 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 216, + 484, + 395, + 497 + ], + "spans": [ + { + "bbox": [ + 216, + 484, + 395, + 497 + ], + "score": 1.0, + "content": "Figure 11: Architecture for the autoencoder.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 13.0 + }, + { + "type": "image", + "bbox": [ + 180, + 522, + 429, + 567 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 180, + 522, + 429, + 567 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 180, + 522, + 429, + 567 + ], + "spans": [ + { + "bbox": [ + 180, + 522, + 429, + 567 + ], + "score": 0.953, + "type": "image", + "image_path": "031a97e555c2460b138d4ed99d26909b098af1b3775c29da264c3560104df77d.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 180, + 522, + 429, + 537.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 180, + 537.0, + 429, + 552.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 180, + 552.0, + 429, + 567.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 216, + 577, + 395, + 589 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 213, + 575, + 397, + 591 + ], + "spans": [ + { + "bbox": [ + 213, + 575, + 397, + 591 + ], + "score": 1.0, + "content": "Figure 12: Architecture for graph regression.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 17.0 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 183, + 82, + 428, + 187 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 183, + 82, + 428, + 187 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 183, + 82, + 428, + 187 + ], + "spans": [ + { + "bbox": [ + 183, + 82, + 428, + 187 + ], + "score": 0.969, + "type": "image", + "image_path": "d9ade9085ab54361a1aa6626d8499f13441a1f331f8a266df6a2b0894a69a6db.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 183, + 82, + 428, + 117.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 183, + 117.0, + 428, + 152.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 183, + 152.0, + 428, + 187.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 216, + 200, + 394, + 213 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 199, + 395, + 216 + ], + "spans": [ + { + "bbox": [ + 215, + 199, + 395, + 216 + ], + "score": 1.0, + "content": "Figure 8: Schema of the minCUTpool layer.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 245, + 241, + 362, + 286 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 245, + 241, + 362, + 286 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 245, + 241, + 362, + 286 + ], + "spans": [ + { + "bbox": [ + 245, + 241, + 362, + 286 + ], + "score": 0.941, + "type": "image", + "image_path": "d710cb546df4f56d39a534e37d809d53060d1d9bc2b6ee1c45fe2367b58e3ef9.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 245, + 241, + 362, + 263.5 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 245, + 263.5, + 362, + 286.0 + ], + "spans": [], + "index": 5 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 203, + 300, + 407, + 312 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 201, + 297, + 409, + 315 + ], + "spans": [ + { + "bbox": [ + 201, + 297, + 409, + 315 + ], + "score": 1.0, + "content": "Figure 9: Architecture for clustering/segmentation.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + } + ], + "index": 5.25 + }, + { + "type": "image", + "bbox": [ + 141, + 338, + 469, + 382 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 141, + 338, + 469, + 382 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 141, + 338, + 469, + 382 + ], + "spans": [ + { + "bbox": [ + 141, + 338, + 469, + 382 + ], + "score": 0.967, + "type": "image", + "image_path": "da255a17d1270f108c2393a35c6f5725d6c431af5f91c472b5369ec2db9913d6.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 141, + 338, + 469, + 352.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 141, + 352.6666666666667, + 469, + 367.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 141, + 367.33333333333337, + 469, + 382.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 209, + 392, + 402, + 404 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 208, + 391, + 402, + 404 + ], + "spans": [ + { + "bbox": [ + 208, + 391, + 402, + 404 + ], + "score": 1.0, + "content": "Figure 10: Architecture for graph classification.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + }, + { + "type": "image", + "bbox": [ + 136, + 429, + 474, + 475 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 136, + 429, + 474, + 475 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 136, + 429, + 474, + 475 + ], + "spans": [ + { + "bbox": [ + 136, + 429, + 474, + 475 + ], + "score": 0.963, + "type": "image", + "image_path": "969e11240083f473509da7280d39c4b88009ae72ad1c88f632f5a4207bc0742d.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 136, + 429, + 474, + 444.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 136, + 444.3333333333333, + 474, + 459.66666666666663 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 136, + 459.66666666666663, + 474, + 474.99999999999994 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 217, + 484, + 394, + 497 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 216, + 484, + 395, + 497 + ], + "spans": [ + { + "bbox": [ + 216, + 484, + 395, + 497 + ], + "score": 1.0, + "content": "Figure 11: Architecture for the autoencoder.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 13.0 + }, + { + "type": "image", + "bbox": [ + 180, + 522, + 429, + 567 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 180, + 522, + 429, + 567 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 180, + 522, + 429, + 567 + ], + "spans": [ + { + "bbox": [ + 180, + 522, + 429, + 567 + ], + "score": 0.953, + "type": "image", + "image_path": "031a97e555c2460b138d4ed99d26909b098af1b3775c29da264c3560104df77d.jpg" + } + ] + } + ], + "index": 16, + "virtual_lines": [ + { + "bbox": [ + 180, + 522, + 429, + 537.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 180, + 537.0, + 429, + 552.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 180, + 552.0, + 429, + 567.0 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 216, + 577, + 395, + 589 + ], + "group_id": 4, + "lines": [ + { + "bbox": [ + 213, + 575, + 397, + 591 + ], + "spans": [ + { + "bbox": [ + 213, + 575, + 397, + 591 + ], + "score": 1.0, + "content": "Figure 12: Architecture for graph regression.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "index": 17.0 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/BkxfshNYwB/BkxfshNYwB_model.json b/parse/train/BkxfshNYwB/BkxfshNYwB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5f468ccc367df690f9d95c92f01da1ac4da66587 --- /dev/null +++ b/parse/train/BkxfshNYwB/BkxfshNYwB_model.json @@ -0,0 +1,18385 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 535, + 1302, + 535, + 1302, + 963, + 398, + 963 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1759, + 1403, + 1759, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1099, + 1403, + 1099, + 1403, + 1344, + 298, + 1344 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1360, + 1403, + 1360, + 1403, + 1545, + 298, + 1545 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1559, + 1402, + 1559, + 1402, + 1742, + 299, + 1742 + ], + "score": 0.977 + }, + { + "category_id": 0, + "poly": [ + 301, + 218, + 1331, + 218, + 1331, + 269, + 301, + 269 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 313, + 322, + 680, + 322, + 680, + 383, + 313, + 383 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 302, + 1028, + 572, + 1028, + 572, + 1063, + 302, + 1063 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 773, + 464, + 927, + 464, + 927, + 498, + 773, + 498 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.608 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 216.0, + 1337.0, + 216.0, + 1337.0, + 275.0, + 293.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1025.0, + 579.0, + 1025.0, + 579.0, + 1072.0, + 293.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 463.0, + 932.0, + 463.0, + 932.0, + 501.0, + 770.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 534.0, + 1305.0, + 534.0, + 1305.0, + 572.0, + 394.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 565.0, + 1307.0, + 565.0, + 1307.0, + 604.0, + 393.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 597.0, + 1306.0, + 597.0, + 1306.0, + 631.0, + 393.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 626.0, + 1307.0, + 626.0, + 1307.0, + 664.0, + 393.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 658.0, + 1304.0, + 658.0, + 1304.0, + 692.0, + 393.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 688.0, + 1305.0, + 688.0, + 1305.0, + 722.0, + 393.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 719.0, + 1305.0, + 719.0, + 1305.0, + 754.0, + 394.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 749.0, + 1305.0, + 749.0, + 1305.0, + 783.0, + 393.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 778.0, + 1308.0, + 778.0, + 1308.0, + 816.0, + 393.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 813.0, + 1307.0, + 813.0, + 1307.0, + 843.0, + 393.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 842.0, + 1306.0, + 842.0, + 1306.0, + 874.0, + 394.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 872.0, + 1306.0, + 872.0, + 1306.0, + 907.0, + 393.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 903.0, + 1305.0, + 903.0, + 1305.0, + 935.0, + 394.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 929.0, + 1201.0, + 929.0, + 1201.0, + 969.0, + 393.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1757.0, + 1404.0, + 1757.0, + 1404.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 1403.0, + 1787.0, + 1403.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1405.0, + 1818.0, + 1405.0, + 1857.0, + 293.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1405.0, + 1848.0, + 1405.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1883.0, + 1404.0, + 1883.0, + 1404.0, + 1916.0, + 296.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1908.0, + 1404.0, + 1908.0, + 1404.0, + 1951.0, + 291.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2008.0, + 296.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 940.0, + 2003.0, + 940.0, + 2036.0, + 296.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1099.0, + 1403.0, + 1099.0, + 1403.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1132.0, + 1405.0, + 1132.0, + 1405.0, + 1165.0, + 294.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1163.0, + 1405.0, + 1163.0, + 1405.0, + 1196.0, + 296.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1193.0, + 1405.0, + 1193.0, + 1405.0, + 1226.0, + 294.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1220.0, + 1407.0, + 1220.0, + 1407.0, + 1261.0, + 292.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1250.0, + 1404.0, + 1250.0, + 1404.0, + 1289.0, + 292.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1284.0, + 1405.0, + 1284.0, + 1405.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1312.0, + 400.0, + 1312.0, + 400.0, + 1346.0, + 293.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1360.0, + 1405.0, + 1360.0, + 1405.0, + 1395.0, + 296.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1391.0, + 1403.0, + 1391.0, + 1403.0, + 1427.0, + 294.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1419.0, + 1406.0, + 1419.0, + 1406.0, + 1458.0, + 292.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 1405.0, + 1452.0, + 1405.0, + 1488.0, + 294.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1514.0, + 971.0, + 1514.0, + 971.0, + 1549.0, + 294.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1560.0, + 1404.0, + 1560.0, + 1404.0, + 1594.0, + 293.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1625.0, + 294.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1619.0, + 1407.0, + 1619.0, + 1407.0, + 1657.0, + 293.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1653.0, + 1403.0, + 1653.0, + 1403.0, + 1685.0, + 295.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1406.0, + 1680.0, + 1406.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1712.0, + 731.0, + 1712.0, + 731.0, + 1744.0, + 297.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 322.0, + 560.0, + 322.0, + 560.0, + 355.0, + 315.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 352.0, + 682.0, + 352.0, + 682.0, + 385.0, + 311.0, + 385.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1589, + 1404, + 1589, + 1404, + 1776, + 297, + 1776 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1787, + 1404, + 1787, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1016, + 1405, + 1016, + 1405, + 1141, + 298, + 1141 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1406, + 1481, + 1406, + 1576, + 298, + 1576 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1265, + 1407, + 1265, + 1407, + 1359, + 299, + 1359 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 798, + 1403, + 798, + 1403, + 895, + 298, + 895 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 527, + 220, + 1170, + 220, + 1170, + 553, + 527, + 553 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 569, + 911, + 1131, + 911, + 1131, + 1004, + 569, + 1004 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 470, + 1158, + 1226, + 1158, + 1226, + 1248, + 470, + 1248 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 462, + 1378, + 1236, + 1378, + 1236, + 1467, + 462, + 1467 + ], + "score": 0.947 + }, + { + "category_id": 4, + "poly": [ + 331, + 588, + 1366, + 588, + 1366, + 623, + 331, + 623 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 299, + 673, + 558, + 673, + 558, + 710, + 299, + 710 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 299, + 743, + 814, + 743, + 814, + 776, + 299, + 776 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1366, + 941, + 1400, + 941, + 1400, + 972, + 1366, + 972 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1188, + 1400, + 1188, + 1400, + 1218, + 1366, + 1218 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1406, + 1400, + 1406, + 1400, + 1437, + 1366, + 1437 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.736 + }, + { + "category_id": 14, + "poly": [ + 567, + 908, + 1131, + 908, + 1131, + 1005, + 567, + 1005 + ], + "score": 0.94, + "latex": "\\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\operatorname* { l i n k s } ( \\mathcal { V } _ { k } ) } { \\deg \\mathrm { r e e } ( \\mathcal { V } _ { k } ) } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { \\sum _ { i , j \\in \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } { \\sum _ { i \\in \\mathcal { V } _ { k } , j \\in \\mathcal { V } \\backslash \\mathcal { V } _ { k } } \\mathcal { E } _ { i , j } } ," + }, + { + "category_id": 13, + "poly": [ + 625, + 1742, + 708, + 1742, + 708, + 1776, + 625, + 1776 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( \\bar { N } ^ { 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1244, + 1711, + 1328, + 1711, + 1328, + 1746, + 1244, + 1746 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( N ^ { 2 } )" + }, + { + "category_id": 14, + "poly": [ + 460, + 1373, + 1241, + 1373, + 1241, + 1470, + 460, + 1470 + ], + "score": 0.93, + "latex": "\\arg \\operatorname* { m a x } _ { \\mathbf { Q } \\in \\mathbb { R } ^ { N \\times K } } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { Q } _ { k } ^ { T } \\mathbf { A } \\mathbf { Q } _ { k } , \\quad \\mathrm { s . t . } \\mathbf { Q } = \\mathbf { C } ( \\mathbf { C } ^ { T } \\mathbf { D } \\mathbf { C } ) ^ { - \\frac { 1 } { 2 } } , \\ \\mathbf { Q } ^ { T } \\mathbf { Q } = \\mathbf { I } _ { K } ." + }, + { + "category_id": 13, + "poly": [ + 729, + 1079, + 832, + 1079, + 832, + 1113, + 729, + 1113 + ], + "score": 0.92, + "latex": "\\mathbf { C } _ { i , j } = 1" + }, + { + "category_id": 13, + "poly": [ + 390, + 1971, + 474, + 1971, + 474, + 2006, + 390, + 2006 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( N ^ { 3 } )" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1046, + 1399, + 1046, + 1399, + 1077, + 1263, + 1077 + ], + "score": 0.92, + "latex": "\\mathbf { C } \\in \\mathbb { R } ^ { N \\times K }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 799, + 1265, + 799, + 1265, + 830, + 1128, + 830 + ], + "score": 0.92, + "latex": "\\mathbf { A } \\in \\mathbb { R } ^ { N \\times N }" + }, + { + "category_id": 13, + "poly": [ + 495, + 1850, + 567, + 1850, + 567, + 1884, + 495, + 1884 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( N )" + }, + { + "category_id": 14, + "poly": [ + 469, + 1154, + 1228, + 1154, + 1228, + 1252, + 469, + 1252 + ], + "score": 0.92, + "latex": "\\mathrm { m a x i m i z e } \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\frac { { \\bf C } _ { k } ^ { T } { \\bf A } { \\bf C } _ { k } } { { \\bf C } _ { k } ^ { T } { \\bf D } { \\bf C } _ { k } } , \\mathrm { s . t . } { \\bf C } \\in \\{ 0 , 1 \\} ^ { N \\times K } , { \\bf C } { \\bf 1 } _ { K } = { \\bf 1 } _ { N } ," + }, + { + "category_id": 13, + "poly": [ + 298, + 1849, + 381, + 1849, + 381, + 1884, + 298, + 1884 + ], + "score": 0.92, + "latex": "\\mathcal { O } ( N ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 1248, + 1482, + 1393, + 1482, + 1393, + 1514, + 1248, + 1514 + ], + "score": 0.91, + "latex": "\\mathbf { Q } ^ { * } = \\mathbf { U } _ { K } \\mathbf { O }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1511, + 542, + 1511, + 542, + 1542, + 374, + 1542 + ], + "score": 0.91, + "latex": "{ \\bf U } _ { K } \\dot { } \\in \\mathbb { R } ^ { N \\times K }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1266, + 572, + 1266, + 572, + 1299, + 373, + 1299 + ], + "score": 0.9, + "latex": "\\mathbf { D } = \\mathrm { d i a g } ( \\mathbf { A } \\mathbf { 1 } _ { N } )" + }, + { + "category_id": 13, + "poly": [ + 345, + 1543, + 483, + 1543, + 483, + 1572, + 345, + 1572 + ], + "score": 0.9, + "latex": "\\mathbf { O } \\in \\mathbb { R } ^ { K \\times K }" + }, + { + "category_id": 13, + "poly": [ + 540, + 1591, + 579, + 1591, + 579, + 1622, + 540, + 1622 + ], + "score": 0.9, + "latex": "\\mathbf { Q } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1621, + 1275, + 1621, + 1275, + 1652, + 1235, + 1652 + ], + "score": 0.89, + "latex": "\\mathbf { Q } ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 1145, + 1515, + 1174, + 1515, + 1174, + 1541, + 1145, + 1541 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1319, + 802, + 1347, + 802, + 1347, + 829, + 1319, + 829 + ], + "score": 0.83, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 669, + 1683, + 687, + 1683, + 687, + 1709, + 669, + 1709 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1159, + 833, + 1188, + 833, + 1188, + 859, + 1159, + 859 + ], + "score": 0.82, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1137, + 1081, + 1153, + 1081, + 1153, + 1111, + 1137, + 1111 + ], + "score": 0.8, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 461, + 801, + 598, + 801, + 598, + 835, + 461, + 835 + ], + "score": 0.79, + "latex": "G = \\{ \\nu , \\mathcal { E } \\}" + }, + { + "category_id": 13, + "poly": [ + 1103, + 833, + 1127, + 833, + 1127, + 860, + 1103, + 860 + ], + "score": 0.78, + "latex": "\\nu" + }, + { + "category_id": 13, + "poly": [ + 919, + 1082, + 932, + 1082, + 932, + 1106, + 919, + 1106 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 610, + 801, + 710, + 801, + 710, + 834, + 610, + 834 + ], + "score": 0.73, + "latex": "| \\nu | = N" + }, + { + "category_id": 13, + "poly": [ + 1090, + 1882, + 1118, + 1882, + 1118, + 1913, + 1090, + 1913 + ], + "score": 0.68, + "latex": "\\mathbf { Q }" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 282.0, + 1119.0, + 282.0, + 1119.0, + 306.0, + 1094.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 299.0, + 831.0, + 299.0, + 831.0, + 311.0, + 818.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 308.0, + 1087.0, + 308.0, + 1087.0, + 320.0, + 1075.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 326.0, + 1020.0, + 326.0, + 1020.0, + 349.0, + 995.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 339.0, + 1117.0, + 339.0, + 1117.0, + 353.0, + 1106.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 357.0, + 834.0, + 357.0, + 834.0, + 375.0, + 815.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 373.0, + 1036.0, + 373.0, + 1036.0, + 386.0, + 1023.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 376.0, + 1113.0, + 376.0, + 1113.0, + 389.0, + 1102.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 403.0, + 1062.0, + 403.0, + 1062.0, + 419.0, + 1047.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 521.0, + 743.0, + 521.0, + 743.0, + 554.0, + 567.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 522.0, + 922.0, + 522.0, + 922.0, + 556.0, + 744.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 517.0, + 1131.0, + 517.0, + 1131.0, + 555.0, + 957.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.75, + 323.0, + 1050.75, + 323.0, + 1050.75, + 334.5, + 1045.75, + 334.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 384.0, + 679.0, + 384.0, + 679.0, + 401.0, + 649.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 582.0, + 1370.0, + 582.0, + 1370.0, + 632.0, + 326.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 670.0, + 561.0, + 670.0, + 561.0, + 717.0, + 291.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 743.0, + 817.0, + 743.0, + 817.0, + 779.0, + 295.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1585.0, + 539.0, + 1585.0, + 539.0, + 1629.0, + 292.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1585.0, + 1406.0, + 1585.0, + 1406.0, + 1629.0, + 580.0, + 1629.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1234.0, + 1620.0, + 1234.0, + 1656.0, + 295.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 1620.0, + 1406.0, + 1620.0, + 1406.0, + 1656.0, + 1276.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1652.0, + 1406.0, + 1652.0, + 1406.0, + 1685.0, + 292.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1680.0, + 668.0, + 1680.0, + 668.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1680.0, + 1406.0, + 1680.0, + 1406.0, + 1715.0, + 688.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1709.0, + 1243.0, + 1709.0, + 1243.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 1709.0, + 1407.0, + 1709.0, + 1407.0, + 1748.0, + 1329.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1740.0, + 624.0, + 1740.0, + 624.0, + 1780.0, + 294.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1740.0, + 1248.0, + 1740.0, + 1248.0, + 1780.0, + 709.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1787.0, + 1406.0, + 1787.0, + 1406.0, + 1826.0, + 292.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1819.0, + 1405.0, + 1819.0, + 1405.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 297.0, + 1847.0, + 297.0, + 1888.0, + 294.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1847.0, + 494.0, + 1847.0, + 494.0, + 1888.0, + 382.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 1847.0, + 1406.0, + 1847.0, + 1406.0, + 1888.0, + 568.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1878.0, + 1089.0, + 1878.0, + 1089.0, + 1918.0, + 291.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 1878.0, + 1406.0, + 1878.0, + 1406.0, + 1918.0, + 1119.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1948.0, + 295.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 389.0, + 1969.0, + 389.0, + 2010.0, + 292.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1969.0, + 1409.0, + 1969.0, + 1409.0, + 2010.0, + 475.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 604.0, + 2003.0, + 604.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1016.0, + 1403.0, + 1016.0, + 1403.0, + 1049.0, + 296.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1045.0, + 1262.0, + 1045.0, + 1262.0, + 1082.0, + 289.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 1045.0, + 1407.0, + 1045.0, + 1407.0, + 1082.0, + 1400.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1077.0, + 728.0, + 1077.0, + 728.0, + 1114.0, + 292.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1077.0, + 918.0, + 1077.0, + 918.0, + 1114.0, + 833.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1077.0, + 1136.0, + 1077.0, + 1136.0, + 1114.0, + 933.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 1077.0, + 1405.0, + 1077.0, + 1405.0, + 1114.0, + 1154.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1110.0, + 718.0, + 1110.0, + 718.0, + 1143.0, + 292.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1479.0, + 1247.0, + 1479.0, + 1247.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1479.0, + 1404.0, + 1479.0, + 1404.0, + 1519.0, + 1394.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1504.0, + 373.0, + 1504.0, + 373.0, + 1554.0, + 291.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1504.0, + 1144.0, + 1504.0, + 1144.0, + 1554.0, + 543.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 1504.0, + 1408.0, + 1504.0, + 1408.0, + 1554.0, + 1175.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1537.0, + 344.0, + 1537.0, + 344.0, + 1580.0, + 292.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1537.0, + 1068.0, + 1537.0, + 1068.0, + 1580.0, + 484.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1266.0, + 372.0, + 1266.0, + 372.0, + 1300.0, + 296.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1266.0, + 1403.0, + 1266.0, + 1403.0, + 1300.0, + 573.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1296.0, + 1407.0, + 1296.0, + 1407.0, + 1331.0, + 292.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1328.0, + 746.0, + 1328.0, + 746.0, + 1360.0, + 294.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 792.0, + 460.0, + 792.0, + 460.0, + 840.0, + 291.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 792.0, + 609.0, + 792.0, + 609.0, + 840.0, + 599.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 792.0, + 1127.0, + 792.0, + 1127.0, + 840.0, + 711.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 792.0, + 1318.0, + 792.0, + 1318.0, + 840.0, + 1266.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 792.0, + 1407.0, + 792.0, + 1407.0, + 840.0, + 1348.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 829.0, + 1102.0, + 829.0, + 1102.0, + 869.0, + 292.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 829.0, + 1158.0, + 829.0, + 1158.0, + 869.0, + 1128.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 829.0, + 1404.0, + 829.0, + 1404.0, + 869.0, + 1189.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 860.0, + 1227.0, + 860.0, + 1227.0, + 899.0, + 294.0, + 899.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1622, + 1405, + 1622, + 1405, + 2035, + 296, + 2035 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 541, + 1404, + 541, + 1404, + 819, + 297, + 819 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1404, + 228, + 1404, + 446, + 297, + 446 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1290, + 1405, + 1290, + 1405, + 1419, + 299, + 1419 + ], + "score": 0.979 + }, + { + "category_id": 8, + "poly": [ + 523, + 1435, + 1176, + 1435, + 1176, + 1557, + 523, + 1557 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 292, + 836, + 1400, + 836, + 1400, + 902, + 292, + 902 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 297, + 976, + 1402, + 976, + 1402, + 1042, + 297, + 1042 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 294, + 1156, + 1404, + 1156, + 1404, + 1222, + 294, + 1222 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 507, + 918, + 1187, + 918, + 1187, + 961, + 507, + 961 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 649, + 1239, + 1049, + 1239, + 1049, + 1276, + 649, + 1276 + ], + "score": 0.933 + }, + { + "category_id": 1, + "poly": [ + 299, + 1576, + 776, + 1576, + 776, + 1610, + 299, + 1610 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 300, + 1088, + 636, + 1088, + 636, + 1124, + 300, + 1124 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 301, + 484, + 707, + 484, + 707, + 516, + 301, + 516 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1243, + 1400, + 1243, + 1400, + 1273, + 1366, + 1273 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1462, + 1400, + 1462, + 1400, + 1492, + 1366, + 1492 + ], + "score": 0.875 + }, + { + "category_id": 9, + "poly": [ + 1366, + 927, + 1400, + 927, + 1400, + 957, + 1366, + 957 + ], + "score": 0.868 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.744 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.647 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.422 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.206 + }, + { + "category_id": 14, + "poly": [ + 521, + 1434, + 1174, + 1434, + 1174, + 1560, + 521, + 1560 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { u } = \\mathcal { L } _ { c } + \\mathcal { L } _ { o } = \\underbrace { - \\frac { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } ) } } _ { \\mathcal { L } _ { c } } + \\underbrace { \\left\\| \\frac { \\mathbf { S } ^ { T } \\mathbf { S } } { \\| \\mathbf { S } ^ { T } \\mathbf { S } \\| _ { F } } - \\frac { \\mathbf { I } _ { K } } { \\sqrt { K } } \\right\\| _ { F } } _ { \\mathcal { L } _ { o } } ," + }, + { + "category_id": 13, + "poly": [ + 373, + 977, + 617, + 977, + 617, + 1013, + 373, + 1013 + ], + "score": 0.93, + "latex": "\\boldsymbol { \\Theta } _ { M P } = \\{ \\mathbf { W } _ { m } , \\mathbf { W } _ { s } \\}" + }, + { + "category_id": 13, + "poly": [ + 1085, + 1156, + 1228, + 1156, + 1228, + 1188, + 1085, + 1188 + ], + "score": 0.92, + "latex": "\\mathbf { S } \\in \\mathbb { R } ^ { N \\times K }" + }, + { + "category_id": 13, + "poly": [ + 917, + 1326, + 1049, + 1326, + 1049, + 1356, + 917, + 1356 + ], + "score": 0.92, + "latex": "\\mathbf { S 1 } _ { K } = \\mathbf { 1 } _ { N }" + }, + { + "category_id": 13, + "poly": [ + 496, + 1686, + 577, + 1686, + 577, + 1719, + 496, + 1719 + ], + "score": 0.92, + "latex": "\\left. \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\right." + }, + { + "category_id": 13, + "poly": [ + 420, + 1785, + 734, + 1785, + 734, + 1823, + 420, + 1823 + ], + "score": 0.92, + "latex": "T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { D } } \\mathbf { S } )" + }, + { + "category_id": 13, + "poly": [ + 340, + 832, + 657, + 832, + 657, + 869, + 340, + 869 + ], + "score": 0.92, + "latex": "\\tilde { \\mathbf { A } } = \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\mathbf { A } \\mathbf { D } ^ { - \\frac { 1 } { 2 } } \\in \\mathbb { R } ^ { N \\times N }" + }, + { + "category_id": 13, + "poly": [ + 421, + 1718, + 847, + 1718, + 847, + 1759, + 421, + 1759 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { T r ( \\mathbf { S } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } ) = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\mathbf { S } _ { k } ^ { T } \\tilde { \\mathbf { A } } \\mathbf { S } _ { k } = 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 373, + 1578, + 440, + 1578, + 440, + 1611, + 373, + 1611 + ], + "score": 0.92, + "latex": "\\| \\cdot \\| _ { F }" + }, + { + "category_id": 14, + "poly": [ + 647, + 1239, + 1049, + 1239, + 1049, + 1277, + 647, + 1277 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\mathbf { S } = s o f t m a x ( \\mathbf { R e L U } ( \\mathbf { X } \\mathbf { W } _ { 1 } ) \\mathbf { W } _ { 2 } ) , } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 669, + 1357, + 739, + 1357, + 739, + 1388, + 669, + 1388 + ], + "score": 0.92, + "latex": "\\Theta _ { M P }" + }, + { + "category_id": 13, + "poly": [ + 795, + 1356, + 870, + 1356, + 870, + 1387, + 795, + 1387 + ], + "score": 0.91, + "latex": "\\Theta _ { P o o l }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1655, + 457, + 1655, + 457, + 1686, + 297, + 1686 + ], + "score": 0.91, + "latex": "- 1 \\leq \\mathcal { L } _ { c } \\leq 0" + }, + { + "category_id": 13, + "poly": [ + 470, + 1328, + 600, + 1328, + 600, + 1359, + 470, + 1359 + ], + "score": 0.91, + "latex": "s _ { i , j } \\in [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 964, + 1753, + 1101, + 1753, + 1101, + 1788, + 964, + 1788 + ], + "score": 0.91, + "latex": "\\langle \\mathbf { s } _ { i } , \\mathbf { s } _ { j } \\rangle = 0 ," + }, + { + "category_id": 13, + "poly": [ + 1265, + 836, + 1400, + 836, + 1400, + 868, + 1265, + 868 + ], + "score": 0.91, + "latex": "\\mathbf { X } \\in \\mathbb { R } ^ { N \\times F }" + }, + { + "category_id": 13, + "poly": [ + 354, + 1755, + 454, + 1755, + 454, + 1788, + 354, + 1788 + ], + "score": 0.9, + "latex": "\\tilde { a } _ { i , j } > 0 \\rangle" + }, + { + "category_id": 14, + "poly": [ + 509, + 917, + 1185, + 917, + 1185, + 962, + 509, + 962 + ], + "score": 0.9, + "latex": "\\mathbf { X } ^ { ( t + 1 ) } = M P ( \\mathbf { X } ^ { ( t ) } , \\tilde { \\mathbf { A } } ) = \\operatorname { R e L U } ( \\tilde { \\mathbf { A } } \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { m } + \\mathbf { X } ^ { ( t ) } \\mathbf { W } _ { s } ) ," + }, + { + "category_id": 13, + "poly": [ + 757, + 1687, + 801, + 1687, + 801, + 1719, + 757, + 1719 + ], + "score": 0.9, + "latex": "\\tilde { a } _ { i , j }" + }, + { + "category_id": 13, + "poly": [ + 850, + 1388, + 886, + 1388, + 886, + 1417, + 850, + 1417 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { u }" + }, + { + "category_id": 13, + "poly": [ + 608, + 1656, + 642, + 1656, + 642, + 1685, + 608, + 1685 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1112, + 1852, + 1146, + 1852, + 1146, + 1881, + 1112, + 1881 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 902, + 1686, + 936, + 1686, + 936, + 1716, + 902, + 1716 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 506, + 1626, + 539, + 1626, + 539, + 1655, + 506, + 1655 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 901, + 1195, + 931, + 1195, + 931, + 1221, + 901, + 1221 + ], + "score": 0.86, + "latex": "\\mathbf { x } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1120, + 1754, + 1155, + 1754, + 1155, + 1785, + 1120, + 1785 + ], + "score": 0.86, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1291, + 861, + 1291, + 861, + 1328, + 373, + 1328 + ], + "score": 0.83, + "latex": "\\boldsymbol { \\Theta } _ { P o o l } = \\{ \\mathbf { W } _ { 1 } \\in \\mathbb { R } ^ { F \\times H } , \\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { H \\times K } \\}" + }, + { + "category_id": 13, + "poly": [ + 1151, + 1790, + 1181, + 1790, + 1181, + 1818, + 1151, + 1818 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 298, + 1791, + 336, + 1791, + 336, + 1820, + 298, + 1820 + ], + "score": 0.79, + "latex": "- 1" + }, + { + "category_id": 13, + "poly": [ + 822, + 262, + 853, + 262, + 853, + 289, + 822, + 289 + ], + "score": 0.75, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 680, + 1291, + 861, + 1291, + 861, + 1329, + 680, + 1329 + ], + "score": 0.42, + "latex": "\\mathbf { W } _ { 2 } \\in \\mathbb { R } ^ { H \\times K } \\}" + }, + { + "category_id": 13, + "poly": [ + 1238, + 234, + 1250, + 234, + 1250, + 258, + 1238, + 258 + ], + "score": 0.31, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1084.0, + 639.0, + 1084.0, + 639.0, + 1129.0, + 291.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 481.0, + 711.0, + 481.0, + 711.0, + 521.0, + 293.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1623.0, + 505.0, + 1623.0, + 505.0, + 1660.0, + 295.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1623.0, + 1403.0, + 1623.0, + 1403.0, + 1660.0, + 540.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1653.0, + 607.0, + 1653.0, + 607.0, + 1692.0, + 458.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1653.0, + 1406.0, + 1653.0, + 1406.0, + 1692.0, + 643.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1684.0, + 495.0, + 1684.0, + 495.0, + 1722.0, + 294.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1684.0, + 756.0, + 1684.0, + 756.0, + 1722.0, + 578.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 1684.0, + 901.0, + 1684.0, + 901.0, + 1722.0, + 802.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1684.0, + 1405.0, + 1684.0, + 1405.0, + 1722.0, + 937.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 282.0, + 1713.0, + 420.0, + 1713.0, + 420.0, + 1765.0, + 282.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1713.0, + 1413.0, + 1713.0, + 1413.0, + 1765.0, + 848.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1755.0, + 353.0, + 1755.0, + 353.0, + 1789.0, + 295.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 1755.0, + 963.0, + 1755.0, + 963.0, + 1789.0, + 455.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1755.0, + 1119.0, + 1755.0, + 1119.0, + 1789.0, + 1102.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1755.0, + 1405.0, + 1755.0, + 1405.0, + 1789.0, + 1156.0, + 1789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1785.0, + 419.0, + 1785.0, + 419.0, + 1827.0, + 337.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1785.0, + 1150.0, + 1785.0, + 1150.0, + 1827.0, + 735.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1785.0, + 1405.0, + 1785.0, + 1405.0, + 1827.0, + 1182.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1820.0, + 1406.0, + 1820.0, + 1406.0, + 1856.0, + 291.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1852.0, + 1111.0, + 1852.0, + 1111.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1852.0, + 1405.0, + 1852.0, + 1405.0, + 1886.0, + 1147.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1407.0, + 1882.0, + 1407.0, + 1916.0, + 294.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1914.0, + 1405.0, + 1914.0, + 1405.0, + 1944.0, + 295.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 291.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1405.0, + 1973.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2002.0, + 1357.0, + 2002.0, + 1357.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 544.0, + 1406.0, + 544.0, + 1406.0, + 580.0, + 295.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 572.0, + 1405.0, + 572.0, + 1405.0, + 610.0, + 294.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 604.0, + 1405.0, + 604.0, + 1405.0, + 641.0, + 295.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 635.0, + 1406.0, + 635.0, + 1406.0, + 671.0, + 292.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 667.0, + 1405.0, + 667.0, + 1405.0, + 700.0, + 296.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 696.0, + 1405.0, + 696.0, + 1405.0, + 733.0, + 295.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 722.0, + 1409.0, + 722.0, + 1409.0, + 765.0, + 294.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 753.0, + 1406.0, + 753.0, + 1406.0, + 793.0, + 292.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 788.0, + 637.0, + 788.0, + 637.0, + 823.0, + 295.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1237.0, + 228.0, + 1237.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 265.0, + 1251.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 260.0, + 821.0, + 260.0, + 821.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 296.0, + 854.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 1404.0, + 290.0, + 1404.0, + 326.0, + 294.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 320.0, + 1404.0, + 320.0, + 1404.0, + 357.0, + 294.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 387.0, + 295.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 418.0, + 292.0, + 418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 413.0, + 1405.0, + 413.0, + 1405.0, + 451.0, + 295.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1288.0, + 372.0, + 1288.0, + 372.0, + 1332.0, + 290.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1288.0, + 1408.0, + 1288.0, + 1408.0, + 1332.0, + 862.0, + 1332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1325.0, + 469.0, + 1325.0, + 469.0, + 1362.0, + 294.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1325.0, + 916.0, + 1325.0, + 916.0, + 1362.0, + 601.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1325.0, + 1406.0, + 1325.0, + 1406.0, + 1362.0, + 1050.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1354.0, + 668.0, + 1354.0, + 668.0, + 1391.0, + 294.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1354.0, + 794.0, + 1354.0, + 794.0, + 1391.0, + 740.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1354.0, + 1405.0, + 1354.0, + 1405.0, + 1391.0, + 871.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1386.0, + 849.0, + 1386.0, + 849.0, + 1422.0, + 295.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1386.0, + 1264.0, + 1386.0, + 1264.0, + 1422.0, + 887.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 828.0, + 339.0, + 828.0, + 339.0, + 878.0, + 289.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 828.0, + 1264.0, + 828.0, + 1264.0, + 878.0, + 658.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 828.0, + 1408.0, + 828.0, + 1408.0, + 878.0, + 1401.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 868.0, + 1080.0, + 868.0, + 1080.0, + 904.0, + 295.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 973.0, + 372.0, + 973.0, + 372.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 973.0, + 1406.0, + 973.0, + 1406.0, + 1013.0, + 618.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1006.0, + 552.0, + 1006.0, + 552.0, + 1047.0, + 294.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1149.0, + 1084.0, + 1149.0, + 1084.0, + 1197.0, + 292.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 1149.0, + 1409.0, + 1149.0, + 1409.0, + 1197.0, + 1229.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1190.0, + 900.0, + 1190.0, + 900.0, + 1223.0, + 294.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1190.0, + 1175.0, + 1190.0, + 1175.0, + 1223.0, + 932.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1575.0, + 372.0, + 1575.0, + 372.0, + 1612.0, + 297.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1575.0, + 780.0, + 1575.0, + 780.0, + 1612.0, + 441.0, + 1612.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1038, + 1404, + 1038, + 1404, + 1351, + 297, + 1351 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1618, + 1403, + 1618, + 1403, + 1804, + 298, + 1804 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1403, + 1819, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1404, + 228, + 1404, + 450, + 297, + 450 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 463, + 1404, + 463, + 1404, + 588, + 300, + 588 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 361, + 611, + 1405, + 611, + 1405, + 851, + 361, + 851 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 299, + 1542, + 1403, + 1542, + 1403, + 1606, + 299, + 1606 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 656, + 984, + 1045, + 984, + 1045, + 1025, + 656, + 1025 + ], + "score": 0.941 + }, + { + "category_id": 1, + "poly": [ + 299, + 940, + 1236, + 940, + 1236, + 975, + 299, + 975 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 302, + 1417, + 771, + 1417, + 771, + 1450, + 302, + 1450 + ], + "score": 0.922 + }, + { + "category_id": 8, + "poly": [ + 535, + 1363, + 1163, + 1363, + 1163, + 1405, + 535, + 1405 + ], + "score": 0.922 + }, + { + "category_id": 0, + "poly": [ + 298, + 884, + 528, + 884, + 528, + 917, + 298, + 917 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1366, + 993, + 1400, + 993, + 1400, + 1023, + 1366, + 1023 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1373, + 1400, + 1373, + 1400, + 1402, + 1366, + 1402 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.866 + }, + { + "category_id": 0, + "poly": [ + 300, + 1487, + 1112, + 1487, + 1112, + 1518, + 300, + 1518 + ], + "score": 0.846 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.76 + }, + { + "category_id": 13, + "poly": [ + 610, + 351, + 741, + 351, + 741, + 388, + 610, + 388 + ], + "score": 0.93, + "latex": "\\mathbf { I } _ { K } = \\hat { \\mathbf { S } } ^ { T } \\hat { \\mathbf { S } }" + }, + { + "category_id": 13, + "poly": [ + 1193, + 1078, + 1252, + 1078, + 1252, + 1121, + 1193, + 1121 + ], + "score": 0.93, + "latex": "a _ { i , i } ^ { p o o l }" + }, + { + "category_id": 13, + "poly": [ + 969, + 1118, + 1029, + 1118, + 1029, + 1163, + 969, + 1163 + ], + "score": 0.92, + "latex": "a _ { i , j } ^ { p o o l }" + }, + { + "category_id": 13, + "poly": [ + 486, + 1037, + 547, + 1037, + 547, + 1081, + 486, + 1081 + ], + "score": 0.92, + "latex": "x _ { i , j } ^ { p o o l }" + }, + { + "category_id": 13, + "poly": [ + 586, + 1040, + 777, + 1040, + 777, + 1073, + 586, + 1073 + ], + "score": 0.92, + "latex": "\\mathbf { X } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times F }" + }, + { + "category_id": 13, + "poly": [ + 735, + 292, + 875, + 292, + 875, + 322, + 735, + 322 + ], + "score": 0.92, + "latex": "0 \\leq \\mathcal { L } _ { o } \\leq 2" + }, + { + "category_id": 14, + "poly": [ + 652, + 984, + 1043, + 984, + 1043, + 1025, + 652, + 1025 + ], + "score": 0.91, + "latex": "{ \\bf A } ^ { p o o l } = { \\bf S } ^ { T } { \\tilde { \\bf A } } { \\bf S } ; ~ { \\bf X } ^ { p o o l } = { \\bf S } ^ { T } { \\bf X } ," + }, + { + "category_id": 13, + "poly": [ + 1030, + 356, + 1096, + 356, + 1096, + 390, + 1030, + 390 + ], + "score": 0.91, + "latex": "N / K" + }, + { + "category_id": 14, + "poly": [ + 533, + 1361, + 1166, + 1361, + 1166, + 1406, + 533, + 1406 + ], + "score": 0.9, + "latex": "\\hat { \\bf A } = { \\bf A } ^ { p o o l } - { \\bf I } _ { K } d i a g ( { \\bf A } ^ { p o o l } ) ; \\quad \\tilde { \\bf A } ^ { p o o l } = \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } \\hat { \\bf A } \\hat { \\bf D } ^ { - \\frac { 1 } { 2 } } ." + }, + { + "category_id": 13, + "poly": [ + 1058, + 1284, + 1126, + 1284, + 1126, + 1317, + 1058, + 1317 + ], + "score": 0.9, + "latex": "\\tilde { \\mathbf { A } } ^ { p o o l }" + }, + { + "category_id": 13, + "poly": [ + 556, + 1080, + 751, + 1080, + 751, + 1113, + 556, + 1113 + ], + "score": 0.9, + "latex": "\\mathbf { A } ^ { p o o l } ~ \\in ~ \\mathbb { R } ^ { K \\times K }" + }, + { + "category_id": 13, + "poly": [ + 964, + 1164, + 997, + 1164, + 997, + 1194, + 964, + 1194 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1248, + 1192, + 1315, + 1192, + 1315, + 1222, + 1248, + 1222 + ], + "score": 0.89, + "latex": "\\mathbf { A } ^ { p o o l }" + }, + { + "category_id": 13, + "poly": [ + 366, + 527, + 400, + 527, + 400, + 556, + 366, + 556 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1066, + 232, + 1100, + 232, + 1100, + 261, + 1066, + 261 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 1320, + 293, + 1353, + 293, + 1353, + 322, + 1320, + 322 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 406, + 497, + 440, + 497, + 440, + 526, + 406, + 526 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { o }" + }, + { + "category_id": 13, + "poly": [ + 535, + 1160, + 603, + 1160, + 603, + 1191, + 535, + 1191 + ], + "score": 0.88, + "latex": "\\mathbf { A } ^ { p o o l }" + }, + { + "category_id": 13, + "poly": [ + 1013, + 293, + 1048, + 293, + 1048, + 322, + 1013, + 322 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { o }" + }, + { + "category_id": 13, + "poly": [ + 1318, + 818, + 1350, + 818, + 1350, + 847, + 1318, + 847 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { c }" + }, + { + "category_id": 13, + "poly": [ + 619, + 231, + 653, + 231, + 653, + 261, + 619, + 261 + ], + "score": 0.88, + "latex": "\\mathcal { L } _ { o }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 322, + 1147, + 322, + 1147, + 353, + 1109, + 353 + ], + "score": 0.87, + "latex": "{ \\mathbf { I } } _ { K }" + }, + { + "category_id": 13, + "poly": [ + 1366, + 262, + 1401, + 262, + 1401, + 292, + 1366, + 292 + ], + "score": 0.87, + "latex": "\\mathcal { L } _ { o }" + }, + { + "category_id": 13, + "poly": [ + 372, + 1417, + 456, + 1417, + 456, + 1451, + 372, + 1451 + ], + "score": 0.86, + "latex": "d i a g ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 705, + 1686, + 735, + 1686, + 735, + 1713, + 705, + 1713 + ], + "score": 0.85, + "latex": "\\mathbf { x } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1685, + 322, + 1685, + 322, + 1713, + 296, + 1713 + ], + "score": 0.84, + "latex": "\\mathbf { s } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1254, + 1046, + 1271, + 1046, + 1271, + 1076, + 1254, + 1076 + ], + "score": 0.83, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 483, + 2004, + 501, + 2004, + 501, + 2030, + 483, + 2030 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 607, + 1883, + 636, + 1883, + 636, + 1909, + 607, + 1909 + ], + "score": 0.79, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 439, + 1163, + 456, + 1163, + 456, + 1195, + 439, + 1195 + ], + "score": 0.79, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 828, + 352, + 851, + 352, + 851, + 385, + 828, + 385 + ], + "score": 0.78, + "latex": "\\hat { \\bf S }" + }, + { + "category_id": 13, + "poly": [ + 872, + 1128, + 885, + 1128, + 885, + 1154, + 872, + 1154 + ], + "score": 0.67, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 377, + 1165, + 391, + 1165, + 391, + 1191, + 377, + 1191 + ], + "score": 0.67, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 522, + 1087, + 535, + 1087, + 535, + 1112, + 522, + 1112 + ], + "score": 0.59, + "latex": "i" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 883.0, + 530.0, + 883.0, + 530.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 1114.0, + 1486.0, + 1114.0, + 1521.0, + 295.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 1033.0, + 585.0, + 1033.0, + 585.0, + 1084.0, + 548.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1033.0, + 1253.0, + 1033.0, + 1253.0, + 1084.0, + 778.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1272.0, + 1033.0, + 1408.0, + 1033.0, + 1408.0, + 1084.0, + 1272.0, + 1084.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1073.0, + 521.0, + 1073.0, + 521.0, + 1130.0, + 290.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 1073.0, + 555.0, + 1073.0, + 555.0, + 1130.0, + 536.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1073.0, + 1192.0, + 1073.0, + 1192.0, + 1130.0, + 752.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 1085.0, + 1405.0, + 1085.0, + 1405.0, + 1116.0, + 1254.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1121.0, + 1406.0, + 1121.0, + 1406.0, + 1161.0, + 1030.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1156.0, + 376.0, + 1156.0, + 376.0, + 1201.0, + 292.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1156.0, + 438.0, + 1156.0, + 438.0, + 1201.0, + 392.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1156.0, + 534.0, + 1156.0, + 534.0, + 1201.0, + 457.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1156.0, + 963.0, + 1156.0, + 963.0, + 1201.0, + 604.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 1156.0, + 1408.0, + 1156.0, + 1408.0, + 1201.0, + 998.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1189.0, + 1247.0, + 1189.0, + 1247.0, + 1227.0, + 292.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1316.0, + 1189.0, + 1407.0, + 1189.0, + 1407.0, + 1227.0, + 1316.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1223.0, + 1405.0, + 1223.0, + 1405.0, + 1258.0, + 294.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1253.0, + 1405.0, + 1253.0, + 1405.0, + 1290.0, + 292.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1284.0, + 1057.0, + 1284.0, + 1057.0, + 1326.0, + 292.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1284.0, + 1406.0, + 1284.0, + 1406.0, + 1326.0, + 1127.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1318.0, + 765.0, + 1318.0, + 765.0, + 1355.0, + 295.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1038.5, + 540.0, + 1038.5, + 540.0, + 1083.5, + 292.0, + 1083.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1078.5, + 1254.0, + 1078.5, + 1254.0, + 1102.5, + 1204.0, + 1102.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.75, + 1115.5, + 1023.75, + 1115.5, + 1023.75, + 1168.5, + 288.75, + 1168.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1615.0, + 1405.0, + 1615.0, + 1405.0, + 1657.0, + 293.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 1405.0, + 1650.0, + 1405.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1679.0, + 704.0, + 1679.0, + 704.0, + 1719.0, + 323.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1679.0, + 1407.0, + 1679.0, + 1407.0, + 1719.0, + 736.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1714.0, + 1404.0, + 1714.0, + 1404.0, + 1745.0, + 296.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1771.0, + 588.0, + 1771.0, + 588.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1818.0, + 1404.0, + 1818.0, + 1404.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1407.0, + 1850.0, + 1407.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1880.0, + 606.0, + 1880.0, + 606.0, + 1917.0, + 293.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1917.0, + 637.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1940.0, + 1407.0, + 1940.0, + 1407.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2006.0, + 294.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2005.0, + 482.0, + 2005.0, + 482.0, + 2036.0, + 297.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 2005.0, + 712.0, + 2005.0, + 712.0, + 2036.0, + 502.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 618.0, + 227.0, + 618.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 227.0, + 1065.0, + 227.0, + 1065.0, + 267.0, + 654.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 227.0, + 1407.0, + 227.0, + 1407.0, + 267.0, + 1101.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 261.0, + 1365.0, + 261.0, + 1365.0, + 299.0, + 292.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 299.0, + 1402.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 290.0, + 734.0, + 290.0, + 734.0, + 329.0, + 292.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 290.0, + 1012.0, + 290.0, + 1012.0, + 329.0, + 876.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 290.0, + 1319.0, + 290.0, + 1319.0, + 329.0, + 1049.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 290.0, + 1406.0, + 290.0, + 1406.0, + 329.0, + 1354.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 323.0, + 1108.0, + 323.0, + 1108.0, + 358.0, + 295.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 323.0, + 1406.0, + 323.0, + 1406.0, + 358.0, + 1148.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 348.0, + 609.0, + 348.0, + 609.0, + 394.0, + 292.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 348.0, + 827.0, + 348.0, + 827.0, + 394.0, + 742.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 348.0, + 1029.0, + 348.0, + 1029.0, + 394.0, + 852.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1097.0, + 348.0, + 1406.0, + 348.0, + 1406.0, + 394.0, + 1097.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 388.0, + 1406.0, + 388.0, + 1406.0, + 423.0, + 295.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 420.0, + 1342.0, + 420.0, + 1342.0, + 451.0, + 296.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 462.0, + 1404.0, + 462.0, + 1404.0, + 497.0, + 295.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 496.0, + 405.0, + 496.0, + 405.0, + 529.0, + 295.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 496.0, + 1404.0, + 496.0, + 1404.0, + 529.0, + 441.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 523.0, + 365.0, + 523.0, + 365.0, + 562.0, + 294.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 523.0, + 1405.0, + 523.0, + 1405.0, + 562.0, + 401.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 557.0, + 744.0, + 557.0, + 744.0, + 589.0, + 295.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 614.0, + 1403.0, + 614.0, + 1403.0, + 644.0, + 363.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 641.0, + 1406.0, + 641.0, + 1406.0, + 679.0, + 392.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 672.0, + 1005.0, + 672.0, + 1005.0, + 711.0, + 392.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 712.0, + 1406.0, + 712.0, + 1406.0, + 749.0, + 358.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 746.0, + 1405.0, + 746.0, + 1405.0, + 779.0, + 394.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 776.0, + 638.0, + 776.0, + 638.0, + 809.0, + 395.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 816.0, + 1317.0, + 816.0, + 1317.0, + 853.0, + 360.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 816.0, + 1361.0, + 816.0, + 1361.0, + 853.0, + 1351.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1541.0, + 1403.0, + 1541.0, + 1403.0, + 1579.0, + 295.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1573.0, + 1287.0, + 1573.0, + 1287.0, + 1609.0, + 294.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 937.0, + 1239.0, + 937.0, + 1239.0, + 981.0, + 293.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1416.0, + 371.0, + 1416.0, + 371.0, + 1452.0, + 298.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1416.0, + 772.0, + 1416.0, + 772.0, + 1452.0, + 457.0, + 1452.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 733, + 1403, + 733, + 1403, + 979, + 299, + 979 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 403, + 1403, + 403, + 1403, + 718, + 298, + 718 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1181, + 1403, + 1181, + 1403, + 1366, + 298, + 1366 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1580, + 1404, + 1580, + 1404, + 1765, + 298, + 1765 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1014, + 1402, + 1014, + 1402, + 1167, + 298, + 1167 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1382, + 1402, + 1382, + 1402, + 1566, + 298, + 1566 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 296, + 1402, + 296, + 1402, + 388, + 299, + 388 + ], + "score": 0.971 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 909, + 226, + 909, + 261, + 299, + 261 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 1810, + 557, + 1810, + 557, + 1846, + 300, + 1846 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.695 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.093 + }, + { + "category_id": 13, + "poly": [ + 1218, + 915, + 1301, + 915, + 1301, + 950, + 1218, + 950 + ], + "score": 0.93, + "latex": "\\mathcal { O } ( N ^ { 2 } )" + }, + { + "category_id": 13, + "poly": [ + 298, + 497, + 455, + 497, + 455, + 533, + 298, + 533 + ], + "score": 0.92, + "latex": "\\| \\mathbf { A } - \\mathbf { S } \\mathbf { \\dot { S } } ^ { T } \\| _ { F }" + }, + { + "category_id": 13, + "poly": [ + 1097, + 1276, + 1167, + 1276, + 1167, + 1310, + 1097, + 1310 + ], + "score": 0.91, + "latex": "k _ { ( l + 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 409, + 559, + 586, + 559, + 586, + 601, + 409, + 601 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } H ( \\mathbf { S } _ { i } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 662, + 435, + 744, + 435, + 744, + 467, + 662, + 467 + ], + "score": 0.91, + "latex": "\\mathbf { X } ^ { ( t + 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 604, + 1276, + 627, + 1276, + 627, + 1306, + 604, + 1306 + ], + "score": 0.85, + "latex": "j _ { l }" + }, + { + "category_id": 13, + "poly": [ + 531, + 1276, + 553, + 1276, + 553, + 1305, + 531, + 1305 + ], + "score": 0.85, + "latex": "i _ { l }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 915, + 1149, + 915, + 1149, + 946, + 1109, + 946 + ], + "score": 0.84, + "latex": "{ \\bf A } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 359, + 915, + 399, + 915, + 399, + 946, + 359, + 946 + ], + "score": 0.81, + "latex": "{ \\bf A } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 612, + 857, + 640, + 857, + 640, + 883, + 612, + 883 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1206, + 766, + 1235, + 766, + 1235, + 792, + 1206, + 792 + ], + "score": 0.79, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 897, + 796, + 926, + 796, + 926, + 824, + 897, + 824 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1022, + 887, + 1050, + 887, + 1050, + 915, + 1022, + 915 + ], + "score": 0.76, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 599, + 736, + 627, + 736, + 627, + 763, + 599, + 763 + ], + "score": 0.69, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1038, + 1045, + 1083, + 1045, + 1083, + 1076, + 1038, + 1076 + ], + "score": 0.53, + "latex": "\\mathbf { \\Pi } ( \\mathbf { X } )" + }, + { + "category_id": 13, + "poly": [ + 358, + 358, + 386, + 358, + 386, + 385, + 358, + 385 + ], + "score": 0.42, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 944, + 1245, + 956, + 1245, + 956, + 1271, + 944, + 1271 + ], + "score": 0.39, + "latex": "l" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 913.0, + 224.0, + 913.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1806.0, + 562.0, + 1806.0, + 562.0, + 1851.0, + 291.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2119.0, + 838.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 735.0, + 598.0, + 735.0, + 598.0, + 768.0, + 295.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 735.0, + 1404.0, + 735.0, + 1404.0, + 768.0, + 628.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 765.0, + 1205.0, + 765.0, + 1205.0, + 798.0, + 295.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 798.0, + 1236.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 796.0, + 896.0, + 796.0, + 896.0, + 830.0, + 294.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 796.0, + 1405.0, + 796.0, + 1405.0, + 830.0, + 927.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 826.0, + 1405.0, + 826.0, + 1405.0, + 860.0, + 295.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 856.0, + 611.0, + 856.0, + 611.0, + 890.0, + 295.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 856.0, + 1405.0, + 856.0, + 1405.0, + 890.0, + 641.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 886.0, + 1021.0, + 886.0, + 1021.0, + 921.0, + 293.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 886.0, + 1407.0, + 886.0, + 1407.0, + 921.0, + 1051.0, + 921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 914.0, + 358.0, + 914.0, + 358.0, + 955.0, + 293.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 914.0, + 1108.0, + 914.0, + 1108.0, + 955.0, + 400.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 914.0, + 1217.0, + 914.0, + 1217.0, + 955.0, + 1150.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1302.0, + 914.0, + 1408.0, + 914.0, + 1408.0, + 955.0, + 1302.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 949.0, + 830.0, + 949.0, + 830.0, + 982.0, + 295.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 404.0, + 1407.0, + 404.0, + 1407.0, + 440.0, + 296.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 428.0, + 661.0, + 428.0, + 661.0, + 476.0, + 289.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 428.0, + 1408.0, + 428.0, + 1408.0, + 476.0, + 745.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 467.0, + 1405.0, + 467.0, + 1405.0, + 503.0, + 293.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 496.0, + 297.0, + 496.0, + 297.0, + 536.0, + 294.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 496.0, + 1407.0, + 496.0, + 1407.0, + 536.0, + 456.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 529.0, + 1405.0, + 529.0, + 1405.0, + 566.0, + 293.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 553.0, + 408.0, + 553.0, + 408.0, + 611.0, + 297.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 553.0, + 1408.0, + 553.0, + 1408.0, + 611.0, + 587.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 594.0, + 1404.0, + 594.0, + 1404.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 627.0, + 1405.0, + 627.0, + 1405.0, + 662.0, + 293.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 658.0, + 1404.0, + 658.0, + 1404.0, + 690.0, + 296.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 688.0, + 1030.0, + 688.0, + 1030.0, + 723.0, + 293.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.25, + 564.0, + 524.25, + 564.0, + 524.25, + 594.0, + 476.25, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1218.0, + 296.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1215.0, + 1405.0, + 1215.0, + 1405.0, + 1247.0, + 296.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1243.0, + 943.0, + 1243.0, + 943.0, + 1279.0, + 294.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 1243.0, + 1404.0, + 1243.0, + 1404.0, + 1279.0, + 957.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1273.0, + 530.0, + 1273.0, + 530.0, + 1313.0, + 293.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1273.0, + 603.0, + 1273.0, + 603.0, + 1313.0, + 554.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1273.0, + 1096.0, + 1273.0, + 1096.0, + 1313.0, + 628.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1273.0, + 1407.0, + 1273.0, + 1407.0, + 1313.0, + 1168.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1304.0, + 1405.0, + 1304.0, + 1405.0, + 1338.0, + 293.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1336.0, + 1285.0, + 1336.0, + 1285.0, + 1371.0, + 294.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1581.0, + 1404.0, + 1581.0, + 1404.0, + 1617.0, + 294.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1613.0, + 1404.0, + 1613.0, + 1404.0, + 1645.0, + 296.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1638.0, + 1404.0, + 1638.0, + 1404.0, + 1680.0, + 292.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1671.0, + 1405.0, + 1671.0, + 1405.0, + 1711.0, + 292.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 1408.0, + 1702.0, + 1408.0, + 1740.0, + 293.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1734.0, + 673.0, + 1734.0, + 673.0, + 1769.0, + 293.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1011.0, + 1404.0, + 1011.0, + 1404.0, + 1050.0, + 294.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1042.0, + 1037.0, + 1042.0, + 1037.0, + 1080.0, + 293.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1042.0, + 1406.0, + 1042.0, + 1406.0, + 1080.0, + 1084.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1071.0, + 1404.0, + 1071.0, + 1404.0, + 1112.0, + 292.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 1404.0, + 1104.0, + 1404.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1135.0, + 875.0, + 1135.0, + 875.0, + 1171.0, + 294.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1380.0, + 1406.0, + 1380.0, + 1406.0, + 1418.0, + 292.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1412.0, + 1406.0, + 1412.0, + 1406.0, + 1447.0, + 293.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1442.0, + 1406.0, + 1442.0, + 1406.0, + 1478.0, + 293.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1473.0, + 1406.0, + 1473.0, + 1406.0, + 1510.0, + 293.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1502.0, + 1406.0, + 1502.0, + 1406.0, + 1540.0, + 293.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1536.0, + 494.0, + 1536.0, + 494.0, + 1569.0, + 294.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1879.0, + 1406.0, + 1879.0, + 1406.0, + 1920.0, + 295.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1911.0, + 1404.0, + 1911.0, + 1404.0, + 1946.0, + 293.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1941.0, + 1406.0, + 1941.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1972.0, + 1406.0, + 1972.0, + 1406.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 1284.0, + 2001.0, + 1284.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 295.0, + 1406.0, + 295.0, + 1406.0, + 332.0, + 295.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 328.0, + 1403.0, + 328.0, + 1403.0, + 361.0, + 294.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 358.0, + 357.0, + 358.0, + 357.0, + 391.0, + 294.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 358.0, + 1094.0, + 358.0, + 1094.0, + 391.0, + 387.0, + 391.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1706, + 1405, + 1706, + 1405, + 1960, + 297, + 1960 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 296, + 1427, + 1405, + 1427, + 1405, + 1674, + 296, + 1674 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 412, + 1404, + 412, + 1404, + 659, + 297, + 659 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 443, + 699, + 1256, + 699, + 1256, + 1313, + 443, + 1313 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 301, + 286, + 1402, + 286, + 1402, + 380, + 301, + 380 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1403, + 1972, + 1403, + 2035, + 297, + 2035 + ], + "score": 0.939 + }, + { + "category_id": 4, + "poly": [ + 364, + 1349, + 1329, + 1349, + 1329, + 1382, + 364, + 1382 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.787 + }, + { + "category_id": 0, + "poly": [ + 300, + 231, + 753, + 231, + 753, + 261, + 300, + 261 + ], + "score": 0.772 + }, + { + "category_id": 13, + "poly": [ + 522, + 1902, + 832, + 1902, + 832, + 1961, + 522, + 1961 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\mathbf { N M I } ( \\tilde { \\mathbf { y } } , \\mathbf { y } ) = \\frac { H ( \\tilde { \\mathbf { y } } ) - H ( \\tilde { \\mathbf { y } } | \\mathbf { y } ) } { \\sqrt { H ( \\tilde { \\mathbf { y } } ) - H ( \\mathbf { y } ) } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 956, + 1859, + 1216, + 1859, + 1216, + 1907, + 956, + 1907 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\mathrm { C S } ( \\tilde { \\bf y } , { \\bf y } ) = 1 - \\frac { H ( \\tilde { \\bf y } | { \\bf y } ) } { H ( \\tilde { \\bf y } ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 919, + 1909, + 976, + 1909, + 976, + 1944, + 919, + 1944 + ], + "score": 0.91, + "latex": "H ( \\cdot )" + }, + { + "category_id": 13, + "poly": [ + 824, + 1582, + 904, + 1582, + 904, + 1609, + 824, + 1609 + ], + "score": 0.9, + "latex": "K = 4" + }, + { + "category_id": 13, + "poly": [ + 432, + 350, + 468, + 350, + 468, + 379, + 432, + 379 + ], + "score": 0.89, + "latex": "\\mathcal { L } _ { u }" + }, + { + "category_id": 13, + "poly": [ + 779, + 1803, + 805, + 1803, + 805, + 1831, + 779, + 1831 + ], + "score": 0.85, + "latex": "y _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1254, + 1350, + 1321, + 1350, + 1321, + 1380, + 1254, + 1380 + ], + "score": 0.79, + "latex": "( K { = } 5 )" + }, + { + "category_id": 13, + "poly": [ + 953, + 1351, + 1017, + 1351, + 1017, + 1380, + 953, + 1380 + ], + "score": 0.79, + "latex": "K { = } 6 )" + }, + { + "category_id": 13, + "poly": [ + 360, + 1801, + 374, + 1801, + 374, + 1826, + 360, + 1826 + ], + "score": 0.74, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 745, + 2004, + 772, + 2004, + 772, + 2030, + 745, + 2030 + ], + "score": 0.68, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 402, + 1768, + 430, + 1768, + 430, + 1796, + 402, + 1796 + ], + "score": 0.39, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 826, + 1769, + 853, + 1769, + 853, + 1796, + 826, + 1796 + ], + "score": 0.33, + "latex": "\\mathbf { A }" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 824.0, + 494.0, + 824.0, + 494.0, + 850.0, + 456.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 825.0, + 1208.0, + 825.0, + 1208.0, + 834.0, + 1198.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 852.0, + 930.0, + 852.0, + 930.0, + 860.0, + 921.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 850.0, + 1212.0, + 850.0, + 1212.0, + 859.0, + 1203.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 859.0, + 581.0, + 859.0, + 581.0, + 929.0, + 527.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 890.0, + 861.0, + 890.0, + 861.0, + 905.0, + 839.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 864.0, + 1135.0, + 864.0, + 1135.0, + 917.0, + 1089.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 902.0, + 530.0, + 902.0, + 530.0, + 914.0, + 517.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 961.0, + 610.0, + 961.0, + 610.0, + 997.0, + 528.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 960.0, + 918.0, + 960.0, + 918.0, + 998.0, + 781.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 960.0, + 1221.0, + 960.0, + 1221.0, + 997.0, + 1039.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1280.0, + 611.0, + 1280.0, + 611.0, + 1315.0, + 526.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1279.0, + 918.0, + 1279.0, + 918.0, + 1318.0, + 783.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1041.0, + 1279.0, + 1220.0, + 1279.0, + 1220.0, + 1317.0, + 1041.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 757.5, + 810.0, + 757.5, + 810.0, + 769.5, + 807.0, + 769.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 814.5, + 1215.0, + 814.5, + 1215.0, + 855.0, + 1209.0, + 855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.25, + 800.5, + 705.25, + 800.5, + 705.25, + 853.5, + 607.25, + 853.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 812.5, + 1229.0, + 812.5, + 1229.0, + 821.0, + 1212.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 724.75, + 814.5, + 811.75, + 814.5, + 811.75, + 854.5, + 724.75, + 854.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 834.0, + 949.0, + 834.0, + 949.0, + 849.5, + 916.0, + 849.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 817.0, + 1085.0, + 817.0, + 1085.0, + 853.5, + 1007.0, + 853.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 883.5, + 629.0, + 883.5, + 629.0, + 919.5, + 583.0, + 919.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 874.5, + 850.0, + 874.5, + 850.0, + 907.5, + 799.0, + 907.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.75, + 869.0, + 937.75, + 869.0, + 937.75, + 942.0, + 840.75, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.75, + 893.0, + 1178.75, + 893.0, + 1178.75, + 911.5, + 1161.75, + 911.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1345.0, + 952.0, + 1345.0, + 952.0, + 1388.0, + 362.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1345.0, + 1253.0, + 1345.0, + 1253.0, + 1388.0, + 1018.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 1345.0, + 1335.0, + 1345.0, + 1335.0, + 1388.0, + 1322.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 756.0, + 230.0, + 756.0, + 264.0, + 295.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1705.0, + 1405.0, + 1705.0, + 1405.0, + 1742.0, + 295.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1736.0, + 1406.0, + 1736.0, + 1406.0, + 1772.0, + 294.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1769.0, + 401.0, + 1769.0, + 401.0, + 1800.0, + 296.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 431.0, + 1769.0, + 825.0, + 1769.0, + 825.0, + 1800.0, + 431.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1769.0, + 1403.0, + 1769.0, + 1403.0, + 1800.0, + 854.0, + 1800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1799.0, + 359.0, + 1799.0, + 359.0, + 1833.0, + 294.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1799.0, + 778.0, + 1799.0, + 778.0, + 1833.0, + 375.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 806.0, + 1799.0, + 1405.0, + 1799.0, + 1405.0, + 1833.0, + 806.0, + 1833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1828.0, + 1406.0, + 1828.0, + 1406.0, + 1861.0, + 294.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1858.0, + 955.0, + 1858.0, + 955.0, + 1911.0, + 292.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1858.0, + 1406.0, + 1858.0, + 1406.0, + 1911.0, + 1217.0, + 1911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1907.0, + 521.0, + 1907.0, + 521.0, + 1951.0, + 291.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1907.0, + 918.0, + 1907.0, + 918.0, + 1949.0, + 840.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1907.0, + 1142.0, + 1907.0, + 1142.0, + 1949.0, + 977.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1430.0, + 1402.0, + 1430.0, + 1402.0, + 1464.0, + 295.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1494.0, + 294.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1490.0, + 1405.0, + 1490.0, + 1405.0, + 1524.0, + 294.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1520.0, + 1403.0, + 1520.0, + 1403.0, + 1554.0, + 294.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1549.0, + 1406.0, + 1549.0, + 1406.0, + 1585.0, + 294.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1578.0, + 823.0, + 1578.0, + 823.0, + 1618.0, + 290.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1578.0, + 1407.0, + 1578.0, + 1407.0, + 1618.0, + 905.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1613.0, + 1405.0, + 1613.0, + 1405.0, + 1647.0, + 294.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 811.0, + 1642.0, + 811.0, + 1677.0, + 295.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 411.0, + 1407.0, + 411.0, + 1407.0, + 448.0, + 294.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 444.0, + 1404.0, + 444.0, + 1404.0, + 476.0, + 294.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 475.0, + 1404.0, + 475.0, + 1404.0, + 509.0, + 295.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 502.0, + 1405.0, + 502.0, + 1405.0, + 543.0, + 292.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 537.0, + 1405.0, + 537.0, + 1405.0, + 570.0, + 295.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 567.0, + 1406.0, + 567.0, + 1406.0, + 601.0, + 295.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 597.0, + 1405.0, + 597.0, + 1405.0, + 631.0, + 295.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 627.0, + 1324.0, + 627.0, + 1324.0, + 661.0, + 295.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 287.0, + 1407.0, + 287.0, + 1407.0, + 321.0, + 296.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 316.0, + 1404.0, + 316.0, + 1404.0, + 355.0, + 294.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 349.0, + 431.0, + 349.0, + 431.0, + 383.0, + 297.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 349.0, + 1381.0, + 349.0, + 1381.0, + 383.0, + 469.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1972.0, + 1405.0, + 1972.0, + 1405.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 744.0, + 2001.0, + 744.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2038.0, + 773.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1849, + 1404, + 1849, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 320, + 1548, + 1380, + 1548, + 1380, + 1726, + 320, + 1726 + ], + "score": 0.978, + "html": "
DatasetKSpectral clusteringDiffpoolminCUTpool
NMIcsNMICSNMICS
Cora70.025 ± 0.0140.126 ± 0.0420.315 ± 0.0050.309 ±0.0050.404 ± 0.0180.392 ± 0.018
Citeseer60.014 ± 0.0030.033 ±0.0000.139 ± 0.0160.153 ± 0.0200.287 ±0.0470.283 ± 0.046
Pubmed30.182 ± 0.0000.261 ± 0.0000.079 ±0.0010.085 ±0.0010.200 ± 0.0200.197 ± 0.019
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1331, + 1403, + 1331, + 1403, + 1426, + 298, + 1426 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 341, + 240, + 1374, + 240, + 1374, + 744, + 341, + 744 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 312, + 859, + 1382, + 859, + 1382, + 1205, + 312, + 1205 + ], + "score": 0.97 + }, + { + "category_id": 6, + "poly": [ + 298, + 1480, + 1399, + 1480, + 1399, + 1544, + 298, + 1544 + ], + "score": 0.93 + }, + { + "category_id": 4, + "poly": [ + 403, + 1239, + 1290, + 1239, + 1290, + 1274, + 403, + 1274 + ], + "score": 0.919 + }, + { + "category_id": 4, + "poly": [ + 363, + 776, + 1332, + 776, + 1332, + 812, + 363, + 812 + ], + "score": 0.918 + }, + { + "category_id": 0, + "poly": [ + 302, + 1792, + 806, + 1792, + 806, + 1823, + 302, + 1823 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 858, + 76, + 858, + 104, + 297, + 104 + ], + "score": 0.888 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.681 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.256 + }, + { + "category_id": 13, + "poly": [ + 1113, + 1850, + 1222, + 1850, + 1222, + 1884, + 1113, + 1884 + ], + "score": 0.93, + "latex": "\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1942, + 1204, + 1942, + 1204, + 1972, + 1150, + 1972 + ], + "score": 0.88, + "latex": "1 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 765, + 1852, + 800, + 1852, + 800, + 1881, + 765, + 1881 + ], + "score": 0.88, + "latex": "N _ { i }" + }, + { + "category_id": 13, + "poly": [ + 602, + 1884, + 631, + 1884, + 631, + 1913, + 602, + 1913 + ], + "score": 0.84, + "latex": "\\mathbf { y } _ { i }" + }, + { + "category_id": 13, + "poly": [ + 560, + 1514, + 589, + 1514, + 589, + 1540, + 560, + 1540 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 472, + 1853, + 485, + 1853, + 485, + 1878, + 472, + 1878 + ], + "score": 0.68, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1249, + 710, + 1326, + 710, + 1326, + 738, + 1249, + 738 + ], + "score": 0.53, + "latex": "X = 4" + }, + { + "category_id": 15, + "poly": [ + 1149.0, + 399.0, + 1175.0, + 399.0, + 1175.0, + 407.0, + 1149.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 442.0, + 600.0, + 442.0, + 600.0, + 483.0, + 395.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 444.0, + 970.0, + 444.0, + 970.0, + 481.0, + 732.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 444.0, + 1352.0, + 444.0, + 1352.0, + 481.0, + 1052.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 707.0, + 540.0, + 707.0, + 540.0, + 744.0, + 452.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 707.0, + 964.0, + 707.0, + 964.0, + 744.0, + 736.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 708.0, + 1248.0, + 708.0, + 1248.0, + 746.0, + 1066.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1327.0, + 708.0, + 1338.0, + 708.0, + 1338.0, + 746.0, + 1327.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 876.0, + 373.0, + 876.0, + 373.0, + 910.0, + 312.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 876.0, + 560.0, + 876.0, + 560.0, + 912.0, + 475.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 874.0, + 635.0, + 874.0, + 635.0, + 905.0, + 591.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 879.0, + 899.0, + 879.0, + 899.0, + 908.0, + 860.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 877.0, + 1091.0, + 877.0, + 1091.0, + 909.0, + 1006.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 886.0, + 1165.0, + 886.0, + 1165.0, + 915.0, + 1125.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 904.0, + 1043.0, + 904.0, + 1043.0, + 941.0, + 1005.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 907.0, + 1196.0, + 907.0, + 1196.0, + 925.0, + 1184.0, + 925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 921.0, + 372.0, + 921.0, + 372.0, + 947.0, + 314.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 932.0, + 634.0, + 932.0, + 634.0, + 961.0, + 593.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 931.0, + 901.0, + 931.0, + 901.0, + 959.0, + 860.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 940.0, + 1165.0, + 940.0, + 1165.0, + 969.0, + 1125.0, + 969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 960.0, + 372.0, + 960.0, + 372.0, + 986.0, + 314.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 999.0, + 372.0, + 999.0, + 372.0, + 1026.0, + 314.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 988.0, + 636.0, + 988.0, + 636.0, + 1019.0, + 592.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 980.0, + 906.0, + 980.0, + 906.0, + 1012.0, + 859.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 991.0, + 1165.0, + 991.0, + 1165.0, + 1021.0, + 1125.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1038.0, + 372.0, + 1038.0, + 372.0, + 1065.0, + 314.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 1046.0, + 634.0, + 1046.0, + 634.0, + 1075.0, + 593.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1032.0, + 901.0, + 1032.0, + 901.0, + 1062.0, + 848.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1042.0, + 1168.0, + 1042.0, + 1168.0, + 1075.0, + 1124.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 1076.0, + 373.0, + 1076.0, + 373.0, + 1107.0, + 313.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1067.0, + 821.0, + 1067.0, + 821.0, + 1099.0, + 769.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1081.0, + 902.0, + 1081.0, + 902.0, + 1113.0, + 848.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 1067.0, + 1356.0, + 1067.0, + 1356.0, + 1098.0, + 1305.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1112.0, + 392.0, + 1112.0, + 392.0, + 1136.0, + 372.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1108.0, + 498.0, + 1108.0, + 498.0, + 1138.0, + 445.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 1108.0, + 591.0, + 1108.0, + 591.0, + 1138.0, + 529.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1113.0, + 652.0, + 1113.0, + 652.0, + 1136.0, + 632.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1108.0, + 759.0, + 1108.0, + 759.0, + 1138.0, + 704.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 1108.0, + 853.0, + 1108.0, + 853.0, + 1138.0, + 788.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1112.0, + 920.0, + 1112.0, + 920.0, + 1136.0, + 901.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1108.0, + 1029.0, + 1108.0, + 1029.0, + 1138.0, + 974.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 1108.0, + 1124.0, + 1108.0, + 1124.0, + 1138.0, + 1059.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1113.0, + 1185.0, + 1113.0, + 1185.0, + 1136.0, + 1165.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1108.0, + 1293.0, + 1108.0, + 1293.0, + 1138.0, + 1237.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 1108.0, + 1387.0, + 1108.0, + 1387.0, + 1138.0, + 1324.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1130.0, + 514.0, + 1130.0, + 514.0, + 1159.0, + 425.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1130.0, + 776.0, + 1130.0, + 776.0, + 1159.0, + 686.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1127.0, + 1047.0, + 1127.0, + 1047.0, + 1163.0, + 955.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1131.0, + 1310.0, + 1131.0, + 1310.0, + 1160.0, + 1221.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1169.0, + 651.0, + 1169.0, + 651.0, + 1209.0, + 514.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1170.0, + 1209.0, + 1170.0, + 1209.0, + 1207.0, + 1025.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 930.5, + 1042.0, + 930.5, + 1042.0, + 964.5, + 1008.0, + 964.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1480.0, + 1403.0, + 1480.0, + 1403.0, + 1516.0, + 296.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1511.0, + 559.0, + 1511.0, + 559.0, + 1545.0, + 296.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1511.0, + 1026.0, + 1511.0, + 1026.0, + 1545.0, + 590.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1234.0, + 1295.0, + 1234.0, + 1295.0, + 1282.0, + 403.0, + 1282.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 773.0, + 1333.0, + 773.0, + 1333.0, + 817.0, + 361.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1791.0, + 810.0, + 1791.0, + 810.0, + 1827.0, + 295.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 471.0, + 1851.0, + 471.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 1851.0, + 764.0, + 1851.0, + 764.0, + 1883.0, + 486.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1851.0, + 1112.0, + 1851.0, + 1112.0, + 1883.0, + 801.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1223.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1883.0, + 1223.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 601.0, + 1882.0, + 601.0, + 1914.0, + 295.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 1882.0, + 1402.0, + 1882.0, + 1402.0, + 1914.0, + 632.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 1402.0, + 1913.0, + 1402.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1939.0, + 1149.0, + 1939.0, + 1149.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1939.0, + 1406.0, + 1939.0, + 1406.0, + 1978.0, + 1205.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2006.0, + 292.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2039.0, + 292.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1331.0, + 1404.0, + 1331.0, + 1404.0, + 1369.0, + 294.0, + 1369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1363.0, + 1405.0, + 1363.0, + 1405.0, + 1398.0, + 292.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1392.0, + 1355.0, + 1392.0, + 1355.0, + 1430.0, + 293.0, + 1430.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 772, + 1405, + 772, + 1405, + 1139, + 296, + 1139 + ], + "score": 0.982 + }, + { + "category_id": 5, + "poly": [ + 298, + 474, + 1410, + 474, + 1410, + 734, + 298, + 734 + ], + "score": 0.979, + "html": "
DatasetWLDenseFlatGraclusDecim.DiffpoolTop-KSAGpoolminCUT
Bench-easy92.629.3±0.398.5±0.397.5±0.597.9±0.598.6±0.482.4±8.984.2±2.399.0±0.0
Bench-hard60.029.4±0.367.6±2.869.0±1.572.6±0.969.9±1.942.7±15.237.7±14.573.8±1.9
Mutagenicity81.7±1.168.4±0.378.0±1.374.4±1.877.8±2.377.6±2.771.9±3.772.4±2.479.9±2.1
Proteins71.2±2.668.7±3.372.6±4.868.6±4.673.3±3.772.7±3.869.6±3.570.5±2.676.5±2.6
DD78.6±2.770.6±5.276.8±1.570.5±4.872.0±3.179.3±2.469.4±7.871.5±4.580.8±2.3
COLLAB74.8±1.379.3±1.682.1±1.877.1±2.179.1±1.581.8±1.479.3±1.879.2±2.083.4±1.7
Reddit-Binary68.2±1.748.5±2.680.3±2.679.2±0.484.3±2.486.8±2.174.7±4.573.9±5.191.4±1.5
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1830, + 1402, + 1830, + 1402, + 1953, + 299, + 1953 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1692, + 1404, + 1692, + 1404, + 1815, + 298, + 1815 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1404, + 228, + 1404, + 384, + 298, + 384 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 305, + 1180, + 783, + 1180, + 783, + 1385, + 305, + 1385 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 820, + 1217, + 1403, + 1217, + 1403, + 1340, + 820, + 1340 + ], + "score": 0.951 + }, + { + "category_id": 6, + "poly": [ + 342, + 438, + 1345, + 438, + 1345, + 471, + 342, + 471 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 300, + 1622, + 560, + 1622, + 560, + 1658, + 300, + 1658 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.789 + }, + { + "category_id": 1, + "poly": [ + 296, + 1424, + 1404, + 1424, + 1404, + 1578, + 296, + 1578 + ], + "score": 0.786 + }, + { + "category_id": 4, + "poly": [ + 296, + 1424, + 1404, + 1424, + 1404, + 1578, + 296, + 1578 + ], + "score": 0.397 + }, + { + "category_id": 13, + "poly": [ + 842, + 292, + 959, + 292, + 959, + 324, + 842, + 324 + ], + "score": 0.9, + "latex": "K = N / 2" + }, + { + "category_id": 13, + "poly": [ + 1092, + 439, + 1210, + 439, + 1210, + 471, + 1092, + 471 + ], + "score": 0.86, + "latex": "( p < 0 . 0 5 )" + }, + { + "category_id": 13, + "poly": [ + 296, + 804, + 387, + 804, + 387, + 834, + 296, + 834 + ], + "score": 0.82, + "latex": "< ~ 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 1042, + 293, + 1070, + 293, + 1070, + 320, + 1042, + 320 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1291, + 1517, + 1318, + 1517, + 1318, + 1543, + 1291, + 1543 + ], + "score": 0.77, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1315, + 778, + 1332, + 778, + 1332, + 805, + 1315, + 805 + ], + "score": 0.74, + "latex": "\\dot { p }" + }, + { + "category_id": 13, + "poly": [ + 350, + 262, + 377, + 262, + 377, + 290, + 350, + 290 + ], + "score": 0.6, + "latex": "K" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1196.0, + 341.0, + 1196.0, + 341.0, + 1209.0, + 328.0, + 1209.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1215.0, + 333.0, + 1215.0, + 333.0, + 1328.0, + 301.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1229.0, + 344.0, + 1229.0, + 344.0, + 1253.0, + 326.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1270.0, + 342.0, + 1270.0, + 342.0, + 1290.0, + 327.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1309.0, + 342.0, + 1309.0, + 342.0, + 1328.0, + 326.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1347.0, + 343.0, + 1347.0, + 343.0, + 1365.0, + 327.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1357.0, + 760.0, + 1357.0, + 760.0, + 1386.0, + 375.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1216.0, + 1402.0, + 1216.0, + 1402.0, + 1252.0, + 820.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1247.0, + 1404.0, + 1247.0, + 1404.0, + 1280.0, + 821.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1278.0, + 1405.0, + 1278.0, + 1405.0, + 1311.0, + 820.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1308.0, + 1261.0, + 1308.0, + 1261.0, + 1339.0, + 821.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 436.0, + 1091.0, + 436.0, + 1091.0, + 474.0, + 352.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 436.0, + 1347.0, + 436.0, + 1347.0, + 474.0, + 1211.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1615.0, + 566.0, + 1615.0, + 566.0, + 1666.0, + 291.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1404.0, + 1423.0, + 1404.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 1402.0, + 1453.0, + 1402.0, + 1489.0, + 295.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1485.0, + 1405.0, + 1485.0, + 1405.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1513.0, + 1290.0, + 1513.0, + 1290.0, + 1554.0, + 293.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1513.0, + 1407.0, + 1513.0, + 1407.0, + 1554.0, + 1319.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1546.0, + 840.0, + 1546.0, + 840.0, + 1583.0, + 296.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 771.0, + 1314.0, + 771.0, + 1314.0, + 810.0, + 293.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1333.0, + 771.0, + 1406.0, + 771.0, + 1406.0, + 810.0, + 1333.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 802.0, + 1405.0, + 802.0, + 1405.0, + 839.0, + 388.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 833.0, + 1403.0, + 833.0, + 1403.0, + 869.0, + 294.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 863.0, + 1405.0, + 863.0, + 1405.0, + 900.0, + 293.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 891.0, + 1405.0, + 891.0, + 1405.0, + 932.0, + 293.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 925.0, + 1405.0, + 925.0, + 1405.0, + 960.0, + 293.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 957.0, + 1405.0, + 957.0, + 1405.0, + 991.0, + 294.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 983.0, + 1406.0, + 983.0, + 1406.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1015.0, + 1407.0, + 1015.0, + 1407.0, + 1053.0, + 291.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1048.0, + 1406.0, + 1048.0, + 1406.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1078.0, + 1407.0, + 1078.0, + 1407.0, + 1112.0, + 295.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1109.0, + 1005.0, + 1109.0, + 1005.0, + 1140.0, + 294.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1826.0, + 1403.0, + 1826.0, + 1403.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1862.0, + 1404.0, + 1862.0, + 1404.0, + 1895.0, + 295.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1895.0, + 1404.0, + 1895.0, + 1404.0, + 1924.0, + 297.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1923.0, + 772.0, + 1923.0, + 772.0, + 1955.0, + 295.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1693.0, + 1404.0, + 1693.0, + 1404.0, + 1726.0, + 297.0, + 1726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1724.0, + 1404.0, + 1724.0, + 1404.0, + 1755.0, + 293.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1755.0, + 1404.0, + 1755.0, + 1404.0, + 1786.0, + 293.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1785.0, + 814.0, + 1785.0, + 814.0, + 1816.0, + 294.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 267.0, + 293.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 262.0, + 349.0, + 262.0, + 349.0, + 295.0, + 297.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 295.0, + 378.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 841.0, + 289.0, + 841.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 289.0, + 1041.0, + 289.0, + 1041.0, + 326.0, + 960.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 289.0, + 1402.0, + 289.0, + 1402.0, + 326.0, + 1071.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 350.0, + 1245.0, + 350.0, + 1245.0, + 390.0, + 293.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1404.0, + 1423.0, + 1404.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 1402.0, + 1453.0, + 1402.0, + 1489.0, + 295.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1485.0, + 1405.0, + 1485.0, + 1405.0, + 1519.0, + 293.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1513.0, + 1290.0, + 1513.0, + 1290.0, + 1554.0, + 293.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1513.0, + 1407.0, + 1513.0, + 1407.0, + 1554.0, + 1319.0, + 1554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1546.0, + 840.0, + 1546.0, + 840.0, + 1583.0, + 296.0, + 1583.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 300, + 1739, + 1403, + 1739, + 1403, + 1834, + 300, + 1834 + ], + "score": 0.848 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 488, + 227, + 488, + 261, + 300, + 261 + ], + "score": 0.828 + }, + { + "category_id": 1, + "poly": [ + 299, + 1455, + 1403, + 1455, + 1403, + 1520, + 299, + 1520 + ], + "score": 0.788 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 299, + 1539, + 1400, + 1539, + 1400, + 1604, + 299, + 1604 + ], + "score": 0.776 + }, + { + "category_id": 1, + "poly": [ + 293, + 708, + 1399, + 708, + 1399, + 774, + 293, + 774 + ], + "score": 0.763 + }, + { + "category_id": 1, + "poly": [ + 294, + 1856, + 1400, + 1856, + 1400, + 1920, + 294, + 1920 + ], + "score": 0.757 + }, + { + "category_id": 1, + "poly": [ + 288, + 1024, + 1400, + 1024, + 1400, + 1089, + 288, + 1089 + ], + "score": 0.746 + }, + { + "category_id": 1, + "poly": [ + 298, + 1224, + 1402, + 1224, + 1402, + 1319, + 298, + 1319 + ], + "score": 0.744 + }, + { + "category_id": 1, + "poly": [ + 300, + 1109, + 1400, + 1109, + 1400, + 1206, + 300, + 1206 + ], + "score": 0.728 + }, + { + "category_id": 1, + "poly": [ + 297, + 1339, + 1405, + 1339, + 1405, + 1435, + 297, + 1435 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 291, + 909, + 1401, + 909, + 1401, + 1004, + 291, + 1004 + ], + "score": 0.721 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1401, + 1941, + 1401, + 2034, + 299, + 2034 + ], + "score": 0.714 + }, + { + "category_id": 1, + "poly": [ + 297, + 279, + 1405, + 279, + 1405, + 402, + 297, + 402 + ], + "score": 0.692 + }, + { + "category_id": 1, + "poly": [ + 293, + 508, + 1401, + 508, + 1401, + 574, + 293, + 574 + ], + "score": 0.68 + }, + { + "category_id": 1, + "poly": [ + 294, + 425, + 1402, + 425, + 1402, + 489, + 294, + 489 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 295, + 793, + 1405, + 793, + 1405, + 889, + 295, + 889 + ], + "score": 0.585 + }, + { + "category_id": 1, + "poly": [ + 295, + 593, + 1404, + 593, + 1404, + 688, + 295, + 688 + ], + "score": 0.349 + }, + { + "category_id": 1, + "poly": [ + 296, + 1624, + 1404, + 1624, + 1404, + 1719, + 296, + 1719 + ], + "score": 0.262 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 227.0, + 490.0, + 227.0, + 490.0, + 264.0, + 296.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1739.0, + 1407.0, + 1739.0, + 1407.0, + 1776.0, + 293.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1808.0, + 320.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1796.0, + 398.0, + 1796.0, + 398.0, + 1837.0, + 320.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1455.0, + 1405.0, + 1455.0, + 1405.0, + 1491.0, + 297.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1486.0, + 1121.0, + 1486.0, + 1121.0, + 1521.0, + 324.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1541.0, + 1404.0, + 1541.0, + 1404.0, + 1577.0, + 295.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1570.0, + 698.0, + 1570.0, + 698.0, + 1606.0, + 322.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 701.0, + 1406.0, + 701.0, + 1406.0, + 751.0, + 292.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 741.0, + 774.0, + 741.0, + 774.0, + 773.0, + 323.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1856.0, + 1404.0, + 1856.0, + 1404.0, + 1892.0, + 296.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1887.0, + 1180.0, + 1887.0, + 1180.0, + 1921.0, + 323.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1022.0, + 1403.0, + 1022.0, + 1403.0, + 1063.0, + 292.0, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1057.0, + 980.0, + 1057.0, + 980.0, + 1090.0, + 321.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 1404.0, + 1224.0, + 1404.0, + 1261.0, + 293.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1253.0, + 1407.0, + 1253.0, + 1407.0, + 1295.0, + 320.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1286.0, + 602.0, + 1286.0, + 602.0, + 1320.0, + 323.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1108.0, + 1403.0, + 1108.0, + 1403.0, + 1146.0, + 295.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1141.0, + 1404.0, + 1141.0, + 1404.0, + 1176.0, + 322.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1169.0, + 1024.0, + 1169.0, + 1024.0, + 1208.0, + 324.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1377.0, + 294.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1372.0, + 1402.0, + 1372.0, + 1402.0, + 1407.0, + 324.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1402.0, + 1054.0, + 1402.0, + 1054.0, + 1440.0, + 323.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 911.0, + 1403.0, + 911.0, + 1403.0, + 946.0, + 295.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 938.0, + 1404.0, + 938.0, + 1404.0, + 978.0, + 320.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 970.0, + 932.0, + 970.0, + 932.0, + 1008.0, + 323.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1938.0, + 1402.0, + 1938.0, + 1402.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2007.0, + 323.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1999.0, + 530.0, + 1999.0, + 530.0, + 2038.0, + 322.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 280.0, + 1403.0, + 280.0, + 1403.0, + 313.0, + 294.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 310.0, + 1405.0, + 310.0, + 1405.0, + 343.0, + 323.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 339.0, + 1406.0, + 339.0, + 1406.0, + 377.0, + 319.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 371.0, + 397.0, + 371.0, + 397.0, + 403.0, + 321.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 506.0, + 1403.0, + 506.0, + 1403.0, + 547.0, + 293.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 541.0, + 1114.0, + 541.0, + 1114.0, + 575.0, + 322.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 425.0, + 1404.0, + 425.0, + 1404.0, + 461.0, + 296.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 454.0, + 1037.0, + 454.0, + 1037.0, + 490.0, + 321.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 794.0, + 1403.0, + 794.0, + 1403.0, + 829.0, + 294.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 823.0, + 1406.0, + 823.0, + 1406.0, + 864.0, + 321.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 856.0, + 676.0, + 856.0, + 676.0, + 890.0, + 323.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 590.0, + 1405.0, + 590.0, + 1405.0, + 633.0, + 293.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 625.0, + 1405.0, + 625.0, + 1405.0, + 662.0, + 323.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 655.0, + 656.0, + 655.0, + 656.0, + 690.0, + 320.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1624.0, + 1405.0, + 1624.0, + 1405.0, + 1661.0, + 296.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1654.0, + 1404.0, + 1654.0, + 1404.0, + 1692.0, + 322.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1687.0, + 505.0, + 1687.0, + 505.0, + 1718.0, + 324.0, + 1718.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 296, + 1118, + 1399, + 1118, + 1399, + 1184, + 296, + 1184 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 296, + 1481, + 1400, + 1481, + 1400, + 1546, + 296, + 1546 + ], + "score": 0.8 + }, + { + "category_id": 1, + "poly": [ + 295, + 1398, + 1401, + 1398, + 1401, + 1464, + 295, + 1464 + ], + "score": 0.798 + }, + { + "category_id": 1, + "poly": [ + 301, + 922, + 1399, + 922, + 1399, + 988, + 301, + 988 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 294, + 787, + 1298, + 787, + 1298, + 823, + 294, + 823 + ], + "score": 0.792 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 1400, + 229, + 1400, + 323, + 297, + 323 + ], + "score": 0.791 + }, + { + "category_id": 1, + "poly": [ + 298, + 704, + 1404, + 704, + 1404, + 769, + 298, + 769 + ], + "score": 0.778 + }, + { + "category_id": 1, + "poly": [ + 295, + 1647, + 1400, + 1647, + 1400, + 1712, + 295, + 1712 + ], + "score": 0.767 + }, + { + "category_id": 1, + "poly": [ + 298, + 839, + 1397, + 839, + 1397, + 905, + 298, + 905 + ], + "score": 0.766 + }, + { + "category_id": 1, + "poly": [ + 296, + 1201, + 1397, + 1201, + 1397, + 1268, + 296, + 1268 + ], + "score": 0.765 + }, + { + "category_id": 1, + "poly": [ + 297, + 1565, + 1401, + 1565, + 1401, + 1628, + 297, + 1628 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 298, + 1005, + 1403, + 1005, + 1403, + 1101, + 298, + 1101 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 295, + 621, + 1401, + 621, + 1401, + 687, + 295, + 687 + ], + "score": 0.718 + }, + { + "category_id": 1, + "poly": [ + 297, + 1284, + 1404, + 1284, + 1404, + 1380, + 297, + 1380 + ], + "score": 0.717 + }, + { + "category_id": 1, + "poly": [ + 298, + 507, + 1400, + 507, + 1400, + 603, + 298, + 603 + ], + "score": 0.657 + }, + { + "category_id": 1, + "poly": [ + 298, + 1957, + 1402, + 1957, + 1402, + 2022, + 298, + 2022 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 294, + 342, + 1402, + 342, + 1402, + 407, + 294, + 407 + ], + "score": 0.639 + }, + { + "category_id": 1, + "poly": [ + 295, + 424, + 1402, + 424, + 1402, + 490, + 295, + 490 + ], + "score": 0.635 + }, + { + "category_id": 1, + "poly": [ + 295, + 1844, + 1400, + 1844, + 1400, + 1939, + 295, + 1939 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 295, + 1729, + 1404, + 1729, + 1404, + 1825, + 295, + 1825 + ], + "score": 0.456 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1113.0, + 1405.0, + 1113.0, + 1405.0, + 1159.0, + 292.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1151.0, + 565.0, + 1151.0, + 565.0, + 1185.0, + 321.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1483.0, + 1402.0, + 1483.0, + 1402.0, + 1519.0, + 296.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1512.0, + 1076.0, + 1512.0, + 1076.0, + 1547.0, + 321.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1397.0, + 1406.0, + 1397.0, + 1406.0, + 1437.0, + 294.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1431.0, + 875.0, + 1431.0, + 875.0, + 1463.0, + 322.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 921.0, + 1403.0, + 921.0, + 1403.0, + 960.0, + 296.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 955.0, + 1308.0, + 955.0, + 1308.0, + 989.0, + 322.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 788.0, + 1296.0, + 788.0, + 1296.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 226.0, + 1404.0, + 226.0, + 1404.0, + 266.0, + 295.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 257.0, + 1405.0, + 257.0, + 1405.0, + 300.0, + 321.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 292.0, + 503.0, + 292.0, + 503.0, + 322.0, + 322.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 703.0, + 1406.0, + 703.0, + 1406.0, + 742.0, + 295.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 737.0, + 966.0, + 737.0, + 966.0, + 770.0, + 318.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1645.0, + 1402.0, + 1645.0, + 1402.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1680.0, + 938.0, + 1680.0, + 938.0, + 1712.0, + 321.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 838.0, + 1403.0, + 838.0, + 1403.0, + 876.0, + 294.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 873.0, + 1375.0, + 873.0, + 1375.0, + 906.0, + 325.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1202.0, + 1401.0, + 1202.0, + 1401.0, + 1238.0, + 296.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1235.0, + 1327.0, + 1235.0, + 1327.0, + 1268.0, + 321.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1602.0, + 293.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1593.0, + 400.0, + 1593.0, + 400.0, + 1631.0, + 318.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1002.0, + 1403.0, + 1002.0, + 1403.0, + 1044.0, + 292.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1035.0, + 1407.0, + 1035.0, + 1407.0, + 1071.0, + 321.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1065.0, + 465.0, + 1065.0, + 465.0, + 1101.0, + 321.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 622.0, + 1405.0, + 622.0, + 1405.0, + 658.0, + 297.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 654.0, + 1264.0, + 654.0, + 1264.0, + 688.0, + 322.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1320.0, + 295.0, + 1320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1352.0, + 322.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1347.0, + 542.0, + 1347.0, + 542.0, + 1382.0, + 324.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 507.0, + 1406.0, + 507.0, + 1406.0, + 545.0, + 294.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 538.0, + 1404.0, + 538.0, + 1404.0, + 577.0, + 323.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 570.0, + 584.0, + 570.0, + 584.0, + 605.0, + 323.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1955.0, + 1404.0, + 1955.0, + 1404.0, + 1995.0, + 297.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1988.0, + 856.0, + 1988.0, + 856.0, + 2023.0, + 324.0, + 2023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 342.0, + 1404.0, + 342.0, + 1404.0, + 378.0, + 296.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 373.0, + 1103.0, + 373.0, + 1103.0, + 409.0, + 323.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 425.0, + 1404.0, + 425.0, + 1404.0, + 461.0, + 297.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 457.0, + 1213.0, + 457.0, + 1213.0, + 491.0, + 320.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1843.0, + 1404.0, + 1843.0, + 1404.0, + 1881.0, + 295.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1874.0, + 1404.0, + 1874.0, + 1404.0, + 1912.0, + 323.0, + 1912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1906.0, + 888.0, + 1906.0, + 888.0, + 1940.0, + 324.0, + 1940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1729.0, + 1404.0, + 1729.0, + 1404.0, + 1767.0, + 293.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1799.0, + 322.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1792.0, + 722.0, + 1792.0, + 722.0, + 1826.0, + 322.0, + 1826.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 433, + 1405, + 433, + 1405, + 589, + 298, + 589 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 735, + 1403, + 735, + 1403, + 891, + 298, + 891 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 346, + 1354, + 1354, + 1354, + 1354, + 1640, + 346, + 1640 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 300, + 1941, + 1403, + 1941, + 1403, + 2035, + 300, + 2035 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 344, + 937, + 1355, + 937, + 1355, + 1224, + 344, + 1224 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 296, + 1741, + 1405, + 1741, + 1405, + 1927, + 296, + 1927 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 648, + 667, + 1050, + 667, + 1050, + 706, + 648, + 706 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 297, + 601, + 1402, + 601, + 1402, + 664, + 297, + 664 + ], + "score": 0.936 + }, + { + "category_id": 4, + "poly": [ + 607, + 1675, + 1094, + 1675, + 1094, + 1709, + 607, + 1709 + ], + "score": 0.92 + }, + { + "category_id": 4, + "poly": [ + 608, + 1256, + 1093, + 1256, + 1093, + 1290, + 608, + 1290 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1366, + 673, + 1399, + 673, + 1399, + 703, + 1366, + 703 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 300, + 373, + 627, + 373, + 627, + 407, + 300, + 407 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.851 + }, + { + "category_id": 0, + "poly": [ + 301, + 227, + 449, + 227, + 449, + 262, + 301, + 262 + ], + "score": 0.678 + }, + { + "category_id": 1, + "poly": [ + 303, + 304, + 754, + 304, + 754, + 339, + 303, + 339 + ], + "score": 0.498 + }, + { + "category_id": 0, + "poly": [ + 303, + 304, + 754, + 304, + 754, + 339, + 303, + 339 + ], + "score": 0.454 + }, + { + "category_id": 1, + "poly": [ + 301, + 227, + 449, + 227, + 449, + 262, + 301, + 262 + ], + "score": 0.221 + }, + { + "category_id": 13, + "poly": [ + 1156, + 524, + 1304, + 524, + 1304, + 559, + 1156, + 559 + ], + "score": 0.93, + "latex": "\\lVert \\bf { X } - \\bf { X } ^ { \\mathrm { { r e c } } } \\rVert ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1233, + 734, + 1400, + 734, + 1400, + 769, + 1233, + 769 + ], + "score": 0.93, + "latex": "\\mathbf { m } = \\{ 0 , 1 \\} ^ { N }" + }, + { + "category_id": 14, + "poly": [ + 646, + 665, + 1054, + 665, + 1054, + 705, + 646, + 705 + ], + "score": 0.9, + "latex": "{ \\bf X } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf X } ^ { \\mathrm { p o o l } } ; ~ { \\bf A } ^ { \\mathrm { r e c } } = { \\bf S } { \\bf A } ^ { \\mathrm { p o o l } } { \\bf S } ^ { T } ." + }, + { + "category_id": 13, + "poly": [ + 1147, + 797, + 1342, + 797, + 1342, + 832, + 1147, + 832 + ], + "score": 0.9, + "latex": "\\bar { \\mathbf { U } } = [ \\mathbf { I } _ { N } ] _ { : , \\mathbf { m } = = 1 }" + }, + { + "category_id": 13, + "poly": [ + 788, + 798, + 825, + 798, + 825, + 829, + 788, + 829 + ], + "score": 0.88, + "latex": "{ \\mathbf { I } } _ { N }" + }, + { + "category_id": 13, + "poly": [ + 462, + 2003, + 561, + 2003, + 561, + 2035, + 462, + 2035 + ], + "score": 0.88, + "latex": "\\mathbf { s } = \\mathbf { X } \\mathbf { p } " + }, + { + "category_id": 13, + "poly": [ + 776, + 556, + 830, + 556, + 830, + 586, + 776, + 586 + ], + "score": 0.88, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 629, + 1774, + 681, + 1774, + 681, + 1802, + 629, + 1802 + ], + "score": 0.88, + "latex": "\\mathbf { X } ^ { \\mathrm { r e c } }" + }, + { + "category_id": 13, + "poly": [ + 847, + 1945, + 874, + 1945, + 874, + 1970, + 847, + 1970 + ], + "score": 0.82, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 430, + 1974, + 460, + 1974, + 460, + 2000, + 430, + 2000 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 350, + 738, + 379, + 738, + 379, + 765, + 350, + 765 + ], + "score": 0.77, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 524, + 1865, + 553, + 1865, + 553, + 1893, + 524, + 1893 + ], + "score": 0.76, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1105, + 801, + 1136, + 801, + 1136, + 826, + 1105, + 826 + ], + "score": 0.47, + "latex": "\\mathbf { m }" + }, + { + "category_id": 13, + "poly": [ + 994, + 466, + 1021, + 466, + 1021, + 493, + 994, + 493 + ], + "score": 0.47, + "latex": "\\mathbf { X }" + }, + { + "category_id": 13, + "poly": [ + 1251, + 769, + 1279, + 769, + 1279, + 795, + 1251, + 795 + ], + "score": 0.47, + "latex": "\\mathbf { U }" + }, + { + "category_id": 13, + "poly": [ + 904, + 829, + 933, + 829, + 933, + 857, + 904, + 857 + ], + "score": 0.38, + "latex": "\\mathbf { U }" + }, + { + "category_id": 13, + "poly": [ + 808, + 1743, + 838, + 1743, + 838, + 1771, + 808, + 1771 + ], + "score": 0.27, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1409.0, + 892.0, + 1409.0, + 892.0, + 1468.0, + 876.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 1460.0, + 1147.0, + 1460.0, + 1147.0, + 1495.0, + 1135.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1486.0, + 890.0, + 1486.0, + 890.0, + 1510.0, + 879.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1499.0, + 1148.0, + 1499.0, + 1148.0, + 1508.0, + 1137.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1491.0, + 1328.0, + 1491.0, + 1328.0, + 1507.0, + 1319.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1505.0, + 724.0, + 1505.0, + 724.0, + 1525.0, + 713.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1505.0, + 1145.0, + 1505.0, + 1145.0, + 1534.0, + 1134.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1566.0, + 1209.0, + 1566.0, + 1209.0, + 1581.0, + 1142.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1605.0, + 532.0, + 1605.0, + 532.0, + 1642.0, + 398.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1602.0, + 781.0, + 1602.0, + 781.0, + 1643.0, + 659.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1602.0, + 1047.0, + 1602.0, + 1047.0, + 1643.0, + 909.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1604.0, + 1326.0, + 1604.0, + 1326.0, + 1643.0, + 1142.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 952.0, + 1259.0, + 952.0, + 1259.0, + 962.0, + 1245.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 998.0, + 887.0, + 998.0, + 887.0, + 1020.0, + 875.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1004.0, + 896.0, + 1004.0, + 896.0, + 1014.0, + 888.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1033.0, + 887.0, + 1033.0, + 887.0, + 1060.0, + 865.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1119.0, + 920.0, + 1119.0, + 920.0, + 1140.0, + 900.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1052.0, + 1123.0, + 1061.0, + 1123.0, + 1061.0, + 1133.0, + 1052.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1134.0, + 937.0, + 1134.0, + 937.0, + 1153.0, + 914.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1185.0, + 534.0, + 1185.0, + 534.0, + 1225.0, + 398.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1183.0, + 784.0, + 1183.0, + 784.0, + 1227.0, + 661.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1185.0, + 1047.0, + 1185.0, + 1047.0, + 1224.0, + 908.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1185.0, + 1326.0, + 1185.0, + 1326.0, + 1227.0, + 1141.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 963.5, + 1299.0, + 963.5, + 1299.0, + 976.5, + 1275.0, + 976.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 955.5, + 980.0, + 955.5, + 980.0, + 984.5, + 875.0, + 984.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1018.5, + 893.0, + 1018.5, + 893.0, + 1041.5, + 884.0, + 1041.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1079.0, + 889.0, + 1079.0, + 889.0, + 1093.0, + 880.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1090.5, + 1096.0, + 1090.5, + 1096.0, + 1099.0, + 1069.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.75, + 1109.0, + 1079.75, + 1109.0, + 1079.75, + 1118.0, + 1058.75, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1669.0, + 1095.0, + 1669.0, + 1095.0, + 1714.0, + 602.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1250.0, + 1096.0, + 1250.0, + 1096.0, + 1295.0, + 601.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 373.0, + 632.0, + 373.0, + 632.0, + 410.0, + 296.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 227.0, + 453.0, + 227.0, + 453.0, + 268.0, + 297.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 306.0, + 755.0, + 306.0, + 755.0, + 343.0, + 297.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 434.0, + 1403.0, + 434.0, + 1403.0, + 467.0, + 297.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 463.0, + 993.0, + 463.0, + 993.0, + 499.0, + 293.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 463.0, + 1406.0, + 463.0, + 1406.0, + 499.0, + 1022.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 494.0, + 1405.0, + 494.0, + 1405.0, + 531.0, + 295.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 524.0, + 1155.0, + 524.0, + 1155.0, + 560.0, + 292.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1305.0, + 524.0, + 1406.0, + 524.0, + 1406.0, + 560.0, + 1305.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 556.0, + 775.0, + 556.0, + 775.0, + 592.0, + 293.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 556.0, + 1075.0, + 556.0, + 1075.0, + 592.0, + 831.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 730.0, + 349.0, + 730.0, + 349.0, + 773.0, + 292.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 730.0, + 1232.0, + 730.0, + 1232.0, + 773.0, + 380.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 730.0, + 1405.0, + 730.0, + 1405.0, + 773.0, + 1401.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 1250.0, + 766.0, + 1250.0, + 804.0, + 294.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 766.0, + 1404.0, + 766.0, + 1404.0, + 804.0, + 1280.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 793.0, + 787.0, + 793.0, + 787.0, + 837.0, + 292.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 793.0, + 1104.0, + 793.0, + 1104.0, + 837.0, + 826.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 793.0, + 1146.0, + 793.0, + 1146.0, + 837.0, + 1137.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 793.0, + 1407.0, + 793.0, + 1407.0, + 837.0, + 1343.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 826.0, + 903.0, + 826.0, + 903.0, + 866.0, + 292.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 826.0, + 1407.0, + 826.0, + 1407.0, + 866.0, + 934.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 858.0, + 1227.0, + 858.0, + 1227.0, + 896.0, + 293.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 846.0, + 1941.0, + 846.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1941.0, + 1407.0, + 1941.0, + 1407.0, + 1978.0, + 875.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 429.0, + 1970.0, + 429.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1970.0, + 1405.0, + 1970.0, + 1405.0, + 2009.0, + 461.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2003.0, + 461.0, + 2003.0, + 461.0, + 2036.0, + 295.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2036.0, + 562.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 807.0, + 1740.0, + 807.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1779.0, + 839.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1769.0, + 628.0, + 1769.0, + 628.0, + 1811.0, + 291.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 1769.0, + 1405.0, + 1769.0, + 1405.0, + 1811.0, + 682.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1839.0, + 294.0, + 1839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1834.0, + 1405.0, + 1834.0, + 1405.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1861.0, + 523.0, + 1861.0, + 523.0, + 1902.0, + 291.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1861.0, + 1406.0, + 1861.0, + 1406.0, + 1902.0, + 554.0, + 1902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1894.0, + 704.0, + 1894.0, + 704.0, + 1930.0, + 294.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 601.0, + 1404.0, + 601.0, + 1404.0, + 640.0, + 292.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 634.0, + 512.0, + 634.0, + 512.0, + 668.0, + 292.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 306.0, + 755.0, + 306.0, + 755.0, + 343.0, + 297.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 227.0, + 453.0, + 227.0, + 453.0, + 268.0, + 297.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 760, + 1403, + 760, + 1403, + 1035, + 297, + 1035 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 530, + 1403, + 530, + 1403, + 744, + 298, + 744 + ], + "score": 0.983 + }, + { + "category_id": 5, + "poly": [ + 443, + 1252, + 1255, + 1252, + 1255, + 1567, + 443, + 1567 + ], + "score": 0.983, + "html": "
PropertyTop-KDiffpoolminCUTpoolFlat baseline
mu0.600±0.0850.651±0.0260.538±0.0120.559±0.007
alpha0.197±0.0870.114±0.0010.078±0.0070.065±0.006
homo0.698±0.1020.712±0.0150.526±0.0210.435±0.013
lumo0.601±0.0500.646±0.0130.540±0.0050.515±0.007
gap0.630±0.0440.698±0.0040.584±0.0070.552±0.008
r20.452±0.0870.440±0.0240.261±0.0060.204±0.006
zpve0.402±0.0320.410±0.0040.328±0.0050.284±0.005
uO_atom0.308±0.0550.245±0.0060.193±0.0020.163±0.001
cv0.291±0.1180.337±0.0180.148±0.0040.127±0.002
" + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 414, + 298, + 414 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1051, + 1402, + 1051, + 1402, + 1205, + 299, + 1205 + ], + "score": 0.975 + }, + { + "category_id": 2, + "poly": [ + 330, + 1943, + 886, + 1943, + 886, + 2035, + 330, + 2035 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 294, + 1830, + 1401, + 1830, + 1401, + 1893, + 294, + 1893 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 302, + 1753, + 709, + 1753, + 709, + 1787, + 302, + 1787 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.859 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.855 + }, + { + "category_id": 1, + "poly": [ + 291, + 1599, + 1405, + 1599, + 1405, + 1662, + 291, + 1662 + ], + "score": 0.84 + }, + { + "category_id": 0, + "poly": [ + 300, + 468, + 1071, + 468, + 1071, + 498, + 300, + 498 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 300, + 468, + 1071, + 468, + 1071, + 498, + 300, + 498 + ], + "score": 0.206 + }, + { + "category_id": 6, + "poly": [ + 291, + 1599, + 1405, + 1599, + 1405, + 1662, + 291, + 1662 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 931, + 592, + 1039, + 592, + 1039, + 626, + 931, + 626 + ], + "score": 0.93, + "latex": "\\{ \\mathbf { A } _ { i } , \\mathbf { X } _ { i } \\}" + }, + { + "category_id": 13, + "poly": [ + 554, + 943, + 608, + 943, + 608, + 973, + 554, + 973 + ], + "score": 0.88, + "latex": "1 \\dot { 0 } \\%" + }, + { + "category_id": 13, + "poly": [ + 823, + 532, + 905, + 532, + 905, + 561, + 823, + 561 + ], + "score": 0.88, + "latex": "{ \\approx } 1 3 5 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 1297, + 1600, + 1402, + 1600, + 1402, + 1631, + 1297, + 1631 + ], + "score": 0.86, + "latex": "p < 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 1275, + 293, + 1303, + 293, + 1303, + 319, + 1275, + 319 + ], + "score": 0.77, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1373, + 322, + 1402, + 322, + 1402, + 349, + 1373, + 349 + ], + "score": 0.76, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 903, + 262, + 933, + 262, + 933, + 290, + 903, + 290 + ], + "score": 0.76, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 350, + 1144, + 379, + 1144, + 379, + 1171, + 350, + 1171 + ], + "score": 0.74, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 693, + 884, + 721, + 884, + 721, + 911, + 693, + 911 + ], + "score": 0.72, + "latex": "K" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1939.0, + 887.0, + 1939.0, + 887.0, + 1977.0, + 331.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1969.0, + 856.0, + 1969.0, + 856.0, + 2007.0, + 329.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1999.0, + 857.0, + 1999.0, + 857.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1751.0, + 714.0, + 1751.0, + 714.0, + 1792.0, + 294.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 466.0, + 1074.0, + 466.0, + 1074.0, + 502.0, + 294.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1599.0, + 1296.0, + 1599.0, + 1296.0, + 1634.0, + 295.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1629.0, + 1326.0, + 1629.0, + 1326.0, + 1662.0, + 294.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 762.0, + 1404.0, + 762.0, + 1404.0, + 795.0, + 296.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 791.0, + 1406.0, + 791.0, + 1406.0, + 827.0, + 293.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 821.0, + 1404.0, + 821.0, + 1404.0, + 858.0, + 292.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 852.0, + 1405.0, + 852.0, + 1405.0, + 888.0, + 292.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 882.0, + 692.0, + 882.0, + 692.0, + 917.0, + 293.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 882.0, + 1406.0, + 882.0, + 1406.0, + 917.0, + 722.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 913.0, + 1405.0, + 913.0, + 1405.0, + 947.0, + 292.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 943.0, + 553.0, + 943.0, + 553.0, + 976.0, + 295.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 943.0, + 1403.0, + 943.0, + 1403.0, + 976.0, + 609.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 973.0, + 1405.0, + 973.0, + 1405.0, + 1009.0, + 295.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1004.0, + 811.0, + 1004.0, + 811.0, + 1041.0, + 295.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 529.0, + 822.0, + 529.0, + 822.0, + 566.0, + 294.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 529.0, + 1403.0, + 529.0, + 1403.0, + 566.0, + 906.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 560.0, + 1403.0, + 560.0, + 1403.0, + 595.0, + 292.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 593.0, + 930.0, + 593.0, + 930.0, + 627.0, + 294.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 593.0, + 1405.0, + 593.0, + 1405.0, + 627.0, + 1040.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 622.0, + 1405.0, + 622.0, + 1405.0, + 657.0, + 293.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 649.0, + 1405.0, + 649.0, + 1405.0, + 690.0, + 292.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 682.0, + 1405.0, + 682.0, + 1405.0, + 719.0, + 293.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 714.0, + 848.0, + 714.0, + 848.0, + 749.0, + 294.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 231.0, + 1403.0, + 231.0, + 1403.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 263.0, + 902.0, + 263.0, + 902.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 295.0, + 934.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1274.0, + 291.0, + 1274.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 291.0, + 1405.0, + 291.0, + 1405.0, + 327.0, + 1304.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 323.0, + 1372.0, + 323.0, + 1372.0, + 355.0, + 296.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 352.0, + 1405.0, + 352.0, + 1405.0, + 388.0, + 294.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 383.0, + 1076.0, + 383.0, + 1076.0, + 416.0, + 293.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1049.0, + 1406.0, + 1049.0, + 1406.0, + 1087.0, + 294.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1080.0, + 1406.0, + 1080.0, + 1406.0, + 1117.0, + 295.0, + 1117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1109.0, + 1406.0, + 1109.0, + 1406.0, + 1150.0, + 293.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1140.0, + 349.0, + 1140.0, + 349.0, + 1178.0, + 294.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1140.0, + 1404.0, + 1140.0, + 1404.0, + 1178.0, + 380.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1173.0, + 1337.0, + 1173.0, + 1337.0, + 1210.0, + 295.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1825.0, + 1406.0, + 1825.0, + 1406.0, + 1869.0, + 292.0, + 1869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1858.0, + 1054.0, + 1858.0, + 1054.0, + 1897.0, + 294.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1599.0, + 1296.0, + 1599.0, + 1296.0, + 1634.0, + 295.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1629.0, + 1326.0, + 1629.0, + 1326.0, + 1662.0, + 294.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 466.0, + 1074.0, + 466.0, + 1074.0, + 502.0, + 294.0, + 502.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 385, + 1162, + 1314, + 1162, + 1314, + 1421, + 385, + 1421 + ], + "score": 0.983, + "html": "
Datasetsamplesclassesavg. nodesavg. edgesnode attr.node labels
Bench-easy18003147.82922.66yes
Bench-hard18003148.32572.32yes
Mutagenicity4337230.3230.77yes
Proteins1113239.0672.821no
DD11782284.32715.661yes
COLLAB5000374.492457.78no
Reddit-2K20002429.63497.751no
" + }, + { + "category_id": 1, + "poly": [ + 297, + 794, + 1406, + 794, + 1406, + 1071, + 297, + 1071 + ], + "score": 0.981 + }, + { + "category_id": 5, + "poly": [ + 531, + 530, + 1165, + 530, + 1165, + 679, + 531, + 679 + ], + "score": 0.976, + "html": "
DatasetNodesEdgesNode featuresNode classes
Cora2708542914337
Citeseer3327922837036
Pubmed19717886515003
" + }, + { + "category_id": 1, + "poly": [ + 298, + 285, + 1403, + 285, + 1403, + 444, + 298, + 444 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 296, + 1558, + 1404, + 1558, + 1404, + 1683, + 296, + 1683 + ], + "score": 0.967 + }, + { + "category_id": 2, + "poly": [ + 298, + 1947, + 1352, + 1947, + 1352, + 2033, + 298, + 2033 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 299, + 738, + 661, + 738, + 661, + 771, + 299, + 771 + ], + "score": 0.916 + }, + { + "category_id": 6, + "poly": [ + 482, + 1125, + 1217, + 1125, + 1217, + 1158, + 482, + 1158 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 301, + 1490, + 761, + 1490, + 761, + 1525, + 301, + 1525 + ], + "score": 0.904 + }, + { + "category_id": 6, + "poly": [ + 577, + 498, + 1121, + 498, + 1121, + 528, + 577, + 528 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 300, + 229, + 826, + 229, + 826, + 262, + 300, + 262 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.862 + }, + { + "category_id": 13, + "poly": [ + 767, + 348, + 823, + 348, + 823, + 380, + 767, + 380 + ], + "score": 0.84, + "latex": "\\mathbf { X } ^ { ( 1 ) }" + }, + { + "category_id": 13, + "poly": [ + 843, + 798, + 877, + 798, + 877, + 828, + 843, + 828 + ], + "score": 0.82, + "latex": "\\mathrm { L } _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 714, + 858, + 745, + 858, + 745, + 886, + 714, + 886 + ], + "score": 0.69, + "latex": "{ \\it 2 k }" + }, + { + "category_id": 13, + "poly": [ + 364, + 828, + 409, + 828, + 409, + 857, + 364, + 857 + ], + "score": 0.59, + "latex": "( H )" + }, + { + "category_id": 13, + "poly": [ + 675, + 918, + 705, + 918, + 705, + 946, + 675, + 946 + ], + "score": 0.3, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1941.0, + 1355.0, + 1941.0, + 1355.0, + 1982.0, + 329.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1971.0, + 1204.0, + 1971.0, + 1204.0, + 2015.0, + 329.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 2005.0, + 514.0, + 2005.0, + 514.0, + 2034.0, + 297.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 737.0, + 664.0, + 737.0, + 664.0, + 773.0, + 294.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 1124.0, + 1216.0, + 1124.0, + 1216.0, + 1160.0, + 484.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1487.0, + 766.0, + 1487.0, + 766.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 494.0, + 1124.0, + 494.0, + 1124.0, + 531.0, + 574.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 229.0, + 830.0, + 229.0, + 830.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 795.0, + 842.0, + 795.0, + 842.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 795.0, + 1406.0, + 795.0, + 1406.0, + 832.0, + 878.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 824.0, + 363.0, + 824.0, + 363.0, + 863.0, + 294.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 824.0, + 1404.0, + 824.0, + 1404.0, + 863.0, + 410.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 857.0, + 713.0, + 857.0, + 713.0, + 892.0, + 295.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 857.0, + 1404.0, + 857.0, + 1404.0, + 892.0, + 746.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 885.0, + 1406.0, + 885.0, + 1406.0, + 922.0, + 294.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 917.0, + 674.0, + 917.0, + 674.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 917.0, + 1407.0, + 917.0, + 1407.0, + 953.0, + 706.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 948.0, + 1406.0, + 948.0, + 1406.0, + 985.0, + 295.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 980.0, + 1404.0, + 980.0, + 1404.0, + 1012.0, + 296.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1007.0, + 1407.0, + 1007.0, + 1407.0, + 1045.0, + 292.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1039.0, + 548.0, + 1039.0, + 548.0, + 1072.0, + 293.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 286.0, + 1405.0, + 286.0, + 1405.0, + 320.0, + 294.0, + 320.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 319.0, + 1404.0, + 319.0, + 1404.0, + 352.0, + 296.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 347.0, + 766.0, + 347.0, + 766.0, + 387.0, + 292.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 347.0, + 1405.0, + 347.0, + 1405.0, + 387.0, + 824.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 378.0, + 1405.0, + 378.0, + 1405.0, + 417.0, + 293.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 412.0, + 828.0, + 412.0, + 828.0, + 445.0, + 296.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1560.0, + 1404.0, + 1560.0, + 1404.0, + 1592.0, + 295.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 1404.0, + 1590.0, + 1404.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1657.0, + 295.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1652.0, + 933.0, + 1652.0, + 933.0, + 1685.0, + 295.0, + 1685.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 509, + 230, + 1190, + 230, + 1190, + 522, + 509, + 522 + ], + "score": 0.969 + }, + { + "category_id": 3, + "poly": [ + 393, + 939, + 1305, + 939, + 1305, + 1063, + 393, + 1063 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 379, + 1194, + 1319, + 1194, + 1319, + 1321, + 379, + 1321 + ], + "score": 0.963 + }, + { + "category_id": 3, + "poly": [ + 502, + 1452, + 1194, + 1452, + 1194, + 1575, + 502, + 1575 + ], + "score": 0.953 + }, + { + "category_id": 3, + "poly": [ + 682, + 671, + 1007, + 671, + 1007, + 797, + 682, + 797 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 602, + 558, + 1097, + 558, + 1097, + 594, + 602, + 594 + ], + "score": 0.92 + }, + { + "category_id": 4, + "poly": [ + 604, + 1347, + 1096, + 1347, + 1096, + 1381, + 604, + 1381 + ], + "score": 0.919 + }, + { + "category_id": 4, + "poly": [ + 583, + 1089, + 1117, + 1089, + 1117, + 1124, + 583, + 1124 + ], + "score": 0.916 + }, + { + "category_id": 4, + "poly": [ + 600, + 1603, + 1098, + 1603, + 1098, + 1637, + 600, + 1637 + ], + "score": 0.916 + }, + { + "category_id": 4, + "poly": [ + 566, + 834, + 1132, + 834, + 1132, + 868, + 566, + 868 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 104, + 297, + 104 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 15, + "poly": [ + 886.0, + 261.0, + 939.0, + 261.0, + 939.0, + 290.0, + 886.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 284.0, + 927.0, + 284.0, + 927.0, + 328.0, + 891.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 329.0, + 636.0, + 329.0, + 636.0, + 380.0, + 556.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 333.0, + 838.0, + 333.0, + 838.0, + 367.0, + 757.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 322.0, + 922.0, + 322.0, + 922.0, + 367.0, + 869.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 335.0, + 1029.0, + 335.0, + 1029.0, + 354.0, + 1012.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 332.0, + 1178.0, + 332.0, + 1178.0, + 367.0, + 1131.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 374.0, + 931.0, + 374.0, + 931.0, + 424.0, + 885.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 422.0, + 544.0, + 422.0, + 544.0, + 457.0, + 512.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 455.0, + 535.0, + 455.0, + 535.0, + 469.0, + 522.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 489.0, + 972.0, + 489.0, + 972.0, + 523.0, + 839.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 233.5, + 955.0, + 233.5, + 955.0, + 279.5, + 861.0, + 279.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 315.0, + 567.0, + 315.0, + 567.0, + 357.0, + 501.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 316.0, + 749.0, + 316.0, + 749.0, + 355.0, + 658.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 411.0, + 940.0, + 411.0, + 940.0, + 451.0, + 877.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 965.0, + 524.0, + 965.0, + 524.0, + 1047.0, + 389.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 984.0, + 630.0, + 984.0, + 630.0, + 1019.0, + 565.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 984.0, + 726.0, + 984.0, + 726.0, + 1019.0, + 632.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 986.0, + 829.0, + 986.0, + 829.0, + 1018.0, + 739.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 983.0, + 929.0, + 983.0, + 929.0, + 1019.0, + 840.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 968.0, + 1059.0, + 968.0, + 1059.0, + 1034.0, + 947.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 988.0, + 1095.0, + 988.0, + 1095.0, + 1014.0, + 1069.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 970.0, + 1216.0, + 970.0, + 1216.0, + 1036.0, + 1102.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 979.0, + 1311.0, + 979.0, + 1311.0, + 1026.0, + 1225.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 987.0, + 578.0, + 987.0, + 578.0, + 1015.5, + 531.0, + 1015.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 1222.0, + 513.0, + 1222.0, + 513.0, + 1303.0, + 376.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1241.0, + 611.0, + 1241.0, + 611.0, + 1275.0, + 555.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1240.0, + 660.0, + 1240.0, + 660.0, + 1277.0, + 614.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1240.0, + 733.0, + 1240.0, + 733.0, + 1276.0, + 663.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 752.0, + 1239.0, + 1074.0, + 1239.0, + 1074.0, + 1278.0, + 752.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1226.0, + 1210.0, + 1226.0, + 1210.0, + 1293.0, + 1089.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1217.0, + 1239.0, + 1257.0, + 1239.0, + 1257.0, + 1277.0, + 1217.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 526.0, + 1242.5, + 544.0, + 1242.5, + 544.0, + 1272.5, + 526.0, + 1272.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1220.0, + 1330.0, + 1220.0, + 1330.0, + 1262.5, + 1247.0, + 1262.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 1481.0, + 580.0, + 1481.0, + 580.0, + 1556.0, + 502.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1497.0, + 635.0, + 1497.0, + 635.0, + 1533.0, + 581.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1499.0, + 746.0, + 1499.0, + 746.0, + 1533.0, + 650.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 759.0, + 1497.0, + 850.0, + 1497.0, + 850.0, + 1533.0, + 759.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1500.0, + 901.0, + 1500.0, + 901.0, + 1525.0, + 878.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1480.0, + 993.0, + 1480.0, + 993.0, + 1549.0, + 902.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1481.0, + 1121.0, + 1481.0, + 1121.0, + 1552.0, + 1004.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 1487.0, + 1205.0, + 1487.0, + 1205.0, + 1533.0, + 1121.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 697.0, + 725.0, + 697.0, + 725.0, + 780.0, + 679.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 701.0, + 1008.0, + 701.0, + 1008.0, + 739.0, + 976.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 713.0, + 761.0, + 713.0, + 761.0, + 752.0, + 716.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 716.0, + 814.0, + 716.0, + 814.0, + 752.0, + 761.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 717.0, + 927.0, + 717.0, + 927.0, + 752.0, + 862.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 717.0, + 975.0, + 717.0, + 975.0, + 750.0, + 940.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.25, + 717.5, + 875.25, + 717.5, + 875.25, + 749.5, + 831.25, + 749.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 553.0, + 1099.0, + 553.0, + 1099.0, + 601.0, + 598.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1346.0, + 1098.0, + 1346.0, + 1098.0, + 1383.0, + 600.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1087.0, + 1117.0, + 1087.0, + 1117.0, + 1124.0, + 580.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1598.0, + 1103.0, + 1598.0, + 1103.0, + 1643.0, + 594.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 827.0, + 1137.0, + 827.0, + 1137.0, + 875.0, + 561.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/Byf5-30qFX/Byf5-30qFX.md b/parse/train/Byf5-30qFX/Byf5-30qFX.md new file mode 100644 index 0000000000000000000000000000000000000000..ffac93e69c376e28a04cb3a5f77f7f2f2d29627b --- /dev/null +++ b/parse/train/Byf5-30qFX/Byf5-30qFX.md @@ -0,0 +1,271 @@ +# DHER: HINDSIGHT EXPERIENCE REPLAY FOR DYNAMIC GOALS + +Meng Fang∗, Cheng Zhou, Bei Shi, Boqing Gong, Jia Xu, Tong Zhang Tencent AI Lab + +# ABSTRACT + +Dealing with sparse rewards is one of the most important challenges in reinforcement learning (RL), especially when a goal is dynamic (e.g., to grasp a moving object). Hindsight experience replay (HER) has been shown an effective solution to handling sparse rewards with fixed goals. However, it does not account for dynamic goals in its vanilla form and, as a result, even degrades the performance of existing off-policy RL algorithms when the goal is changing over time. + +In this paper, we present Dynamic Hindsight Experience Replay (DHER), a novel approach for tasks with dynamic goals in the presence of sparse rewards. DHER automatically assembles successful experiences from two relevant failures and can be used to enhance an arbitrary off-policy RL algorithm when the tasks’ goals are dynamic. We evaluate DHER on tasks of robotic manipulation and moving object tracking, and transfer the polices from simulation to physical robots. Extensive comparison and ablation studies demonstrate the superiority of our approach, showing that DHER is a crucial ingredient to enable RL to solve tasks with dynamic goals in manipulation and grid world domains. + +# 1 INTRODUCTION + +Deep reinforcement learning has been shown an effective framework for solving a rich repertoire of complex control problems. In simulated domains, agents have been trained to perform a diverse array of challenging tasks (Mnih et al., 2015; Lillicrap et al., 2015; Duan et al., 2016). In order to train such agents, it is often the case that one has to design a reward function that not only reflects the task at hand but also is carefully shaped $\mathrm { N g }$ et al., 1999) to guide the policy optimization. Unfortunately, many of the capabilities demonstrated by reward engineering are often limited to specific tasks. Moreover, it requires both RL expertise and domain-specific knowledge to reshape the reward functions. For situations where we do not know what admissible behavior may look like, for example, using LEGO bricks to build a desired architecture, it is difficult to apply reward engineering. Therefore, it is essential to develop algorithms which can learn from unshaped and usually sparse reward signals. + +Learning with sparse rewards is challenging, especially when a goal is dynamic. Dynamic goals are common in games and planning problems, often addressed using reward shaping or search (Kaelbling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). However, the difficulty posed by a sparse reward is exacerbated by the complicated environment dynamics in robotics (Andrychowicz et al., 2017). For instance, system dynamics around contacts are difficult to model and induce sensitivity in the system to small errors. Many robotic tasks also need executing multiple steps successfully over a long horizon, involve enormous search space, and require generalization to varying task instances. Policy gradient methods are breakthroughs in the challenging environments, such as PPO (Heess et al., 2017; Schulman et al., 2017), ACER (Wang et al., 2016), TRPO (Schulman et al., 2015) and so on. They are used in environments, where an agent tries to reach a target, learns to walk, runs, and so on. Recently, sampling-efficient learning is introduced and demonstrates a significant increase in performance for off-policy actor-critic DQN (Mnih et al., 2015) and DDPG (Lillicrap et al., 2015) algorithms. Hindsight experience replay (HER) is very effective for improving the performance of off-policy RL algorithms in solving goal-based tasks with sparse rewards (Andrychowicz et al., 2017). Similar to UVFA (Schaul et al., 2015a), it takes a goal state as part of input. However, it assumes the goal is fixed. As a result, this assumption actually impedes the learning of RL agents in the environments of moving goals. + +![](images/a34ff9c3bf9148b8d43e21b58e1ea68fe1bcb07b8ac12d22cc17bf9a91fd4060.jpg) +Figure 1: The framework of DHER. DHER is a kind of experience replay method. It searches relevant failed experiences and then assembles them into successful experiences. + +In this paper, we address this challenge with a new method, Dynamic Hindsight Experience Replay (DHER), for accomplishing tasks with moving goals. We follow the multi-goal setting in UVFA (Schaul et al., 2015a) and HER (Andrychowicz et al., 2017). It assumes that the goal being pursued does not influence the environment dynamics. We also need to have the knowledge of goal similarity. For example, in manipulation or grid world domains, we can use Euclidean distance between positions to measure the goal similarity. HER turns a failed episode to a success by composing a new task whose goal is achieved by that episode. Our idea allows an agent to learn from the failure one step further than HER: the agent not only sets a new goal but also hallucinates how to reach the original goal from the new one. Take playing frisbee for instance. When an agent jumps to catch the frisbee and yet misses it, the agent receives no positive feedback under the sparse reward setting. Using HER, the agent could set the end of its episode as the new goal — position of the frisbee; with DHER, however, the agent finds a trajectory from its past experiences as the imagined path of the frisbee, and thereby extrapolates towards the original goal. + +In particular, we do the following for DHER. To finish the tasks with dynamic goals needs to explore experience and understand multiple goals. DHER uses replay buffers to allow the agent to learn from a couple of failures by assembling new ‘experience’ from different episodes. The proposed method retrieves memories to find the connection between the experience of different episodes. It largely improves the sample efficiency in dynamic goal task settings. More importantly, this strategy makes it possible to learn in the setting that both sparse rewards and dynamic goals exist. + +We evaluate our method along with the state-of-the-art baselines on new environments and manipulation tasks, which have sparse rewards and dynamic goals. Our results demonstrate that DHER is clearly better than others for these tasks. We also transfer policies trained in our simulation based on DHER to a physical robot and show that DHER can be applied to solving real-world robotics problems. + +We summarize our main contributions as follows: (1) We demonstrate that, both in simulation and real worlds, DHER succeeds in continuous control with a moving target. To our knowledge, this is the first empirical result on manipulation tasks that demonstrates model-free learning methods can tackle tasks of this complexity. (2) We show that with assembling new experience from two failures, the sample complexity can be reduced dramatically. We attribute this to global knowledge learning in a set of failed experience which breaks the constraint of local one-episode experience towards more robust strategies. (3) We design and implement a set of new environments and continuous tasks with dynamic goals, which would be of interest to researchers at the intersection of robotics and reinforcement learning.1 + +# 2 RELATED WORK + +Recent works in deep RL have shown impressive results in different domains, such as games (Bellemare et al., 2013; Silver et al., 2016), simulated control tasks (Brockman et al., 2016) and so on. There have been several proposed RL methods for playing Atari games, including DQN (Mnih et al., 2015), UVFA (Schaul et al., 2015b) and so on. UVFA trains a single neural network approximating multiple value functions for state and goal. For the continuous control, DDPG (Lillicrap et al., 2015) is a popular actor-critic algorithm that has shown impressive results in continuous control tasks. Dynamic goals appear in games and planning, often addressed by rewards shaping or search (Kaelbling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). When the rewards are sparse, there is few work studying the dynamic goals to the best of our knowledge. + +Curriculum learning is also used for reinforcement learning scenarios. The idea is that solving easier problems first has advantages to learn more complex goals later and thus that learning can be optimized by presenting the problems in an optimal order, a curriculum (Bengio et al., 2009). Narvekar et al. (2017) and Florensa et al. (2018) proposed methods to automatically produce subtasks or subgoals for a given target. Narvekar et al. (2017) produce subtasks according to predefined tasks of a given domain problem. Florensa et al. (2018) use a Generative Adversarial Network (GAN) to produce goals with different difficulties. Different from these methods, our approach produces a series of goals for an episode from failed experience. + +Experience replay is an important technique and introduced to break temporal correlations by mixing more and less recent experience for updating policies (Lin, 1992). It was demonstrated for its efficiency in DQN (Mnih et al., 2015). Prioritized experience replay improves the speed of training by considering prioritizing transitions in the replay buffer (Schaul et al., 2015b). HER considers modifying experience in the replay buffer for continuous control (Andrychowicz et al., 2017). By contrary, our approach assembles successful experience from two failures. Comparing with these methods, which do not consider dynamic goals, our approach uses a series of goals to assemble successful experience. As a result, our method is able to accomplish the tasks with sparse rewards and dynamic goals. + +# 3 METHODOLOGY + +We first review how HER works (Andrychowicz et al., 2017), followed by details of the proposed DHER for dealing with dynamic goals. + +HER is a simple and effective method of manipulating the replay buffer used in off-policy RL algorithms that allows it to learn policies more efficiently from sparse rewards. It assumes the goal being pursued does not influence the environment dynamics. After experiencing an episode $\bar { \{ } s _ { 0 } , s _ { 1 } , \dotsb , s _ { T } \}$ , every transition $s _ { t } \to s _ { t + 1 }$ along with the goal for this episode is usually stored in the replay buffer. Some of the saved episodes fail to reach the goal, providing no positive feedback to the agent. However, with HER, the failed experience is modified and also stored in the replay buffer in the following manner. The idea is to replace the original goal with a state visited by the failed episode. As the reward function remains unchanged, this change of goals hints the agent how to achieve the new goal in the environment. HER assumes that the mechanism of reaching the new goal helps the agent learn for the original goal. + +# 3.1 DYNAMIC GOALS + +Dynamic goals are not static and change at every timestep. We follow the multi-goal setting of Andrychowicz et al. (2017). The goals are part of the environment and do not influence the environment dynamics. We also assume that a dynamic goal $g _ { t } \in \mathcal G$ moves by following some law $g _ { t } = g ( t ; \gamma )$ , where $\gamma$ parameterizes the law (e.g., acceleration in Newton’s law of motion), and yet its underlying moving law is unknown to the agent. + +Moreover, we need to have some basic knowledge of goals, i.e., the measure of goal similarity on $\mathcal { G }$ . We assume that $g _ { t } \in \mathcal G$ corresponds to some predicate $f _ { g _ { t } } : S \{ 0 , 1 \}$ and that the agent’s goal is to achieve any state $s$ that satisfies $f _ { g _ { t } } ( s ) = 1$ . We use $S = \mathcal { G }$ and define $f _ { g _ { t } } ( s ) : = [ s = g _ { t } ]$ , which can be considered as a measure of goal similarity between $g _ { t }$ and $s$ . The goals can also specify only some properties of the state. Take manipulation tasks for instance: $\mathcal { G } = \mathbb { R } ^ { 3 }$ corresponds to the 3D + +# Algorithm 1 Dynamic Hindsight Experience Replay with Experience Assembling + +Require: an off-policy RL algorithm A, replay buffer $R$ , a reward function $r$ + +1: Initialize A and replay buffer $R$ +2: for episode $\mathbf { \Omega } = 1 , 2 , \cdots , M$ do +3: Sample an initial goal $g _ { 0 }$ and an initial state $s _ { 0 }$ +4: for $t = 0 , \cdots , T - 1$ do +5: Sample an action $a _ { t }$ using the behavioral policy from A: +6: $a _ { t } \pi ( s _ { t } | g _ { t } )$ +7: Execute the action $a _ { t }$ and observe a new state $s _ { t + 1 }$ and a new goal $g _ { t + 1 }$ +8: end for +9: for $t = 0 , \cdots , T - 1$ do +10: $r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )$ +11: Store the transitions $\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \right)$ in $R$ (Standard experience replay) +12: end for +13: Collect failed episodes to $\mathcal { E }$ +14: 15: for S $E _ { i } \in \mathcal { E }$ donother $E _ { j } ( i \neq j ) \in \mathcal { E }$ where $g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }$ +16: if $E _ { j } \neq \boldsymbol { \mathcal { O } }$ then +17: Clone a goal trajectory $\{ g _ { 0 } ^ { \prime } , \cdot \cdot \cdot , g _ { m } ^ { \prime } \} _ { m = \operatorname* { m i n } \{ p , q \} }$ in which $g _ { t } ^ { \prime } = g _ { j , q - m + t } ^ { d e }$ from $E _ { j }$ +18: for $t = \{ 0 , \cdots , m - 1 \}$ do +19: $r _ { t } ^ { \prime } : = r ( s _ { i , p - m + t } , a _ { i , p - m + t } , g _ { t + 1 } ^ { \prime } )$ +20: Store the transition $\left( s _ { i , p - m + t } | g _ { t } ^ { \prime } , a _ { i , p - m + t } , r _ { t } ^ { \prime } , s _ { i , p - m + t + 1 } | g _ { t + 1 } ^ { \prime } \right) \mathrm { i n } R \left( \mathrm { D H E R } \right)$ +21: end for +22: end if +23: end for +24: for $t = 1 , \cdots , N$ do +25: Sample a minibatch $B$ from the replay buffer $R$ +26: Optimize A using the minibatch $B$ +27: end for +28: end for + +positions of an object $s ^ { \mathrm { o b j } }$ about which the observation could include additional properties of the object. For a more concrete example, consider pushing a block towards a moving target position. The success of a task is defined as $f \big ( s _ { t } , g _ { t } \big ) = \mathbf { 1 } _ { \mathrm { c o n d i t i o n } } \big ( \| s _ { t } ^ { \mathrm { o b j } } - g _ { t } \| \leq \epsilon \big )$ , where $s _ { t } ^ { \mathrm { o b j } }$ is the position of the object in the state $s _ { t }$ and $\epsilon$ denotes a tolerance by the environment. $\mathbf { 1 } _ { \mathrm { c o n d i t i o n } }$ is an indicator function. The agent aims to achieve any state $s _ { t }$ that satisfies $f ( s _ { t } , g _ { t } ) = 1$ . It receives a sparse reward $r _ { t } : = r ( \tilde { s } _ { t } , a _ { t } , g _ { t + 1 } ) = - \mathbf { 1 } _ { \mathrm { c o n d i t i o n } } ( f ( s _ { t + 1 } , g _ { t + 1 } ) \neq 1 )$ upon making an action $a _ { t }$ . + +It is worth discussing the main difference between the implications of the dynamic goals and the static ones. Expressing a static goal in the following way, $g _ { t } ^ { \mathrm { s t a t i c } } = g ^ { \mathrm { s t a t i c } } , \forall t$ , highlights the key challenge of dealing with the dynamic goal. Namely, the agent has no access to the underlying law of the dynamic goal in our setting, whereas the law of being static is known to the agent in Andrychowicz et al. (2017). In other words, the agent has no clue at all how to construct a new dynamic goal that is admissible by the environment. + +# 3.2 DYNAMIC HINDSIGHT EXPERIENCE REPLAY + +At the first glance, we shall compose a new dynamic goal $g ^ { \mathrm { d y n a m i c } } = \{ s _ { t _ { 0 } } ^ { \mathrm { o b j } } , s _ { t _ { 1 } } ^ { \mathrm { o b j } } , \cdot \cdot \cdot , s _ { t _ { T ^ { \prime } } } ^ { \mathrm { o b j } } \}$ from a in order to apply HER (Andrychowicz et al., 2017) to our problem setting. However, per the discussion above, this new dynamic goal $g ^ { \mathrm { d y n a m i c } }$ may be inadmissible by the environment, leading to no positive feedback to the agent at all. + +We tackle the challenge by drawing the following two observations. One is that many episodes fail in the replay buffer, implying that the agent can actually build a new goal upon more than one episodes. The other is that the more failed experience the agent has, the more possible for the agent to use the connection between achieved goals in an episode and desired goals in some other episode. Take the example of a moving object that the agent must reach, desired goals are the positions of the moving object and achieved goals are the positions of a gripper (controlled by the agent). There may exist some positions that both the object and the gripper have ever reached respectively. + +![](images/685259757917f2907f387b49a9224aabd57ab150ad7074e7a37b24515868d9cf.jpg) +Figure 2: The proposed tasks with dynamic goals (red objects). Arrow indicates the movement of a goal. The first row indicates initial states. The second row indicates final states. + +After experiencing some episode $s _ { 0 } , s _ { 1 } , \cdots , s _ { T }$ , we store in the replay buffer every transition $s _ { t } \ \to \ s _ { t + 1 }$ for this episode, defined as $\left( { { s _ { t } } , { a _ { t } } , { r _ { t } } , { s _ { t + 1 } } } \right)$ , where $s _ { t }$ indicates a state $s _ { t }$ at timestep $t$ , and $a _ { t }$ indicates an action and $r _ { t }$ indicates a reward. Thus before $t$ , there are a series of records $\left\{ \left( s _ { 0 } , a _ { 0 } , r _ { 0 } , s _ { 1 } \right) \cdot \cdot \cdot , \left( s _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } \right) \right\}$ . A state consists of three parts: observation $o _ { t }$ , desired goal $g _ { t } ^ { d e }$ and achieved goal $g _ { t } ^ { a c }$ , define as $\dot { s } _ { t } = \langle o _ { t } , g _ { t } ^ { a c } , g _ { t } ^ { d e } \rangle$ , where normally $g _ { t } ^ { d e } = g _ { t }$ and $g _ { t } ^ { a c }$ indicates goals that the agent has achieved. + +We reuse the failed experience from the replay buffer with inverse simulation to create successful rewards for the agent, as shown in Algorithm 1 (lines 13-23). Our inverse simulation contains two main steps: First, given a failed episode, for its achieved goal trajectory, we try to find a desired goal trajectory from other episodes that could match it (line 15). Second, we assemble a new episode by matching the achieved goal trajectory of the given episode to the desired goal trajectory of the founded one (lines 17-21). Let $g _ { i , q } ^ { a c }$ indicate the achieved goal of the agent at timestep $q$ in episodxperience $i$ $g _ { j , p } ^ { d e }$ te the desired goal at timestep , we search and draw two faile $p$ in episoepisodes $j$ . Gand n of failed), where $\{ E _ { 0 } , E _ { 1 } , \stackrel { \cdot } { E } _ { 2 } , \cdot \cdot \cdot \}$ $E _ { i }$ $E _ { j }$ $( i \neq j )$ $\exists i , j , p , q ,$ s.t. $g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }$ . If we find two such failed episodes, we combine the two experience by replacing the desired goals in $E _ { i }$ by $\{ g _ { j , t } ^ { d e } \}$ , where $j$ indicates $E _ { j }$ and $t \leq \operatorname* { m i n } \{ p , q \}$ . Based on this, we end up assembling a new experience $E _ { i } ^ { \prime }$ based on $E _ { i }$ with a new “imagined” goal trajectory $\{ g _ { j , 0 } ^ { d e } , \cdot \cdot \cdot , g _ { j , t } ^ { d e } \}$ where $t \leq \operatorname* { m i n } \{ p , q \}$ . + +More details of the complete $\mathrm { R L + D H E R }$ method are shown in Algorithm 1. Unlike HER, our DHER needs to search all failed experiences to compose a “imagined” goal trajectory. Hence, the efficiency of searching the memory is important. In our implementation, we use two hash tables to store the trajectories of achieved goals and desired goals, respectively. + +# 4 EXPERIMENT + +We run extensive experiments to examine the proposed DHER for moving goals and compare it with some competing baselines. We first introduce the environments and tasks that we want to address, followed by the experimental results. Demo videos from our experiments are available at https://sites.google.com/view/dher. + +![](images/868b5ca50a241e6883855868e5cdc6d4e2e4a7d4f491d103d018de67c19012a7.jpg) +Figure 3: Results on different environments. + +# 4.1 ENVIRONMENTS + +Whereas grasping moving targets are fairly common in robotic applications, there are rarely existing environments featuring dynamic goals. We modify the robotic manipulation environments created by OpenAI (Brockman et al., 2016) for our experiments. As shown in Figure 2, we assign certain rules to the goals so that they accordingly move in the environments while an agent is required to control the robotic arm’s grippers to reach the goal that moves along a straight line (Dy-Reaching), to reach the goal that moves in a circle (Dy-Circling), or to push a block to the goal that moves along a straight line (Dy-Pushing). In addition, we also develop a new GREEDY SNAKE environment (DySnake), in which the goal moves from one discrete cell to another. The greedy snake aims to reach the goal (and eat it). No matter what actions are taken by the agent, the underlying rule of changing the goals’ positions remain the same. We assume that the law of the goal’s motion is unknown to the agent. More details of the tasks are described in the appendix. + +For the first three tasks, we follow the basic settings of OpenAI robotics environments (Brockman et al., 2016). States are read from the MuJoCo physics engine. An observation consists of relative positions of the object and the target (grippers are blocked) Goals are positions in the 3D world coordinate system with a fixed tolerance (we use $\epsilon = 0 . 0 1$ for the tolerance). Note the goals are able to move in our tasks. The velocity we use is $v = 0 . 0 1 1$ . 1 epoch indicates 100 episodes. The start positions of the goals are randomly chosen. Rewards are binary and sparse: $r ( s _ { t } , a _ { t } , g _ { t } ) =$ $- \mathbf { 1 } _ { \mathrm { c o n d i t i o n } } \big ( | s _ { t + 1 } ^ { \mathrm { o b j } } - g _ { t + 1 } | \geq \epsilon \big )$ where $s _ { t + 1 }$ and $g _ { t + 1 }$ are respectively the state and dynamic goal after the execution of the action $a _ { t }$ in the state $s _ { t }$ . Two positions are overlapped if they are close within the tolerance $\epsilon$ . We use 3-dimensional actions, in which three dimensions correspond to the desired relative gripper position at the next timestep. + +For the last task, it is very similar to the standard snake game. There is a 2D plane grid whose size is $3 0 \times 4 0$ . A snake and a goal (food) both move in this grid. We use a $1 \times 1$ square as the body of the snake and do not allow the snake to grow any longer as it moves. The states of the system are represented by using the positions of the snake and food. Goals are the positions of the food. The snake has to move itself such that it resides in the same cell as the food at a certain timestep. The velocity of the goal is set to $( 0 , 1 )$ . The start positions of the goal are set randomly in different episodes of the game, so are the start positions of the snake. Rewards are binary and sparse: $r \big ( s _ { t } , \bar { a } _ { t } , g _ { t + 1 } \big ) = - \bar { \mathbf { 1 } } _ { \mathrm { c o n d i t i o n } } \big ( \big | s _ { t + 1 } ^ { \mathrm { s n a k e } } - g _ { t + 1 } \big | \bar { \neq } 0 \big )$ where $s _ { t + 1 }$ and $g _ { t + 1 }$ are the environment’s state and goal after the agent executes action $a _ { t }$ in the state $s _ { t }$ . Observations are represented by the positions of the snake and food. We also add to the state the distance between the snake and food. Actions allowed in the game are the following: move up, move down, move left, and move right, for one cell per timestep. + +# 4.2 BASELINES + +For the first three tasks which call for continuous control, we consider two competing baselines:2 • DDPG, which is a model-free RL algorithm for continuous control (Lillicrap et al., 2015). It learns a deterministic policy by using a stochastic counterpart to explore in the training. + +![](images/732a387a80c212738b2fb410b777a5c71d62a4dacb48b29dd438871f23abaf26.jpg) +Figure 4: Low velocity: 0.001. + +![](images/f8c0fd7014d07719a2c477eef04492742d4a9d1ff91065b3d850a3053e6e0ee6.jpg) +Figure 5: High velocity: 0.016. + +• $\mathrm { D D P G } + \mathrm { H E R }$ , which improves the replay buffer of DDPG by the hindsight experience replay (Andrychowicz et al., 2017). +DDPG (dense), which employs the negative distance $( - d )$ as dense rewards. +DDPG (dense-2), which employs the negative distance $( - d )$ as dense rewards if $d \geq \epsilon$ . However, if $d < \epsilon$ (i.e., success) it uses $( - d + 1 . 0 )$ as rewards. 1.0 is a bonus. + +For the last task of discrete control, we use the following baselines in the experiments: DQN (Mnih et al., 2015) and $\mathrm { D Q N + H E R }$ , which uses HER to enhance the replay in DQN. + +• DQN, which is a powerful model-free RL algorithm for discrete action spaces (Mnih et al., 2015). +• $\mathrm { D Q N + H E R }$ , which uses HER to enhance the replay in DQN. +• DQN (dense), which uses the negative distance $( - d )$ as dense rewards. +• DQN (dense-2), which uses the negative distance $( - d )$ as dense rewards. However, if $d = 0 . 0$ (i.e., success) it uses $( - d + 1 . 0 )$ as rewards instead. 1.0 is a bonus. + +# 4.3 COMPARISON RESULTS ON THE ROBOTIC ENVIRONMENTS + +For the continuous control, we present three sets of comparison results in Figure 3 for the first three tasks, respectively. Consistently, the results show our DHER algorithm outperforms the others. The two baselines are not able to catch up even after we train them for thousands of iterations. Vanilla DDPG is slightly better than the version with HER. HER does not benefit DDPG in these tasks because the goals in HER are fixed, fundamentally misleading the agent in the attempt of solving the tasks with dynamic goals. The results of DDPG (dense) and DDPG (dense-2) suggest that even the dense rewards do not work well as they are agnostic to the task of interest. A good reward shaping may give rise to better performance by carefully tuning it for the task of dynamic goals. + +Comparing Figure 3a and Figure 3b with Figure 3c, we find that DHER learns faster in Dy-Reaching and Dy-Circling than Dy-Pushing probably because Dy-Reaching and Dy-Circling are easier tasks than Dy-Pushing. In Dy-Pushing, all the algorithms take a fairly big amount of time to explore without receiving any positive feedback. However, the more failed experiences the agent encounters, the better change our algorithm is able to identify relevant episodes from them for assembling useful dynamic goals. As a result, DHER is able to pick up the momentum and learns faster and better than the baselines after a certain point. In Figure 3a, the performance decreases a little. The reason may be that as successful experience increases, some assembled experience is inconsistent with these successful experience. + +# 4.3.1 COMPARISON USING DIFFERENT VELOCITIES OF GOALS + +To show the performance of our method on more complex tasks with different velocities, we study different methods in Dy-Reaching environment as shown in Figures 4 and 5 with the same physical properties as the previous experiments. Overall the results show our method is much better than DDPG and HER. As a reminder, the threshold for calculating rewards we used is 0.01. In Figure 4, the task becomes an easier task because of the slower velocity. It shows that our method quickly achieves to a good result around 5 epoch. Comparing with Figure 3a, it shows that the performance with $v = 0 . 0 0 1$ is better than the performance with $v = 0 . 0 1 1$ and get $1 5 \%$ improvements. However, it also shows the performance with $v = 0 . 0 1 6$ is worse than the performance with $v = 0 . 0 1 1$ . Both DDPG and $\mathrm { \Delta D D P G + H E R }$ failed when $v = 0 . 0 1 6$ and their performance is 0. This performance is consistent because the task becomes more difficult when the velocity increases. + +![](images/14dd0b3e1878d3f49fb1b7d44ca84006cca0ac15133dc036aef116910149f401.jpg) +Figure 6: Snake with dynamic goals. + +![](images/97b49eb4c50a91b591b5c09b7c8417f90a64b3336f8b06a8631418b78f185b01.jpg) +Figure 7: Special: $\mathrm { H E R + }$ knows how goals move. + +# 4.4 COMPARISON ON THE DYNAMIC SNAKE ENVIRONMENT + +In the Dy-Snake environment, which is a discrete control environment, we present the results of the chasing food task in Figure 6. The results show that the proposed algorithm works best. DQN is better than HER. HER fails for this task. That HER fails for this task shows just using achieved goals is not enough for the tasks with dynamic goals and can lead wrong direction. The results also show that around 800 episodes, the performance of DQN and DHER is close and DHER is slightly better than DQN. This is because the chasing food task is a simple task. The action space is very small and just 4 types of actions. After enough exploration, DQN also has competitive performance. However, at the beginning, DHER quickly achieves very good performance. It shows that assembling experience from two failures improves the performance very efficiently in this task. DQN (dense) and DQN (dense-2) help learn the policy at the early stage. However, in the long run, it does not lead to any particular benefits. + +# 4.5 SIM TO REAL ADAPTATION + +We used policies for Dy-Circling task and a new Pouring task trained in our simulator 3 to deploy them on a physical robot. As shown in Figure 8, the policies were trained by using DHER and adapted to the real robot without any finetuning. However, the policy requires accurate localizations of the gripper and the goal. For Dy-Circling task, the robot’s gripper was blocked. There were a toy turntable, whose speed is unknown, and a blue block on the turntable. We set the position 1cm above the block as the target position. The position of the block was predicted based on traditional contour shape analysis using camera images. For Pouring task, the robot gripped a can. A man held a cup and moved it. The cup was set as the target and with a green marker. We used the marker to estimate its position. + +Our policies were transferred successfully for both tasks. With the accurate positions, we have $100 \%$ success rate for 5 trials. It was observed that the robot had learned to not only follow the current target but also step forward to the future target position. Demo videos about the experiments are available at https://sites.google.com/view/dher. + +![](images/ff03243dba37e8e374b610dab429ef50349cfd9bf2cd37e4c19e2d40e16ac98c.jpg) +Figure 8: Adapting the policies trained based on DHER from our simulation to a real robotic arm. + +4.6 SPECIAL CASE: HOW ABOUT IF HER KNOWS THE LAW OF THE MOTION OF A GOAL + +We use Dy-Snake to demonstrate the experimental results for a special case that the law of the motion of the target (food) is known to agents. Because the motion of the food is very simple and controllable in Dy-Snake environment. We develop a direct extension of HER, called $\mathrm { H E R + }$ , that modifies desired goals at every timestep based on the law of the motion of the food to create successful experience. More details of $\mathrm { H E R + }$ are described in the appendix. + +We show the results in Figure 7. DHER and $\mathrm { H E R + }$ are both better than DQN at the beginning. DHER is slightly better than $\mathrm { H E R + }$ , which shows the efficiency of DHER is comparable in this simple task. + +# 5 CONCLUSION + +We introduced a novel technique that assembles successful experience from a couple of failures. With this technique, our proposed algorithm called DHER ( Dynamic Hindsight Experience Replay) is able to address the tasks with sparse rewards and dynamic goals. Our technique can be combined with an arbitrary off-policy RL algorithm and we experimentally demonstrated that with DQN and DDPG. As far as we know, it is the first time that an agent is allowed to learn from assembled experience from two failures. + +# ACKNOWLEDGMENTS + +We would like to thank Weitao Xi, Tianzhou Wang, Tingguang Li for performing some additional transfer experiments. We would also like to thank Han Liu and the whole RL team for fruitful discussions as well as the anonymous reviewers for their comments. + +# REFERENCES + +Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI. Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. In Advances in Neural Information Processing Systems, pp. 5048–5058, 2017. +Marc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279, 2013. ISSN 1076-9757. +Yoshua Bengio, Jrme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th International Conference on Machine Learning, pp. 41–48. ACM, 2009. +Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. +Maurizio Di Rocco, Federico Pecora, Prasanna Kumar Sivakumar, and Alessandro Saffiotti. Configuration planning with multiple dynamic goals. In AAAI Spring Symposium: Designing Intelligent Robots, 2013. +Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control. In Proceedings of the 33rd International Conference on Machine Learning, pp. 1329–1338, 2016. + +Carlos Florensa, David Held, Xinyang Geng, and Pieter Abbeel. Automatic goal generation for reinforcement learning agents. In Proceedings of the 35th International Conference on Machine Learning, volume 80, pp. 1515–1528, 2018. + +Nicolas Heess, Dhruva TB, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, Tom Erez, Ziyu Wang, S. M. Ali Eslami, Martin A. Riedmiller, and David Silver. Emergence of locomotion behaviours in rich environments. CoRR, abs/1707.02286, 2017. + +Leslie Pack Kaelbling. Learning to achieve goals. In Proceedings of the 13th International Joint Conference on Artificial Intelligence, pp. 1094–1099. Citeseer, 1993. + +Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. + +Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine Learning, 8(3-4):293–321, 1992. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518:529, 2015. + +Sanmit Narvekar, Jivko Sinapov, and Peter Stone. Autonomous task sequencing for customized curriculum design in reinforcement learning. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, volume 147, pp. 149, 2017. + +Andrew Y. Ng, Daishi Harada, and Stuart Russell. Policy invariance under reward transformations: Theory and application to reward shaping. In Proceedings of the 16th International Conference on Machine Learning, volume 99, pp. 278–287, 1999. + +Tom Schaul, Daniel Horgan, Karol Gregor, and David Silver. Universal value function approximators. In Proceedings of the 32nd International Conference on Machine Learning, pp. 1312–1320, 2015a. + +Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015b. + +John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel. Trust region policy optimization. CoRR, abs/1502.05477, 2015. + +John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347, 2017. + +David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484, 2016. + +Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033. IEEE, 2012. + +Ziyu Wang, Victor Bapst, Nicolas Heess, Volodymyr Mnih, Remi Munos, Koray Kavukcuoglu, and ´ Nando de Freitas. Sample efficient actor-critic with experience replay. CoRR, abs/1611.01224, 2016. + +# APPENDIX + +# TASKS WITH DYNAMIC GOALS + +We have created four tasks as described below and illustrated by Figure 2. The first three tasks are based on the robotic environments. The last is based on the GREEDY SNAKE environment. + +• Dy-Reaching: The task is to control the robotic arm so that its gripper can reach the target position. The target position moves from one point to another along a straight line with a constant velocity. +• Dy-Circling: The task is to control the robotic arm so that its gripper can reach the target position. The target position moves along a circle with a constant velocity. Dy-Pushing: in this task, a box is placed on a table in front of the robotic arm. The robot is required to move the box to the target location on the table. The target location moves from one position to another with a fixed velocity. Note that the robot’s grippers are locked to prevent it from grasping. The learned behavior is actually a mixture of pushing and rolling. +Dy-Snake: in this task, a snake and food are placed on a rectangle map and the task is to control the snake to eat the food. The food is moving from one position to another position with a fixed velocity. + +# HARDWARE + +We use the Universal Robots UR10 with a gripper. We use a RealSense Camera SR300 to track the position of objects. The gripper is blocked. We use a marker on the gripper for camera calibration. During adapting the robot, for the position of the target, we use the camera to estimate it. + +# SIMULATION + +We simulate the physical system using the MuJoCo physics engine (Todorov et al., 2012) and also use MuJoCo to render the images. In our tasks, we use positions provided by MuJoCo for training policies in the simulation. + +In Figure 9, we demonstrate our simulation and the physics environment. + +![](images/0ab0a04331e1ed7ee05503beb83d72a71ddd08962ea6776f4dbc386edf14fc2c.jpg) +Figure 9: Our simulation and the physics environment. + +A SPECIAL AND SIMPLER CASE - HER+ + +The paper considers a general situation that the law of the motion of goals is invisible to an agent. However, we relax this assumption and assume that in some situations the law of the motion of goals is easy to obtain, i.e., the parameters of $g ( t ; \gamma )$ are known. For example, it is able to calculate the velocity of an object by observing the trajectories of the object. + +In this situation, we do not need to search failed experience. With the knowledge of the velocity of an object, it is straightforward that we can calculate any trajectory of desired goals. Thus for every failed experience, at any time step $t$ , based on the achieved goal $g _ { t } ^ { a c }$ , we calculate new desired goals $g _ { t } ^ { d e } = g ( { \dot { t } } )$ corresponding to the achieved goals in order to construct successful experience. + +The details of $\mathrm { H E R + }$ are described in Algorithm 2. $\mathrm { H E R + }$ can be seen as a direct extension of HER. HER and $\mathrm { H E R + }$ can modify any failed experience to successful experience because they both assume the law of the motion of goals is known. + +# Algorithm 2 Hindsight Experience Replay Plus + +Require: an off-policy RL algorithm A, replay buffer $R$ , a reward function $r$ +1: Initialize A and replay buffer $R$ +2: for episode $\mathbf { \Omega } = 1 , 2 , \cdots , M$ do +3: Sample an initial goal $g _ { 0 }$ and an initial state $s _ { 0 }$ +4: for $t = 0 , \cdots , T - 1$ do +5: Sample an action $a _ { t }$ using the behavioral policy from A: +6: $a _ { t } \doteq \pi ( s _ { t } | g _ { t } )$ +7: Execute the action $a _ { t }$ and observe a new state $s _ { t + 1 }$ and a new goal $g _ { t + 1 }$ +8: end for +9: for $t = 0 , \cdots , T - 1$ do +10: $r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )$ +11: Store the transitions $\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \right)$ in $R$ (Standard experience replay) +12: Sample a set of the achieved goals of $E _ { i }$ as additional goals for reply $G ^ { \prime }$ +13: for $\bar { g _ { p } ^ { \prime } } \in G ^ { \prime }$ do +14: Calculate a goal trajectory $\{ g _ { 0 } ^ { \prime } , \cdots , g _ { p } ^ { \prime } \}$ where $g _ { t } ^ { \prime } = g ( t ; \gamma )$ $\mathrm { ( H E R + ) }$ +15: for $t = \{ 0 , \cdots , p - 1 \}$ do +16: $r _ { t } ^ { \prime } : = r ( s _ { i , t } , a _ { i , t } , g _ { t + 1 } ^ { \prime } )$ +17: Store the transition $\left( s _ { i , t } | g _ { t } ^ { \prime } , a _ { i , t } , r _ { t } ^ { \prime } , s _ { i , t + 1 } | g _ { t + 1 } ^ { \prime } \right) \mathrm { i n } R$ +18: end for +19: end for +20: end for +21: end for \ No newline at end of file diff --git a/parse/train/Byf5-30qFX/Byf5-30qFX_content_list.json b/parse/train/Byf5-30qFX/Byf5-30qFX_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..90c37aadf82796b8272914ad011ce89dc24a6868 --- /dev/null +++ b/parse/train/Byf5-30qFX/Byf5-30qFX_content_list.json @@ -0,0 +1,1238 @@ +[ + { + "type": "text", + "text": "DHER: HINDSIGHT EXPERIENCE REPLAY FOR DYNAMIC GOALS ", + "text_level": 1, + "bbox": [ + 174, + 98, + 678, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meng Fang∗, Cheng Zhou, Bei Shi, Boqing Gong, Jia Xu, Tong Zhang Tencent AI Lab ", + "bbox": [ + 184, + 169, + 665, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dealing with sparse rewards is one of the most important challenges in reinforcement learning (RL), especially when a goal is dynamic (e.g., to grasp a moving object). Hindsight experience replay (HER) has been shown an effective solution to handling sparse rewards with fixed goals. However, it does not account for dynamic goals in its vanilla form and, as a result, even degrades the performance of existing off-policy RL algorithms when the goal is changing over time. ", + "bbox": [ + 233, + 266, + 764, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present Dynamic Hindsight Experience Replay (DHER), a novel approach for tasks with dynamic goals in the presence of sparse rewards. DHER automatically assembles successful experiences from two relevant failures and can be used to enhance an arbitrary off-policy RL algorithm when the tasks’ goals are dynamic. We evaluate DHER on tasks of robotic manipulation and moving object tracking, and transfer the polices from simulation to physical robots. Extensive comparison and ablation studies demonstrate the superiority of our approach, showing that DHER is a crucial ingredient to enable RL to solve tasks with dynamic goals in manipulation and grid world domains. ", + "bbox": [ + 233, + 352, + 764, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 502, + 336, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep reinforcement learning has been shown an effective framework for solving a rich repertoire of complex control problems. In simulated domains, agents have been trained to perform a diverse array of challenging tasks (Mnih et al., 2015; Lillicrap et al., 2015; Duan et al., 2016). In order to train such agents, it is often the case that one has to design a reward function that not only reflects the task at hand but also is carefully shaped $\\mathrm { N g }$ et al., 1999) to guide the policy optimization. Unfortunately, many of the capabilities demonstrated by reward engineering are often limited to specific tasks. Moreover, it requires both RL expertise and domain-specific knowledge to reshape the reward functions. For situations where we do not know what admissible behavior may look like, for example, using LEGO bricks to build a desired architecture, it is difficult to apply reward engineering. Therefore, it is essential to develop algorithms which can learn from unshaped and usually sparse reward signals. ", + "bbox": [ + 174, + 534, + 825, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning with sparse rewards is challenging, especially when a goal is dynamic. Dynamic goals are common in games and planning problems, often addressed using reward shaping or search (Kaelbling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). However, the difficulty posed by a sparse reward is exacerbated by the complicated environment dynamics in robotics (Andrychowicz et al., 2017). For instance, system dynamics around contacts are difficult to model and induce sensitivity in the system to small errors. Many robotic tasks also need executing multiple steps successfully over a long horizon, involve enormous search space, and require generalization to varying task instances. Policy gradient methods are breakthroughs in the challenging environments, such as PPO (Heess et al., 2017; Schulman et al., 2017), ACER (Wang et al., 2016), TRPO (Schulman et al., 2015) and so on. They are used in environments, where an agent tries to reach a target, learns to walk, runs, and so on. Recently, sampling-efficient learning is introduced and demonstrates a significant increase in performance for off-policy actor-critic DQN (Mnih et al., 2015) and DDPG (Lillicrap et al., 2015) algorithms. Hindsight experience replay (HER) is very effective for improving the performance of off-policy RL algorithms in solving goal-based tasks with sparse rewards (Andrychowicz et al., 2017). Similar to UVFA (Schaul et al., 2015a), it takes a goal state as part of input. However, it assumes the goal is fixed. As a result, this assumption actually impedes the learning of RL agents in the environments of moving goals. ", + "bbox": [ + 173, + 694, + 825, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a34ff9c3bf9148b8d43e21b58e1ea68fe1bcb07b8ac12d22cc17bf9a91fd4060.jpg", + "image_caption": [ + "Figure 1: The framework of DHER. DHER is a kind of experience replay method. It searches relevant failed experiences and then assembles them into successful experiences. " + ], + "image_footnote": [], + "bbox": [ + 179, + 103, + 823, + 261 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 176, + 347, + 823, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we address this challenge with a new method, Dynamic Hindsight Experience Replay (DHER), for accomplishing tasks with moving goals. We follow the multi-goal setting in UVFA (Schaul et al., 2015a) and HER (Andrychowicz et al., 2017). It assumes that the goal being pursued does not influence the environment dynamics. We also need to have the knowledge of goal similarity. For example, in manipulation or grid world domains, we can use Euclidean distance between positions to measure the goal similarity. HER turns a failed episode to a success by composing a new task whose goal is achieved by that episode. Our idea allows an agent to learn from the failure one step further than HER: the agent not only sets a new goal but also hallucinates how to reach the original goal from the new one. Take playing frisbee for instance. When an agent jumps to catch the frisbee and yet misses it, the agent receives no positive feedback under the sparse reward setting. Using HER, the agent could set the end of its episode as the new goal — position of the frisbee; with DHER, however, the agent finds a trajectory from its past experiences as the imagined path of the frisbee, and thereby extrapolates towards the original goal. ", + "bbox": [ + 174, + 382, + 825, + 563 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, we do the following for DHER. To finish the tasks with dynamic goals needs to explore experience and understand multiple goals. DHER uses replay buffers to allow the agent to learn from a couple of failures by assembling new ‘experience’ from different episodes. The proposed method retrieves memories to find the connection between the experience of different episodes. It largely improves the sample efficiency in dynamic goal task settings. More importantly, this strategy makes it possible to learn in the setting that both sparse rewards and dynamic goals exist. ", + "bbox": [ + 174, + 570, + 825, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our method along with the state-of-the-art baselines on new environments and manipulation tasks, which have sparse rewards and dynamic goals. Our results demonstrate that DHER is clearly better than others for these tasks. We also transfer policies trained in our simulation based on DHER to a physical robot and show that DHER can be applied to solving real-world robotics problems. ", + "bbox": [ + 174, + 660, + 823, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows: (1) We demonstrate that, both in simulation and real worlds, DHER succeeds in continuous control with a moving target. To our knowledge, this is the first empirical result on manipulation tasks that demonstrates model-free learning methods can tackle tasks of this complexity. (2) We show that with assembling new experience from two failures, the sample complexity can be reduced dramatically. We attribute this to global knowledge learning in a set of failed experience which breaks the constraint of local one-episode experience towards more robust strategies. (3) We design and implement a set of new environments and continuous tasks with dynamic goals, which would be of interest to researchers at the intersection of robotics and reinforcement learning.1 ", + "bbox": [ + 174, + 737, + 825, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 339, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works in deep RL have shown impressive results in different domains, such as games (Bellemare et al., 2013; Silver et al., 2016), simulated control tasks (Brockman et al., 2016) and so on. There have been several proposed RL methods for playing Atari games, including DQN (Mnih et al., 2015), UVFA (Schaul et al., 2015b) and so on. UVFA trains a single neural network approximating multiple value functions for state and goal. For the continuous control, DDPG (Lillicrap et al., 2015) is a popular actor-critic algorithm that has shown impressive results in continuous control tasks. Dynamic goals appear in games and planning, often addressed by rewards shaping or search (Kaelbling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). When the rewards are sparse, there is few work studying the dynamic goals to the best of our knowledge. ", + "bbox": [ + 174, + 132, + 825, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Curriculum learning is also used for reinforcement learning scenarios. The idea is that solving easier problems first has advantages to learn more complex goals later and thus that learning can be optimized by presenting the problems in an optimal order, a curriculum (Bengio et al., 2009). Narvekar et al. (2017) and Florensa et al. (2018) proposed methods to automatically produce subtasks or subgoals for a given target. Narvekar et al. (2017) produce subtasks according to predefined tasks of a given domain problem. Florensa et al. (2018) use a Generative Adversarial Network (GAN) to produce goals with different difficulties. Different from these methods, our approach produces a series of goals for an episode from failed experience. ", + "bbox": [ + 174, + 265, + 825, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experience replay is an important technique and introduced to break temporal correlations by mixing more and less recent experience for updating policies (Lin, 1992). It was demonstrated for its efficiency in DQN (Mnih et al., 2015). Prioritized experience replay improves the speed of training by considering prioritizing transitions in the replay buffer (Schaul et al., 2015b). HER considers modifying experience in the replay buffer for continuous control (Andrychowicz et al., 2017). By contrary, our approach assembles successful experience from two failures. Comparing with these methods, which do not consider dynamic goals, our approach uses a series of goals to assemble successful experience. As a result, our method is able to accomplish the tasks with sparse rewards and dynamic goals. ", + "bbox": [ + 174, + 383, + 825, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 METHODOLOGY ", + "text_level": 1, + "bbox": [ + 176, + 529, + 341, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first review how HER works (Andrychowicz et al., 2017), followed by details of the proposed DHER for dealing with dynamic goals. ", + "bbox": [ + 176, + 559, + 823, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "HER is a simple and effective method of manipulating the replay buffer used in off-policy RL algorithms that allows it to learn policies more efficiently from sparse rewards. It assumes the goal being pursued does not influence the environment dynamics. After experiencing an episode $\\bar { \\{ } s _ { 0 } , s _ { 1 } , \\dotsb , s _ { T } \\}$ , every transition $s _ { t } \\to s _ { t + 1 }$ along with the goal for this episode is usually stored in the replay buffer. Some of the saved episodes fail to reach the goal, providing no positive feedback to the agent. However, with HER, the failed experience is modified and also stored in the replay buffer in the following manner. The idea is to replace the original goal with a state visited by the failed episode. As the reward function remains unchanged, this change of goals hints the agent how to achieve the new goal in the environment. HER assumes that the mechanism of reaching the new goal helps the agent learn for the original goal. ", + "bbox": [ + 174, + 594, + 825, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 DYNAMIC GOALS", + "text_level": 1, + "bbox": [ + 176, + 751, + 338, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic goals are not static and change at every timestep. We follow the multi-goal setting of Andrychowicz et al. (2017). The goals are part of the environment and do not influence the environment dynamics. We also assume that a dynamic goal $g _ { t } \\in \\mathcal G$ moves by following some law $g _ { t } = g ( t ; \\gamma )$ , where $\\gamma$ parameterizes the law (e.g., acceleration in Newton’s law of motion), and yet its underlying moving law is unknown to the agent. ", + "bbox": [ + 174, + 776, + 825, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, we need to have some basic knowledge of goals, i.e., the measure of goal similarity on $\\mathcal { G }$ . We assume that $g _ { t } \\in \\mathcal G$ corresponds to some predicate $f _ { g _ { t } } : S \\{ 0 , 1 \\}$ and that the agent’s goal is to achieve any state $s$ that satisfies $f _ { g _ { t } } ( s ) = 1$ . We use $S = \\mathcal { G }$ and define $f _ { g _ { t } } ( s ) : = [ s = g _ { t } ]$ , which can be considered as a measure of goal similarity between $g _ { t }$ and $s$ . The goals can also specify only some properties of the state. Take manipulation tasks for instance: $\\mathcal { G } = \\mathbb { R } ^ { 3 }$ corresponds to the 3D ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 Dynamic Hindsight Experience Replay with Experience Assembling ", + "text_level": 1, + "bbox": [ + 173, + 102, + 717, + 118 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Require: an off-policy RL algorithm A, replay buffer $R$ , a reward function $r$ ", + "bbox": [ + 176, + 123, + 679, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Initialize A and replay buffer $R$ \n2: for episode $\\mathbf { \\Omega } = 1 , 2 , \\cdots , M$ do \n3: Sample an initial goal $g _ { 0 }$ and an initial state $s _ { 0 }$ \n4: for $t = 0 , \\cdots , T - 1$ do \n5: Sample an action $a _ { t }$ using the behavioral policy from A: \n6: $a _ { t } \\pi ( s _ { t } | g _ { t } )$ \n7: Execute the action $a _ { t }$ and observe a new state $s _ { t + 1 }$ and a new goal $g _ { t + 1 }$ \n8: end for \n9: for $t = 0 , \\cdots , T - 1$ do \n10: $r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )$ \n11: Store the transitions $\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)$ in $R$ (Standard experience replay) \n12: end for \n13: Collect failed episodes to $\\mathcal { E }$ \n14: 15: for S $E _ { i } \\in \\mathcal { E }$ donother $E _ { j } ( i \\neq j ) \\in \\mathcal { E }$ where $g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }$ \n16: if $E _ { j } \\neq \\boldsymbol { \\mathcal { O } }$ then \n17: Clone a goal trajectory $\\{ g _ { 0 } ^ { \\prime } , \\cdot \\cdot \\cdot , g _ { m } ^ { \\prime } \\} _ { m = \\operatorname* { m i n } \\{ p , q \\} }$ in which $g _ { t } ^ { \\prime } = g _ { j , q - m + t } ^ { d e }$ from $E _ { j }$ \n18: for $t = \\{ 0 , \\cdots , m - 1 \\}$ do \n19: $r _ { t } ^ { \\prime } : = r ( s _ { i , p - m + t } , a _ { i , p - m + t } , g _ { t + 1 } ^ { \\prime } )$ \n20: Store the transition $\\left( s _ { i , p - m + t } | g _ { t } ^ { \\prime } , a _ { i , p - m + t } , r _ { t } ^ { \\prime } , s _ { i , p - m + t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R \\left( \\mathrm { D H E R } \\right)$ \n21: end for \n22: end if \n23: end for \n24: for $t = 1 , \\cdots , N$ do \n25: Sample a minibatch $B$ from the replay buffer $R$ \n26: Optimize A using the minibatch $B$ \n27: end for \n28: end for ", + "bbox": [ + 173, + 127, + 799, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "positions of an object $s ^ { \\mathrm { o b j } }$ about which the observation could include additional properties of the object. For a more concrete example, consider pushing a block towards a moving target position. The success of a task is defined as $f \\big ( s _ { t } , g _ { t } \\big ) = \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( \\| s _ { t } ^ { \\mathrm { o b j } } - g _ { t } \\| \\leq \\epsilon \\big )$ , where $s _ { t } ^ { \\mathrm { o b j } }$ is the position of the object in the state $s _ { t }$ and $\\epsilon$ denotes a tolerance by the environment. $\\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } }$ is an indicator function. The agent aims to achieve any state $s _ { t }$ that satisfies $f ( s _ { t } , g _ { t } ) = 1$ . It receives a sparse reward $r _ { t } : = r ( \\tilde { s } _ { t } , a _ { t } , g _ { t + 1 } ) = - \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } ( f ( s _ { t + 1 } , g _ { t + 1 } ) \\neq 1 )$ upon making an action $a _ { t }$ . ", + "bbox": [ + 173, + 563, + 825, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth discussing the main difference between the implications of the dynamic goals and the static ones. Expressing a static goal in the following way, $g _ { t } ^ { \\mathrm { s t a t i c } } = g ^ { \\mathrm { s t a t i c } } , \\forall t$ , highlights the key challenge of dealing with the dynamic goal. Namely, the agent has no access to the underlying law of the dynamic goal in our setting, whereas the law of being static is known to the agent in Andrychowicz et al. (2017). In other words, the agent has no clue at all how to construct a new dynamic goal that is admissible by the environment. ", + "bbox": [ + 174, + 655, + 825, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 DYNAMIC HINDSIGHT EXPERIENCE REPLAY", + "text_level": 1, + "bbox": [ + 174, + 761, + 522, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the first glance, we shall compose a new dynamic goal $g ^ { \\mathrm { d y n a m i c } } = \\{ s _ { t _ { 0 } } ^ { \\mathrm { o b j } } , s _ { t _ { 1 } } ^ { \\mathrm { o b j } } , \\cdot \\cdot \\cdot , s _ { t _ { T ^ { \\prime } } } ^ { \\mathrm { o b j } } \\}$ from a in order to apply HER (Andrychowicz et al., 2017) to our problem setting. However, per the discussion above, this new dynamic goal $g ^ { \\mathrm { d y n a m i c } }$ may be inadmissible by the environment, leading to no positive feedback to the agent at all. ", + "bbox": [ + 174, + 787, + 825, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We tackle the challenge by drawing the following two observations. One is that many episodes fail in the replay buffer, implying that the agent can actually build a new goal upon more than one episodes. The other is that the more failed experience the agent has, the more possible for the agent to use the connection between achieved goals in an episode and desired goals in some other episode. Take the example of a moving object that the agent must reach, desired goals are the positions of the moving object and achieved goals are the positions of a gripper (controlled by the agent). There may exist some positions that both the object and the gripper have ever reached respectively. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/685259757917f2907f387b49a9224aabd57ab150ad7074e7a37b24515868d9cf.jpg", + "image_caption": [ + "Figure 2: The proposed tasks with dynamic goals (red objects). Arrow indicates the movement of a goal. The first row indicates initial states. The second row indicates final states. " + ], + "image_footnote": [], + "bbox": [ + 212, + 107, + 779, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 409, + 821, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After experiencing some episode $s _ { 0 } , s _ { 1 } , \\cdots , s _ { T }$ , we store in the replay buffer every transition $s _ { t } \\ \\to \\ s _ { t + 1 }$ for this episode, defined as $\\left( { { s _ { t } } , { a _ { t } } , { r _ { t } } , { s _ { t + 1 } } } \\right)$ , where $s _ { t }$ indicates a state $s _ { t }$ at timestep $t$ , and $a _ { t }$ indicates an action and $r _ { t }$ indicates a reward. Thus before $t$ , there are a series of records $\\left\\{ \\left( s _ { 0 } , a _ { 0 } , r _ { 0 } , s _ { 1 } \\right) \\cdot \\cdot \\cdot , \\left( s _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } \\right) \\right\\}$ . A state consists of three parts: observation $o _ { t }$ , desired goal $g _ { t } ^ { d e }$ and achieved goal $g _ { t } ^ { a c }$ , define as $\\dot { s } _ { t } = \\langle o _ { t } , g _ { t } ^ { a c } , g _ { t } ^ { d e } \\rangle$ , where normally $g _ { t } ^ { d e } = g _ { t }$ and $g _ { t } ^ { a c }$ indicates goals that the agent has achieved. ", + "bbox": [ + 173, + 445, + 825, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We reuse the failed experience from the replay buffer with inverse simulation to create successful rewards for the agent, as shown in Algorithm 1 (lines 13-23). Our inverse simulation contains two main steps: First, given a failed episode, for its achieved goal trajectory, we try to find a desired goal trajectory from other episodes that could match it (line 15). Second, we assemble a new episode by matching the achieved goal trajectory of the given episode to the desired goal trajectory of the founded one (lines 17-21). Let $g _ { i , q } ^ { a c }$ indicate the achieved goal of the agent at timestep $q$ in episodxperience $i$ $g _ { j , p } ^ { d e }$ te the desired goal at timestep , we search and draw two faile $p$ in episoepisodes $j$ . Gand n of failed), where $\\{ E _ { 0 } , E _ { 1 } , \\stackrel { \\cdot } { E } _ { 2 } , \\cdot \\cdot \\cdot \\}$ $E _ { i }$ $E _ { j }$ $( i \\neq j )$ $\\exists i , j , p , q ,$ s.t. $g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }$ . If we find two such failed episodes, we combine the two experience by replacing the desired goals in $E _ { i }$ by $\\{ g _ { j , t } ^ { d e } \\}$ , where $j$ indicates $E _ { j }$ and $t \\leq \\operatorname* { m i n } \\{ p , q \\}$ . Based on this, we end up assembling a new experience $E _ { i } ^ { \\prime }$ based on $E _ { i }$ with a new “imagined” goal trajectory $\\{ g _ { j , 0 } ^ { d e } , \\cdot \\cdot \\cdot , g _ { j , t } ^ { d e } \\}$ where $t \\leq \\operatorname* { m i n } \\{ p , q \\}$ . ", + "bbox": [ + 173, + 536, + 825, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "More details of the complete $\\mathrm { R L + D H E R }$ method are shown in Algorithm 1. Unlike HER, our DHER needs to search all failed experiences to compose a “imagined” goal trajectory. Hence, the efficiency of searching the memory is important. In our implementation, we use two hash tables to store the trajectories of achieved goals and desired goals, respectively. ", + "bbox": [ + 174, + 723, + 823, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENT ", + "text_level": 1, + "bbox": [ + 174, + 821, + 316, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We run extensive experiments to examine the proposed DHER for moving goals and compare it with some competing baselines. We first introduce the environments and tasks that we want to address, followed by the experimental results. Demo videos from our experiments are available at https://sites.google.com/view/dher. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/868b5ca50a241e6883855868e5cdc6d4e2e4a7d4f491d103d018de67c19012a7.jpg", + "image_caption": [ + "Figure 3: Results on different environments. " + ], + "image_footnote": [], + "bbox": [ + 181, + 99, + 813, + 263 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 ENVIRONMENTS ", + "text_level": 1, + "bbox": [ + 174, + 320, + 328, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Whereas grasping moving targets are fairly common in robotic applications, there are rarely existing environments featuring dynamic goals. We modify the robotic manipulation environments created by OpenAI (Brockman et al., 2016) for our experiments. As shown in Figure 2, we assign certain rules to the goals so that they accordingly move in the environments while an agent is required to control the robotic arm’s grippers to reach the goal that moves along a straight line (Dy-Reaching), to reach the goal that moves in a circle (Dy-Circling), or to push a block to the goal that moves along a straight line (Dy-Pushing). In addition, we also develop a new GREEDY SNAKE environment (DySnake), in which the goal moves from one discrete cell to another. The greedy snake aims to reach the goal (and eat it). No matter what actions are taken by the agent, the underlying rule of changing the goals’ positions remain the same. We assume that the law of the goal’s motion is unknown to the agent. More details of the tasks are described in the appendix. ", + "bbox": [ + 173, + 348, + 825, + 502 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the first three tasks, we follow the basic settings of OpenAI robotics environments (Brockman et al., 2016). States are read from the MuJoCo physics engine. An observation consists of relative positions of the object and the target (grippers are blocked) Goals are positions in the 3D world coordinate system with a fixed tolerance (we use $\\epsilon = 0 . 0 1$ for the tolerance). Note the goals are able to move in our tasks. The velocity we use is $v = 0 . 0 1 1$ . 1 epoch indicates 100 episodes. The start positions of the goals are randomly chosen. Rewards are binary and sparse: $r ( s _ { t } , a _ { t } , g _ { t } ) =$ $- \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( | s _ { t + 1 } ^ { \\mathrm { o b j } } - g _ { t + 1 } | \\geq \\epsilon \\big )$ where $s _ { t + 1 }$ and $g _ { t + 1 }$ are respectively the state and dynamic goal after the execution of the action $a _ { t }$ in the state $s _ { t }$ . Two positions are overlapped if they are close within the tolerance $\\epsilon$ . We use 3-dimensional actions, in which three dimensions correspond to the desired relative gripper position at the next timestep. ", + "bbox": [ + 173, + 508, + 825, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the last task, it is very similar to the standard snake game. There is a 2D plane grid whose size is $3 0 \\times 4 0$ . A snake and a goal (food) both move in this grid. We use a $1 \\times 1$ square as the body of the snake and do not allow the snake to grow any longer as it moves. The states of the system are represented by using the positions of the snake and food. Goals are the positions of the food. The snake has to move itself such that it resides in the same cell as the food at a certain timestep. The velocity of the goal is set to $( 0 , 1 )$ . The start positions of the goal are set randomly in different episodes of the game, so are the start positions of the snake. Rewards are binary and sparse: $r \\big ( s _ { t } , \\bar { a } _ { t } , g _ { t + 1 } \\big ) = - \\bar { \\mathbf { 1 } } _ { \\mathrm { c o n d i t i o n } } \\big ( \\big | s _ { t + 1 } ^ { \\mathrm { s n a k e } } - g _ { t + 1 } \\big | \\bar { \\neq } 0 \\big )$ where $s _ { t + 1 }$ and $g _ { t + 1 }$ are the environment’s state and goal after the agent executes action $a _ { t }$ in the state $s _ { t }$ . Observations are represented by the positions of the snake and food. We also add to the state the distance between the snake and food. Actions allowed in the game are the following: move up, move down, move left, and move right, for one cell per timestep. ", + "bbox": [ + 173, + 657, + 825, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2 BASELINES", + "text_level": 1, + "bbox": [ + 174, + 847, + 294, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the first three tasks which call for continuous control, we consider two competing baselines:2 • DDPG, which is a model-free RL algorithm for continuous control (Lillicrap et al., 2015). It learns a deterministic policy by using a stochastic counterpart to explore in the training. ", + "bbox": [ + 176, + 876, + 805, + 890 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/732a387a80c212738b2fb410b777a5c71d62a4dacb48b29dd438871f23abaf26.jpg", + "image_caption": [ + "Figure 4: Low velocity: 0.001. " + ], + "image_footnote": [], + "bbox": [ + 220, + 102, + 449, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8c0fd7014d07719a2c477eef04492742d4a9d1ff91065b3d850a3053e6e0ee6.jpg", + "image_caption": [ + "Figure 5: High velocity: 0.016. " + ], + "image_footnote": [], + "bbox": [ + 545, + 102, + 772, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 217, + 299, + 823, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "• $\\mathrm { D D P G } + \\mathrm { H E R }$ , which improves the replay buffer of DDPG by the hindsight experience replay (Andrychowicz et al., 2017). \nDDPG (dense), which employs the negative distance $( - d )$ as dense rewards. \nDDPG (dense-2), which employs the negative distance $( - d )$ as dense rewards if $d \\geq \\epsilon$ . However, if $d < \\epsilon$ (i.e., success) it uses $( - d + 1 . 0 )$ as rewards. 1.0 is a bonus. ", + "bbox": [ + 215, + 333, + 825, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the last task of discrete control, we use the following baselines in the experiments: DQN (Mnih et al., 2015) and $\\mathrm { D Q N + H E R }$ , which uses HER to enhance the replay in DQN. ", + "bbox": [ + 171, + 426, + 823, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "• DQN, which is a powerful model-free RL algorithm for discrete action spaces (Mnih et al., 2015). \n• $\\mathrm { D Q N + H E R }$ , which uses HER to enhance the replay in DQN. \n• DQN (dense), which uses the negative distance $( - d )$ as dense rewards. \n• DQN (dense-2), which uses the negative distance $( - d )$ as dense rewards. However, if $d = 0 . 0$ (i.e., success) it uses $( - d + 1 . 0 )$ as rewards instead. 1.0 is a bonus. ", + "bbox": [ + 215, + 465, + 825, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 COMPARISON RESULTS ON THE ROBOTIC ENVIRONMENTS ", + "text_level": 1, + "bbox": [ + 174, + 585, + 612, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the continuous control, we present three sets of comparison results in Figure 3 for the first three tasks, respectively. Consistently, the results show our DHER algorithm outperforms the others. The two baselines are not able to catch up even after we train them for thousands of iterations. Vanilla DDPG is slightly better than the version with HER. HER does not benefit DDPG in these tasks because the goals in HER are fixed, fundamentally misleading the agent in the attempt of solving the tasks with dynamic goals. The results of DDPG (dense) and DDPG (dense-2) suggest that even the dense rewards do not work well as they are agnostic to the task of interest. A good reward shaping may give rise to better performance by carefully tuning it for the task of dynamic goals. ", + "bbox": [ + 173, + 611, + 825, + 723 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing Figure 3a and Figure 3b with Figure 3c, we find that DHER learns faster in Dy-Reaching and Dy-Circling than Dy-Pushing probably because Dy-Reaching and Dy-Circling are easier tasks than Dy-Pushing. In Dy-Pushing, all the algorithms take a fairly big amount of time to explore without receiving any positive feedback. However, the more failed experiences the agent encounters, the better change our algorithm is able to identify relevant episodes from them for assembling useful dynamic goals. As a result, DHER is able to pick up the momentum and learns faster and better than the baselines after a certain point. In Figure 3a, the performance decreases a little. The reason may be that as successful experience increases, some assembled experience is inconsistent with these successful experience. ", + "bbox": [ + 174, + 729, + 825, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.1 COMPARISON USING DIFFERENT VELOCITIES OF GOALS", + "text_level": 1, + "bbox": [ + 174, + 872, + 612, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show the performance of our method on more complex tasks with different velocities, we study different methods in Dy-Reaching environment as shown in Figures 4 and 5 with the same physical properties as the previous experiments. Overall the results show our method is much better than DDPG and HER. As a reminder, the threshold for calculating rewards we used is 0.01. In Figure 4, the task becomes an easier task because of the slower velocity. It shows that our method quickly achieves to a good result around 5 epoch. Comparing with Figure 3a, it shows that the performance with $v = 0 . 0 0 1$ is better than the performance with $v = 0 . 0 1 1$ and get $1 5 \\%$ improvements. However, it also shows the performance with $v = 0 . 0 1 6$ is worse than the performance with $v = 0 . 0 1 1$ . Both DDPG and $\\mathrm { \\Delta D D P G + H E R }$ failed when $v = 0 . 0 1 6$ and their performance is 0. This performance is consistent because the task becomes more difficult when the velocity increases. ", + "bbox": [ + 174, + 895, + 820, + 924 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/14dd0b3e1878d3f49fb1b7d44ca84006cca0ac15133dc036aef116910149f401.jpg", + "image_caption": [ + "Figure 6: Snake with dynamic goals. " + ], + "image_footnote": [], + "bbox": [ + 220, + 103, + 450, + 242 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97b49eb4c50a91b591b5c09b7c8417f90a64b3336f8b06a8631418b78f185b01.jpg", + "image_caption": [ + "Figure 7: Special: $\\mathrm { H E R + }$ knows how goals move. " + ], + "image_footnote": [], + "bbox": [ + 545, + 102, + 774, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 306, + 825, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4 COMPARISON ON THE DYNAMIC SNAKE ENVIRONMENT ", + "text_level": 1, + "bbox": [ + 174, + 445, + 594, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the Dy-Snake environment, which is a discrete control environment, we present the results of the chasing food task in Figure 6. The results show that the proposed algorithm works best. DQN is better than HER. HER fails for this task. That HER fails for this task shows just using achieved goals is not enough for the tasks with dynamic goals and can lead wrong direction. The results also show that around 800 episodes, the performance of DQN and DHER is close and DHER is slightly better than DQN. This is because the chasing food task is a simple task. The action space is very small and just 4 types of actions. After enough exploration, DQN also has competitive performance. However, at the beginning, DHER quickly achieves very good performance. It shows that assembling experience from two failures improves the performance very efficiently in this task. DQN (dense) and DQN (dense-2) help learn the policy at the early stage. However, in the long run, it does not lead to any particular benefits. ", + "bbox": [ + 173, + 473, + 825, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5 SIM TO REAL ADAPTATION ", + "text_level": 1, + "bbox": [ + 176, + 654, + 398, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We used policies for Dy-Circling task and a new Pouring task trained in our simulator 3 to deploy them on a physical robot. As shown in Figure 8, the policies were trained by using DHER and adapted to the real robot without any finetuning. However, the policy requires accurate localizations of the gripper and the goal. For Dy-Circling task, the robot’s gripper was blocked. There were a toy turntable, whose speed is unknown, and a blue block on the turntable. We set the position 1cm above the block as the target position. The position of the block was predicted based on traditional contour shape analysis using camera images. For Pouring task, the robot gripped a can. A man held a cup and moved it. The cup was set as the target and with a green marker. We used the marker to estimate its position. ", + "bbox": [ + 174, + 683, + 825, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our policies were transferred successfully for both tasks. With the accurate positions, we have $100 \\%$ success rate for 5 trials. It was observed that the robot had learned to not only follow the current target but also step forward to the future target position. Demo videos about the experiments are available at https://sites.google.com/view/dher. ", + "bbox": [ + 174, + 815, + 823, + 871 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ff03243dba37e8e374b610dab429ef50349cfd9bf2cd37e4c19e2d40e16ac98c.jpg", + "image_caption": [ + "Figure 8: Adapting the policies trained based on DHER from our simulation to a real robotic arm. " + ], + "image_footnote": [], + "bbox": [ + 179, + 106, + 820, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "4.6 SPECIAL CASE: HOW ABOUT IF HER KNOWS THE LAW OF THE MOTION OF A GOAL", + "bbox": [ + 171, + 237, + 787, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We use Dy-Snake to demonstrate the experimental results for a special case that the law of the motion of the target (food) is known to agents. Because the motion of the food is very simple and controllable in Dy-Snake environment. We develop a direct extension of HER, called $\\mathrm { H E R + }$ , that modifies desired goals at every timestep based on the law of the motion of the food to create successful experience. More details of $\\mathrm { H E R + }$ are described in the appendix. ", + "bbox": [ + 174, + 262, + 825, + 332 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We show the results in Figure 7. DHER and $\\mathrm { H E R + }$ are both better than DQN at the beginning. DHER is slightly better than $\\mathrm { H E R + }$ , which shows the efficiency of DHER is comparable in this simple task. ", + "bbox": [ + 176, + 338, + 823, + 381 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 401, + 318, + 417 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduced a novel technique that assembles successful experience from a couple of failures. With this technique, our proposed algorithm called DHER ( Dynamic Hindsight Experience Replay) is able to address the tasks with sparse rewards and dynamic goals. Our technique can be combined with an arbitrary off-policy RL algorithm and we experimentally demonstrated that with DQN and DDPG. As far as we know, it is the first time that an agent is allowed to learn from assembled experience from two failures. ", + "bbox": [ + 174, + 433, + 825, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGMENTS ", + "text_level": 1, + "bbox": [ + 176, + 537, + 356, + 553 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Weitao Xi, Tianzhou Wang, Tingguang Li for performing some additional transfer experiments. We would also like to thank Han Liu and the whole RL team for fruitful discussions as well as the anonymous reviewers for their comments. ", + "bbox": [ + 174, + 568, + 825, + 609 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 631, + 285, + 646 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob McGrew, Josh Tobin, OpenAI. Pieter Abbeel, and Wojciech Zaremba. Hindsight experience replay. In Advances in Neural Information Processing Systems, pp. 5048–5058, 2017. \nMarc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning environment: An evaluation platform for general agents. Journal of Artificial Intelligence Research, 47:253–279, 2013. ISSN 1076-9757. \nYoshua Bengio, Jrme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th International Conference on Machine Learning, pp. 41–48. ACM, 2009. \nGreg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym, 2016. \nMaurizio Di Rocco, Federico Pecora, Prasanna Kumar Sivakumar, and Alessandro Saffiotti. Configuration planning with multiple dynamic goals. In AAAI Spring Symposium: Designing Intelligent Robots, 2013. \nYan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep reinforcement learning for continuous control. In Proceedings of the 33rd International Conference on Machine Learning, pp. 1329–1338, 2016. ", + "bbox": [ + 171, + 654, + 826, + 925 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Carlos Florensa, David Held, Xinyang Geng, and Pieter Abbeel. Automatic goal generation for reinforcement learning agents. In Proceedings of the 35th International Conference on Machine Learning, volume 80, pp. 1515–1528, 2018. ", + "bbox": [ + 176, + 103, + 823, + 146 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nicolas Heess, Dhruva TB, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa, Tom Erez, Ziyu Wang, S. M. Ali Eslami, Martin A. Riedmiller, and David Silver. Emergence of locomotion behaviours in rich environments. CoRR, abs/1707.02286, 2017. ", + "bbox": [ + 176, + 155, + 820, + 198 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Leslie Pack Kaelbling. Learning to achieve goals. In Proceedings of the 13th International Joint Conference on Artificial Intelligence, pp. 1094–1099. Citeseer, 1993. ", + "bbox": [ + 173, + 207, + 821, + 236 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. ", + "bbox": [ + 173, + 244, + 823, + 286 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching. Machine Learning, 8(3-4):293–321, 1992. ", + "bbox": [ + 173, + 295, + 820, + 325 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning. Nature, 518:529, 2015. ", + "bbox": [ + 173, + 333, + 825, + 404 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sanmit Narvekar, Jivko Sinapov, and Peter Stone. Autonomous task sequencing for customized curriculum design in reinforcement learning. In Proceedings of the 27th International Joint Conference on Artificial Intelligence, volume 147, pp. 149, 2017. ", + "bbox": [ + 171, + 412, + 825, + 457 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andrew Y. Ng, Daishi Harada, and Stuart Russell. Policy invariance under reward transformations: Theory and application to reward shaping. In Proceedings of the 16th International Conference on Machine Learning, volume 99, pp. 278–287, 1999. ", + "bbox": [ + 174, + 463, + 823, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tom Schaul, Daniel Horgan, Karol Gregor, and David Silver. Universal value function approximators. In Proceedings of the 32nd International Conference on Machine Learning, pp. 1312–1320, 2015a. ", + "bbox": [ + 174, + 515, + 825, + 558 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015b. ", + "bbox": [ + 171, + 566, + 823, + 597 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel. Trust region policy optimization. CoRR, abs/1502.05477, 2015. ", + "bbox": [ + 171, + 604, + 825, + 635 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. CoRR, abs/1707.06347, 2017. ", + "bbox": [ + 171, + 642, + 823, + 672 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484, 2016. ", + "bbox": [ + 176, + 680, + 823, + 723 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033. IEEE, 2012. ", + "bbox": [ + 173, + 732, + 825, + 775 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ziyu Wang, Victor Bapst, Nicolas Heess, Volodymyr Mnih, Remi Munos, Koray Kavukcuoglu, and ´ Nando de Freitas. Sample efficient actor-critic with experience replay. CoRR, abs/1611.01224, 2016. ", + "bbox": [ + 174, + 784, + 825, + 827 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 103, + 263, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "TASKS WITH DYNAMIC GOALS ", + "text_level": 1, + "bbox": [ + 176, + 135, + 385, + 148 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We have created four tasks as described below and illustrated by Figure 2. The first three tasks are based on the robotic environments. The last is based on the GREEDY SNAKE environment. ", + "bbox": [ + 173, + 161, + 825, + 189 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "• Dy-Reaching: The task is to control the robotic arm so that its gripper can reach the target position. The target position moves from one point to another along a straight line with a constant velocity. \n• Dy-Circling: The task is to control the robotic arm so that its gripper can reach the target position. The target position moves along a circle with a constant velocity. Dy-Pushing: in this task, a box is placed on a table in front of the robotic arm. The robot is required to move the box to the target location on the table. The target location moves from one position to another with a fixed velocity. Note that the robot’s grippers are locked to prevent it from grasping. The learned behavior is actually a mixture of pushing and rolling. \nDy-Snake: in this task, a snake and food are placed on a rectangle map and the task is to control the snake to eat the food. The food is moving from one position to another position with a fixed velocity. ", + "bbox": [ + 215, + 202, + 825, + 388 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "HARDWARE ", + "text_level": 1, + "bbox": [ + 174, + 407, + 258, + 421 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We use the Universal Robots UR10 with a gripper. We use a RealSense Camera SR300 to track the position of objects. The gripper is blocked. We use a marker on the gripper for camera calibration. During adapting the robot, for the position of the target, we use the camera to estimate it. ", + "bbox": [ + 176, + 434, + 825, + 477 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "SIMULATION ", + "text_level": 1, + "bbox": [ + 176, + 496, + 264, + 508 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "We simulate the physical system using the MuJoCo physics engine (Todorov et al., 2012) and also use MuJoCo to render the images. In our tasks, we use positions provided by MuJoCo for training policies in the simulation. ", + "bbox": [ + 176, + 522, + 825, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "In Figure 9, we demonstrate our simulation and the physics environment. ", + "bbox": [ + 173, + 570, + 651, + 585 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/0ab0a04331e1ed7ee05503beb83d72a71ddd08962ea6776f4dbc386edf14fc2c.jpg", + "image_caption": [ + "Figure 9: Our simulation and the physics environment. " + ], + "image_footnote": [], + "bbox": [ + 215, + 604, + 782, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A SPECIAL AND SIMPLER CASE - HER+", + "bbox": [ + 176, + 868, + 455, + 882 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "The paper considers a general situation that the law of the motion of goals is invisible to an agent. However, we relax this assumption and assume that in some situations the law of the motion of goals is easy to obtain, i.e., the parameters of $g ( t ; \\gamma )$ are known. For example, it is able to calculate the velocity of an object by observing the trajectories of the object. ", + "bbox": [ + 173, + 895, + 823, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "In this situation, we do not need to search failed experience. With the knowledge of the velocity of an object, it is straightforward that we can calculate any trajectory of desired goals. Thus for every failed experience, at any time step $t$ , based on the achieved goal $g _ { t } ^ { a c }$ , we calculate new desired goals $g _ { t } ^ { d e } = g ( { \\dot { t } } )$ corresponding to the achieved goals in order to construct successful experience. ", + "bbox": [ + 174, + 138, + 825, + 195 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "The details of $\\mathrm { H E R + }$ are described in Algorithm 2. $\\mathrm { H E R + }$ can be seen as a direct extension of HER. HER and $\\mathrm { H E R + }$ can modify any failed experience to successful experience because they both assume the law of the motion of goals is known. ", + "bbox": [ + 176, + 202, + 821, + 244 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Algorithm 2 Hindsight Experience Replay Plus ", + "text_level": 1, + "bbox": [ + 176, + 260, + 490, + 275 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Require: an off-policy RL algorithm A, replay buffer $R$ , a reward function $r$ \n1: Initialize A and replay buffer $R$ \n2: for episode $\\mathbf { \\Omega } = 1 , 2 , \\cdots , M$ do \n3: Sample an initial goal $g _ { 0 }$ and an initial state $s _ { 0 }$ \n4: for $t = 0 , \\cdots , T - 1$ do \n5: Sample an action $a _ { t }$ using the behavioral policy from A: \n6: $a _ { t } \\doteq \\pi ( s _ { t } | g _ { t } )$ \n7: Execute the action $a _ { t }$ and observe a new state $s _ { t + 1 }$ and a new goal $g _ { t + 1 }$ \n8: end for \n9: for $t = 0 , \\cdots , T - 1$ do \n10: $r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )$ \n11: Store the transitions $\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)$ in $R$ (Standard experience replay) \n12: Sample a set of the achieved goals of $E _ { i }$ as additional goals for reply $G ^ { \\prime }$ \n13: for $\\bar { g _ { p } ^ { \\prime } } \\in G ^ { \\prime }$ do \n14: Calculate a goal trajectory $\\{ g _ { 0 } ^ { \\prime } , \\cdots , g _ { p } ^ { \\prime } \\}$ where $g _ { t } ^ { \\prime } = g ( t ; \\gamma )$ $\\mathrm { ( H E R + ) }$ \n15: for $t = \\{ 0 , \\cdots , p - 1 \\}$ do \n16: $r _ { t } ^ { \\prime } : = r ( s _ { i , t } , a _ { i , t } , g _ { t + 1 } ^ { \\prime } )$ \n17: Store the transition $\\left( s _ { i , t } | g _ { t } ^ { \\prime } , a _ { i , t } , r _ { t } ^ { \\prime } , s _ { i , t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R$ \n18: end for \n19: end for \n20: end for \n21: end for ", + "bbox": [ + 176, + 280, + 766, + 589 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/Byf5-30qFX/Byf5-30qFX_middle.json b/parse/train/Byf5-30qFX/Byf5-30qFX_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..7a8da440177896c112385223d2048de612e59450 --- /dev/null +++ b/parse/train/Byf5-30qFX/Byf5-30qFX_middle.json @@ -0,0 +1,33613 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 415, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 417, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 417, + 97 + ], + "score": 1.0, + "content": "DHER: HINDSIGHT EXPERIENCE REPLAY", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 272, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 272, + 118 + ], + "score": 1.0, + "content": "FOR DYNAMIC GOALS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 407, + 156 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 407, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 407, + 149 + ], + "score": 1.0, + "content": "Meng Fang∗, Cheng Zhou, Bei Shi, Boqing Gong, Jia Xu, Tong Zhang", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 177, + 157 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 177, + 157 + ], + "score": 1.0, + "content": "Tencent AI Lab", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 211, + 468, + 277 + ], + "lines": [ + { + "bbox": [ + 142, + 211, + 469, + 223 + ], + "spans": [ + { + "bbox": [ + 142, + 211, + 469, + 223 + ], + "score": 1.0, + "content": "Dealing with sparse rewards is one of the most important challenges in reinforce-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "score": 1.0, + "content": "ment learning (RL), especially when a goal is dynamic (e.g., to grasp a moving", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "score": 1.0, + "content": "object). Hindsight experience replay (HER) has been shown an effective solution", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "score": 1.0, + "content": "to handling sparse rewards with fixed goals. However, it does not account for dy-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 470, + 267 + ], + "score": 1.0, + "content": "namic goals in its vanilla form and, as a result, even degrades the performance of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 266, + 428, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 266, + 428, + 279 + ], + "score": 1.0, + "content": "existing off-policy RL algorithms when the goal is changing over time.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 143, + 279, + 468, + 377 + ], + "lines": [ + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "In this paper, we present Dynamic Hindsight Experience Replay (DHER), a novel", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 302 + ], + "score": 1.0, + "content": "approach for tasks with dynamic goals in the presence of sparse rewards. DHER", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 302, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 469, + 312 + ], + "score": 1.0, + "content": "automatically assembles successful experiences from two relevant failures and can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "be used to enhance an arbitrary off-policy RL algorithm when the tasks’ goals are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "dynamic. We evaluate DHER on tasks of robotic manipulation and moving ob-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 140, + 334, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 140, + 334, + 470, + 346 + ], + "score": 1.0, + "content": "ject tracking, and transfer the polices from simulation to physical robots. Exten-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 469, + 357 + ], + "score": 1.0, + "content": "sive comparison and ablation studies demonstrate the superiority of our approach,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 354, + 469, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 369 + ], + "score": 1.0, + "content": "showing that DHER is a crucial ingredient to enable RL to solve tasks with dy-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 357, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 357, + 378 + ], + "score": 1.0, + "content": "namic goals in manipulation and grid world domains.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15 + }, + { + "type": "title", + "bbox": [ + 108, + 398, + 206, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 208, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 208, + 414 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "Deep reinforcement learning has been shown an effective framework for solving a rich repertoire", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "of complex control problems. In simulated domains, agents have been trained to perform a diverse", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "score": 1.0, + "content": "array of challenging tasks (Mnih et al., 2015; Lillicrap et al., 2015; Duan et al., 2016). In order to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "train such agents, it is often the case that one has to design a reward function that not only reflects", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 297, + 480 + ], + "score": 1.0, + "content": "the task at hand but also is carefully shaped", + "type": "text" + }, + { + "bbox": [ + 297, + 468, + 312, + 479 + ], + "score": 0.43, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "et al., 1999) to guide the policy optimization.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "Unfortunately, many of the capabilities demonstrated by reward engineering are often limited to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "specific tasks. Moreover, it requires both RL expertise and domain-specific knowledge to reshape", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 500, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 511 + ], + "score": 1.0, + "content": "the reward functions. For situations where we do not know what admissible behavior may look", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "like, for example, using LEGO bricks to build a desired architecture, it is difficult to apply reward", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "score": 1.0, + "content": "engineering. Therefore, it is essential to develop algorithms which can learn from unshaped and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 534, + 227, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 227, + 545 + ], + "score": 1.0, + "content": "usually sparse reward signals.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Learning with sparse rewards is challenging, especially when a goal is dynamic. Dynamic goals are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "common in games and planning problems, often addressed using reward shaping or search (Kael-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "bling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). However, the difficulty posed by a sparse", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 583, + 504, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 504, + 595 + ], + "score": 1.0, + "content": "reward is exacerbated by the complicated environment dynamics in robotics (Andrychowicz et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 607 + ], + "score": 1.0, + "content": "2017). For instance, system dynamics around contacts are difficult to model and induce sensitivity in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "the system to small errors. Many robotic tasks also need executing multiple steps successfully over", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "a long horizon, involve enormous search space, and require generalization to varying task instances.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "Policy gradient methods are breakthroughs in the challenging environments, such as PPO (Heess", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "et al., 2017; Schulman et al., 2017), ACER (Wang et al., 2016), TRPO (Schulman et al., 2015) and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "so on. They are used in environments, where an agent tries to reach a target, learns to walk, runs, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "so on. Recently, sampling-efficient learning is introduced and demonstrates a significant increase in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "performance for off-policy actor-critic DQN (Mnih et al., 2015) and DDPG (Lillicrap et al., 2015)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "algorithms. Hindsight experience replay (HER) is very effective for improving the performance", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 691, + 506, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 506, + 706 + ], + "score": 1.0, + "content": "of off-policy RL algorithms in solving goal-based tasks with sparse rewards (Andrychowicz et al.,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "2017). Similar to UVFA (Schaul et al., 2015a), it takes a goal state as part of input. However, it", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 39 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 722, + 331, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 333, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 333, + 734 + ], + "score": 1.0, + "content": "∗Correspondence to: Meng Fang .", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 78, + 415, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 417, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 417, + 97 + ], + "score": 1.0, + "content": "DHER: HINDSIGHT EXPERIENCE REPLAY", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 272, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 272, + 118 + ], + "score": 1.0, + "content": "FOR DYNAMIC GOALS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 113, + 134, + 407, + 156 + ], + "lines": [ + { + "bbox": [ + 111, + 133, + 407, + 149 + ], + "spans": [ + { + "bbox": [ + 111, + 133, + 407, + 149 + ], + "score": 1.0, + "content": "Meng Fang∗, Cheng Zhou, Bei Shi, Boqing Gong, Jia Xu, Tong Zhang", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 177, + 157 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 177, + 157 + ], + "score": 1.0, + "content": "Tencent AI Lab", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 133, + 407, + 157 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 211, + 468, + 277 + ], + "lines": [ + { + "bbox": [ + 142, + 211, + 469, + 223 + ], + "spans": [ + { + "bbox": [ + 142, + 211, + 469, + 223 + ], + "score": 1.0, + "content": "Dealing with sparse rewards is one of the most important challenges in reinforce-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 222, + 469, + 235 + ], + "score": 1.0, + "content": "ment learning (RL), especially when a goal is dynamic (e.g., to grasp a moving", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 470, + 246 + ], + "score": 1.0, + "content": "object). Hindsight experience replay (HER) has been shown an effective solution", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 244, + 470, + 257 + ], + "score": 1.0, + "content": "to handling sparse rewards with fixed goals. However, it does not account for dy-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 255, + 470, + 267 + ], + "spans": [ + { + "bbox": [ + 141, + 255, + 470, + 267 + ], + "score": 1.0, + "content": "namic goals in its vanilla form and, as a result, even degrades the performance of", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 266, + 428, + 279 + ], + "spans": [ + { + "bbox": [ + 142, + 266, + 428, + 279 + ], + "score": 1.0, + "content": "existing off-policy RL algorithms when the goal is changing over time.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5, + "bbox_fs": [ + 141, + 211, + 470, + 279 + ] + }, + { + "type": "text", + "bbox": [ + 143, + 279, + 468, + 377 + ], + "lines": [ + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 141, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "In this paper, we present Dynamic Hindsight Experience Replay (DHER), a novel", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 289, + 470, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 289, + 470, + 302 + ], + "score": 1.0, + "content": "approach for tasks with dynamic goals in the presence of sparse rewards. DHER", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 302, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 469, + 312 + ], + "score": 1.0, + "content": "automatically assembles successful experiences from two relevant failures and can", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "be used to enhance an arbitrary off-policy RL algorithm when the tasks’ goals are", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 469, + 335 + ], + "score": 1.0, + "content": "dynamic. We evaluate DHER on tasks of robotic manipulation and moving ob-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 140, + 334, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 140, + 334, + 470, + 346 + ], + "score": 1.0, + "content": "ject tracking, and transfer the polices from simulation to physical robots. Exten-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 469, + 357 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 469, + 357 + ], + "score": 1.0, + "content": "sive comparison and ablation studies demonstrate the superiority of our approach,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 354, + 469, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 354, + 469, + 369 + ], + "score": 1.0, + "content": "showing that DHER is a crucial ingredient to enable RL to solve tasks with dy-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 366, + 357, + 378 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 357, + 378 + ], + "score": 1.0, + "content": "namic goals in manipulation and grid world domains.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 15, + "bbox_fs": [ + 140, + 279, + 470, + 378 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 398, + 206, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 397, + 208, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 208, + 414 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 423, + 505, + 544 + ], + "lines": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "Deep reinforcement learning has been shown an effective framework for solving a rich repertoire", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "of complex control problems. In simulated domains, agents have been trained to perform a diverse", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "score": 1.0, + "content": "array of challenging tasks (Mnih et al., 2015; Lillicrap et al., 2015; Duan et al., 2016). In order to", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 455, + 505, + 470 + ], + "score": 1.0, + "content": "train such agents, it is often the case that one has to design a reward function that not only reflects", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 297, + 480 + ], + "score": 1.0, + "content": "the task at hand but also is carefully shaped", + "type": "text" + }, + { + "bbox": [ + 297, + 468, + 312, + 479 + ], + "score": 0.43, + "content": "\\mathrm { N g }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "et al., 1999) to guide the policy optimization.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 506, + 491 + ], + "score": 1.0, + "content": "Unfortunately, many of the capabilities demonstrated by reward engineering are often limited to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "specific tasks. Moreover, it requires both RL expertise and domain-specific knowledge to reshape", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 500, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 511 + ], + "score": 1.0, + "content": "the reward functions. For situations where we do not know what admissible behavior may look", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "like, for example, using LEGO bricks to build a desired architecture, it is difficult to apply reward", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 522, + 506, + 534 + ], + "score": 1.0, + "content": "engineering. Therefore, it is essential to develop algorithms which can learn from unshaped and", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 534, + 227, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 227, + 545 + ], + "score": 1.0, + "content": "usually sparse reward signals.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 423, + 506, + 545 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 550, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Learning with sparse rewards is challenging, especially when a goal is dynamic. Dynamic goals are", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 505, + 574 + ], + "score": 1.0, + "content": "common in games and planning problems, often addressed using reward shaping or search (Kael-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 571, + 505, + 585 + ], + "score": 1.0, + "content": "bling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). However, the difficulty posed by a sparse", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 583, + 504, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 504, + 595 + ], + "score": 1.0, + "content": "reward is exacerbated by the complicated environment dynamics in robotics (Andrychowicz et al.,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 592, + 506, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 506, + 607 + ], + "score": 1.0, + "content": "2017). For instance, system dynamics around contacts are difficult to model and induce sensitivity in", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 617 + ], + "score": 1.0, + "content": "the system to small errors. Many robotic tasks also need executing multiple steps successfully over", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "a long horizon, involve enormous search space, and require generalization to varying task instances.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "Policy gradient methods are breakthroughs in the challenging environments, such as PPO (Heess", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 505, + 650 + ], + "score": 1.0, + "content": "et al., 2017; Schulman et al., 2017), ACER (Wang et al., 2016), TRPO (Schulman et al., 2015) and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 661 + ], + "score": 1.0, + "content": "so on. They are used in environments, where an agent tries to reach a target, learns to walk, runs, and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 672 + ], + "score": 1.0, + "content": "so on. Recently, sampling-efficient learning is introduced and demonstrates a significant increase in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 684 + ], + "score": 1.0, + "content": "performance for off-policy actor-critic DQN (Mnih et al., 2015) and DDPG (Lillicrap et al., 2015)", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "algorithms. Hindsight experience replay (HER) is very effective for improving the performance", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 691, + 506, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 506, + 706 + ], + "score": 1.0, + "content": "of off-policy RL algorithms in solving goal-based tasks with sparse rewards (Andrychowicz et al.,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 506, + 716 + ], + "score": 1.0, + "content": "2017). Similar to UVFA (Schaul et al., 2015a), it takes a goal state as part of input. However, it", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 274, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 289 + ], + "score": 1.0, + "content": "assumes the goal is fixed. As a result, this assumption actually impedes the learning of RL agents in", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 287, + 246, + 298 + ], + "spans": [ + { + "bbox": [ + 107, + 287, + 246, + 298 + ], + "score": 1.0, + "content": "the environments of moving goals.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 550, + 506, + 716 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 82, + 504, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 82, + 504, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 82, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 110, + 82, + 504, + 207 + ], + "score": 0.966, + "type": "image", + "image_path": "a34ff9c3bf9148b8d43e21b58e1ea68fe1bcb07b8ac12d22cc17bf9a91fd4060.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 82, + 504, + 123.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 123.66666666666666, + 504, + 165.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 165.33333333333331, + 504, + 206.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 505, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "score": 1.0, + "content": "Figure 1: The framework of DHER. DHER is a kind of experience replay method. It searches", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 227, + 430, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 430, + 240 + ], + "score": 1.0, + "content": "relevant failed experiences and then assembles them into successful experiences.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 108, + 275, + 504, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 506, + 289 + ], + "score": 1.0, + "content": "assumes the goal is fixed. As a result, this assumption actually impedes the learning of RL agents in", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 287, + 246, + 298 + ], + "spans": [ + { + "bbox": [ + 107, + 287, + 246, + 298 + ], + "score": 1.0, + "content": "the environments of moving goals.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "In this paper, we address this challenge with a new method, Dynamic Hindsight Experience Re-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 314, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 327 + ], + "score": 1.0, + "content": "play (DHER), for accomplishing tasks with moving goals. We follow the multi-goal setting in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 323, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 339 + ], + "score": 1.0, + "content": "UVFA (Schaul et al., 2015a) and HER (Andrychowicz et al., 2017). It assumes that the goal being", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "pursued does not influence the environment dynamics. We also need to have the knowledge of goal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "similarity. For example, in manipulation or grid world domains, we can use Euclidean distance", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "between positions to measure the goal similarity. HER turns a failed episode to a success by com-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "posing a new task whose goal is achieved by that episode. Our idea allows an agent to learn from", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "the failure one step further than HER: the agent not only sets a new goal but also hallucinates how to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "reach the original goal from the new one. Take playing frisbee for instance. When an agent jumps to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "score": 1.0, + "content": "catch the frisbee and yet misses it, the agent receives no positive feedback under the sparse reward", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "setting. Using HER, the agent could set the end of its episode as the new goal — position of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "frisbee; with DHER, however, the agent finds a trajectory from its past experiences as the imagined", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 435, + 387, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 387, + 448 + ], + "score": 1.0, + "content": "path of the frisbee, and thereby extrapolates towards the original goal.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "In particular, we do the following for DHER. To finish the tasks with dynamic goals needs to explore", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "experience and understand multiple goals. DHER uses replay buffers to allow the agent to learn from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "a couple of failures by assembling new ‘experience’ from different episodes. The proposed method", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "retrieves memories to find the connection between the experience of different episodes. It largely", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "improves the sample efficiency in dynamic goal task settings. More importantly, this strategy makes", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 507, + 436, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 436, + 519 + ], + "score": 1.0, + "content": "it possible to learn in the setting that both sparse rewards and dynamic goals exist.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 504, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "We evaluate our method along with the state-of-the-art baselines on new environments and manipu-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "lation tasks, which have sparse rewards and dynamic goals. Our results demonstrate that DHER is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "clearly better than others for these tasks. We also transfer policies trained in our simulation based", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "on DHER to a physical robot and show that DHER can be applied to solving real-world robotics", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 567, + 149, + 580 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 149, + 580 + ], + "score": 1.0, + "content": "problems.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 107, + 584, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 596 + ], + "score": 1.0, + "content": "We summarize our main contributions as follows: (1) We demonstrate that, both in simulation and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "real worlds, DHER succeeds in continuous control with a moving target. To our knowledge, this is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "the first empirical result on manipulation tasks that demonstrates model-free learning methods can", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 631 + ], + "score": 1.0, + "content": "tackle tasks of this complexity. (2) We show that with assembling new experience from two failures,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "the sample complexity can be reduced dramatically. We attribute this to global knowledge learning", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "in a set of failed experience which breaks the constraint of local one-episode experience towards", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "more robust strategies. (3) We design and implement a set of new environments and continuous", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "score": 1.0, + "content": "tasks with dynamic goals, which would be of interest to researchers at the intersection of robotics", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 671, + 221, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 221, + 685 + ], + "score": 1.0, + "content": "and reinforcement learning.1", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 721, + 444, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 446, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 446, + 734 + ], + "score": 1.0, + "content": "1Our code and environments are available at https://github.com/mengf1/DHER.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 82, + 504, + 207 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 82, + 504, + 207 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 82, + 504, + 207 + ], + "spans": [ + { + "bbox": [ + 110, + 82, + 504, + 207 + ], + "score": 0.966, + "type": "image", + "image_path": "a34ff9c3bf9148b8d43e21b58e1ea68fe1bcb07b8ac12d22cc17bf9a91fd4060.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 82, + 504, + 123.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 123.66666666666666, + 504, + 165.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 165.33333333333331, + 504, + 206.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 505, + 239 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 229 + ], + "score": 1.0, + "content": "Figure 1: The framework of DHER. DHER is a kind of experience replay method. It searches", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 227, + 430, + 240 + ], + "spans": [ + { + "bbox": [ + 106, + 227, + 430, + 240 + ], + "score": 1.0, + "content": "relevant failed experiences and then assembles them into successful experiences.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 108, + 275, + 504, + 298 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 274, + 506, + 298 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 316 + ], + "score": 1.0, + "content": "In this paper, we address this challenge with a new method, Dynamic Hindsight Experience Re-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 314, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 505, + 327 + ], + "score": 1.0, + "content": "play (DHER), for accomplishing tasks with moving goals. We follow the multi-goal setting in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 323, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 506, + 339 + ], + "score": 1.0, + "content": "UVFA (Schaul et al., 2015a) and HER (Andrychowicz et al., 2017). It assumes that the goal being", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "pursued does not influence the environment dynamics. We also need to have the knowledge of goal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "similarity. For example, in manipulation or grid world domains, we can use Euclidean distance", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 505, + 371 + ], + "score": 1.0, + "content": "between positions to measure the goal similarity. HER turns a failed episode to a success by com-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "posing a new task whose goal is achieved by that episode. Our idea allows an agent to learn from", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "the failure one step further than HER: the agent not only sets a new goal but also hallucinates how to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 506, + 404 + ], + "score": 1.0, + "content": "reach the original goal from the new one. Take playing frisbee for instance. When an agent jumps to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 416 + ], + "score": 1.0, + "content": "catch the frisbee and yet misses it, the agent receives no positive feedback under the sparse reward", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 426 + ], + "score": 1.0, + "content": "setting. Using HER, the agent could set the end of its episode as the new goal — position of the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 437 + ], + "score": 1.0, + "content": "frisbee; with DHER, however, the agent finds a trajectory from its past experiences as the imagined", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 435, + 387, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 387, + 448 + ], + "score": 1.0, + "content": "path of the frisbee, and thereby extrapolates towards the original goal.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 303, + 506, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 505, + 464 + ], + "score": 1.0, + "content": "In particular, we do the following for DHER. To finish the tasks with dynamic goals needs to explore", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 506, + 475 + ], + "score": 1.0, + "content": "experience and understand multiple goals. DHER uses replay buffers to allow the agent to learn from", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "a couple of failures by assembling new ‘experience’ from different episodes. The proposed method", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "retrieves memories to find the connection between the experience of different episodes. It largely", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "improves the sample efficiency in dynamic goal task settings. More importantly, this strategy makes", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 507, + 436, + 519 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 436, + 519 + ], + "score": 1.0, + "content": "it possible to learn in the setting that both sparse rewards and dynamic goals exist.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 452, + 506, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 504, + 578 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "We evaluate our method along with the state-of-the-art baselines on new environments and manipu-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "lation tasks, which have sparse rewards and dynamic goals. Our results demonstrate that DHER is", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 558 + ], + "score": 1.0, + "content": "clearly better than others for these tasks. We also transfer policies trained in our simulation based", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "on DHER to a physical robot and show that DHER can be applied to solving real-world robotics", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 567, + 149, + 580 + ], + "spans": [ + { + "bbox": [ + 104, + 567, + 149, + 580 + ], + "score": 1.0, + "content": "problems.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 104, + 523, + 506, + 580 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 584, + 505, + 684 + ], + "lines": [ + { + "bbox": [ + 106, + 585, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 596 + ], + "score": 1.0, + "content": "We summarize our main contributions as follows: (1) We demonstrate that, both in simulation and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 595, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 607 + ], + "score": 1.0, + "content": "real worlds, DHER succeeds in continuous control with a moving target. To our knowledge, this is", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "the first empirical result on manipulation tasks that demonstrates model-free learning methods can", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 615, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 631 + ], + "score": 1.0, + "content": "tackle tasks of this complexity. (2) We show that with assembling new experience from two failures,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 506, + 642 + ], + "score": 1.0, + "content": "the sample complexity can be reduced dramatically. We attribute this to global knowledge learning", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "in a set of failed experience which breaks the constraint of local one-episode experience towards", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 663 + ], + "score": 1.0, + "content": "more robust strategies. (3) We design and implement a set of new environments and continuous", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 662, + 505, + 673 + ], + "score": 1.0, + "content": "tasks with dynamic goals, which would be of interest to researchers at the intersection of robotics", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 671, + 221, + 685 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 221, + 685 + ], + "score": 1.0, + "content": "and reinforcement learning.1", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 585, + 506, + 685 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 208, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 211, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 211, + 96 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "Recent works in deep RL have shown impressive results in different domains, such as games (Belle-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 130 + ], + "score": 1.0, + "content": "mare et al., 2013; Silver et al., 2016), simulated control tasks (Brockman et al., 2016) and so on.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "There have been several proposed RL methods for playing Atari games, including DQN (Mnih", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 153 + ], + "score": 1.0, + "content": "et al., 2015), UVFA (Schaul et al., 2015b) and so on. UVFA trains a single neural network approx-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 149, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 104, + 149, + 505, + 164 + ], + "score": 1.0, + "content": "imating multiple value functions for state and goal. For the continuous control, DDPG (Lillicrap", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "et al., 2015) is a popular actor-critic algorithm that has shown impressive results in continuous con-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "trol tasks. Dynamic goals appear in games and planning, often addressed by rewards shaping or", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "score": 1.0, + "content": "search (Kaelbling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). When the rewards are sparse,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 408, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 408, + 207 + ], + "score": 1.0, + "content": "there is few work studying the dynamic goals to the best of our knowledge.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 210, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "Curriculum learning is also used for reinforcement learning scenarios. The idea is that solving easier", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "score": 1.0, + "content": "problems first has advantages to learn more complex goals later and thus that learning can be opti-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "mized by presenting the problems in an optimal order, a curriculum (Bengio et al., 2009). Narvekar", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 244, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 504, + 256 + ], + "score": 1.0, + "content": "et al. (2017) and Florensa et al. (2018) proposed methods to automatically produce subtasks or sub-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 267 + ], + "score": 1.0, + "content": "goals for a given target. Narvekar et al. (2017) produce subtasks according to predefined tasks of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 265, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 104, + 265, + 506, + 278 + ], + "score": 1.0, + "content": "a given domain problem. Florensa et al. (2018) use a Generative Adversarial Network (GAN) to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 507, + 289 + ], + "score": 1.0, + "content": "produce goals with different difficulties. Different from these methods, our approach produces a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 320, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 320, + 301 + ], + "score": 1.0, + "content": "series of goals for an episode from failed experience.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Experience replay is an important technique and introduced to break temporal correlations by mix-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "ing more and less recent experience for updating policies (Lin, 1992). It was demonstrated for its", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 324, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 324, + 506, + 340 + ], + "score": 1.0, + "content": "efficiency in DQN (Mnih et al., 2015). Prioritized experience replay improves the speed of training", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "by considering prioritizing transitions in the replay buffer (Schaul et al., 2015b). HER considers", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "modifying experience in the replay buffer for continuous control (Andrychowicz et al., 2017). By", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "contrary, our approach assembles successful experience from two failures. Comparing with these", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "methods, which do not consider dynamic goals, our approach uses a series of goals to assemble", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "successful experience. As a result, our method is able to accomplish the tasks with sparse rewards", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 392, + 186, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 186, + 406 + ], + "score": 1.0, + "content": "and dynamic goals.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 419, + 209, + 432 + ], + "lines": [ + { + "bbox": [ + 104, + 417, + 211, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 211, + 435 + ], + "score": 1.0, + "content": "3 METHODOLOGY", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 443, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "We first review how HER works (Andrychowicz et al., 2017), followed by details of the proposed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 455, + 265, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 265, + 468 + ], + "score": 1.0, + "content": "DHER for dealing with dynamic goals.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "HER is a simple and effective method of manipulating the replay buffer used in off-policy RL", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "algorithms that allows it to learn policies more efficiently from sparse rewards. It assumes the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "goal being pursued does not influence the environment dynamics. After experiencing an episode", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 107, + 505, + 175, + 517 + ], + "score": 0.9, + "content": "\\bar { \\{ } s _ { 0 } , s _ { 1 } , \\dotsb , s _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 505, + 242, + 518 + ], + "score": 1.0, + "content": ", every transition", + "type": "text" + }, + { + "bbox": [ + 243, + 506, + 286, + 516 + ], + "score": 0.9, + "content": "s _ { t } \\to s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "along with the goal for this episode is usually stored in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "the replay buffer. Some of the saved episodes fail to reach the goal, providing no positive feedback", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "to the agent. However, with HER, the failed experience is modified and also stored in the replay", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "buffer in the following manner. The idea is to replace the original goal with a state visited by the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "failed episode. As the reward function remains unchanged, this change of goals hints the agent how", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 560, + 504, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 504, + 572 + ], + "score": 1.0, + "content": "to achieve the new goal in the environment. HER assumes that the mechanism of reaching the new", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 570, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 295, + 583 + ], + "score": 1.0, + "content": "goal helps the agent learn for the original goal.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 595, + 207, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 208, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 208, + 609 + ], + "score": 1.0, + "content": "3.1 DYNAMIC GOALS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Dynamic goals are not static and change at every timestep. We follow the multi-goal setting of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "Andrychowicz et al. (2017). The goals are part of the environment and do not influence the en-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 347, + 651 + ], + "score": 1.0, + "content": "vironment dynamics. We also assume that a dynamic goal", + "type": "text" + }, + { + "bbox": [ + 348, + 638, + 379, + 649 + ], + "score": 0.91, + "content": "g _ { t } \\in \\mathcal G", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "moves by following some law", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 649, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 156, + 661 + ], + "score": 0.93, + "content": "g _ { t } = g ( t ; \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 649, + 186, + 663 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 187, + 651, + 194, + 660 + ], + "score": 0.81, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 649, + 506, + 663 + ], + "score": 1.0, + "content": "parameterizes the law (e.g., acceleration in Newton’s law of motion), and yet", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 660, + 313, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 313, + 673 + ], + "score": 1.0, + "content": "its underlying moving law is unknown to the agent.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 493, + 690 + ], + "score": 1.0, + "content": "Moreover, we need to have some basic knowledge of goals, i.e., the measure of goal similarity on", + "type": "text" + }, + { + "bbox": [ + 493, + 677, + 501, + 687 + ], + "score": 0.79, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 676, + 505, + 690 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 171, + 702 + ], + "score": 1.0, + "content": "We assume that", + "type": "text" + }, + { + "bbox": [ + 172, + 688, + 200, + 699 + ], + "score": 0.92, + "content": "g _ { t } \\in \\mathcal G", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 687, + 325, + 702 + ], + "score": 1.0, + "content": "corresponds to some predicate", + "type": "text" + }, + { + "bbox": [ + 326, + 687, + 394, + 700 + ], + "score": 0.93, + "content": "f _ { g _ { t } } : S \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "and that the agent’s goal is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 187, + 712 + ], + "score": 1.0, + "content": "to achieve any state", + "type": "text" + }, + { + "bbox": [ + 187, + 701, + 194, + 709 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 698, + 246, + 712 + ], + "score": 1.0, + "content": "that satisfies", + "type": "text" + }, + { + "bbox": [ + 247, + 699, + 288, + 711 + ], + "score": 0.92, + "content": "f _ { g _ { t } } ( s ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 698, + 325, + 712 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 325, + 700, + 353, + 710 + ], + "score": 0.9, + "content": "S = \\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 698, + 399, + 712 + ], + "score": 1.0, + "content": "and define", + "type": "text" + }, + { + "bbox": [ + 399, + 699, + 474, + 711 + ], + "score": 0.91, + "content": "f _ { g _ { t } } ( s ) : = [ s = g _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 698, + 506, + 712 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 339, + 723 + ], + "score": 1.0, + "content": "can be considered as a measure of goal similarity between", + "type": "text" + }, + { + "bbox": [ + 340, + 712, + 349, + 721 + ], + "score": 0.84, + "content": "g _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 709, + 367, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 712, + 373, + 720 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 709, + 505, + 723 + ], + "score": 1.0, + "content": ". The goals can also specify only", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 378, + 733 + ], + "score": 1.0, + "content": "some properties of the state. Take manipulation tasks for instance:", + "type": "text" + }, + { + "bbox": [ + 378, + 720, + 412, + 731 + ], + "score": 0.92, + "content": "\\mathcal { G } = \\mathbb { R } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "corresponds to the 3D", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 208, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 211, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 211, + 96 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 505, + 118 + ], + "score": 1.0, + "content": "Recent works in deep RL have shown impressive results in different domains, such as games (Belle-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 116, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 116, + 505, + 130 + ], + "score": 1.0, + "content": "mare et al., 2013; Silver et al., 2016), simulated control tasks (Brockman et al., 2016) and so on.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "There have been several proposed RL methods for playing Atari games, including DQN (Mnih", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 138, + 505, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 153 + ], + "score": 1.0, + "content": "et al., 2015), UVFA (Schaul et al., 2015b) and so on. UVFA trains a single neural network approx-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 149, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 104, + 149, + 505, + 164 + ], + "score": 1.0, + "content": "imating multiple value functions for state and goal. For the continuous control, DDPG (Lillicrap", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "et al., 2015) is a popular actor-critic algorithm that has shown impressive results in continuous con-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "trol tasks. Dynamic goals appear in games and planning, often addressed by rewards shaping or", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 197 + ], + "score": 1.0, + "content": "search (Kaelbling, 1993; Mnih et al., 2015; Di Rocco et al., 2013). When the rewards are sparse,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 194, + 408, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 194, + 408, + 207 + ], + "score": 1.0, + "content": "there is few work studying the dynamic goals to the best of our knowledge.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 106, + 506, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 210, + 505, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 223 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 505, + 223 + ], + "score": 1.0, + "content": "Curriculum learning is also used for reinforcement learning scenarios. The idea is that solving easier", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 235 + ], + "score": 1.0, + "content": "problems first has advantages to learn more complex goals later and thus that learning can be opti-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "mized by presenting the problems in an optimal order, a curriculum (Bengio et al., 2009). Narvekar", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 244, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 504, + 256 + ], + "score": 1.0, + "content": "et al. (2017) and Florensa et al. (2018) proposed methods to automatically produce subtasks or sub-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 255, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 267 + ], + "score": 1.0, + "content": "goals for a given target. Narvekar et al. (2017) produce subtasks according to predefined tasks of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 265, + 506, + 278 + ], + "spans": [ + { + "bbox": [ + 104, + 265, + 506, + 278 + ], + "score": 1.0, + "content": "a given domain problem. Florensa et al. (2018) use a Generative Adversarial Network (GAN) to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 507, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 507, + 289 + ], + "score": 1.0, + "content": "produce goals with different difficulties. Different from these methods, our approach produces a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 287, + 320, + 301 + ], + "spans": [ + { + "bbox": [ + 106, + 287, + 320, + 301 + ], + "score": 1.0, + "content": "series of goals for an episode from failed experience.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 210, + 507, + 301 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 304, + 505, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 505, + 317 + ], + "score": 1.0, + "content": "Experience replay is an important technique and introduced to break temporal correlations by mix-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 506, + 328 + ], + "score": 1.0, + "content": "ing more and less recent experience for updating policies (Lin, 1992). It was demonstrated for its", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 324, + 506, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 324, + 506, + 340 + ], + "score": 1.0, + "content": "efficiency in DQN (Mnih et al., 2015). Prioritized experience replay improves the speed of training", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 350 + ], + "score": 1.0, + "content": "by considering prioritizing transitions in the replay buffer (Schaul et al., 2015b). HER considers", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 505, + 361 + ], + "score": 1.0, + "content": "modifying experience in the replay buffer for continuous control (Andrychowicz et al., 2017). By", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 505, + 372 + ], + "score": 1.0, + "content": "contrary, our approach assembles successful experience from two failures. Comparing with these", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "methods, which do not consider dynamic goals, our approach uses a series of goals to assemble", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "successful experience. As a result, our method is able to accomplish the tasks with sparse rewards", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 392, + 186, + 406 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 186, + 406 + ], + "score": 1.0, + "content": "and dynamic goals.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 303, + 506, + 406 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 419, + 209, + 432 + ], + "lines": [ + { + "bbox": [ + 104, + 417, + 211, + 435 + ], + "spans": [ + { + "bbox": [ + 104, + 417, + 211, + 435 + ], + "score": 1.0, + "content": "3 METHODOLOGY", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 108, + 443, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "We first review how HER works (Andrychowicz et al., 2017), followed by details of the proposed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 455, + 265, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 265, + 468 + ], + "score": 1.0, + "content": "DHER for dealing with dynamic goals.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 106, + 443, + 505, + 468 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "HER is a simple and effective method of manipulating the replay buffer used in off-policy RL", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "algorithms that allows it to learn policies more efficiently from sparse rewards. It assumes the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 506, + 507 + ], + "score": 1.0, + "content": "goal being pursued does not influence the environment dynamics. After experiencing an episode", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 505, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 107, + 505, + 175, + 517 + ], + "score": 0.9, + "content": "\\bar { \\{ } s _ { 0 } , s _ { 1 } , \\dotsb , s _ { T } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 505, + 242, + 518 + ], + "score": 1.0, + "content": ", every transition", + "type": "text" + }, + { + "bbox": [ + 243, + 506, + 286, + 516 + ], + "score": 0.9, + "content": "s _ { t } \\to s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 505, + 505, + 518 + ], + "score": 1.0, + "content": "along with the goal for this episode is usually stored in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 506, + 529 + ], + "score": 1.0, + "content": "the replay buffer. Some of the saved episodes fail to reach the goal, providing no positive feedback", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "to the agent. However, with HER, the failed experience is modified and also stored in the replay", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "buffer in the following manner. The idea is to replace the original goal with a state visited by the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 505, + 561 + ], + "score": 1.0, + "content": "failed episode. As the reward function remains unchanged, this change of goals hints the agent how", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 560, + 504, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 504, + 572 + ], + "score": 1.0, + "content": "to achieve the new goal in the environment. HER assumes that the mechanism of reaching the new", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 570, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 295, + 583 + ], + "score": 1.0, + "content": "goal helps the agent learn for the original goal.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 471, + 506, + 583 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 595, + 207, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 208, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 208, + 609 + ], + "score": 1.0, + "content": "3.1 DYNAMIC GOALS", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Dynamic goals are not static and change at every timestep. We follow the multi-goal setting of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 640 + ], + "score": 1.0, + "content": "Andrychowicz et al. (2017). The goals are part of the environment and do not influence the en-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 347, + 651 + ], + "score": 1.0, + "content": "vironment dynamics. We also assume that a dynamic goal", + "type": "text" + }, + { + "bbox": [ + 348, + 638, + 379, + 649 + ], + "score": 0.91, + "content": "g _ { t } \\in \\mathcal G", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "moves by following some law", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 107, + 649, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 107, + 649, + 156, + 661 + ], + "score": 0.93, + "content": "g _ { t } = g ( t ; \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 649, + 186, + 663 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 187, + 651, + 194, + 660 + ], + "score": 0.81, + "content": "\\gamma", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 649, + 506, + 663 + ], + "score": 1.0, + "content": "parameterizes the law (e.g., acceleration in Newton’s law of motion), and yet", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 660, + 313, + 673 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 313, + 673 + ], + "score": 1.0, + "content": "its underlying moving law is unknown to the agent.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 616, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 493, + 690 + ], + "score": 1.0, + "content": "Moreover, we need to have some basic knowledge of goals, i.e., the measure of goal similarity on", + "type": "text" + }, + { + "bbox": [ + 493, + 677, + 501, + 687 + ], + "score": 0.79, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 676, + 505, + 690 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 171, + 702 + ], + "score": 1.0, + "content": "We assume that", + "type": "text" + }, + { + "bbox": [ + 172, + 688, + 200, + 699 + ], + "score": 0.92, + "content": "g _ { t } \\in \\mathcal G", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 687, + 325, + 702 + ], + "score": 1.0, + "content": "corresponds to some predicate", + "type": "text" + }, + { + "bbox": [ + 326, + 687, + 394, + 700 + ], + "score": 0.93, + "content": "f _ { g _ { t } } : S \\{ 0 , 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "and that the agent’s goal is", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 187, + 712 + ], + "score": 1.0, + "content": "to achieve any state", + "type": "text" + }, + { + "bbox": [ + 187, + 701, + 194, + 709 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 698, + 246, + 712 + ], + "score": 1.0, + "content": "that satisfies", + "type": "text" + }, + { + "bbox": [ + 247, + 699, + 288, + 711 + ], + "score": 0.92, + "content": "f _ { g _ { t } } ( s ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 698, + 325, + 712 + ], + "score": 1.0, + "content": ". We use", + "type": "text" + }, + { + "bbox": [ + 325, + 700, + 353, + 710 + ], + "score": 0.9, + "content": "S = \\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 698, + 399, + 712 + ], + "score": 1.0, + "content": "and define", + "type": "text" + }, + { + "bbox": [ + 399, + 699, + 474, + 711 + ], + "score": 0.91, + "content": "f _ { g _ { t } } ( s ) : = [ s = g _ { t } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 698, + 506, + 712 + ], + "score": 1.0, + "content": ", which", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 339, + 723 + ], + "score": 1.0, + "content": "can be considered as a measure of goal similarity between", + "type": "text" + }, + { + "bbox": [ + 340, + 712, + 349, + 721 + ], + "score": 0.84, + "content": "g _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 709, + 367, + 723 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 367, + 712, + 373, + 720 + ], + "score": 0.71, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 374, + 709, + 505, + 723 + ], + "score": 1.0, + "content": ". The goals can also specify only", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 378, + 733 + ], + "score": 1.0, + "content": "some properties of the state. Take manipulation tasks for instance:", + "type": "text" + }, + { + "bbox": [ + 378, + 720, + 412, + 731 + ], + "score": 0.92, + "content": "\\mathcal { G } = \\mathbb { R } ^ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "corresponds to the 3D", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 676, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 439, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 441, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 441, + 99 + ], + "score": 1.0, + "content": "Algorithm 1 Dynamic Hindsight Experience Replay with Experience Assembling", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 98, + 416, + 109 + ], + "lines": [ + { + "bbox": [ + 106, + 97, + 416, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 325, + 111 + ], + "score": 1.0, + "content": "Require: an off-policy RL algorithm A, replay buffer", + "type": "text" + }, + { + "bbox": [ + 325, + 98, + 334, + 108 + ], + "score": 0.7, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 97, + 410, + 111 + ], + "score": 1.0, + "content": ", a reward function", + "type": "text" + }, + { + "bbox": [ + 410, + 100, + 416, + 108 + ], + "score": 0.46, + "content": "r", + "type": "inline_equation" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 101, + 489, + 422 + ], + "lines": [ + { + "bbox": [ + 111, + 108, + 251, + 119 + ], + "spans": [ + { + "bbox": [ + 111, + 108, + 241, + 119 + ], + "score": 1.0, + "content": "1: Initialize A and replay buffer", + "type": "text" + }, + { + "bbox": [ + 242, + 109, + 251, + 118 + ], + "score": 0.74, + "content": "R", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 119, + 244, + 131 + ], + "spans": [ + { + "bbox": [ + 110, + 119, + 170, + 131 + ], + "score": 1.0, + "content": "2: for episode", + "type": "text" + }, + { + "bbox": [ + 171, + 120, + 230, + 131 + ], + "score": 0.76, + "content": "\\mathbf { \\Omega } = 1 , 2 , \\cdots , M", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 119, + 244, + 131 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 130, + 320, + 143 + ], + "spans": [ + { + "bbox": [ + 110, + 131, + 122, + 141 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 132, + 130, + 223, + 143 + ], + "score": 1.0, + "content": "Sample an initial goal", + "type": "text" + }, + { + "bbox": [ + 223, + 132, + 233, + 142 + ], + "score": 0.84, + "content": "g _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 130, + 310, + 143 + ], + "score": 1.0, + "content": "and an initial state", + "type": "text" + }, + { + "bbox": [ + 310, + 133, + 320, + 141 + ], + "score": 0.81, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 142, + 234, + 153 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 122, + 153 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 132, + 142, + 148, + 153 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 142, + 219, + 153 + ], + "score": 0.89, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 142, + 234, + 153 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 152, + 370, + 164 + ], + "spans": [ + { + "bbox": [ + 110, + 152, + 122, + 164 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 142, + 152, + 214, + 164 + ], + "score": 1.0, + "content": "Sample an action", + "type": "text" + }, + { + "bbox": [ + 214, + 154, + 224, + 163 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 152, + 370, + 164 + ], + "score": 1.0, + "content": "using the behavioral policy from A:", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 163, + 202, + 176 + ], + "spans": [ + { + "bbox": [ + 110, + 164, + 122, + 175 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 143, + 163, + 202, + 176 + ], + "score": 0.6, + "content": "a _ { t } \\pi ( s _ { t } | g _ { t } )", + "type": "inline_equation", + "image_path": "cb0ea2fac3b31734b1dd8861ef711a8a066cf4997fff4534e33dc8452e6d4e5f.jpg" + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 172, + 430, + 188 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 122, + 186 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 142, + 172, + 219, + 188 + ], + "score": 1.0, + "content": "Execute the action", + "type": "text" + }, + { + "bbox": [ + 219, + 176, + 229, + 185 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 172, + 327, + 188 + ], + "score": 1.0, + "content": "and observe a new state", + "type": "text" + }, + { + "bbox": [ + 327, + 177, + 347, + 186 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 172, + 410, + 188 + ], + "score": 1.0, + "content": "and a new goal", + "type": "text" + }, + { + "bbox": [ + 410, + 176, + 430, + 186 + ], + "score": 0.84, + "content": "g _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 185, + 167, + 197 + ], + "spans": [ + { + "bbox": [ + 110, + 186, + 122, + 197 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 133, + 185, + 167, + 196 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 109, + 195, + 235, + 208 + ], + "spans": [ + { + "bbox": [ + 109, + 196, + 122, + 208 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 132, + 195, + 148, + 208 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 196, + 220, + 207 + ], + "score": 0.85, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 195, + 235, + 208 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 207, + 226, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 122, + 219 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 144, + 208, + 226, + 219 + ], + "score": 0.8, + "content": "r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 217, + 467, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 122, + 230 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 142, + 217, + 226, + 231 + ], + "score": 1.0, + "content": "Store the transitions", + "type": "text" + }, + { + "bbox": [ + 226, + 218, + 322, + 230 + ], + "score": 0.83, + "content": "\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 217, + 334, + 231 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 335, + 219, + 344, + 228 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 217, + 467, + 231 + ], + "score": 1.0, + "content": "(Standard experience replay)", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 229, + 167, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 122, + 241 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 132, + 229, + 167, + 240 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 240, + 244, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 122, + 251 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 133, + 240, + 236, + 252 + ], + "score": 1.0, + "content": "Collect failed episodes to", + "type": "text" + }, + { + "bbox": [ + 237, + 241, + 244, + 250 + ], + "score": 0.77, + "content": "\\mathcal { E }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 251, + 338, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 122, + 275 + ], + "score": 1.0, + "content": "14: 15:", + "type": "text" + }, + { + "bbox": [ + 132, + 251, + 148, + 281 + ], + "score": 1.0, + "content": "for S", + "type": "text" + }, + { + "bbox": [ + 148, + 251, + 179, + 262 + ], + "score": 0.91, + "content": "E _ { i } \\in \\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 251, + 205, + 281 + ], + "score": 1.0, + "content": "donother", + "type": "text" + }, + { + "bbox": [ + 205, + 262, + 266, + 275 + ], + "score": 0.91, + "content": "E _ { j } ( i \\neq j ) \\in \\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 251, + 293, + 281 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 294, + 261, + 338, + 276 + ], + "score": 0.88, + "content": "g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 273, + 208, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 122, + 286 + ], + "score": 1.0, + "content": "16:", + "type": "text" + }, + { + "bbox": [ + 142, + 273, + 152, + 286 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 152, + 274, + 186, + 286 + ], + "score": 0.92, + "content": "E _ { j } \\neq \\boldsymbol { \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 273, + 208, + 286 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 483, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 122, + 299 + ], + "score": 1.0, + "content": "17:", + "type": "text" + }, + { + "bbox": [ + 149, + 281, + 247, + 304 + ], + "score": 1.0, + "content": "Clone a goal trajectory", + "type": "text" + }, + { + "bbox": [ + 247, + 286, + 348, + 299 + ], + "score": 0.92, + "content": "\\{ g _ { 0 } ^ { \\prime } , \\cdot \\cdot \\cdot , g _ { m } ^ { \\prime } \\} _ { m = \\operatorname* { m i n } \\{ p , q \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 281, + 387, + 304 + ], + "score": 1.0, + "content": "in which", + "type": "text" + }, + { + "bbox": [ + 387, + 286, + 447, + 300 + ], + "score": 0.93, + "content": "g _ { t } ^ { \\prime } = g _ { j , q - m + t } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 281, + 470, + 304 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 470, + 287, + 483, + 299 + ], + "score": 0.87, + "content": "E _ { j }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 299, + 266, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 122, + 311 + ], + "score": 1.0, + "content": "18:", + "type": "text" + }, + { + "bbox": [ + 153, + 299, + 168, + 311 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 168, + 299, + 251, + 311 + ], + "score": 0.87, + "content": "t = \\{ 0 , \\cdots , m - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 299, + 266, + 311 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 310, + 303, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 122, + 321 + ], + "score": 1.0, + "content": "19:", + "type": "text" + }, + { + "bbox": [ + 163, + 311, + 303, + 322 + ], + "score": 0.86, + "content": "r _ { t } ^ { \\prime } : = r ( s _ { i , p - m + t } , a _ { i , p - m + t } , g _ { t + 1 } ^ { \\prime } )", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 318, + 488, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 122, + 333 + ], + "score": 1.0, + "content": "20:", + "type": "text" + }, + { + "bbox": [ + 161, + 318, + 242, + 335 + ], + "score": 1.0, + "content": "Store the transition", + "type": "text" + }, + { + "bbox": [ + 243, + 320, + 488, + 334 + ], + "score": 0.44, + "content": "\\left( s _ { i , p - m + t } | g _ { t } ^ { \\prime } , a _ { i , p - m + t } , r _ { t } ^ { \\prime } , s _ { i , p - m + t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R \\left( \\mathrm { D H E R } \\right)", + "type": "inline_equation", + "image_path": "8615dd8c5d8516ecf10366e1bf2105dce7c5629ddab4d913ed1c209049fbfc7c.jpg" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 331, + 187, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 122, + 344 + ], + "score": 1.0, + "content": "21:", + "type": "text" + }, + { + "bbox": [ + 153, + 331, + 187, + 343 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 343, + 171, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 121, + 354 + ], + "score": 1.0, + "content": "22:", + "type": "text" + }, + { + "bbox": [ + 141, + 343, + 171, + 353 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 353, + 167, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 121, + 365 + ], + "score": 1.0, + "content": "23:", + "type": "text" + }, + { + "bbox": [ + 132, + 353, + 167, + 364 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 364, + 219, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 121, + 376 + ], + "score": 1.0, + "content": "24:", + "type": "text" + }, + { + "bbox": [ + 132, + 364, + 148, + 376 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 365, + 204, + 376 + ], + "score": 0.89, + "content": "t = 1 , \\cdots , N", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 364, + 219, + 376 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 375, + 335, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 122, + 388 + ], + "score": 1.0, + "content": "25:", + "type": "text" + }, + { + "bbox": [ + 143, + 375, + 225, + 387 + ], + "score": 1.0, + "content": "Sample a minibatch", + "type": "text" + }, + { + "bbox": [ + 225, + 376, + 234, + 385 + ], + "score": 0.77, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 375, + 326, + 387 + ], + "score": 1.0, + "content": "from the replay buffer", + "type": "text" + }, + { + "bbox": [ + 326, + 376, + 335, + 385 + ], + "score": 0.75, + "content": "R", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 387, + 283, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 121, + 398 + ], + "score": 1.0, + "content": "26:", + "type": "text" + }, + { + "bbox": [ + 143, + 387, + 273, + 398 + ], + "score": 1.0, + "content": "Optimize A using the minibatch", + "type": "text" + }, + { + "bbox": [ + 274, + 387, + 283, + 397 + ], + "score": 0.75, + "content": "B", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 397, + 167, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 122, + 409 + ], + "score": 1.0, + "content": "27:", + "type": "text" + }, + { + "bbox": [ + 132, + 397, + 167, + 408 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 408, + 157, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 157, + 420 + ], + "score": 1.0, + "content": "28: end for", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 198, + 459 + ], + "score": 1.0, + "content": "positions of an object", + "type": "text" + }, + { + "bbox": [ + 198, + 446, + 213, + 457 + ], + "score": 0.88, + "content": "s ^ { \\mathrm { o b j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "about which the observation could include additional properties of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 457, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 504, + 469 + ], + "score": 1.0, + "content": "object. For a more concrete example, consider pushing a block towards a moving target position.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 102, + 462, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 102, + 462, + 244, + 487 + ], + "score": 1.0, + "content": "The success of a task is defined as", + "type": "text" + }, + { + "bbox": [ + 244, + 469, + 400, + 482 + ], + "score": 0.91, + "content": "f \\big ( s _ { t } , g _ { t } \\big ) = \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( \\| s _ { t } ^ { \\mathrm { o b j } } - g _ { t } \\| \\leq \\epsilon \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 462, + 430, + 487 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 430, + 468, + 446, + 482 + ], + "score": 0.9, + "content": "s _ { t } ^ { \\mathrm { o b j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 462, + 505, + 487 + ], + "score": 1.0, + "content": "is the position", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 480, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 205, + 495 + ], + "score": 1.0, + "content": "of the object in the state", + "type": "text" + }, + { + "bbox": [ + 206, + 483, + 216, + 492 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 480, + 234, + 495 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 234, + 483, + 240, + 491 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 480, + 404, + 495 + ], + "score": 1.0, + "content": "denotes a tolerance by the environment.", + "type": "text" + }, + { + "bbox": [ + 405, + 482, + 444, + 492 + ], + "score": 0.85, + "content": "\\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 480, + 507, + 495 + ], + "score": 1.0, + "content": "is an indicator", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 297, + 505 + ], + "score": 1.0, + "content": "function. The agent aims to achieve any state", + "type": "text" + }, + { + "bbox": [ + 297, + 494, + 306, + 503 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 491, + 361, + 505 + ], + "score": 1.0, + "content": "that satisfies", + "type": "text" + }, + { + "bbox": [ + 362, + 492, + 418, + 504 + ], + "score": 0.93, + "content": "f ( s _ { t } , g _ { t } ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 491, + 506, + 505 + ], + "score": 1.0, + "content": ". It receives a sparse", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 467, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 136, + 516 + ], + "score": 1.0, + "content": "reward", + "type": "text" + }, + { + "bbox": [ + 136, + 502, + 357, + 515 + ], + "score": 0.41, + "content": "r _ { t } : = r ( \\tilde { s } _ { t } , a _ { t } , g _ { t + 1 } ) = - \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } ( f ( s _ { t + 1 } , g _ { t + 1 } ) \\neq 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 503, + 453, + 516 + ], + "score": 1.0, + "content": "upon making an action", + "type": "text" + }, + { + "bbox": [ + 453, + 505, + 463, + 514 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 503, + 467, + 516 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "It is worth discussing the main difference between the implications of the dynamic goals and the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 103, + 529, + 507, + 545 + ], + "spans": [ + { + "bbox": [ + 103, + 529, + 336, + 545 + ], + "score": 1.0, + "content": "static ones. Expressing a static goal in the following way,", + "type": "text" + }, + { + "bbox": [ + 337, + 531, + 405, + 543 + ], + "score": 0.9, + "content": "g _ { t } ^ { \\mathrm { s t a t i c } } = g ^ { \\mathrm { s t a t i c } } , \\forall t", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 529, + 507, + 545 + ], + "score": 1.0, + "content": ", highlights the key chal-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "lenge of dealing with the dynamic goal. Namely, the agent has no access to the underlying law of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "dynamic goal in our setting, whereas the law of being static is known to the agent in Andrychowicz", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "et al. (2017). In other words, the agent has no clue at all how to construct a new dynamic goal that", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 576, + 243, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 243, + 587 + ], + "score": 1.0, + "content": "is admissible by the environment.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 107, + 603, + 320, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 321, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 321, + 616 + ], + "score": 1.0, + "content": "3.2 DYNAMIC HINDSIGHT EXPERIENCE REPLAY", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 100, + 616, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 100, + 616, + 345, + 647 + ], + "score": 1.0, + "content": "At the first glance, we shall compose a new dynamic goal", + "type": "text" + }, + { + "bbox": [ + 345, + 623, + 474, + 639 + ], + "score": 0.92, + "content": "g ^ { \\mathrm { d y n a m i c } } = \\{ s _ { t _ { 0 } } ^ { \\mathrm { o b j } } , s _ { t _ { 1 } } ^ { \\mathrm { o b j } } , \\cdot \\cdot \\cdot , s _ { t _ { T ^ { \\prime } } } ^ { \\mathrm { o b j } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "from a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 226, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 226, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "in order to apply HER (Andrychowicz et al., 2017) to our problem", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 646, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 646, + 375, + 663 + ], + "score": 1.0, + "content": "setting. However, per the discussion above, this new dynamic goal", + "type": "text" + }, + { + "bbox": [ + 376, + 649, + 406, + 660 + ], + "score": 0.8, + "content": "g ^ { \\mathrm { d y n a m i c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 646, + 506, + 663 + ], + "score": 1.0, + "content": "may be inadmissible by", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 376, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 376, + 673 + ], + "score": 1.0, + "content": "the environment, leading to no positive feedback to the agent at all.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "We tackle the challenge by drawing the following two observations. One is that many episodes", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "fail in the replay buffer, implying that the agent can actually build a new goal upon more than one", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "episodes. The other is that the more failed experience the agent has, the more possible for the agent", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "to use the connection between achieved goals in an episode and desired goals in some other episode.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "Take the example of a moving object that the agent must reach, desired goals are the positions of", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 81, + 439, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 441, + 99 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 441, + 99 + ], + "score": 1.0, + "content": "Algorithm 1 Dynamic Hindsight Experience Replay with Experience Assembling", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 98, + 416, + 109 + ], + "lines": [ + { + "bbox": [ + 106, + 97, + 416, + 111 + ], + "spans": [ + { + "bbox": [ + 106, + 97, + 325, + 111 + ], + "score": 1.0, + "content": "Require: an off-policy RL algorithm A, replay buffer", + "type": "text" + }, + { + "bbox": [ + 325, + 98, + 334, + 108 + ], + "score": 0.7, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 97, + 410, + 111 + ], + "score": 1.0, + "content": ", a reward function", + "type": "text" + }, + { + "bbox": [ + 410, + 100, + 416, + 108 + ], + "score": 0.46, + "content": "r", + "type": "inline_equation" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 97, + 416, + 111 + ] + }, + { + "type": "index", + "bbox": [ + 106, + 101, + 489, + 422 + ], + "lines": [ + { + "bbox": [ + 111, + 108, + 251, + 119 + ], + "spans": [ + { + "bbox": [ + 111, + 108, + 241, + 119 + ], + "score": 1.0, + "content": "1: Initialize A and replay buffer", + "type": "text" + }, + { + "bbox": [ + 242, + 109, + 251, + 118 + ], + "score": 0.74, + "content": "R", + "type": "inline_equation" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 119, + 244, + 131 + ], + "spans": [ + { + "bbox": [ + 110, + 119, + 170, + 131 + ], + "score": 1.0, + "content": "2: for episode", + "type": "text" + }, + { + "bbox": [ + 171, + 120, + 230, + 131 + ], + "score": 0.76, + "content": "\\mathbf { \\Omega } = 1 , 2 , \\cdots , M", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 119, + 244, + 131 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 130, + 320, + 143 + ], + "spans": [ + { + "bbox": [ + 110, + 131, + 122, + 141 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 132, + 130, + 223, + 143 + ], + "score": 1.0, + "content": "Sample an initial goal", + "type": "text" + }, + { + "bbox": [ + 223, + 132, + 233, + 142 + ], + "score": 0.84, + "content": "g _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 130, + 310, + 143 + ], + "score": 1.0, + "content": "and an initial state", + "type": "text" + }, + { + "bbox": [ + 310, + 133, + 320, + 141 + ], + "score": 0.81, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 142, + 234, + 153 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 122, + 153 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 132, + 142, + 148, + 153 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 142, + 219, + 153 + ], + "score": 0.89, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 142, + 234, + 153 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 152, + 370, + 164 + ], + "spans": [ + { + "bbox": [ + 110, + 152, + 122, + 164 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 142, + 152, + 214, + 164 + ], + "score": 1.0, + "content": "Sample an action", + "type": "text" + }, + { + "bbox": [ + 214, + 154, + 224, + 163 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 152, + 370, + 164 + ], + "score": 1.0, + "content": "using the behavioral policy from A:", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 163, + 202, + 176 + ], + "spans": [ + { + "bbox": [ + 110, + 164, + 122, + 175 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 143, + 163, + 202, + 176 + ], + "score": 0.6, + "content": "a _ { t } \\pi ( s _ { t } | g _ { t } )", + "type": "inline_equation", + "image_path": "cb0ea2fac3b31734b1dd8861ef711a8a066cf4997fff4534e33dc8452e6d4e5f.jpg" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 172, + 430, + 188 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 122, + 186 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 142, + 172, + 219, + 188 + ], + "score": 1.0, + "content": "Execute the action", + "type": "text" + }, + { + "bbox": [ + 219, + 176, + 229, + 185 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 172, + 327, + 188 + ], + "score": 1.0, + "content": "and observe a new state", + "type": "text" + }, + { + "bbox": [ + 327, + 177, + 347, + 186 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 172, + 410, + 188 + ], + "score": 1.0, + "content": "and a new goal", + "type": "text" + }, + { + "bbox": [ + 410, + 176, + 430, + 186 + ], + "score": 0.84, + "content": "g _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 185, + 167, + 197 + ], + "spans": [ + { + "bbox": [ + 110, + 186, + 122, + 197 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 133, + 185, + 167, + 196 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 109, + 195, + 235, + 208 + ], + "spans": [ + { + "bbox": [ + 109, + 196, + 122, + 208 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 132, + 195, + 148, + 208 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 196, + 220, + 207 + ], + "score": 0.85, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 195, + 235, + 208 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 207, + 226, + 219 + ], + "spans": [ + { + "bbox": [ + 106, + 207, + 122, + 219 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 144, + 208, + 226, + 219 + ], + "score": 0.8, + "content": "r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )", + "type": "inline_equation" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 217, + 467, + 231 + ], + "spans": [ + { + "bbox": [ + 106, + 218, + 122, + 230 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 142, + 217, + 226, + 231 + ], + "score": 1.0, + "content": "Store the transitions", + "type": "text" + }, + { + "bbox": [ + 226, + 218, + 322, + 230 + ], + "score": 0.83, + "content": "\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 217, + 334, + 231 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 335, + 219, + 344, + 228 + ], + "score": 0.78, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 217, + 467, + 231 + ], + "score": 1.0, + "content": "(Standard experience replay)", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 229, + 167, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 229, + 122, + 241 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 132, + 229, + 167, + 240 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 240, + 244, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 122, + 251 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 133, + 240, + 236, + 252 + ], + "score": 1.0, + "content": "Collect failed episodes to", + "type": "text" + }, + { + "bbox": [ + 237, + 241, + 244, + 250 + ], + "score": 0.77, + "content": "\\mathcal { E }", + "type": "inline_equation" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 251, + 338, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 251, + 122, + 275 + ], + "score": 1.0, + "content": "14: 15:", + "type": "text" + }, + { + "bbox": [ + 132, + 251, + 148, + 281 + ], + "score": 1.0, + "content": "for S", + "type": "text" + }, + { + "bbox": [ + 148, + 251, + 179, + 262 + ], + "score": 0.91, + "content": "E _ { i } \\in \\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 251, + 205, + 281 + ], + "score": 1.0, + "content": "donother", + "type": "text" + }, + { + "bbox": [ + 205, + 262, + 266, + 275 + ], + "score": 0.91, + "content": "E _ { j } ( i \\neq j ) \\in \\mathcal { E }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 251, + 293, + 281 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 294, + 261, + 338, + 276 + ], + "score": 0.88, + "content": "g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }", + "type": "inline_equation" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 273, + 208, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 122, + 286 + ], + "score": 1.0, + "content": "16:", + "type": "text" + }, + { + "bbox": [ + 142, + 273, + 152, + 286 + ], + "score": 1.0, + "content": "if", + "type": "text" + }, + { + "bbox": [ + 152, + 274, + 186, + 286 + ], + "score": 0.92, + "content": "E _ { j } \\neq \\boldsymbol { \\mathcal { O } }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 273, + 208, + 286 + ], + "score": 1.0, + "content": "then", + "type": "text" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 281, + 483, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 122, + 299 + ], + "score": 1.0, + "content": "17:", + "type": "text" + }, + { + "bbox": [ + 149, + 281, + 247, + 304 + ], + "score": 1.0, + "content": "Clone a goal trajectory", + "type": "text" + }, + { + "bbox": [ + 247, + 286, + 348, + 299 + ], + "score": 0.92, + "content": "\\{ g _ { 0 } ^ { \\prime } , \\cdot \\cdot \\cdot , g _ { m } ^ { \\prime } \\} _ { m = \\operatorname* { m i n } \\{ p , q \\} }", + "type": "inline_equation" + }, + { + "bbox": [ + 349, + 281, + 387, + 304 + ], + "score": 1.0, + "content": "in which", + "type": "text" + }, + { + "bbox": [ + 387, + 286, + 447, + 300 + ], + "score": 0.93, + "content": "g _ { t } ^ { \\prime } = g _ { j , q - m + t } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 281, + 470, + 304 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 470, + 287, + 483, + 299 + ], + "score": 0.87, + "content": "E _ { j }", + "type": "inline_equation" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 299, + 266, + 311 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 122, + 311 + ], + "score": 1.0, + "content": "18:", + "type": "text" + }, + { + "bbox": [ + 153, + 299, + 168, + 311 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 168, + 299, + 251, + 311 + ], + "score": 0.87, + "content": "t = \\{ 0 , \\cdots , m - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 299, + 266, + 311 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 310, + 303, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 122, + 321 + ], + "score": 1.0, + "content": "19:", + "type": "text" + }, + { + "bbox": [ + 163, + 311, + 303, + 322 + ], + "score": 0.86, + "content": "r _ { t } ^ { \\prime } : = r ( s _ { i , p - m + t } , a _ { i , p - m + t } , g _ { t + 1 } ^ { \\prime } )", + "type": "inline_equation" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 318, + 488, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 122, + 333 + ], + "score": 1.0, + "content": "20:", + "type": "text" + }, + { + "bbox": [ + 161, + 318, + 242, + 335 + ], + "score": 1.0, + "content": "Store the transition", + "type": "text" + }, + { + "bbox": [ + 243, + 320, + 488, + 334 + ], + "score": 0.44, + "content": "\\left( s _ { i , p - m + t } | g _ { t } ^ { \\prime } , a _ { i , p - m + t } , r _ { t } ^ { \\prime } , s _ { i , p - m + t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R \\left( \\mathrm { D H E R } \\right)", + "type": "inline_equation", + "image_path": "8615dd8c5d8516ecf10366e1bf2105dce7c5629ddab4d913ed1c209049fbfc7c.jpg" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 331, + 187, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 122, + 344 + ], + "score": 1.0, + "content": "21:", + "type": "text" + }, + { + "bbox": [ + 153, + 331, + 187, + 343 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 343, + 171, + 354 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 121, + 354 + ], + "score": 1.0, + "content": "22:", + "type": "text" + }, + { + "bbox": [ + 141, + 343, + 171, + 353 + ], + "score": 1.0, + "content": "end if", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 353, + 167, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 121, + 365 + ], + "score": 1.0, + "content": "23:", + "type": "text" + }, + { + "bbox": [ + 132, + 353, + 167, + 364 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 364, + 219, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 121, + 376 + ], + "score": 1.0, + "content": "24:", + "type": "text" + }, + { + "bbox": [ + 132, + 364, + 148, + 376 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 365, + 204, + 376 + ], + "score": 0.89, + "content": "t = 1 , \\cdots , N", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 364, + 219, + 376 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 375, + 335, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 122, + 388 + ], + "score": 1.0, + "content": "25:", + "type": "text" + }, + { + "bbox": [ + 143, + 375, + 225, + 387 + ], + "score": 1.0, + "content": "Sample a minibatch", + "type": "text" + }, + { + "bbox": [ + 225, + 376, + 234, + 385 + ], + "score": 0.77, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 375, + 326, + 387 + ], + "score": 1.0, + "content": "from the replay buffer", + "type": "text" + }, + { + "bbox": [ + 326, + 376, + 335, + 385 + ], + "score": 0.75, + "content": "R", + "type": "inline_equation" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 387, + 283, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 121, + 398 + ], + "score": 1.0, + "content": "26:", + "type": "text" + }, + { + "bbox": [ + 143, + 387, + 273, + 398 + ], + "score": 1.0, + "content": "Optimize A using the minibatch", + "type": "text" + }, + { + "bbox": [ + 274, + 387, + 283, + 397 + ], + "score": 0.75, + "content": "B", + "type": "inline_equation" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 397, + 167, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 122, + 409 + ], + "score": 1.0, + "content": "27:", + "type": "text" + }, + { + "bbox": [ + 132, + 397, + 167, + 408 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 408, + 157, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 157, + 420 + ], + "score": 1.0, + "content": "28: end for", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + } + ], + "index": 15, + "bbox_fs": [ + 105, + 108, + 488, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 446, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 198, + 459 + ], + "score": 1.0, + "content": "positions of an object", + "type": "text" + }, + { + "bbox": [ + 198, + 446, + 213, + 457 + ], + "score": 0.88, + "content": "s ^ { \\mathrm { o b j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 444, + 505, + 459 + ], + "score": 1.0, + "content": "about which the observation could include additional properties of the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 457, + 504, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 504, + 469 + ], + "score": 1.0, + "content": "object. For a more concrete example, consider pushing a block towards a moving target position.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 102, + 462, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 102, + 462, + 244, + 487 + ], + "score": 1.0, + "content": "The success of a task is defined as", + "type": "text" + }, + { + "bbox": [ + 244, + 469, + 400, + 482 + ], + "score": 0.91, + "content": "f \\big ( s _ { t } , g _ { t } \\big ) = \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( \\| s _ { t } ^ { \\mathrm { o b j } } - g _ { t } \\| \\leq \\epsilon \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 462, + 430, + 487 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 430, + 468, + 446, + 482 + ], + "score": 0.9, + "content": "s _ { t } ^ { \\mathrm { o b j } }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 462, + 505, + 487 + ], + "score": 1.0, + "content": "is the position", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 480, + 507, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 480, + 205, + 495 + ], + "score": 1.0, + "content": "of the object in the state", + "type": "text" + }, + { + "bbox": [ + 206, + 483, + 216, + 492 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 480, + 234, + 495 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 234, + 483, + 240, + 491 + ], + "score": 0.72, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 480, + 404, + 495 + ], + "score": 1.0, + "content": "denotes a tolerance by the environment.", + "type": "text" + }, + { + "bbox": [ + 405, + 482, + 444, + 492 + ], + "score": 0.85, + "content": "\\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 480, + 507, + 495 + ], + "score": 1.0, + "content": "is an indicator", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 297, + 505 + ], + "score": 1.0, + "content": "function. The agent aims to achieve any state", + "type": "text" + }, + { + "bbox": [ + 297, + 494, + 306, + 503 + ], + "score": 0.83, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 491, + 361, + 505 + ], + "score": 1.0, + "content": "that satisfies", + "type": "text" + }, + { + "bbox": [ + 362, + 492, + 418, + 504 + ], + "score": 0.93, + "content": "f ( s _ { t } , g _ { t } ) = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 491, + 506, + 505 + ], + "score": 1.0, + "content": ". It receives a sparse", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 467, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 136, + 516 + ], + "score": 1.0, + "content": "reward", + "type": "text" + }, + { + "bbox": [ + 136, + 502, + 357, + 515 + ], + "score": 0.41, + "content": "r _ { t } : = r ( \\tilde { s } _ { t } , a _ { t } , g _ { t + 1 } ) = - \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } ( f ( s _ { t + 1 } , g _ { t + 1 } ) \\neq 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 503, + 453, + 516 + ], + "score": 1.0, + "content": "upon making an action", + "type": "text" + }, + { + "bbox": [ + 453, + 505, + 463, + 514 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 503, + 467, + 516 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 102, + 444, + 507, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 519, + 505, + 586 + ], + "lines": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "It is worth discussing the main difference between the implications of the dynamic goals and the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 103, + 529, + 507, + 545 + ], + "spans": [ + { + "bbox": [ + 103, + 529, + 336, + 545 + ], + "score": 1.0, + "content": "static ones. Expressing a static goal in the following way,", + "type": "text" + }, + { + "bbox": [ + 337, + 531, + 405, + 543 + ], + "score": 0.9, + "content": "g _ { t } ^ { \\mathrm { s t a t i c } } = g ^ { \\mathrm { s t a t i c } } , \\forall t", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 529, + 507, + 545 + ], + "score": 1.0, + "content": ", highlights the key chal-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 555 + ], + "score": 1.0, + "content": "lenge of dealing with the dynamic goal. Namely, the agent has no access to the underlying law of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 566 + ], + "score": 1.0, + "content": "dynamic goal in our setting, whereas the law of being static is known to the agent in Andrychowicz", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 506, + 577 + ], + "score": 1.0, + "content": "et al. (2017). In other words, the agent has no clue at all how to construct a new dynamic goal that", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 576, + 243, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 243, + 587 + ], + "score": 1.0, + "content": "is admissible by the environment.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 103, + 519, + 507, + 587 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 603, + 320, + 615 + ], + "lines": [ + { + "bbox": [ + 106, + 603, + 321, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 603, + 321, + 616 + ], + "score": 1.0, + "content": "3.2 DYNAMIC HINDSIGHT EXPERIENCE REPLAY", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 624, + 505, + 671 + ], + "lines": [ + { + "bbox": [ + 100, + 616, + 506, + 647 + ], + "spans": [ + { + "bbox": [ + 100, + 616, + 345, + 647 + ], + "score": 1.0, + "content": "At the first glance, we shall compose a new dynamic goal", + "type": "text" + }, + { + "bbox": [ + 345, + 623, + 474, + 639 + ], + "score": 0.92, + "content": "g ^ { \\mathrm { d y n a m i c } } = \\{ s _ { t _ { 0 } } ^ { \\mathrm { o b j } } , s _ { t _ { 1 } } ^ { \\mathrm { o b j } } , \\cdot \\cdot \\cdot , s _ { t _ { T ^ { \\prime } } } ^ { \\mathrm { o b j } } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 474, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "from a", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 226, + 637, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 226, + 637, + 506, + 650 + ], + "score": 1.0, + "content": "in order to apply HER (Andrychowicz et al., 2017) to our problem", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 646, + 506, + 663 + ], + "spans": [ + { + "bbox": [ + 104, + 646, + 375, + 663 + ], + "score": 1.0, + "content": "setting. However, per the discussion above, this new dynamic goal", + "type": "text" + }, + { + "bbox": [ + 376, + 649, + 406, + 660 + ], + "score": 0.8, + "content": "g ^ { \\mathrm { d y n a m i c } }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 646, + 506, + 663 + ], + "score": 1.0, + "content": "may be inadmissible by", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 376, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 376, + 673 + ], + "score": 1.0, + "content": "the environment, leading to no positive feedback to the agent at all.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 100, + 616, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "We tackle the challenge by drawing the following two observations. One is that many episodes", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "fail in the replay buffer, implying that the agent can actually build a new goal upon more than one", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "episodes. The other is that the more failed experience the agent has, the more possible for the agent", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "to use the connection between achieved goals in an episode and desired goals in some other episode.", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "Take the example of a moving object that the agent must reach, desired goals are the positions of", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "score": 1.0, + "content": "the moving object and achieved goals are the positions of a gripper (controlled by the agent). There", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 335, + 477, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 477, + 349 + ], + "score": 1.0, + "content": "may exist some positions that both the object and the gripper have ever reached respectively.", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 48, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 85, + 477, + 262 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 130, + 85, + 477, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 85, + 477, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 85, + 477, + 262 + ], + "score": 0.973, + "type": "image", + "image_path": "685259757917f2907f387b49a9224aabd57ab150ad7074e7a37b24515868d9cf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 130, + 85, + 477, + 144.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 130, + 144.0, + 477, + 203.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 130, + 203.0, + 477, + 262.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 265, + 504, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "Figure 2: The proposed tasks with dynamic goals (red objects). Arrow indicates the movement of a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 275, + 426, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 426, + 289 + ], + "score": 1.0, + "content": "goal. The first row indicates initial states. The second row indicates final states.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 503, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 338 + ], + "score": 1.0, + "content": "the moving object and achieved goals are the positions of a gripper (controlled by the agent). There", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 335, + 477, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 477, + 349 + ], + "score": 1.0, + "content": "may exist some positions that both the object and the gripper have ever reached respectively.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 353, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 248, + 366 + ], + "score": 1.0, + "content": "After experiencing some episode", + "type": "text" + }, + { + "bbox": [ + 248, + 356, + 306, + 364 + ], + "score": 0.87, + "content": "s _ { 0 } , s _ { 1 } , \\cdots , s _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 352, + 505, + 366 + ], + "score": 1.0, + "content": ", we store in the replay buffer every transition", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 107, + 365, + 153, + 375 + ], + "score": 0.89, + "content": "s _ { t } \\ \\to \\ s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 363, + 269, + 377 + ], + "score": 1.0, + "content": "for this episode, defined as", + "type": "text" + }, + { + "bbox": [ + 270, + 364, + 334, + 376 + ], + "score": 0.91, + "content": "\\left( { { s _ { t } } , { a _ { t } } , { r _ { t } } , { s _ { t + 1 } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 363, + 366, + 377 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 367, + 365, + 376, + 375 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 363, + 446, + 377 + ], + "score": 1.0, + "content": "indicates a state", + "type": "text" + }, + { + "bbox": [ + 447, + 365, + 456, + 375 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "at timestep", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 375, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 107, + 376, + 111, + 385 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 375, + 133, + 386 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 133, + 377, + 143, + 386 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 375, + 239, + 386 + ], + "score": 1.0, + "content": "indicates an action and", + "type": "text" + }, + { + "bbox": [ + 240, + 376, + 249, + 385 + ], + "score": 0.82, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 375, + 381, + 386 + ], + "score": 1.0, + "content": "indicates a reward. Thus before", + "type": "text" + }, + { + "bbox": [ + 382, + 376, + 387, + 385 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 375, + 505, + 386 + ], + "score": 1.0, + "content": ", there are a series of records", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 261, + 398 + ], + "score": 0.92, + "content": "\\left\\{ \\left( s _ { 0 } , a _ { 0 } , r _ { 0 } , s _ { 1 } \\right) \\cdot \\cdot \\cdot , \\left( s _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } \\right) \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 385, + 440, + 398 + ], + "score": 1.0, + "content": ". A state consists of three parts: observation", + "type": "text" + }, + { + "bbox": [ + 440, + 388, + 450, + 397 + ], + "score": 0.84, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 385, + 505, + 398 + ], + "score": 1.0, + "content": ", desired goal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 395, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 121, + 410 + ], + "score": 0.89, + "content": "g _ { t } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 395, + 196, + 412 + ], + "score": 1.0, + "content": "and achieved goal", + "type": "text" + }, + { + "bbox": [ + 196, + 398, + 210, + 410 + ], + "score": 0.89, + "content": "g _ { t } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 395, + 252, + 412 + ], + "score": 1.0, + "content": ", define as", + "type": "text" + }, + { + "bbox": [ + 252, + 397, + 328, + 411 + ], + "score": 0.93, + "content": "\\dot { s } _ { t } = \\langle o _ { t } , g _ { t } ^ { a c } , g _ { t } ^ { d e } \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 395, + 397, + 412 + ], + "score": 1.0, + "content": ", where normally", + "type": "text" + }, + { + "bbox": [ + 397, + 397, + 433, + 410 + ], + "score": 0.93, + "content": "g _ { t } ^ { d e } = g _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 395, + 451, + 412 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 451, + 398, + 466, + 410 + ], + "score": 0.9, + "content": "g _ { t } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 395, + 507, + 412 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 408, + 242, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 242, + 420 + ], + "score": 1.0, + "content": "goals that the agent has achieved.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 107, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 107, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "We reuse the failed experience from the replay buffer with inverse simulation to create success-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "ful rewards for the agent, as shown in Algorithm 1 (lines 13-23). Our inverse simulation contains", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 448, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 504, + 460 + ], + "score": 1.0, + "content": "two main steps: First, given a failed episode, for its achieved goal trajectory, we try to find a de-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "sired goal trajectory from other episodes that could match it (line 15). Second, we assemble a new", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "episode by matching the achieved goal trajectory of the given episode to the desired goal trajec-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 479, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 104, + 479, + 281, + 496 + ], + "score": 1.0, + "content": "tory of the founded one (lines 17-21). Let", + "type": "text" + }, + { + "bbox": [ + 281, + 481, + 297, + 494 + ], + "score": 0.91, + "content": "g _ { i , q } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 479, + 506, + 496 + ], + "score": 1.0, + "content": "indicate the achieved goal of the agent at timestep", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 488, + 510, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 113, + 505 + ], + "score": 0.74, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 488, + 153, + 519 + ], + "score": 1.0, + "content": "in episodxperience", + "type": "text" + }, + { + "bbox": [ + 159, + 494, + 165, + 504 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 492, + 200, + 507 + ], + "score": 0.91, + "content": "g _ { j , p } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 488, + 352, + 519 + ], + "score": 1.0, + "content": "te the desired goal at timestep , we search and draw two faile", + "type": "text" + }, + { + "bbox": [ + 353, + 496, + 360, + 505 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 488, + 395, + 519 + ], + "score": 1.0, + "content": "in episoepisodes", + "type": "text" + }, + { + "bbox": [ + 406, + 494, + 412, + 505 + ], + "score": 0.76, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 488, + 426, + 519 + ], + "score": 1.0, + "content": ". Gand", + "type": "text" + }, + { + "bbox": [ + 440, + 488, + 443, + 519 + ], + "score": 1.0, + "content": "n", + "type": "text" + }, + { + "bbox": [ + 472, + 488, + 510, + 519 + ], + "score": 1.0, + "content": "of failed), where", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 153, + 506, + 472, + 519 + ], + "spans": [ + { + "bbox": [ + 153, + 506, + 226, + 518 + ], + "score": 0.9, + "content": "\\{ E _ { 0 } , E _ { 1 } , \\stackrel { \\cdot } { E } _ { 2 } , \\cdot \\cdot \\cdot \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 506, + 408, + 517 + ], + "score": 0.88, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 506, + 440, + 519 + ], + "score": 0.81, + "content": "E _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 506, + 472, + 518 + ], + "score": 0.68, + "content": "( i \\neq j )", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 514, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 145, + 531 + ], + "score": 0.78, + "content": "\\exists i , j , p , q ,", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 514, + 162, + 536 + ], + "score": 1.0, + "content": "s.t.", + "type": "text" + }, + { + "bbox": [ + 162, + 518, + 208, + 532 + ], + "score": 0.92, + "content": "g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 514, + 507, + 536 + ], + "score": 1.0, + "content": ". If we find two such failed episodes, we combine the two experience by", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 102, + 527, + 508, + 550 + ], + "spans": [ + { + "bbox": [ + 102, + 527, + 230, + 550 + ], + "score": 1.0, + "content": "replacing the desired goals in", + "type": "text" + }, + { + "bbox": [ + 230, + 532, + 242, + 543 + ], + "score": 0.86, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 527, + 258, + 550 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 258, + 531, + 282, + 545 + ], + "score": 0.93, + "content": "\\{ g _ { j , t } ^ { d e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 527, + 315, + 550 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 315, + 532, + 322, + 543 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 527, + 362, + 550 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 362, + 532, + 375, + 544 + ], + "score": 0.89, + "content": "E _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 527, + 395, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 531, + 457, + 544 + ], + "score": 0.93, + "content": "t \\leq \\operatorname* { m i n } \\{ p , q \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 527, + 508, + 550 + ], + "score": 1.0, + "content": ". Based on", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 287, + 558 + ], + "score": 1.0, + "content": "this, we end up assembling a new experience", + "type": "text" + }, + { + "bbox": [ + 287, + 545, + 299, + 556 + ], + "score": 0.9, + "content": "E _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 543, + 338, + 558 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 338, + 545, + 349, + 555 + ], + "score": 0.88, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "with a new “imagined” goal trajectory", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 551, + 264, + 574 + ], + "spans": [ + { + "bbox": [ + 107, + 555, + 170, + 570 + ], + "score": 0.93, + "content": "\\{ g _ { j , 0 } ^ { d e } , \\cdot \\cdot \\cdot , g _ { j , t } ^ { d e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 551, + 198, + 574 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 198, + 556, + 257, + 568 + ], + "score": 0.93, + "content": "t \\leq \\operatorname* { m i n } \\{ p , q \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 551, + 264, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 504, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 221, + 585 + ], + "score": 1.0, + "content": "More details of the complete", + "type": "text" + }, + { + "bbox": [ + 221, + 573, + 268, + 584 + ], + "score": 0.67, + "content": "\\mathrm { R L + D H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "method are shown in Algorithm 1. Unlike HER, our DHER", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "needs to search all failed experiences to compose a “imagined” goal trajectory. Hence, the efficiency", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "of searching the memory is important. In our implementation, we use two hash tables to store the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 606, + 352, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 352, + 619 + ], + "score": 1.0, + "content": "trajectories of achieved goals and desired goals, respectively.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5 + }, + { + "type": "title", + "bbox": [ + 107, + 651, + 194, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 196, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 196, + 666 + ], + "score": 1.0, + "content": "4 EXPERIMENT", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "We run extensive experiments to examine the proposed DHER for moving goals and compare it", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "with some competing baselines. We first introduce the environments and tasks that we want to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "address, followed by the experimental results. Demo videos from our experiments are available at", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 721, + 315, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 315, + 733 + ], + "score": 1.0, + "content": "https://sites.google.com/view/dher.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 85, + 477, + 262 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 130, + 85, + 477, + 262 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 130, + 85, + 477, + 262 + ], + "spans": [ + { + "bbox": [ + 130, + 85, + 477, + 262 + ], + "score": 0.973, + "type": "image", + "image_path": "685259757917f2907f387b49a9224aabd57ab150ad7074e7a37b24515868d9cf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 130, + 85, + 477, + 144.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 130, + 144.0, + 477, + 203.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 130, + 203.0, + 477, + 262.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 265, + 504, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 264, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 506, + 277 + ], + "score": 1.0, + "content": "Figure 2: The proposed tasks with dynamic goals (red objects). Arrow indicates the movement of a", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 275, + 426, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 426, + 289 + ], + "score": 1.0, + "content": "goal. The first row indicates initial states. The second row indicates final states.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 324, + 503, + 347 + ], + "lines": [], + "index": 5.5, + "bbox_fs": [ + 105, + 323, + 505, + 349 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 353, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 248, + 366 + ], + "score": 1.0, + "content": "After experiencing some episode", + "type": "text" + }, + { + "bbox": [ + 248, + 356, + 306, + 364 + ], + "score": 0.87, + "content": "s _ { 0 } , s _ { 1 } , \\cdots , s _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 352, + 505, + 366 + ], + "score": 1.0, + "content": ", we store in the replay buffer every transition", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 363, + 505, + 377 + ], + "spans": [ + { + "bbox": [ + 107, + 365, + 153, + 375 + ], + "score": 0.89, + "content": "s _ { t } \\ \\to \\ s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 153, + 363, + 269, + 377 + ], + "score": 1.0, + "content": "for this episode, defined as", + "type": "text" + }, + { + "bbox": [ + 270, + 364, + 334, + 376 + ], + "score": 0.91, + "content": "\\left( { { s _ { t } } , { a _ { t } } , { r _ { t } } , { s _ { t + 1 } } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 363, + 366, + 377 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 367, + 365, + 376, + 375 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 363, + 446, + 377 + ], + "score": 1.0, + "content": "indicates a state", + "type": "text" + }, + { + "bbox": [ + 447, + 365, + 456, + 375 + ], + "score": 0.84, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 457, + 363, + 505, + 377 + ], + "score": 1.0, + "content": "at timestep", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 375, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 107, + 376, + 111, + 385 + ], + "score": 0.73, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 375, + 133, + 386 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 133, + 377, + 143, + 386 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 375, + 239, + 386 + ], + "score": 1.0, + "content": "indicates an action and", + "type": "text" + }, + { + "bbox": [ + 240, + 376, + 249, + 385 + ], + "score": 0.82, + "content": "r _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 375, + 381, + 386 + ], + "score": 1.0, + "content": "indicates a reward. Thus before", + "type": "text" + }, + { + "bbox": [ + 382, + 376, + 387, + 385 + ], + "score": 0.75, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 375, + 505, + 386 + ], + "score": 1.0, + "content": ", there are a series of records", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 261, + 398 + ], + "score": 0.92, + "content": "\\left\\{ \\left( s _ { 0 } , a _ { 0 } , r _ { 0 } , s _ { 1 } \\right) \\cdot \\cdot \\cdot , \\left( s _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } \\right) \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 385, + 440, + 398 + ], + "score": 1.0, + "content": ". A state consists of three parts: observation", + "type": "text" + }, + { + "bbox": [ + 440, + 388, + 450, + 397 + ], + "score": 0.84, + "content": "o _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 385, + 505, + 398 + ], + "score": 1.0, + "content": ", desired goal", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 395, + 507, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 121, + 410 + ], + "score": 0.89, + "content": "g _ { t } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 122, + 395, + 196, + 412 + ], + "score": 1.0, + "content": "and achieved goal", + "type": "text" + }, + { + "bbox": [ + 196, + 398, + 210, + 410 + ], + "score": 0.89, + "content": "g _ { t } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 211, + 395, + 252, + 412 + ], + "score": 1.0, + "content": ", define as", + "type": "text" + }, + { + "bbox": [ + 252, + 397, + 328, + 411 + ], + "score": 0.93, + "content": "\\dot { s } _ { t } = \\langle o _ { t } , g _ { t } ^ { a c } , g _ { t } ^ { d e } \\rangle", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 395, + 397, + 412 + ], + "score": 1.0, + "content": ", where normally", + "type": "text" + }, + { + "bbox": [ + 397, + 397, + 433, + 410 + ], + "score": 0.93, + "content": "g _ { t } ^ { d e } = g _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 433, + 395, + 451, + 412 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 451, + 398, + 466, + 410 + ], + "score": 0.9, + "content": "g _ { t } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 395, + 507, + 412 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 408, + 242, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 242, + 420 + ], + "score": 1.0, + "content": "goals that the agent has achieved.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 352, + 507, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 425, + 505, + 569 + ], + "lines": [ + { + "bbox": [ + 107, + 426, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 107, + 426, + 505, + 438 + ], + "score": 1.0, + "content": "We reuse the failed experience from the replay buffer with inverse simulation to create success-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 505, + 448 + ], + "score": 1.0, + "content": "ful rewards for the agent, as shown in Algorithm 1 (lines 13-23). Our inverse simulation contains", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 448, + 504, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 448, + 504, + 460 + ], + "score": 1.0, + "content": "two main steps: First, given a failed episode, for its achieved goal trajectory, we try to find a de-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 470 + ], + "score": 1.0, + "content": "sired goal trajectory from other episodes that could match it (line 15). Second, we assemble a new", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "episode by matching the achieved goal trajectory of the given episode to the desired goal trajec-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 479, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 104, + 479, + 281, + 496 + ], + "score": 1.0, + "content": "tory of the founded one (lines 17-21). Let", + "type": "text" + }, + { + "bbox": [ + 281, + 481, + 297, + 494 + ], + "score": 0.91, + "content": "g _ { i , q } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 479, + 506, + 496 + ], + "score": 1.0, + "content": "indicate the achieved goal of the agent at timestep", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 488, + 510, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 113, + 505 + ], + "score": 0.74, + "content": "q", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 488, + 153, + 519 + ], + "score": 1.0, + "content": "in episodxperience", + "type": "text" + }, + { + "bbox": [ + 159, + 494, + 165, + 504 + ], + "score": 0.72, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 492, + 200, + 507 + ], + "score": 0.91, + "content": "g _ { j , p } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 488, + 352, + 519 + ], + "score": 1.0, + "content": "te the desired goal at timestep , we search and draw two faile", + "type": "text" + }, + { + "bbox": [ + 353, + 496, + 360, + 505 + ], + "score": 0.77, + "content": "p", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 488, + 395, + 519 + ], + "score": 1.0, + "content": "in episoepisodes", + "type": "text" + }, + { + "bbox": [ + 406, + 494, + 412, + 505 + ], + "score": 0.76, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 488, + 426, + 519 + ], + "score": 1.0, + "content": ". Gand", + "type": "text" + }, + { + "bbox": [ + 440, + 488, + 443, + 519 + ], + "score": 1.0, + "content": "n", + "type": "text" + }, + { + "bbox": [ + 472, + 488, + 510, + 519 + ], + "score": 1.0, + "content": "of failed), where", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 153, + 506, + 472, + 519 + ], + "spans": [ + { + "bbox": [ + 153, + 506, + 226, + 518 + ], + "score": 0.9, + "content": "\\{ E _ { 0 } , E _ { 1 } , \\stackrel { \\cdot } { E } _ { 2 } , \\cdot \\cdot \\cdot \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 506, + 408, + 517 + ], + "score": 0.88, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 506, + 440, + 519 + ], + "score": 0.81, + "content": "E _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 506, + 472, + 518 + ], + "score": 0.68, + "content": "( i \\neq j )", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 514, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 145, + 531 + ], + "score": 0.78, + "content": "\\exists i , j , p , q ,", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 514, + 162, + 536 + ], + "score": 1.0, + "content": "s.t.", + "type": "text" + }, + { + "bbox": [ + 162, + 518, + 208, + 532 + ], + "score": 0.92, + "content": "g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 514, + 507, + 536 + ], + "score": 1.0, + "content": ". If we find two such failed episodes, we combine the two experience by", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 102, + 527, + 508, + 550 + ], + "spans": [ + { + "bbox": [ + 102, + 527, + 230, + 550 + ], + "score": 1.0, + "content": "replacing the desired goals in", + "type": "text" + }, + { + "bbox": [ + 230, + 532, + 242, + 543 + ], + "score": 0.86, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 527, + 258, + 550 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 258, + 531, + 282, + 545 + ], + "score": 0.93, + "content": "\\{ g _ { j , t } ^ { d e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 527, + 315, + 550 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 315, + 532, + 322, + 543 + ], + "score": 0.81, + "content": "j", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 527, + 362, + 550 + ], + "score": 1.0, + "content": "indicates", + "type": "text" + }, + { + "bbox": [ + 362, + 532, + 375, + 544 + ], + "score": 0.89, + "content": "E _ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 527, + 395, + 550 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 395, + 531, + 457, + 544 + ], + "score": 0.93, + "content": "t \\leq \\operatorname* { m i n } \\{ p , q \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 527, + 508, + 550 + ], + "score": 1.0, + "content": ". Based on", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 543, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 287, + 558 + ], + "score": 1.0, + "content": "this, we end up assembling a new experience", + "type": "text" + }, + { + "bbox": [ + 287, + 545, + 299, + 556 + ], + "score": 0.9, + "content": "E _ { i } ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 299, + 543, + 338, + 558 + ], + "score": 1.0, + "content": "based on", + "type": "text" + }, + { + "bbox": [ + 338, + 545, + 349, + 555 + ], + "score": 0.88, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 543, + 505, + 558 + ], + "score": 1.0, + "content": "with a new “imagined” goal trajectory", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 551, + 264, + 574 + ], + "spans": [ + { + "bbox": [ + 107, + 555, + 170, + 570 + ], + "score": 0.93, + "content": "\\{ g _ { j , 0 } ^ { d e } , \\cdot \\cdot \\cdot , g _ { j , t } ^ { d e } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 551, + 198, + 574 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 198, + 556, + 257, + 568 + ], + "score": 0.93, + "content": "t \\leq \\operatorname* { m i n } \\{ p , q \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 551, + 264, + 574 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 18.5, + "bbox_fs": [ + 102, + 426, + 510, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 504, + 618 + ], + "lines": [ + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 221, + 585 + ], + "score": 1.0, + "content": "More details of the complete", + "type": "text" + }, + { + "bbox": [ + 221, + 573, + 268, + 584 + ], + "score": 0.67, + "content": "\\mathrm { R L + D H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "method are shown in Algorithm 1. Unlike HER, our DHER", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 584, + 505, + 597 + ], + "score": 1.0, + "content": "needs to search all failed experiences to compose a “imagined” goal trajectory. Hence, the efficiency", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "of searching the memory is important. In our implementation, we use two hash tables to store the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 606, + 352, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 352, + 619 + ], + "score": 1.0, + "content": "trajectories of achieved goals and desired goals, respectively.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26.5, + "bbox_fs": [ + 105, + 573, + 505, + 619 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 651, + 194, + 664 + ], + "lines": [ + { + "bbox": [ + 105, + 651, + 196, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 196, + 666 + ], + "score": 1.0, + "content": "4 EXPERIMENT", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "We run extensive experiments to examine the proposed DHER for moving goals and compare it", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "with some competing baselines. We first introduce the environments and tasks that we want to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "address, followed by the experimental results. Demo videos from our experiments are available at", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 721, + 315, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 315, + 733 + ], + "score": 1.0, + "content": "https://sites.google.com/view/dher.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 79, + 498, + 209 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 79, + 498, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 79, + 498, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 79, + 498, + 209 + ], + "score": 0.968, + "type": "image", + "image_path": "868b5ca50a241e6883855868e5cdc6d4e2e4a7d4f491d103d018de67c19012a7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 79, + 498, + 122.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 122.33333333333334, + 498, + 165.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 165.66666666666669, + 498, + 209.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 217, + 217, + 394, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 217, + 395, + 230 + ], + "spans": [ + { + "bbox": [ + 215, + 217, + 395, + 230 + ], + "score": 1.0, + "content": "Figure 3: Results on different environments.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 254, + 201, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 202, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 202, + 266 + ], + "score": 1.0, + "content": "4.1 ENVIRONMENTS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 291 + ], + "score": 1.0, + "content": "Whereas grasping moving targets are fairly common in robotic applications, there are rarely existing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "score": 1.0, + "content": "environments featuring dynamic goals. We modify the robotic manipulation environments created", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "by OpenAI (Brockman et al., 2016) for our experiments. As shown in Figure 2, we assign certain", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "rules to the goals so that they accordingly move in the environments while an agent is required to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 320, + 505, + 335 + ], + "score": 1.0, + "content": "control the robotic arm’s grippers to reach the goal that moves along a straight line (Dy-Reaching),", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "to reach the goal that moves in a circle (Dy-Circling), or to push a block to the goal that moves along", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "a straight line (Dy-Pushing). In addition, we also develop a new GREEDY SNAKE environment (Dy-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 354, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 365 + ], + "score": 1.0, + "content": "Snake), in which the goal moves from one discrete cell to another. The greedy snake aims to reach", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 363, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 379 + ], + "score": 1.0, + "content": "the goal (and eat it). No matter what actions are taken by the agent, the underlying rule of changing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "the goals’ positions remain the same. We assume that the law of the goal’s motion is unknown to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 387, + 371, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 371, + 399 + ], + "score": 1.0, + "content": "the agent. More details of the tasks are described in the appendix.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "For the first three tasks, we follow the basic settings of OpenAI robotics environments (Brockman", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "et al., 2016). States are read from the MuJoCo physics engine. An observation consists of relative", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "positions of the object and the target (grippers are blocked) Goals are positions in the 3D world", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 308, + 449 + ], + "score": 1.0, + "content": "coordinate system with a fixed tolerance (we use", + "type": "text" + }, + { + "bbox": [ + 309, + 437, + 348, + 447 + ], + "score": 0.87, + "content": "\\epsilon = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "for the tolerance). Note the goals are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 307, + 460 + ], + "score": 1.0, + "content": "able to move in our tasks. The velocity we use is", + "type": "text" + }, + { + "bbox": [ + 307, + 448, + 350, + 458 + ], + "score": 0.82, + "content": "v = 0 . 0 1 1", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 447, + 506, + 460 + ], + "score": 1.0, + "content": ". 1 epoch indicates 100 episodes. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 444, + 471 + ], + "score": 1.0, + "content": "start positions of the goals are randomly chosen. Rewards are binary and sparse:", + "type": "text" + }, + { + "bbox": [ + 444, + 458, + 505, + 471 + ], + "score": 0.92, + "content": "r ( s _ { t } , a _ { t } , g _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 469, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 237, + 485 + ], + "score": 0.9, + "content": "- \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( | s _ { t + 1 } ^ { \\mathrm { o b j } } - g _ { t + 1 } | \\geq \\epsilon \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 469, + 266, + 488 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 267, + 473, + 286, + 483 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 469, + 305, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 306, + 474, + 325, + 483 + ], + "score": 0.89, + "content": "g _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 469, + 507, + 488 + ], + "score": 1.0, + "content": "are respectively the state and dynamic goal", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 239, + 495 + ], + "score": 1.0, + "content": "after the execution of the action", + "type": "text" + }, + { + "bbox": [ + 239, + 484, + 249, + 493 + ], + "score": 0.85, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 482, + 299, + 495 + ], + "score": 1.0, + "content": "in the state", + "type": "text" + }, + { + "bbox": [ + 299, + 485, + 308, + 493 + ], + "score": 0.86, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 482, + 505, + 495 + ], + "score": 1.0, + "content": ". Two positions are overlapped if they are close", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 188, + 506 + ], + "score": 1.0, + "content": "within the tolerance", + "type": "text" + }, + { + "bbox": [ + 189, + 495, + 194, + 503 + ], + "score": 0.53, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 492, + 505, + 506 + ], + "score": 1.0, + "content": ". We use 3-dimensional actions, in which three dimensions correspond to the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 504, + 318, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 318, + 518 + ], + "score": 1.0, + "content": "desired relative gripper position at the next timestep.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 104, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "For the last task, it is very similar to the standard snake game. There is a 2D plane grid whose", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 135, + 544 + ], + "score": 1.0, + "content": "size is", + "type": "text" + }, + { + "bbox": [ + 135, + 532, + 169, + 543 + ], + "score": 0.9, + "content": "3 0 \\times 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 532, + 423, + 544 + ], + "score": 1.0, + "content": ". A snake and a goal (food) both move in this grid. We use a", + "type": "text" + }, + { + "bbox": [ + 423, + 533, + 448, + 543 + ], + "score": 0.9, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "square as the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "body of the snake and do not allow the snake to grow any longer as it moves. The states of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 554, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 566 + ], + "score": 1.0, + "content": "system are represented by using the positions of the snake and food. Goals are the positions of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "the food. The snake has to move itself such that it resides in the same cell as the food at a certain", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 576, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 282, + 588 + ], + "score": 1.0, + "content": "timestep. The velocity of the goal is set to", + "type": "text" + }, + { + "bbox": [ + 282, + 576, + 305, + 588 + ], + "score": 0.91, + "content": "( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 576, + 504, + 588 + ], + "score": 1.0, + "content": ". The start positions of the goal are set randomly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "in different episodes of the game, so are the start positions of the snake. Rewards are binary and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 102, + 594, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 102, + 594, + 136, + 616 + ], + "score": 1.0, + "content": "sparse:", + "type": "text" + }, + { + "bbox": [ + 137, + 598, + 334, + 611 + ], + "score": 0.88, + "content": "r \\big ( s _ { t } , \\bar { a } _ { t } , g _ { t + 1 } \\big ) = - \\bar { \\mathbf { 1 } } _ { \\mathrm { c o n d i t i o n } } \\big ( \\big | s _ { t + 1 } ^ { \\mathrm { s n a k e } } - g _ { t + 1 } \\big | \\bar { \\neq } 0 \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 594, + 360, + 616 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 361, + 599, + 380, + 610 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 594, + 397, + 616 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 599, + 417, + 610 + ], + "score": 0.92, + "content": "g _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 594, + 506, + 616 + ], + "score": 1.0, + "content": "are the environment’s", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 288, + 621 + ], + "score": 1.0, + "content": "state and goal after the agent executes action", + "type": "text" + }, + { + "bbox": [ + 288, + 611, + 298, + 620 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 609, + 345, + 621 + ], + "score": 1.0, + "content": "in the state", + "type": "text" + }, + { + "bbox": [ + 345, + 611, + 354, + 620 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 609, + 505, + 621 + ], + "score": 1.0, + "content": ". Observations are represented by the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "positions of the snake and food. We also add to the state the distance between the snake and food.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "Actions allowed in the game are the following: move up, move down, move left, and move right, for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 641, + 194, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 194, + 655 + ], + "score": 1.0, + "content": "one cell per timestep.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 180, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 181, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 181, + 684 + ], + "score": 1.0, + "content": "4.2 BASELINES", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 694, + 493, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 495, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 495, + 707 + ], + "score": 1.0, + "content": "For the first three tasks which call for continuous control, we consider two competing baselines:2", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 721, + 252, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 252, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 252, + 734 + ], + "score": 1.0, + "content": "2https://github.com/openai/baselines", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 79, + 498, + 209 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 111, + 79, + 498, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 111, + 79, + 498, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 79, + 498, + 209 + ], + "score": 0.968, + "type": "image", + "image_path": "868b5ca50a241e6883855868e5cdc6d4e2e4a7d4f491d103d018de67c19012a7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 111, + 79, + 498, + 122.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 111, + 122.33333333333334, + 498, + 165.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 111, + 165.66666666666669, + 498, + 209.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 217, + 217, + 394, + 229 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 217, + 395, + 230 + ], + "spans": [ + { + "bbox": [ + 215, + 217, + 395, + 230 + ], + "score": 1.0, + "content": "Figure 3: Results on different environments.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "title", + "bbox": [ + 107, + 254, + 201, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 254, + 202, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 202, + 266 + ], + "score": 1.0, + "content": "4.1 ENVIRONMENTS", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 276, + 505, + 398 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 505, + 291 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 291 + ], + "score": 1.0, + "content": "Whereas grasping moving targets are fairly common in robotic applications, there are rarely existing", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "score": 1.0, + "content": "environments featuring dynamic goals. We modify the robotic manipulation environments created", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "by OpenAI (Brockman et al., 2016) for our experiments. As shown in Figure 2, we assign certain", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 505, + 323 + ], + "score": 1.0, + "content": "rules to the goals so that they accordingly move in the environments while an agent is required to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 320, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 104, + 320, + 505, + 335 + ], + "score": 1.0, + "content": "control the robotic arm’s grippers to reach the goal that moves along a straight line (Dy-Reaching),", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "to reach the goal that moves in a circle (Dy-Circling), or to push a block to the goal that moves along", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 104, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "a straight line (Dy-Pushing). In addition, we also develop a new GREEDY SNAKE environment (Dy-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 354, + 505, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 365 + ], + "score": 1.0, + "content": "Snake), in which the goal moves from one discrete cell to another. The greedy snake aims to reach", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 363, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 379 + ], + "score": 1.0, + "content": "the goal (and eat it). No matter what actions are taken by the agent, the underlying rule of changing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 376, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 506, + 388 + ], + "score": 1.0, + "content": "the goals’ positions remain the same. We assume that the law of the goal’s motion is unknown to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 387, + 371, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 371, + 399 + ], + "score": 1.0, + "content": "the agent. More details of the tasks are described in the appendix.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10, + "bbox_fs": [ + 104, + 276, + 506, + 399 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 403, + 505, + 516 + ], + "lines": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "For the first three tasks, we follow the basic settings of OpenAI robotics environments (Brockman", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "et al., 2016). States are read from the MuJoCo physics engine. An observation consists of relative", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 425, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 104, + 425, + 506, + 438 + ], + "score": 1.0, + "content": "positions of the object and the target (grippers are blocked) Goals are positions in the 3D world", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 308, + 449 + ], + "score": 1.0, + "content": "coordinate system with a fixed tolerance (we use", + "type": "text" + }, + { + "bbox": [ + 309, + 437, + 348, + 447 + ], + "score": 0.87, + "content": "\\epsilon = 0 . 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "for the tolerance). Note the goals are", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 447, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 307, + 460 + ], + "score": 1.0, + "content": "able to move in our tasks. The velocity we use is", + "type": "text" + }, + { + "bbox": [ + 307, + 448, + 350, + 458 + ], + "score": 0.82, + "content": "v = 0 . 0 1 1", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 447, + 506, + 460 + ], + "score": 1.0, + "content": ". 1 epoch indicates 100 episodes. The", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 444, + 471 + ], + "score": 1.0, + "content": "start positions of the goals are randomly chosen. Rewards are binary and sparse:", + "type": "text" + }, + { + "bbox": [ + 444, + 458, + 505, + 471 + ], + "score": 0.92, + "content": "r ( s _ { t } , a _ { t } , g _ { t } ) =", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 469, + 507, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 469, + 237, + 485 + ], + "score": 0.9, + "content": "- \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( | s _ { t + 1 } ^ { \\mathrm { o b j } } - g _ { t + 1 } | \\geq \\epsilon \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 469, + 266, + 488 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 267, + 473, + 286, + 483 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 469, + 305, + 488 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 306, + 474, + 325, + 483 + ], + "score": 0.89, + "content": "g _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 469, + 507, + 488 + ], + "score": 1.0, + "content": "are respectively the state and dynamic goal", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 482, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 239, + 495 + ], + "score": 1.0, + "content": "after the execution of the action", + "type": "text" + }, + { + "bbox": [ + 239, + 484, + 249, + 493 + ], + "score": 0.85, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 482, + 299, + 495 + ], + "score": 1.0, + "content": "in the state", + "type": "text" + }, + { + "bbox": [ + 299, + 485, + 308, + 493 + ], + "score": 0.86, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 482, + 505, + 495 + ], + "score": 1.0, + "content": ". Two positions are overlapped if they are close", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 188, + 506 + ], + "score": 1.0, + "content": "within the tolerance", + "type": "text" + }, + { + "bbox": [ + 189, + 495, + 194, + 503 + ], + "score": 0.53, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 492, + 505, + 506 + ], + "score": 1.0, + "content": ". We use 3-dimensional actions, in which three dimensions correspond to the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 504, + 318, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 318, + 518 + ], + "score": 1.0, + "content": "desired relative gripper position at the next timestep.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20.5, + "bbox_fs": [ + 104, + 403, + 507, + 518 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 521, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 104, + 520, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 505, + 534 + ], + "score": 1.0, + "content": "For the last task, it is very similar to the standard snake game. There is a 2D plane grid whose", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 532, + 505, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 135, + 544 + ], + "score": 1.0, + "content": "size is", + "type": "text" + }, + { + "bbox": [ + 135, + 532, + 169, + 543 + ], + "score": 0.9, + "content": "3 0 \\times 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 532, + 423, + 544 + ], + "score": 1.0, + "content": ". A snake and a goal (food) both move in this grid. We use a", + "type": "text" + }, + { + "bbox": [ + 423, + 533, + 448, + 543 + ], + "score": 0.9, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 532, + 505, + 544 + ], + "score": 1.0, + "content": "square as the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "body of the snake and do not allow the snake to grow any longer as it moves. The states of the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 554, + 506, + 566 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 506, + 566 + ], + "score": 1.0, + "content": "system are represented by using the positions of the snake and food. Goals are the positions of", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 505, + 577 + ], + "score": 1.0, + "content": "the food. The snake has to move itself such that it resides in the same cell as the food at a certain", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 576, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 106, + 576, + 282, + 588 + ], + "score": 1.0, + "content": "timestep. The velocity of the goal is set to", + "type": "text" + }, + { + "bbox": [ + 282, + 576, + 305, + 588 + ], + "score": 0.91, + "content": "( 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 576, + 504, + 588 + ], + "score": 1.0, + "content": ". The start positions of the goal are set randomly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 600 + ], + "score": 1.0, + "content": "in different episodes of the game, so are the start positions of the snake. Rewards are binary and", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 102, + 594, + 506, + 616 + ], + "spans": [ + { + "bbox": [ + 102, + 594, + 136, + 616 + ], + "score": 1.0, + "content": "sparse:", + "type": "text" + }, + { + "bbox": [ + 137, + 598, + 334, + 611 + ], + "score": 0.88, + "content": "r \\big ( s _ { t } , \\bar { a } _ { t } , g _ { t + 1 } \\big ) = - \\bar { \\mathbf { 1 } } _ { \\mathrm { c o n d i t i o n } } \\big ( \\big | s _ { t + 1 } ^ { \\mathrm { s n a k e } } - g _ { t + 1 } \\big | \\bar { \\neq } 0 \\big )", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 594, + 360, + 616 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 361, + 599, + 380, + 610 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 594, + 397, + 616 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 599, + 417, + 610 + ], + "score": 0.92, + "content": "g _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 594, + 506, + 616 + ], + "score": 1.0, + "content": "are the environment’s", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 609, + 505, + 621 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 288, + 621 + ], + "score": 1.0, + "content": "state and goal after the agent executes action", + "type": "text" + }, + { + "bbox": [ + 288, + 611, + 298, + 620 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 609, + 345, + 621 + ], + "score": 1.0, + "content": "in the state", + "type": "text" + }, + { + "bbox": [ + 345, + 611, + 354, + 620 + ], + "score": 0.85, + "content": "s _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 609, + 505, + 621 + ], + "score": 1.0, + "content": ". Observations are represented by the", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 619, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 505, + 632 + ], + "score": 1.0, + "content": "positions of the snake and food. We also add to the state the distance between the snake and food.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 106, + 631, + 505, + 643 + ], + "score": 1.0, + "content": "Actions allowed in the game are the following: move up, move down, move left, and move right, for", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 641, + 194, + 655 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 194, + 655 + ], + "score": 1.0, + "content": "one cell per timestep.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31.5, + "bbox_fs": [ + 102, + 520, + 506, + 655 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 671, + 180, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 671, + 181, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 181, + 684 + ], + "score": 1.0, + "content": "4.2 BASELINES", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 694, + 493, + 705 + ], + "lines": [ + { + "bbox": [ + 105, + 691, + 495, + 707 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 495, + 707 + ], + "score": 1.0, + "content": "For the first three tasks which call for continuous control, we consider two competing baselines:2", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 133, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "• DDPG, which is a model-free RL algorithm for continuous control (Lillicrap et al., 2015).", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 247, + 503, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 247, + 503, + 262 + ], + "score": 1.0, + "content": "It learns a deterministic policy by using a stochastic counterpart to explore in the training.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 691, + 495, + 707 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 81, + 275, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 81, + 275, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 81, + 275, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 81, + 275, + 194 + ], + "score": 0.966, + "type": "image", + "image_path": "732a387a80c212738b2fb410b777a5c71d62a4dacb48b29dd438871f23abaf26.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 135, + 81, + 275, + 137.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 135, + 137.5, + 275, + 194.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 144, + 204, + 268, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 203, + 269, + 216 + ], + "spans": [ + { + "bbox": [ + 143, + 203, + 269, + 216 + ], + "score": 1.0, + "content": "Figure 4: Low velocity: 0.001.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 334, + 81, + 473, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 81, + 473, + 194 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 81, + 473, + 194 + ], + "spans": [ + { + "bbox": [ + 334, + 81, + 473, + 194 + ], + "score": 0.966, + "type": "image", + "image_path": "f8c0fd7014d07719a2c477eef04492742d4a9d1ff91065b3d850a3053e6e0ee6.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 334, + 81, + 473, + 137.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 334, + 137.5, + 473, + 194.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 342, + 204, + 469, + 216 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 341, + 203, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 341, + 203, + 468, + 217 + ], + "score": 1.0, + "content": "Figure 5: High velocity: 0.016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 133, + 237, + 504, + 260 + ], + "lines": [ + { + "bbox": [ + 133, + 236, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 133, + 236, + 505, + 250 + ], + "score": 1.0, + "content": "• DDPG, which is a model-free RL algorithm for continuous control (Lillicrap et al., 2015).", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 247, + 503, + 262 + ], + "spans": [ + { + "bbox": [ + 141, + 247, + 503, + 262 + ], + "score": 1.0, + "content": "It learns a deterministic policy by using a stochastic counterpart to explore in the training.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 132, + 264, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 134, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 263, + 142, + 277 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 264, + 203, + 275 + ], + "score": 0.35, + "content": "\\mathrm { D D P G } + \\mathrm { H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 263, + 505, + 277 + ], + "score": 1.0, + "content": ", which improves the replay buffer of DDPG by the hindsight experience", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 275, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 286, + 287 + ], + "score": 1.0, + "content": "replay (Andrychowicz et al., 2017).", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 290, + 448, + 303 + ], + "spans": [ + { + "bbox": [ + 137, + 290, + 355, + 303 + ], + "score": 1.0, + "content": "DDPG (dense), which employs the negative distance", + "type": "text" + }, + { + "bbox": [ + 355, + 291, + 375, + 302 + ], + "score": 0.86, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 290, + 448, + 303 + ], + "score": 1.0, + "content": "as dense rewards.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 139, + 305, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 139, + 305, + 370, + 318 + ], + "score": 1.0, + "content": "DDPG (dense-2), which employs the negative distance", + "type": "text" + }, + { + "bbox": [ + 370, + 306, + 390, + 317 + ], + "score": 0.87, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 305, + 474, + 318 + ], + "score": 1.0, + "content": "as dense rewards if", + "type": "text" + }, + { + "bbox": [ + 474, + 306, + 501, + 317 + ], + "score": 0.9, + "content": "d \\geq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 305, + 506, + 318 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 317, + 455, + 329 + ], + "spans": [ + { + "bbox": [ + 142, + 317, + 191, + 329 + ], + "score": 1.0, + "content": "However, if", + "type": "text" + }, + { + "bbox": [ + 191, + 317, + 216, + 327 + ], + "score": 0.9, + "content": "d < \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 317, + 300, + 329 + ], + "score": 1.0, + "content": "(i.e., success) it uses", + "type": "text" + }, + { + "bbox": [ + 301, + 317, + 346, + 328 + ], + "score": 0.9, + "content": "( - d + 1 . 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 317, + 455, + 329 + ], + "score": 1.0, + "content": "as rewards. 1.0 is a bonus.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 105, + 338, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "score": 1.0, + "content": "For the last task of discrete control, we use the following baselines in the experiments: DQN (Mnih", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 349, + 424, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 173, + 362 + ], + "score": 1.0, + "content": "et al., 2015) and", + "type": "text" + }, + { + "bbox": [ + 173, + 349, + 226, + 360 + ], + "score": 0.8, + "content": "\\mathrm { D Q N + H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 349, + 424, + 362 + ], + "score": 1.0, + "content": ", which uses HER to enhance the replay in DQN.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 132, + 369, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 131, + 368, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 506, + 383 + ], + "score": 1.0, + "content": "• DQN, which is a powerful model-free RL algorithm for discrete action spaces (Mnih et al.,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 140, + 380, + 172, + 394 + ], + "spans": [ + { + "bbox": [ + 140, + 380, + 172, + 394 + ], + "score": 1.0, + "content": "2015).", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 131, + 396, + 393, + 408 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 141, + 408 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 396, + 195, + 408 + ], + "score": 0.82, + "content": "\\mathrm { D Q N + H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 396, + 393, + 408 + ], + "score": 1.0, + "content": ", which uses HER to enhance the replay in DQN.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 132, + 412, + 427, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 333, + 424 + ], + "score": 1.0, + "content": "• DQN (dense), which uses the negative distance", + "type": "text" + }, + { + "bbox": [ + 333, + 412, + 353, + 423 + ], + "score": 0.86, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 412, + 427, + 424 + ], + "score": 1.0, + "content": "as dense rewards.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 132, + 427, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 351, + 439 + ], + "score": 1.0, + "content": "• DQN (dense-2), which uses the negative distance", + "type": "text" + }, + { + "bbox": [ + 352, + 428, + 372, + 439 + ], + "score": 0.89, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 427, + 506, + 439 + ], + "score": 1.0, + "content": "as dense rewards. However, if", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 142, + 438, + 445, + 450 + ], + "spans": [ + { + "bbox": [ + 142, + 438, + 175, + 449 + ], + "score": 0.88, + "content": "d = 0 . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 439, + 261, + 450 + ], + "score": 1.0, + "content": "(i.e., success) it uses", + "type": "text" + }, + { + "bbox": [ + 261, + 438, + 306, + 450 + ], + "score": 0.9, + "content": "( - d + 1 . 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 439, + 445, + 450 + ], + "score": 1.0, + "content": "as rewards instead. 1.0 is a bonus.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 464, + 375, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 376, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 376, + 476 + ], + "score": 1.0, + "content": "4.3 COMPARISON RESULTS ON THE ROBOTIC ENVIRONMENTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "For the continuous control, we present three sets of comparison results in Figure 3 for the first three", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "tasks, respectively. Consistently, the results show our DHER algorithm outperforms the others. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "two baselines are not able to catch up even after we train them for thousands of iterations. Vanilla", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "DDPG is slightly better than the version with HER. HER does not benefit DDPG in these tasks", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "because the goals in HER are fixed, fundamentally misleading the agent in the attempt of solving", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "the tasks with dynamic goals. The results of DDPG (dense) and DDPG (dense-2) suggest that even", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "the dense rewards do not work well as they are agnostic to the task of interest. A good reward", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 561, + 491, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 491, + 574 + ], + "score": 1.0, + "content": "shaping may give rise to better performance by carefully tuning it for the task of dynamic goals.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "Comparing Figure 3a and Figure 3b with Figure 3c, we find that DHER learns faster in Dy-Reaching", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 589, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 603 + ], + "score": 1.0, + "content": "and Dy-Circling than Dy-Pushing probably because Dy-Reaching and Dy-Circling are easier tasks", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "than Dy-Pushing. In Dy-Pushing, all the algorithms take a fairly big amount of time to explore", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "without receiving any positive feedback. However, the more failed experiences the agent encounters,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "score": 1.0, + "content": "the better change our algorithm is able to identify relevant episodes from them for assembling useful", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "dynamic goals. As a result, DHER is able to pick up the momentum and learns faster and better than", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "the baselines after a certain point. In Figure 3a, the performance decreases a little. The reason may", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "be that as successful experience increases, some assembled experience is inconsistent with these", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 666, + 198, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 198, + 679 + ], + "score": 1.0, + "content": "successful experience.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 107, + 691, + 375, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 376, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 376, + 703 + ], + "score": 1.0, + "content": "4.3.1 COMPARISON USING DIFFERENT VELOCITIES OF GOALS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "To show the performance of our method on more complex tasks with different velocities, we study", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "score": 1.0, + "content": "different methods in Dy-Reaching environment as shown in Figures 4 and 5 with the same physical", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 81, + 275, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 81, + 275, + 194 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 81, + 275, + 194 + ], + "spans": [ + { + "bbox": [ + 135, + 81, + 275, + 194 + ], + "score": 0.966, + "type": "image", + "image_path": "732a387a80c212738b2fb410b777a5c71d62a4dacb48b29dd438871f23abaf26.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 135, + 81, + 275, + 137.5 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 135, + 137.5, + 275, + 194.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 144, + 204, + 268, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 203, + 269, + 216 + ], + "spans": [ + { + "bbox": [ + 143, + 203, + 269, + 216 + ], + "score": 1.0, + "content": "Figure 4: Low velocity: 0.001.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 334, + 81, + 473, + 194 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 81, + 473, + 194 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 81, + 473, + 194 + ], + "spans": [ + { + "bbox": [ + 334, + 81, + 473, + 194 + ], + "score": 0.966, + "type": "image", + "image_path": "f8c0fd7014d07719a2c477eef04492742d4a9d1ff91065b3d850a3053e6e0ee6.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 334, + 81, + 473, + 137.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 334, + 137.5, + 473, + 194.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 342, + 204, + 469, + 216 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 341, + 203, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 341, + 203, + 468, + 217 + ], + "score": 1.0, + "content": "Figure 5: High velocity: 0.016.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 133, + 237, + 504, + 260 + ], + "lines": [], + "index": 6.5, + "bbox_fs": [ + 133, + 236, + 505, + 262 + ], + "lines_deleted": true + }, + { + "type": "list", + "bbox": [ + 132, + 264, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 134, + 263, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 134, + 263, + 142, + 277 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 264, + 203, + 275 + ], + "score": 0.35, + "content": "\\mathrm { D D P G } + \\mathrm { H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 263, + 505, + 277 + ], + "score": 1.0, + "content": ", which improves the replay buffer of DDPG by the hindsight experience", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 275, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 141, + 275, + 286, + 287 + ], + "score": 1.0, + "content": "replay (Andrychowicz et al., 2017).", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 137, + 290, + 448, + 303 + ], + "spans": [ + { + "bbox": [ + 137, + 290, + 355, + 303 + ], + "score": 1.0, + "content": "DDPG (dense), which employs the negative distance", + "type": "text" + }, + { + "bbox": [ + 355, + 291, + 375, + 302 + ], + "score": 0.86, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 290, + 448, + 303 + ], + "score": 1.0, + "content": "as dense rewards.", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 139, + 305, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 139, + 305, + 370, + 318 + ], + "score": 1.0, + "content": "DDPG (dense-2), which employs the negative distance", + "type": "text" + }, + { + "bbox": [ + 370, + 306, + 390, + 317 + ], + "score": 0.87, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 305, + 474, + 318 + ], + "score": 1.0, + "content": "as dense rewards if", + "type": "text" + }, + { + "bbox": [ + 474, + 306, + 501, + 317 + ], + "score": 0.9, + "content": "d \\geq \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 305, + 506, + 318 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 317, + 455, + 329 + ], + "spans": [ + { + "bbox": [ + 142, + 317, + 191, + 329 + ], + "score": 1.0, + "content": "However, if", + "type": "text" + }, + { + "bbox": [ + 191, + 317, + 216, + 327 + ], + "score": 0.9, + "content": "d < \\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 317, + 300, + 329 + ], + "score": 1.0, + "content": "(i.e., success) it uses", + "type": "text" + }, + { + "bbox": [ + 301, + 317, + 346, + 328 + ], + "score": 0.9, + "content": "( - d + 1 . 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 346, + 317, + 455, + 329 + ], + "score": 1.0, + "content": "as rewards. 1.0 is a bonus.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + } + ], + "index": 10, + "bbox_fs": [ + 134, + 263, + 506, + 329 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 338, + 504, + 361 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 505, + 350 + ], + "score": 1.0, + "content": "For the last task of discrete control, we use the following baselines in the experiments: DQN (Mnih", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 349, + 424, + 362 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 173, + 362 + ], + "score": 1.0, + "content": "et al., 2015) and", + "type": "text" + }, + { + "bbox": [ + 173, + 349, + 226, + 360 + ], + "score": 0.8, + "content": "\\mathrm { D Q N + H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 349, + 424, + 362 + ], + "score": 1.0, + "content": ", which uses HER to enhance the replay in DQN.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 336, + 505, + 362 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 369, + 505, + 450 + ], + "lines": [ + { + "bbox": [ + 131, + 368, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 131, + 368, + 506, + 383 + ], + "score": 1.0, + "content": "• DQN, which is a powerful model-free RL algorithm for discrete action spaces (Mnih et al.,", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 140, + 380, + 172, + 394 + ], + "spans": [ + { + "bbox": [ + 140, + 380, + 172, + 394 + ], + "score": 1.0, + "content": "2015).", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 131, + 396, + 393, + 408 + ], + "spans": [ + { + "bbox": [ + 131, + 396, + 141, + 408 + ], + "score": 1.0, + "content": "•", + "type": "text" + }, + { + "bbox": [ + 142, + 396, + 195, + 408 + ], + "score": 0.82, + "content": "\\mathrm { D Q N + H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 396, + 393, + 408 + ], + "score": 1.0, + "content": ", which uses HER to enhance the replay in DQN.", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 412, + 427, + 424 + ], + "spans": [ + { + "bbox": [ + 132, + 412, + 333, + 424 + ], + "score": 1.0, + "content": "• DQN (dense), which uses the negative distance", + "type": "text" + }, + { + "bbox": [ + 333, + 412, + 353, + 423 + ], + "score": 0.86, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 412, + 427, + 424 + ], + "score": 1.0, + "content": "as dense rewards.", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 427, + 506, + 439 + ], + "spans": [ + { + "bbox": [ + 132, + 427, + 351, + 439 + ], + "score": 1.0, + "content": "• DQN (dense-2), which uses the negative distance", + "type": "text" + }, + { + "bbox": [ + 352, + 428, + 372, + 439 + ], + "score": 0.89, + "content": "( - d )", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 427, + 506, + 439 + ], + "score": 1.0, + "content": "as dense rewards. However, if", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 438, + 445, + 450 + ], + "spans": [ + { + "bbox": [ + 142, + 438, + 175, + 449 + ], + "score": 0.88, + "content": "d = 0 . 0", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 439, + 261, + 450 + ], + "score": 1.0, + "content": "(i.e., success) it uses", + "type": "text" + }, + { + "bbox": [ + 261, + 438, + 306, + 450 + ], + "score": 0.9, + "content": "( - d + 1 . 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 439, + 445, + 450 + ], + "score": 1.0, + "content": "as rewards instead. 1.0 is a bonus.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + } + ], + "index": 17.5, + "bbox_fs": [ + 131, + 368, + 506, + 450 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 464, + 375, + 475 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 376, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 376, + 476 + ], + "score": 1.0, + "content": "4.3 COMPARISON RESULTS ON THE ROBOTIC ENVIRONMENTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 106, + 484, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 497 + ], + "score": 1.0, + "content": "For the continuous control, we present three sets of comparison results in Figure 3 for the first three", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "tasks, respectively. Consistently, the results show our DHER algorithm outperforms the others. The", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "two baselines are not able to catch up even after we train them for thousands of iterations. Vanilla", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "DDPG is slightly better than the version with HER. HER does not benefit DDPG in these tasks", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 527, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 505, + 542 + ], + "score": 1.0, + "content": "because the goals in HER are fixed, fundamentally misleading the agent in the attempt of solving", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "the tasks with dynamic goals. The results of DDPG (dense) and DDPG (dense-2) suggest that even", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "the dense rewards do not work well as they are agnostic to the task of interest. A good reward", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 561, + 491, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 491, + 574 + ], + "score": 1.0, + "content": "shaping may give rise to better performance by carefully tuning it for the task of dynamic goals.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 484, + 505, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "Comparing Figure 3a and Figure 3b with Figure 3c, we find that DHER learns faster in Dy-Reaching", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 589, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 505, + 603 + ], + "score": 1.0, + "content": "and Dy-Circling than Dy-Pushing probably because Dy-Reaching and Dy-Circling are easier tasks", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "than Dy-Pushing. In Dy-Pushing, all the algorithms take a fairly big amount of time to explore", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 505, + 624 + ], + "score": 1.0, + "content": "without receiving any positive feedback. However, the more failed experiences the agent encounters,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 636 + ], + "score": 1.0, + "content": "the better change our algorithm is able to identify relevant episodes from them for assembling useful", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "dynamic goals. As a result, DHER is able to pick up the momentum and learns faster and better than", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "the baselines after a certain point. In Figure 3a, the performance decreases a little. The reason may", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "be that as successful experience increases, some assembled experience is inconsistent with these", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 666, + 198, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 198, + 679 + ], + "score": 1.0, + "content": "successful experience.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 577, + 506, + 679 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 691, + 375, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 689, + 376, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 376, + 703 + ], + "score": 1.0, + "content": "4.3.1 COMPARISON USING DIFFERENT VELOCITIES OF GOALS", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "To show the performance of our method on more complex tasks with different velocities, we study", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 504, + 733 + ], + "score": 1.0, + "content": "different methods in Dy-Reaching environment as shown in Figures 4 and 5 with the same physical", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "score": 1.0, + "content": "properties as the previous experiments. Overall the results show our method is much better than", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "DDPG and HER. As a reminder, the threshold for calculating rewards we used is 0.01. In Figure 4,", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 263, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 278 + ], + "score": 1.0, + "content": "the task becomes an easier task because of the slower velocity. It shows that our method quickly", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "achieves to a good result around 5 epoch. Comparing with Figure 3a, it shows that the performance", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 126, + 300 + ], + "score": 1.0, + "content": "with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 126, + 287, + 169, + 297 + ], + "score": 0.89, + "content": "v = 0 . 0 0 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 169, + 286, + 308, + 300 + ], + "score": 1.0, + "content": "is better than the performance with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 308, + 287, + 351, + 297 + ], + "score": 0.9, + "content": "v = 0 . 0 1 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 351, + 286, + 382, + 300 + ], + "score": 1.0, + "content": "and get", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 382, + 286, + 402, + 298 + ], + "score": 0.89, + "content": "1 5 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 402, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "improvements. However,", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 248, + 309 + ], + "score": 1.0, + "content": "it also shows the performance with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 249, + 298, + 291, + 308 + ], + "score": 0.89, + "content": "v = 0 . 0 1 6", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 292, + 298, + 435, + 309 + ], + "score": 1.0, + "content": "is worse than the performance with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 436, + 298, + 478, + 308 + ], + "score": 0.89, + "content": "v = 0 . 0 1 1", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 479, + 298, + 505, + 309 + ], + "score": 1.0, + "content": ". Both", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 154, + 321 + ], + "score": 1.0, + "content": "DDPG and", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 154, + 309, + 208, + 319 + ], + "score": 0.42, + "content": "\\mathrm { \\Delta D D P G + H E R }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 209, + 308, + 260, + 321 + ], + "score": 1.0, + "content": "failed when", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 261, + 309, + 306, + 319 + ], + "score": 0.88, + "content": "v = 0 . 0 1 6", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 306, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "and their performance is 0. This performance is", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 319, + 425, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 425, + 331 + ], + "score": 1.0, + "content": "consistent because the task becomes more difficult when the velocity increases.", + "type": "text", + "cross_page": true + } + ], + "index": 13 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 709, + 504, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 82, + 276, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 82, + 276, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 82, + 276, + 192 + ], + "spans": [ + { + "bbox": [ + 135, + 82, + 276, + 192 + ], + "score": 0.966, + "type": "image", + "image_path": "14dd0b3e1878d3f49fb1b7d44ca84006cca0ac15133dc036aef116910149f401.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 135, + 82, + 276, + 137.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 135, + 137.0, + 276, + 192.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 132, + 203, + 281, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 281, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 281, + 217 + ], + "score": 1.0, + "content": "Figure 6: Snake with dynamic goals.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 334, + 81, + 474, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 81, + 474, + 192 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 81, + 474, + 192 + ], + "spans": [ + { + "bbox": [ + 334, + 81, + 474, + 192 + ], + "score": 0.967, + "type": "image", + "image_path": "97b49eb4c50a91b591b5c09b7c8417f90a64b3336f8b06a8631418b78f185b01.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 334, + 81, + 474, + 136.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 334, + 136.5, + 474, + 192.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 308, + 203, + 502, + 214 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 306, + 202, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 379, + 216 + ], + "score": 1.0, + "content": "Figure 7: Special:", + "type": "text" + }, + { + "bbox": [ + 379, + 203, + 407, + 213 + ], + "score": 0.74, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 202, + 505, + 216 + ], + "score": 1.0, + "content": "knows how goals move.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 506, + 255 + ], + "score": 1.0, + "content": "properties as the previous experiments. Overall the results show our method is much better than", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "DDPG and HER. As a reminder, the threshold for calculating rewards we used is 0.01. In Figure 4,", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 263, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 505, + 278 + ], + "score": 1.0, + "content": "the task becomes an easier task because of the slower velocity. It shows that our method quickly", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 506, + 289 + ], + "score": 1.0, + "content": "achieves to a good result around 5 epoch. Comparing with Figure 3a, it shows that the performance", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 286, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 126, + 300 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 126, + 287, + 169, + 297 + ], + "score": 0.89, + "content": "v = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 286, + 308, + 300 + ], + "score": 1.0, + "content": "is better than the performance with", + "type": "text" + }, + { + "bbox": [ + 308, + 287, + 351, + 297 + ], + "score": 0.9, + "content": "v = 0 . 0 1 1", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 286, + 382, + 300 + ], + "score": 1.0, + "content": "and get", + "type": "text" + }, + { + "bbox": [ + 382, + 286, + 402, + 298 + ], + "score": 0.89, + "content": "1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 286, + 505, + 300 + ], + "score": 1.0, + "content": "improvements. However,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 298, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 248, + 309 + ], + "score": 1.0, + "content": "it also shows the performance with", + "type": "text" + }, + { + "bbox": [ + 249, + 298, + 291, + 308 + ], + "score": 0.89, + "content": "v = 0 . 0 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 298, + 435, + 309 + ], + "score": 1.0, + "content": "is worse than the performance with", + "type": "text" + }, + { + "bbox": [ + 436, + 298, + 478, + 308 + ], + "score": 0.89, + "content": "v = 0 . 0 1 1", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 298, + 505, + 309 + ], + "score": 1.0, + "content": ". Both", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 154, + 321 + ], + "score": 1.0, + "content": "DDPG and", + "type": "text" + }, + { + "bbox": [ + 154, + 309, + 208, + 319 + ], + "score": 0.42, + "content": "\\mathrm { \\Delta D D P G + H E R }", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 308, + 260, + 321 + ], + "score": 1.0, + "content": "failed when", + "type": "text" + }, + { + "bbox": [ + 261, + 309, + 306, + 319 + ], + "score": 0.88, + "content": "v = 0 . 0 1 6", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "and their performance is 0. This performance is", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 319, + 425, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 425, + 331 + ], + "score": 1.0, + "content": "consistent because the task becomes more difficult when the velocity increases.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 353, + 364, + 363 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 366, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 366, + 364 + ], + "score": 1.0, + "content": "4.4 COMPARISON ON THE DYNAMIC SNAKE ENVIRONMENT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "In the Dy-Snake environment, which is a discrete control environment, we present the results of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "chasing food task in Figure 6. The results show that the proposed algorithm works best. DQN is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "better than HER. HER fails for this task. That HER fails for this task shows just using achieved", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "goals is not enough for the tasks with dynamic goals and can lead wrong direction. The results", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "score": 1.0, + "content": "also show that around 800 episodes, the performance of DQN and DHER is close and DHER is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "slightly better than DQN. This is because the chasing food task is a simple task. The action space", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "is very small and just 4 types of actions. After enough exploration, DQN also has competitive", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "performance. However, at the beginning, DHER quickly achieves very good performance. It shows", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "that assembling experience from two failures improves the performance very efficiently in this task.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 473, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 104, + 473, + 505, + 488 + ], + "score": 1.0, + "content": "DQN (dense) and DQN (dense-2) help learn the policy at the early stage. However, in the long run,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 485, + 273, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 273, + 498 + ], + "score": 1.0, + "content": "it does not lead to any particular benefits.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 518, + 244, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 246, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 246, + 530 + ], + "score": 1.0, + "content": "4.5 SIM TO REAL ADAPTATION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "We used policies for Dy-Circling task and a new Pouring task trained in our simulator 3 to deploy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "them on a physical robot. As shown in Figure 8, the policies were trained by using DHER and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "adapted to the real robot without any finetuning. However, the policy requires accurate localizations", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "of the gripper and the goal. For Dy-Circling task, the robot’s gripper was blocked. There were a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "toy turntable, whose speed is unknown, and a blue block on the turntable. We set the position 1cm", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "above the block as the target position. The position of the block was predicted based on traditional", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "contour shape analysis using camera images. For Pouring task, the robot gripped a can. A man held", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "a cup and moved it. The cup was set as the target and with a green marker. We used the marker to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 630, + 190, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 190, + 641 + ], + "score": 1.0, + "content": "estimate its position.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 504, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 504, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 479, + 658 + ], + "score": 1.0, + "content": "Our policies were transferred successfully for both tasks. With the accurate positions, we have", + "type": "text" + }, + { + "bbox": [ + 480, + 646, + 504, + 657 + ], + "score": 0.85, + "content": "100 \\%", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 657, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 668 + ], + "score": 1.0, + "content": "success rate for 5 trials. It was observed that the robot had learned to not only follow the current", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "target but also step forward to the future target position. Demo videos about the experiments are", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 679, + 363, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 363, + 691 + ], + "score": 1.0, + "content": "available at https://sites.google.com/view/dher.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 712, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 117, + 708, + 506, + 725 + ], + "spans": [ + { + "bbox": [ + 117, + 708, + 506, + 725 + ], + "score": 1.0, + "content": "3We developed a new simulator according to our hardware. More details on our hardware setup are available", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 720, + 167, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 167, + 734 + ], + "score": 1.0, + "content": "in the appendix.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 135, + 82, + 276, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 135, + 82, + 276, + 192 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 135, + 82, + 276, + 192 + ], + "spans": [ + { + "bbox": [ + 135, + 82, + 276, + 192 + ], + "score": 0.966, + "type": "image", + "image_path": "14dd0b3e1878d3f49fb1b7d44ca84006cca0ac15133dc036aef116910149f401.jpg" + } + ] + } + ], + "index": 1.0, + "virtual_lines": [ + { + "bbox": [ + 135, + 82, + 276, + 137.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 135, + 137.0, + 276, + 192.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 132, + 203, + 281, + 214 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 281, + 217 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 281, + 217 + ], + "score": 1.0, + "content": "Figure 6: Snake with dynamic goals.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 334, + 81, + 474, + 192 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 334, + 81, + 474, + 192 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 334, + 81, + 474, + 192 + ], + "spans": [ + { + "bbox": [ + 334, + 81, + 474, + 192 + ], + "score": 0.967, + "type": "image", + "image_path": "97b49eb4c50a91b591b5c09b7c8417f90a64b3336f8b06a8631418b78f185b01.jpg" + } + ] + } + ], + "index": 2.0, + "virtual_lines": [ + { + "bbox": [ + 334, + 81, + 474, + 136.5 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 334, + 136.5, + 474, + 192.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 308, + 203, + 502, + 214 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 306, + 202, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 202, + 379, + 216 + ], + "score": 1.0, + "content": "Figure 7: Special:", + "type": "text" + }, + { + "bbox": [ + 379, + 203, + 407, + 213 + ], + "score": 0.74, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 202, + 505, + 216 + ], + "score": 1.0, + "content": "knows how goals move.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 107, + 243, + 505, + 330 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 105, + 243, + 506, + 331 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 353, + 364, + 363 + ], + "lines": [ + { + "bbox": [ + 106, + 352, + 366, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 366, + 364 + ], + "score": 1.0, + "content": "4.4 COMPARISON ON THE DYNAMIC SNAKE ENVIRONMENT", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 375, + 505, + 496 + ], + "lines": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 388 + ], + "score": 1.0, + "content": "In the Dy-Snake environment, which is a discrete control environment, we present the results of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 387, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 399 + ], + "score": 1.0, + "content": "chasing food task in Figure 6. The results show that the proposed algorithm works best. DQN is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "better than HER. HER fails for this task. That HER fails for this task shows just using achieved", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 421 + ], + "score": 1.0, + "content": "goals is not enough for the tasks with dynamic goals and can lead wrong direction. The results", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 432 + ], + "score": 1.0, + "content": "also show that around 800 episodes, the performance of DQN and DHER is close and DHER is", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 506, + 443 + ], + "score": 1.0, + "content": "slightly better than DQN. This is because the chasing food task is a simple task. The action space", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 454 + ], + "score": 1.0, + "content": "is very small and just 4 types of actions. After enough exploration, DQN also has competitive", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 452, + 505, + 465 + ], + "score": 1.0, + "content": "performance. However, at the beginning, DHER quickly achieves very good performance. It shows", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 476 + ], + "score": 1.0, + "content": "that assembling experience from two failures improves the performance very efficiently in this task.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 473, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 104, + 473, + 505, + 488 + ], + "score": 1.0, + "content": "DQN (dense) and DQN (dense-2) help learn the policy at the early stage. However, in the long run,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 485, + 273, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 273, + 498 + ], + "score": 1.0, + "content": "it does not lead to any particular benefits.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 375, + 506, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 518, + 244, + 528 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 246, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 246, + 530 + ], + "score": 1.0, + "content": "4.5 SIM TO REAL ADAPTATION", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 541, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "We used policies for Dy-Circling task and a new Pouring task trained in our simulator 3 to deploy", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 565 + ], + "score": 1.0, + "content": "them on a physical robot. As shown in Figure 8, the policies were trained by using DHER and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "adapted to the real robot without any finetuning. However, the policy requires accurate localizations", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 587 + ], + "score": 1.0, + "content": "of the gripper and the goal. For Dy-Circling task, the robot’s gripper was blocked. There were a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 506, + 597 + ], + "score": 1.0, + "content": "toy turntable, whose speed is unknown, and a blue block on the turntable. We set the position 1cm", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 505, + 608 + ], + "score": 1.0, + "content": "above the block as the target position. The position of the block was predicted based on traditional", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 607, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 607, + 506, + 619 + ], + "score": 1.0, + "content": "contour shape analysis using camera images. For Pouring task, the robot gripped a can. A man held", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "spans": [ + { + "bbox": [ + 104, + 617, + 506, + 631 + ], + "score": 1.0, + "content": "a cup and moved it. The cup was set as the target and with a green marker. We used the marker to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 630, + 190, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 190, + 641 + ], + "score": 1.0, + "content": "estimate its position.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 541, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 646, + 504, + 690 + ], + "lines": [ + { + "bbox": [ + 105, + 645, + 504, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 479, + 658 + ], + "score": 1.0, + "content": "Our policies were transferred successfully for both tasks. With the accurate positions, we have", + "type": "text" + }, + { + "bbox": [ + 480, + 646, + 504, + 657 + ], + "score": 0.85, + "content": "100 \\%", + "type": "inline_equation" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 657, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 506, + 668 + ], + "score": 1.0, + "content": "success rate for 5 trials. It was observed that the robot had learned to not only follow the current", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 506, + 680 + ], + "score": 1.0, + "content": "target but also step forward to the future target position. Demo videos about the experiments are", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 679, + 363, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 363, + 691 + ], + "score": 1.0, + "content": "available at https://sites.google.com/view/dher.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 645, + 506, + 691 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 84, + 502, + 153 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 84, + 502, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 84, + 502, + 153 + ], + "spans": [ + { + "bbox": [ + 110, + 84, + 502, + 153 + ], + "score": 0.953, + "type": "image", + "image_path": "ff03243dba37e8e374b610dab429ef50349cfd9bf2cd37e4c19e2d40e16ac98c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 84, + 502, + 107.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 107.0, + 502, + 130.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 130.0, + 502, + 153.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 155, + 501, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 154, + 502, + 169 + ], + "spans": [ + { + "bbox": [ + 108, + 154, + 502, + 169 + ], + "score": 1.0, + "content": "Figure 8: Adapting the policies trained based on DHER from our simulation to a real robotic arm.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 105, + 188, + 482, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 483, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 483, + 200 + ], + "score": 1.0, + "content": "4.6 SPECIAL CASE: HOW ABOUT IF HER KNOWS THE LAW OF THE MOTION OF A GOAL", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "We use Dy-Snake to demonstrate the experimental results for a special case that the law of the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "motion of the target (food) is known to agents. Because the motion of the food is very simple", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 230, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 474, + 241 + ], + "score": 1.0, + "content": "and controllable in Dy-Snake environment. We develop a direct extension of HER, called", + "type": "text" + }, + { + "bbox": [ + 474, + 230, + 501, + 241 + ], + "score": 0.8, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 230, + 505, + 241 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "that modifies desired goals at every timestep based on the law of the motion of the food to create", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 252, + 412, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 262, + 265 + ], + "score": 1.0, + "content": "successful experience. More details of", + "type": "text" + }, + { + "bbox": [ + 262, + 253, + 289, + 262 + ], + "score": 0.63, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 252, + 412, + 265 + ], + "score": 1.0, + "content": "are described in the appendix.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 295, + 282 + ], + "score": 1.0, + "content": "We show the results in Figure 7. DHER and", + "type": "text" + }, + { + "bbox": [ + 296, + 270, + 323, + 280 + ], + "score": 0.72, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 267, + 505, + 282 + ], + "score": 1.0, + "content": "are both better than DQN at the beginning.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 227, + 292 + ], + "score": 1.0, + "content": "DHER is slightly better than", + "type": "text" + }, + { + "bbox": [ + 228, + 280, + 255, + 291 + ], + "score": 0.5, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 279, + 505, + 292 + ], + "score": 1.0, + "content": ", which shows the efficiency of DHER is comparable in this", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 291, + 158, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 158, + 303 + ], + "score": 1.0, + "content": "simple task.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 108, + 318, + 195, + 331 + ], + "lines": [ + { + "bbox": [ + 104, + 317, + 197, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 317, + 197, + 334 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "We introduced a novel technique that assembles successful experience from a couple of failures.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "With this technique, our proposed algorithm called DHER ( Dynamic Hindsight Experience Replay)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "is able to address the tasks with sparse rewards and dynamic goals. Our technique can be combined", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "with an arbitrary off-policy RL algorithm and we experimentally demonstrated that with DQN and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "DDPG. As far as we know, it is the first time that an agent is allowed to learn from assembled", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 398, + 226, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 226, + 410 + ], + "score": 1.0, + "content": "experience from two failures.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 426, + 218, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 220, + 441 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "We would like to thank Weitao Xi, Tianzhou Wang, Tingguang Li for performing some additional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "transfer experiments. We would also like to thank Han Liu and the whole RL team for fruitful", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 471, + 381, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 381, + 484 + ], + "score": 1.0, + "content": "discussions as well as the anonymous reviewers for their comments.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 175, + 512 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 176, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 176, + 513 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 518, + 506, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 528, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 115, + 528, + 505, + 543 + ], + "score": 1.0, + "content": "McGrew, Josh Tobin, OpenAI. Pieter Abbeel, and Wojciech Zaremba. Hindsight experience", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 541, + 461, + 554 + ], + "spans": [ + { + "bbox": [ + 116, + 541, + 461, + 554 + ], + "score": 1.0, + "content": "replay. In Advances in Neural Information Processing Systems, pp. 5048–5058, 2017.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "Marc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning envi-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "ronment: An evaluation platform for general agents. Journal of Artificial Intelligence Research,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 581, + 268, + 592 + ], + "spans": [ + { + "bbox": [ + 116, + 581, + 268, + 592 + ], + "score": 1.0, + "content": "47:253–279, 2013. ISSN 1076-9757.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "Yoshua Bengio, Jrme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 116, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "Proceedings of the 26th International Conference on Machine Learning, pp. 41–48. ACM, 2009.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 117, + 640, + 277, + 653 + ], + "spans": [ + { + "bbox": [ + 117, + 640, + 277, + 653 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym, 2016.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "score": 1.0, + "content": "Maurizio Di Rocco, Federico Pecora, Prasanna Kumar Sivakumar, and Alessandro Saffiotti. Config-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "uration planning with multiple dynamic goals. In AAAI Spring Symposium: Designing Intelligent", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 680, + 174, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 174, + 693 + ], + "score": 1.0, + "content": "Robots, 2013.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "score": 1.0, + "content": "Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep rein-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "forcement learning for continuous control. In Proceedings of the 33rd International Conference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 721, + 297, + 733 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 297, + 733 + ], + "score": 1.0, + "content": "on Machine Learning, pp. 1329–1338, 2016.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 32.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 84, + 502, + 153 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 84, + 502, + 153 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 84, + 502, + 153 + ], + "spans": [ + { + "bbox": [ + 110, + 84, + 502, + 153 + ], + "score": 0.953, + "type": "image", + "image_path": "ff03243dba37e8e374b610dab429ef50349cfd9bf2cd37e4c19e2d40e16ac98c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 110, + 84, + 502, + 107.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 107.0, + 502, + 130.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 130.0, + 502, + 153.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 155, + 501, + 167 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 154, + 502, + 169 + ], + "spans": [ + { + "bbox": [ + 108, + 154, + 502, + 169 + ], + "score": 1.0, + "content": "Figure 8: Adapting the policies trained based on DHER from our simulation to a real robotic arm.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 105, + 188, + 482, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 483, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 483, + 200 + ], + "score": 1.0, + "content": "4.6 SPECIAL CASE: HOW ABOUT IF HER KNOWS THE LAW OF THE MOTION OF A GOAL", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 187, + 483, + 200 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "We use Dy-Snake to demonstrate the experimental results for a special case that the law of the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 219, + 505, + 232 + ], + "score": 1.0, + "content": "motion of the target (food) is known to agents. Because the motion of the food is very simple", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 230, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 474, + 241 + ], + "score": 1.0, + "content": "and controllable in Dy-Snake environment. We develop a direct extension of HER, called", + "type": "text" + }, + { + "bbox": [ + 474, + 230, + 501, + 241 + ], + "score": 0.8, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 230, + 505, + 241 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 505, + 253 + ], + "score": 1.0, + "content": "that modifies desired goals at every timestep based on the law of the motion of the food to create", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 252, + 412, + 265 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 262, + 265 + ], + "score": 1.0, + "content": "successful experience. More details of", + "type": "text" + }, + { + "bbox": [ + 262, + 253, + 289, + 262 + ], + "score": 0.63, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 252, + 412, + 265 + ], + "score": 1.0, + "content": "are described in the appendix.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 208, + 505, + 265 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 504, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 267, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 267, + 295, + 282 + ], + "score": 1.0, + "content": "We show the results in Figure 7. DHER and", + "type": "text" + }, + { + "bbox": [ + 296, + 270, + 323, + 280 + ], + "score": 0.72, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 267, + 505, + 282 + ], + "score": 1.0, + "content": "are both better than DQN at the beginning.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 227, + 292 + ], + "score": 1.0, + "content": "DHER is slightly better than", + "type": "text" + }, + { + "bbox": [ + 228, + 280, + 255, + 291 + ], + "score": 0.5, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 279, + 505, + 292 + ], + "score": 1.0, + "content": ", which shows the efficiency of DHER is comparable in this", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 291, + 158, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 291, + 158, + 303 + ], + "score": 1.0, + "content": "simple task.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 267, + 505, + 303 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 318, + 195, + 331 + ], + "lines": [ + { + "bbox": [ + 104, + 317, + 197, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 317, + 197, + 334 + ], + "score": 1.0, + "content": "5 CONCLUSION", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 356 + ], + "score": 1.0, + "content": "We introduced a novel technique that assembles successful experience from a couple of failures.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "With this technique, our proposed algorithm called DHER ( Dynamic Hindsight Experience Replay)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 378 + ], + "score": 1.0, + "content": "is able to address the tasks with sparse rewards and dynamic goals. Our technique can be combined", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 505, + 388 + ], + "score": 1.0, + "content": "with an arbitrary off-policy RL algorithm and we experimentally demonstrated that with DQN and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 506, + 399 + ], + "score": 1.0, + "content": "DDPG. As far as we know, it is the first time that an agent is allowed to learn from assembled", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 398, + 226, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 226, + 410 + ], + "score": 1.0, + "content": "experience from two failures.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 342, + 506, + 410 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 426, + 218, + 438 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 220, + 441 + ], + "score": 1.0, + "content": "ACKNOWLEDGMENTS", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 450, + 505, + 483 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 505, + 462 + ], + "score": 1.0, + "content": "We would like to thank Weitao Xi, Tianzhou Wang, Tingguang Li for performing some additional", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 505, + 473 + ], + "score": 1.0, + "content": "transfer experiments. We would also like to thank Han Liu and the whole RL team for fruitful", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 471, + 381, + 484 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 381, + 484 + ], + "score": 1.0, + "content": "discussions as well as the anonymous reviewers for their comments.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 450, + 505, + 484 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 500, + 175, + 512 + ], + "lines": [ + { + "bbox": [ + 106, + 501, + 176, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 176, + 513 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "list", + "bbox": [ + 105, + 518, + 506, + 733 + ], + "lines": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "Marcin Andrychowicz, Filip Wolski, Alex Ray, Jonas Schneider, Rachel Fong, Peter Welinder, Bob", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 528, + 505, + 543 + ], + "spans": [ + { + "bbox": [ + 115, + 528, + 505, + 543 + ], + "score": 1.0, + "content": "McGrew, Josh Tobin, OpenAI. Pieter Abbeel, and Wojciech Zaremba. Hindsight experience", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 541, + 461, + 554 + ], + "spans": [ + { + "bbox": [ + 116, + 541, + 461, + 554 + ], + "score": 1.0, + "content": "replay. In Advances in Neural Information Processing Systems, pp. 5048–5058, 2017.", + "type": "text" + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 559, + 505, + 572 + ], + "score": 1.0, + "content": "Marc G. Bellemare, Yavar Naddaf, Joel Veness, and Michael Bowling. The arcade learning envi-", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 570, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 116, + 570, + 505, + 582 + ], + "score": 1.0, + "content": "ronment: An evaluation platform for general agents. Journal of Artificial Intelligence Research,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 581, + 268, + 592 + ], + "spans": [ + { + "bbox": [ + 116, + 581, + 268, + 592 + ], + "score": 1.0, + "content": "47:253–279, 2013. ISSN 1076-9757.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "Yoshua Bengio, Jrme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 116, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "Proceedings of the 26th International Conference on Machine Learning, pp. 41–48. ACM, 2009.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 117, + 640, + 277, + 653 + ], + "spans": [ + { + "bbox": [ + 117, + 640, + 277, + 653 + ], + "score": 1.0, + "content": "Wojciech Zaremba. Openai gym, 2016.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 505, + 672 + ], + "score": 1.0, + "content": "Maurizio Di Rocco, Federico Pecora, Prasanna Kumar Sivakumar, and Alessandro Saffiotti. Config-", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 669, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 116, + 669, + 505, + 683 + ], + "score": 1.0, + "content": "uration planning with multiple dynamic goals. In AAAI Spring Symposium: Designing Intelligent", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 680, + 174, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 174, + 693 + ], + "score": 1.0, + "content": "Robots, 2013.", + "type": "text" + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 712 + ], + "score": 1.0, + "content": "Yan Duan, Xi Chen, Rein Houthooft, John Schulman, and Pieter Abbeel. Benchmarking deep rein-", + "type": "text" + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "forcement learning for continuous control. In Proceedings of the 33rd International Conference", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 117, + 721, + 297, + 733 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 297, + 733 + ], + "score": 1.0, + "content": "on Machine Learning, pp. 1329–1338, 2016.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 519, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Carlos Florensa, David Held, Xinyang Geng, and Pieter Abbeel. Automatic goal generation for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "reinforcement learning agents. In Proceedings of the 35th International Conference on Machine", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 295, + 117 + ], + "score": 1.0, + "content": "Learning, volume 80, pp. 1515–1528, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 123, + 502, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 504, + 136 + ], + "score": 1.0, + "content": "Nicolas Heess, Dhruva TB, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Tom Erez, Ziyu Wang, S. M. Ali Eslami, Martin A. Riedmiller, and David Silver. Emergence of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 146, + 420, + 156 + ], + "spans": [ + { + "bbox": [ + 117, + 146, + 420, + 156 + ], + "score": 1.0, + "content": "locomotion behaviours in rich environments. CoRR, abs/1707.02286, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 503, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Leslie Pack Kaelbling. Learning to achieve goals. In Proceedings of the 13th International Joint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 175, + 393, + 187 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 393, + 187 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence, pp. 1094–1099. Citeseer, 1993.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 504, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "score": 1.0, + "content": "Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 217, + 254, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 254, + 227 + ], + "score": 1.0, + "content": "preprint arXiv:1509.02971, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 502, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 504, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 504, + 249 + ], + "score": 1.0, + "content": "Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 288, + 258 + ], + "score": 1.0, + "content": "Machine Learning, 8(3-4):293–321, 1992.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 107, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 107, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Belle-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wier-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 295, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 114, + 295, + 505, + 312 + ], + "score": 1.0, + "content": "stra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 307, + 211, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 307, + 211, + 321 + ], + "score": 1.0, + "content": "Nature, 518:529, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 105, + 327, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "Sanmit Narvekar, Jivko Sinapov, and Peter Stone. Autonomous task sequencing for customized", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "curriculum design in reinforcement learning. In Proceedings of the 27th International Joint Con-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 349, + 362, + 362 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 362, + 362 + ], + "score": 1.0, + "content": "ference on Artificial Intelligence, volume 147, pp. 149, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "Andrew Y. Ng, Daishi Harada, and Stuart Russell. Policy invariance under reward transformations:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 117, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "Theory and application to reward shaping. In Proceedings of the 16th International Conference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 390, + 334, + 402 + ], + "spans": [ + { + "bbox": [ + 116, + 390, + 334, + 402 + ], + "score": 1.0, + "content": "on Machine Learning, volume 99, pp. 278–287, 1999.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "Tom Schaul, Daniel Horgan, Karol Gregor, and David Silver. Universal value function approxima-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 116, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "tors. In Proceedings of the 32nd International Conference on Machine Learning, pp. 1312–1320,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 429, + 147, + 444 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 147, + 444 + ], + "score": 1.0, + "content": "2015a.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 105, + 449, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 461, + 259, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 259, + 473 + ], + "score": 1.0, + "content": "preprint arXiv:1511.05952, 2015b.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 105, + 479, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel. Trust region", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 491, + 322, + 503 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 322, + 503 + ], + "score": 1.0, + "content": "policy optimization. CoRR, abs/1502.05477, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 105, + 509, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 524 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 521, + 340, + 533 + ], + "spans": [ + { + "bbox": [ + 116, + 521, + 340, + 533 + ], + "score": 1.0, + "content": "optimization algorithms. CoRR, abs/1707.06347, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 108, + 539, + 504, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 562, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 562, + 473, + 574 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 580, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 594 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 603, + 168, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 603, + 168, + 614 + ], + "score": 1.0, + "content": "IEEE, 2012.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Ziyu Wang, Victor Bapst, Nicolas Heess, Volodymyr Mnih, Remi Munos, Koray Kavukcuoglu, and ´", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 114, + 631, + 506, + 646 + ], + "score": 1.0, + "content": "Nando de Freitas. Sample efficient actor-critic with experience replay. CoRR, abs/1611.01224,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 641, + 143, + 656 + ], + "spans": [ + { + "bbox": [ + 115, + 641, + 143, + 656 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 108, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Carlos Florensa, David Held, Xinyang Geng, and Pieter Abbeel. Automatic goal generation for", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "reinforcement learning agents. In Proceedings of the 35th International Conference on Machine", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 295, + 117 + ], + "score": 1.0, + "content": "Learning, volume 80, pp. 1515–1528, 2018.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 106, + 82, + 505, + 117 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 123, + 502, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 504, + 136 + ], + "score": 1.0, + "content": "Nicolas Heess, Dhruva TB, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 147 + ], + "score": 1.0, + "content": "Tom Erez, Ziyu Wang, S. M. Ali Eslami, Martin A. Riedmiller, and David Silver. Emergence of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 146, + 420, + 156 + ], + "spans": [ + { + "bbox": [ + 117, + 146, + 420, + 156 + ], + "score": 1.0, + "content": "locomotion behaviours in rich environments. CoRR, abs/1707.02286, 2017.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 122, + 505, + 156 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 164, + 503, + 187 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Leslie Pack Kaelbling. Learning to achieve goals. In Proceedings of the 13th International Joint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 175, + 393, + 187 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 393, + 187 + ], + "score": 1.0, + "content": "Conference on Artificial Intelligence, pp. 1094–1099. Citeseer, 1993.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 164, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 504, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 207 + ], + "score": 1.0, + "content": "Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 217 + ], + "score": 1.0, + "content": "David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 217, + 254, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 254, + 227 + ], + "score": 1.0, + "content": "preprint arXiv:1509.02971, 2015.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 192, + 505, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 234, + 502, + 258 + ], + "lines": [ + { + "bbox": [ + 105, + 233, + 504, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 233, + 504, + 249 + ], + "score": 1.0, + "content": "Long-Ji Lin. Self-improving reactive agents based on reinforcement learning, planning and teaching.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 288, + 258 + ], + "score": 1.0, + "content": "Machine Learning, 8(3-4):293–321, 1992.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 233, + 504, + 258 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 264, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 107, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 107, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Belle-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 275, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 275, + 505, + 289 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 285, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 115, + 285, + 505, + 299 + ], + "score": 1.0, + "content": "Charles Beattie, Amir Sadik, Ioannis Antonoglou, Helen King, Dharshan Kumaran, Daan Wier-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 114, + 295, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 114, + 295, + 505, + 312 + ], + "score": 1.0, + "content": "stra, Shane Legg, and Demis Hassabis. Human-level control through deep reinforcement learning.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 307, + 211, + 321 + ], + "spans": [ + { + "bbox": [ + 115, + 307, + 211, + 321 + ], + "score": 1.0, + "content": "Nature, 518:529, 2015.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 107, + 265, + 505, + 321 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 327, + 505, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 340 + ], + "score": 1.0, + "content": "Sanmit Narvekar, Jivko Sinapov, and Peter Stone. Autonomous task sequencing for customized", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 338, + 505, + 351 + ], + "spans": [ + { + "bbox": [ + 115, + 338, + 505, + 351 + ], + "score": 1.0, + "content": "curriculum design in reinforcement learning. In Proceedings of the 27th International Joint Con-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 349, + 362, + 362 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 362, + 362 + ], + "score": 1.0, + "content": "ference on Artificial Intelligence, volume 147, pp. 149, 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 106, + 327, + 505, + 362 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 380 + ], + "score": 1.0, + "content": "Andrew Y. Ng, Daishi Harada, and Stuart Russell. Policy invariance under reward transformations:", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 117, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "Theory and application to reward shaping. In Proceedings of the 16th International Conference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 116, + 390, + 334, + 402 + ], + "spans": [ + { + "bbox": [ + 116, + 390, + 334, + 402 + ], + "score": 1.0, + "content": "on Machine Learning, volume 99, pp. 278–287, 1999.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 368, + 505, + 402 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 442 + ], + "lines": [ + { + "bbox": [ + 106, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "Tom Schaul, Daniel Horgan, Karol Gregor, and David Silver. Universal value function approxima-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 420, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 116, + 420, + 505, + 432 + ], + "score": 1.0, + "content": "tors. In Proceedings of the 32nd International Conference on Machine Learning, pp. 1312–1320,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 429, + 147, + 444 + ], + "spans": [ + { + "bbox": [ + 115, + 429, + 147, + 444 + ], + "score": 1.0, + "content": "2015a.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 408, + 505, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 449, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 461, + 259, + 473 + ], + "spans": [ + { + "bbox": [ + 115, + 461, + 259, + 473 + ], + "score": 1.0, + "content": "preprint arXiv:1511.05952, 2015b.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 448, + 505, + 473 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 479, + 505, + 503 + ], + "lines": [ + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 505, + 493 + ], + "score": 1.0, + "content": "John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel. Trust region", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 491, + 322, + 503 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 322, + 503 + ], + "score": 1.0, + "content": "policy optimization. CoRR, abs/1502.05477, 2015.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 104, + 478, + 505, + 503 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 509, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 507, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 524 + ], + "score": 1.0, + "content": "John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 521, + 340, + 533 + ], + "spans": [ + { + "bbox": [ + 116, + 521, + 340, + 533 + ], + "score": 1.0, + "content": "optimization algorithms. CoRR, abs/1707.06347, 2017.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 507, + 505, + 533 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 539, + 504, + 573 + ], + "lines": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 506, + 552 + ], + "score": 1.0, + "content": "David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Mastering", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 562, + 473, + 574 + ], + "spans": [ + { + "bbox": [ + 116, + 562, + 473, + 574 + ], + "score": 1.0, + "content": "the game of go with deep neural networks and tree search. Nature, 529(7587):484, 2016.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 539, + 506, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 580, + 505, + 614 + ], + "lines": [ + { + "bbox": [ + 105, + 579, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 594 + ], + "score": 1.0, + "content": "Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A physics engine for model-based control.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 115, + 591, + 505, + 604 + ], + "score": 1.0, + "content": "In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026–5033.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 603, + 168, + 614 + ], + "spans": [ + { + "bbox": [ + 115, + 603, + 168, + 614 + ], + "score": 1.0, + "content": "IEEE, 2012.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 579, + 505, + 614 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 634 + ], + "score": 1.0, + "content": "Ziyu Wang, Victor Bapst, Nicolas Heess, Volodymyr Mnih, Remi Munos, Koray Kavukcuoglu, and ´", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 114, + 631, + 506, + 646 + ], + "score": 1.0, + "content": "Nando de Freitas. Sample efficient actor-critic with experience replay. CoRR, abs/1611.01224,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 641, + 143, + 656 + ], + "spans": [ + { + "bbox": [ + 115, + 641, + 143, + 656 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 621, + 506, + 656 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 161, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 107, + 236, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 236, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 236, + 119 + ], + "score": 1.0, + "content": "TASKS WITH DYNAMIC GOALS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 128, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "We have created four tasks as described below and illustrated by Figure 2. The first three tasks are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 470, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 470, + 151 + ], + "score": 1.0, + "content": "based on the robotic environments. The last is based on the GREEDY SNAKE environment.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 132, + 160, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 133, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "• Dy-Reaching: The task is to control the robotic arm so that its gripper can reach the target", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 142, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 142, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "position. The target position moves from one point to another along a straight line with a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 183, + 214, + 195 + ], + "spans": [ + { + "bbox": [ + 141, + 183, + 214, + 195 + ], + "score": 1.0, + "content": "constant velocity.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "• Dy-Circling: The task is to control the robotic arm so that its gripper can reach the target", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 210, + 441, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 210, + 441, + 222 + ], + "score": 1.0, + "content": "position. The target position moves along a circle with a constant velocity.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 140, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "Dy-Pushing: in this task, a box is placed on a table in front of the robotic arm. The robot is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "required to move the box to the target location on the table. The target location moves from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "one position to another with a fixed velocity. Note that the robot’s grippers are locked to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 140, + 257, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 140, + 257, + 506, + 273 + ], + "score": 1.0, + "content": "prevent it from grasping. The learned behavior is actually a mixture of pushing and rolling.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "Dy-Snake: in this task, a snake and food are placed on a rectangle map and the task is to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "control the snake to eat the food. The food is moving from one position to another position", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 295, + 228, + 310 + ], + "spans": [ + { + "bbox": [ + 142, + 295, + 228, + 310 + ], + "score": 1.0, + "content": "with a fixed velocity.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 107, + 323, + 158, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 323, + 159, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 159, + 335 + ], + "score": 1.0, + "content": "HARDWARE", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 344, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 506, + 357 + ], + "score": 1.0, + "content": "We use the Universal Robots UR10 with a gripper. We use a RealSense Camera SR300 to track the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "score": 1.0, + "content": "position of objects. The gripper is blocked. We use a marker on the gripper for camera calibration.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 366, + 464, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 464, + 379 + ], + "score": 1.0, + "content": "During adapting the robot, for the position of the target, we use the camera to estimate it.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 162, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 164, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 164, + 405 + ], + "score": 1.0, + "content": "SIMULATION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 414, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "We simulate the physical system using the MuJoCo physics engine (Todorov et al., 2012) and also", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "score": 1.0, + "content": "use MuJoCo to render the images. In our tasks, we use positions provided by MuJoCo for training", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 435, + 212, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 212, + 447 + ], + "score": 1.0, + "content": "policies in the simulation.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 399, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 400, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 400, + 466 + ], + "score": 1.0, + "content": "In Figure 9, we demonstrate our simulation and the physics environment.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 132, + 479, + 479, + 634 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 132, + 479, + 479, + 634 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 479, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 479, + 634 + ], + "score": 0.977, + "type": "image", + "image_path": "0ab0a04331e1ed7ee05503beb83d72a71ddd08962ea6776f4dbc386edf14fc2c.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 132, + 479, + 479, + 530.6666666666666 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 132, + 530.6666666666666, + 479, + 582.3333333333333 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 132, + 582.3333333333333, + 479, + 633.9999999999999 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 195, + 648, + 415, + 659 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 194, + 646, + 415, + 661 + ], + "spans": [ + { + "bbox": [ + 194, + 646, + 415, + 661 + ], + "score": 1.0, + "content": "Figure 9: Our simulation and the physics environment.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + } + ], + "index": 27.0 + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 279, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 280, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 280, + 701 + ], + "score": 1.0, + "content": "A SPECIAL AND SIMPLER CASE - HER+", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "The paper considers a general situation that the law of the motion of goals is invisible to an agent.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "However, we relax this assumption and assume that in some situations the law of the motion of goals", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 161, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 163, + 96 + ], + "score": 1.0, + "content": "APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 107, + 236, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 107, + 236, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 107, + 236, + 119 + ], + "score": 1.0, + "content": "TASKS WITH DYNAMIC GOALS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 128, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 505, + 140 + ], + "score": 1.0, + "content": "We have created four tasks as described below and illustrated by Figure 2. The first three tasks are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 138, + 470, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 470, + 151 + ], + "score": 1.0, + "content": "based on the robotic environments. The last is based on the GREEDY SNAKE environment.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 127, + 505, + 151 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 160, + 505, + 308 + ], + "lines": [ + { + "bbox": [ + 133, + 160, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 505, + 173 + ], + "score": 1.0, + "content": "• Dy-Reaching: The task is to control the robotic arm so that its gripper can reach the target", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 142, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "position. The target position moves from one point to another along a straight line with a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 183, + 214, + 195 + ], + "spans": [ + { + "bbox": [ + 141, + 183, + 214, + 195 + ], + "score": 1.0, + "content": "constant velocity.", + "type": "text" + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 132, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 132, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "• Dy-Circling: The task is to control the robotic arm so that its gripper can reach the target", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 210, + 441, + 222 + ], + "spans": [ + { + "bbox": [ + 141, + 210, + 441, + 222 + ], + "score": 1.0, + "content": "position. The target position moves along a circle with a constant velocity.", + "type": "text" + } + ], + "index": 8, + "is_list_end_line": true + }, + { + "bbox": [ + 140, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 140, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "Dy-Pushing: in this task, a box is placed on a table in front of the robotic arm. The robot is", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 237, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 237, + 505, + 249 + ], + "score": 1.0, + "content": "required to move the box to the target location on the table. The target location moves from", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "one position to another with a fixed velocity. Note that the robot’s grippers are locked to", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 140, + 257, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 140, + 257, + 506, + 273 + ], + "score": 1.0, + "content": "prevent it from grasping. The learned behavior is actually a mixture of pushing and rolling.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 138, + 274, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 138, + 274, + 505, + 286 + ], + "score": 1.0, + "content": "Dy-Snake: in this task, a snake and food are placed on a rectangle map and the task is to", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 142, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 142, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "control the snake to eat the food. The food is moving from one position to another position", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 142, + 295, + 228, + 310 + ], + "spans": [ + { + "bbox": [ + 142, + 295, + 228, + 310 + ], + "score": 1.0, + "content": "with a fixed velocity.", + "type": "text" + } + ], + "index": 15, + "is_list_end_line": true + } + ], + "index": 9.5, + "bbox_fs": [ + 132, + 160, + 506, + 310 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 323, + 158, + 334 + ], + "lines": [ + { + "bbox": [ + 106, + 323, + 159, + 335 + ], + "spans": [ + { + "bbox": [ + 106, + 323, + 159, + 335 + ], + "score": 1.0, + "content": "HARDWARE", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 108, + 344, + 505, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 506, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 506, + 357 + ], + "score": 1.0, + "content": "We use the Universal Robots UR10 with a gripper. We use a RealSense Camera SR300 to track the", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 367 + ], + "score": 1.0, + "content": "position of objects. The gripper is blocked. We use a marker on the gripper for camera calibration.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 366, + 464, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 464, + 379 + ], + "score": 1.0, + "content": "During adapting the robot, for the position of the target, we use the camera to estimate it.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 343, + 506, + 379 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 393, + 162, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 392, + 164, + 405 + ], + "spans": [ + { + "bbox": [ + 106, + 392, + 164, + 405 + ], + "score": 1.0, + "content": "SIMULATION", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 108, + 414, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 413, + 505, + 426 + ], + "score": 1.0, + "content": "We simulate the physical system using the MuJoCo physics engine (Todorov et al., 2012) and also", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 505, + 438 + ], + "score": 1.0, + "content": "use MuJoCo to render the images. In our tasks, we use positions provided by MuJoCo for training", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 435, + 212, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 212, + 447 + ], + "score": 1.0, + "content": "policies in the simulation.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 413, + 505, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 399, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 451, + 400, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 400, + 466 + ], + "score": 1.0, + "content": "In Figure 9, we demonstrate our simulation and the physics environment.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 451, + 400, + 466 + ] + }, + { + "type": "image", + "bbox": [ + 132, + 479, + 479, + 634 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 132, + 479, + 479, + 634 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 132, + 479, + 479, + 634 + ], + "spans": [ + { + "bbox": [ + 132, + 479, + 479, + 634 + ], + "score": 0.977, + "type": "image", + "image_path": "0ab0a04331e1ed7ee05503beb83d72a71ddd08962ea6776f4dbc386edf14fc2c.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 132, + 479, + 479, + 530.6666666666666 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 132, + 530.6666666666666, + 479, + 582.3333333333333 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 132, + 582.3333333333333, + 479, + 633.9999999999999 + ], + "spans": [], + "index": 27 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 195, + 648, + 415, + 659 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 194, + 646, + 415, + 661 + ], + "spans": [ + { + "bbox": [ + 194, + 646, + 415, + 661 + ], + "score": 1.0, + "content": "Figure 9: Our simulation and the physics environment.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 28 + } + ], + "index": 27.0 + }, + { + "type": "text", + "bbox": [ + 108, + 688, + 279, + 699 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 280, + 701 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 280, + 701 + ], + "score": 1.0, + "content": "A SPECIAL AND SIMPLER CASE - HER+", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 688, + 280, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "The paper considers a general situation that the law of the motion of goals is invisible to an agent.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "However, we relax this assumption and assume that in some situations the law of the motion of goals", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 268, + 95 + ], + "score": 1.0, + "content": "is easy to obtain, i.e., the parameters of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 269, + 82, + 297, + 95 + ], + "score": 0.93, + "content": "g ( t ; \\gamma )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 297, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "are known. For example, it is able to calculate the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 361, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 361, + 106 + ], + "score": 1.0, + "content": "velocity of an object by observing the trajectories of the object.", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 268, + 95 + ], + "score": 1.0, + "content": "is easy to obtain, i.e., the parameters of", + "type": "text" + }, + { + "bbox": [ + 269, + 82, + 297, + 95 + ], + "score": 0.93, + "content": "g ( t ; \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 82, + 504, + 95 + ], + "score": 1.0, + "content": "are known. For example, it is able to calculate the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 93, + 361, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 361, + 106 + ], + "score": 1.0, + "content": "velocity of an object by observing the trajectories of the object.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 123 + ], + "score": 1.0, + "content": "In this situation, we do not need to search failed experience. With the knowledge of the velocity of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "an object, it is straightforward that we can calculate any trajectory of desired goals. Thus for every", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 131, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 244, + 146 + ], + "score": 1.0, + "content": "failed experience, at any time step", + "type": "text" + }, + { + "bbox": [ + 244, + 133, + 249, + 142 + ], + "score": 0.59, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 131, + 361, + 146 + ], + "score": 1.0, + "content": ", based on the achieved goal", + "type": "text" + }, + { + "bbox": [ + 362, + 133, + 376, + 144 + ], + "score": 0.9, + "content": "g _ { t } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 131, + 506, + 146 + ], + "score": 1.0, + "content": ", we calculate new desired goals", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 473, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 151, + 156 + ], + "score": 0.93, + "content": "g _ { t } ^ { d e } = g ( { \\dot { t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 142, + 473, + 158 + ], + "score": 1.0, + "content": "corresponding to the achieved goals in order to construct successful experience.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 108, + 160, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 107, + 160, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 167, + 172 + ], + "score": 1.0, + "content": "The details of", + "type": "text" + }, + { + "bbox": [ + 167, + 160, + 195, + 171 + ], + "score": 0.84, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 160, + 324, + 172 + ], + "score": 1.0, + "content": "are described in Algorithm 2.", + "type": "text" + }, + { + "bbox": [ + 325, + 160, + 352, + 171 + ], + "score": 0.83, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 160, + 505, + 172 + ], + "score": 1.0, + "content": "can be seen as a direct extension of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 170, + 184 + ], + "score": 1.0, + "content": "HER. HER and", + "type": "text" + }, + { + "bbox": [ + 171, + 172, + 198, + 182 + ], + "score": 0.79, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "can modify any failed experience to successful experience because they both", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 302, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 302, + 195 + ], + "score": 1.0, + "content": "assume the law of the motion of goals is known.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "title", + "bbox": [ + 108, + 206, + 300, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 303, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 303, + 220 + ], + "score": 1.0, + "content": "Algorithm 2 Hindsight Experience Replay Plus", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 222, + 469, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 416, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 325, + 234 + ], + "score": 1.0, + "content": "Require: an off-policy RL algorithm A, replay buffer", + "type": "text" + }, + { + "bbox": [ + 325, + 222, + 334, + 231 + ], + "score": 0.74, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 221, + 410, + 234 + ], + "score": 1.0, + "content": ", a reward function", + "type": "text" + }, + { + "bbox": [ + 410, + 224, + 416, + 231 + ], + "score": 0.41, + "content": "r", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 111, + 232, + 251, + 243 + ], + "spans": [ + { + "bbox": [ + 111, + 232, + 241, + 243 + ], + "score": 1.0, + "content": "1: Initialize A and replay buffer", + "type": "text" + }, + { + "bbox": [ + 242, + 233, + 251, + 242 + ], + "score": 0.75, + "content": "R", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 110, + 242, + 244, + 255 + ], + "spans": [ + { + "bbox": [ + 110, + 242, + 170, + 255 + ], + "score": 1.0, + "content": "2: for episode", + "type": "text" + }, + { + "bbox": [ + 171, + 243, + 230, + 254 + ], + "score": 0.82, + "content": "\\mathbf { \\Omega } = 1 , 2 , \\cdots , M", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 242, + 244, + 255 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 110, + 253, + 320, + 267 + ], + "spans": [ + { + "bbox": [ + 110, + 254, + 122, + 266 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 132, + 253, + 223, + 267 + ], + "score": 1.0, + "content": "Sample an initial goal", + "type": "text" + }, + { + "bbox": [ + 223, + 256, + 234, + 266 + ], + "score": 0.84, + "content": "g _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 253, + 310, + 267 + ], + "score": 1.0, + "content": "and an initial state", + "type": "text" + }, + { + "bbox": [ + 310, + 257, + 320, + 265 + ], + "score": 0.81, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 13 + }, + { + "bbox": [ + 110, + 264, + 235, + 277 + ], + "spans": [ + { + "bbox": [ + 110, + 265, + 122, + 277 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 132, + 264, + 148, + 276 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 266, + 219, + 276 + ], + "score": 0.92, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 264, + 235, + 276 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 110, + 276, + 370, + 289 + ], + "spans": [ + { + "bbox": [ + 110, + 276, + 122, + 288 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 142, + 276, + 214, + 289 + ], + "score": 1.0, + "content": "Sample an action", + "type": "text" + }, + { + "bbox": [ + 214, + 278, + 224, + 287 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 276, + 370, + 289 + ], + "score": 1.0, + "content": "using the behavioral policy from A:", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 110, + 287, + 202, + 299 + ], + "spans": [ + { + "bbox": [ + 110, + 288, + 121, + 298 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 144, + 287, + 202, + 299 + ], + "score": 0.74, + "content": "a _ { t } \\doteq \\pi ( s _ { t } | g _ { t } )", + "type": "inline_equation", + "image_path": "5564f863a11df421373d7a352ee560dd043f7174cc986db89f6acbd09a84f667.jpg" + } + ], + "index": 16 + }, + { + "bbox": [ + 110, + 296, + 430, + 312 + ], + "spans": [ + { + "bbox": [ + 110, + 298, + 122, + 310 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 141, + 296, + 219, + 312 + ], + "score": 1.0, + "content": "Execute the action", + "type": "text" + }, + { + "bbox": [ + 219, + 300, + 229, + 309 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 296, + 327, + 312 + ], + "score": 1.0, + "content": "and observe a new state", + "type": "text" + }, + { + "bbox": [ + 327, + 300, + 347, + 310 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 296, + 410, + 312 + ], + "score": 1.0, + "content": "and a new goal", + "type": "text" + }, + { + "bbox": [ + 411, + 300, + 430, + 310 + ], + "score": 0.85, + "content": "g _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 110, + 308, + 168, + 321 + ], + "spans": [ + { + "bbox": [ + 110, + 309, + 121, + 320 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 132, + 308, + 168, + 321 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 110, + 320, + 234, + 332 + ], + "spans": [ + { + "bbox": [ + 110, + 320, + 122, + 332 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 132, + 320, + 148, + 332 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 321, + 220, + 331 + ], + "score": 0.85, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 320, + 234, + 332 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 107, + 331, + 225, + 343 + ], + "spans": [ + { + "bbox": [ + 107, + 331, + 122, + 342 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 143, + 331, + 225, + 343 + ], + "score": 0.84, + "content": "r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 340, + 468, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 122, + 353 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 142, + 340, + 226, + 355 + ], + "score": 1.0, + "content": "Store the transitions", + "type": "text" + }, + { + "bbox": [ + 226, + 342, + 323, + 354 + ], + "score": 0.87, + "content": "\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 340, + 335, + 355 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 335, + 342, + 344, + 352 + ], + "score": 0.8, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 340, + 468, + 355 + ], + "score": 1.0, + "content": "(Standard experience replay)", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 432, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 122, + 364 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 143, + 352, + 294, + 365 + ], + "score": 1.0, + "content": "Sample a set of the achieved goals of", + "type": "text" + }, + { + "bbox": [ + 294, + 354, + 305, + 364 + ], + "score": 0.88, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 352, + 420, + 365 + ], + "score": 1.0, + "content": "as additional goals for reply", + "type": "text" + }, + { + "bbox": [ + 421, + 353, + 432, + 363 + ], + "score": 0.85, + "content": "G ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 362, + 207, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 123, + 376 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 142, + 363, + 158, + 376 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 158, + 363, + 191, + 377 + ], + "score": 0.92, + "content": "\\bar { g _ { p } ^ { \\prime } } \\in G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 363, + 207, + 376 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 374, + 431, + 389 + ], + "spans": [ + { + "bbox": [ + 107, + 376, + 122, + 388 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 152, + 374, + 261, + 389 + ], + "score": 1.0, + "content": "Calculate a goal trajectory", + "type": "text" + }, + { + "bbox": [ + 261, + 376, + 314, + 389 + ], + "score": 0.93, + "content": "\\{ g _ { 0 } ^ { \\prime } , \\cdots , g _ { p } ^ { \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 374, + 342, + 389 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 343, + 376, + 392, + 388 + ], + "score": 0.56, + "content": "g _ { t } ^ { \\prime } = g ( t ; \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 376, + 431, + 388 + ], + "score": 0.4, + "content": "\\mathrm { ( H E R + ) }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 387, + 263, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 122, + 400 + ], + "score": 1.0, + "content": "15:", + "type": "text" + }, + { + "bbox": [ + 151, + 387, + 168, + 402 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 168, + 388, + 247, + 400 + ], + "score": 0.87, + "content": "t = \\{ 0 , \\cdots , p - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 387, + 263, + 402 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 400, + 256, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 121, + 411 + ], + "score": 1.0, + "content": "16:", + "type": "text" + }, + { + "bbox": [ + 163, + 400, + 256, + 412 + ], + "score": 0.83, + "content": "r _ { t } ^ { \\prime } : = r ( s _ { i , t } , a _ { i , t } , g _ { t + 1 } ^ { \\prime } )", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 409, + 374, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 122, + 421 + ], + "score": 1.0, + "content": "17:", + "type": "text" + }, + { + "bbox": [ + 162, + 409, + 242, + 423 + ], + "score": 1.0, + "content": "Store the transition", + "type": "text" + }, + { + "bbox": [ + 242, + 410, + 374, + 423 + ], + "score": 0.82, + "content": "\\left( s _ { i , t } | g _ { t } ^ { \\prime } , a _ { i , t } , r _ { t } ^ { \\prime } , s _ { i , t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R", + "type": "inline_equation" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 421, + 187, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 122, + 433 + ], + "score": 1.0, + "content": "18:", + "type": "text" + }, + { + "bbox": [ + 152, + 421, + 187, + 433 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 432, + 178, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 122, + 444 + ], + "score": 1.0, + "content": "19:", + "type": "text" + }, + { + "bbox": [ + 142, + 432, + 178, + 444 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 442, + 167, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 122, + 455 + ], + "score": 1.0, + "content": "20:", + "type": "text" + }, + { + "bbox": [ + 132, + 442, + 167, + 455 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 454, + 157, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 157, + 466 + ], + "score": 1.0, + "content": "21: end for", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 20.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2019", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 504, + 105 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 82, + 504, + 106 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 110, + 505, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 111, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 506, + 123 + ], + "score": 1.0, + "content": "In this situation, we do not need to search failed experience. With the knowledge of the velocity of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "an object, it is straightforward that we can calculate any trajectory of desired goals. Thus for every", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 131, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 244, + 146 + ], + "score": 1.0, + "content": "failed experience, at any time step", + "type": "text" + }, + { + "bbox": [ + 244, + 133, + 249, + 142 + ], + "score": 0.59, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 131, + 361, + 146 + ], + "score": 1.0, + "content": ", based on the achieved goal", + "type": "text" + }, + { + "bbox": [ + 362, + 133, + 376, + 144 + ], + "score": 0.9, + "content": "g _ { t } ^ { a c }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 131, + 506, + 146 + ], + "score": 1.0, + "content": ", we calculate new desired goals", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 142, + 473, + 158 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 151, + 156 + ], + "score": 0.93, + "content": "g _ { t } ^ { d e } = g ( { \\dot { t } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 142, + 473, + 158 + ], + "score": 1.0, + "content": "corresponding to the achieved goals in order to construct successful experience.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 104, + 111, + 506, + 158 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 160, + 503, + 194 + ], + "lines": [ + { + "bbox": [ + 107, + 160, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 167, + 172 + ], + "score": 1.0, + "content": "The details of", + "type": "text" + }, + { + "bbox": [ + 167, + 160, + 195, + 171 + ], + "score": 0.84, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 160, + 324, + 172 + ], + "score": 1.0, + "content": "are described in Algorithm 2.", + "type": "text" + }, + { + "bbox": [ + 325, + 160, + 352, + 171 + ], + "score": 0.83, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 160, + 505, + 172 + ], + "score": 1.0, + "content": "can be seen as a direct extension of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 171, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 170, + 184 + ], + "score": 1.0, + "content": "HER. HER and", + "type": "text" + }, + { + "bbox": [ + 171, + 172, + 198, + 182 + ], + "score": 0.79, + "content": "\\mathrm { H E R + }", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 171, + 505, + 184 + ], + "score": 1.0, + "content": "can modify any failed experience to successful experience because they both", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 302, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 302, + 195 + ], + "score": 1.0, + "content": "assume the law of the motion of goals is known.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 106, + 160, + 505, + 195 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 206, + 300, + 218 + ], + "lines": [ + { + "bbox": [ + 106, + 204, + 303, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 303, + 220 + ], + "score": 1.0, + "content": "Algorithm 2 Hindsight Experience Replay Plus", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "index", + "bbox": [ + 108, + 222, + 469, + 467 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 416, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 325, + 234 + ], + "score": 1.0, + "content": "Require: an off-policy RL algorithm A, replay buffer", + "type": "text" + }, + { + "bbox": [ + 325, + 222, + 334, + 231 + ], + "score": 0.74, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 221, + 410, + 234 + ], + "score": 1.0, + "content": ", a reward function", + "type": "text" + }, + { + "bbox": [ + 410, + 224, + 416, + 231 + ], + "score": 0.41, + "content": "r", + "type": "inline_equation" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 232, + 251, + 243 + ], + "spans": [ + { + "bbox": [ + 111, + 232, + 241, + 243 + ], + "score": 1.0, + "content": "1: Initialize A and replay buffer", + "type": "text" + }, + { + "bbox": [ + 242, + 233, + 251, + 242 + ], + "score": 0.75, + "content": "R", + "type": "inline_equation" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 242, + 244, + 255 + ], + "spans": [ + { + "bbox": [ + 110, + 242, + 170, + 255 + ], + "score": 1.0, + "content": "2: for episode", + "type": "text" + }, + { + "bbox": [ + 171, + 243, + 230, + 254 + ], + "score": 0.82, + "content": "\\mathbf { \\Omega } = 1 , 2 , \\cdots , M", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 242, + 244, + 255 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 253, + 320, + 267 + ], + "spans": [ + { + "bbox": [ + 110, + 254, + 122, + 266 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 132, + 253, + 223, + 267 + ], + "score": 1.0, + "content": "Sample an initial goal", + "type": "text" + }, + { + "bbox": [ + 223, + 256, + 234, + 266 + ], + "score": 0.84, + "content": "g _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 234, + 253, + 310, + 267 + ], + "score": 1.0, + "content": "and an initial state", + "type": "text" + }, + { + "bbox": [ + 310, + 257, + 320, + 265 + ], + "score": 0.81, + "content": "s _ { 0 }", + "type": "inline_equation" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 264, + 235, + 277 + ], + "spans": [ + { + "bbox": [ + 110, + 265, + 122, + 277 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 132, + 264, + 148, + 276 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 266, + 219, + 276 + ], + "score": 0.92, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 264, + 235, + 276 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 276, + 370, + 289 + ], + "spans": [ + { + "bbox": [ + 110, + 276, + 122, + 288 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 142, + 276, + 214, + 289 + ], + "score": 1.0, + "content": "Sample an action", + "type": "text" + }, + { + "bbox": [ + 214, + 278, + 224, + 287 + ], + "score": 0.83, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 225, + 276, + 370, + 289 + ], + "score": 1.0, + "content": "using the behavioral policy from A:", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 287, + 202, + 299 + ], + "spans": [ + { + "bbox": [ + 110, + 288, + 121, + 298 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 144, + 287, + 202, + 299 + ], + "score": 0.74, + "content": "a _ { t } \\doteq \\pi ( s _ { t } | g _ { t } )", + "type": "inline_equation", + "image_path": "5564f863a11df421373d7a352ee560dd043f7174cc986db89f6acbd09a84f667.jpg" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 296, + 430, + 312 + ], + "spans": [ + { + "bbox": [ + 110, + 298, + 122, + 310 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 141, + 296, + 219, + 312 + ], + "score": 1.0, + "content": "Execute the action", + "type": "text" + }, + { + "bbox": [ + 219, + 300, + 229, + 309 + ], + "score": 0.84, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 296, + 327, + 312 + ], + "score": 1.0, + "content": "and observe a new state", + "type": "text" + }, + { + "bbox": [ + 327, + 300, + 347, + 310 + ], + "score": 0.89, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 296, + 410, + 312 + ], + "score": 1.0, + "content": "and a new goal", + "type": "text" + }, + { + "bbox": [ + 411, + 300, + 430, + 310 + ], + "score": 0.85, + "content": "g _ { t + 1 }", + "type": "inline_equation" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 308, + 168, + 321 + ], + "spans": [ + { + "bbox": [ + 110, + 309, + 121, + 320 + ], + "score": 1.0, + "content": "8:", + "type": "text" + }, + { + "bbox": [ + 132, + 308, + 168, + 321 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 320, + 234, + 332 + ], + "spans": [ + { + "bbox": [ + 110, + 320, + 122, + 332 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 132, + 320, + 148, + 332 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 148, + 321, + 220, + 331 + ], + "score": 0.85, + "content": "t = 0 , \\cdots , T - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 320, + 234, + 332 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 107, + 331, + 225, + 343 + ], + "spans": [ + { + "bbox": [ + 107, + 331, + 122, + 342 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 143, + 331, + 225, + 343 + ], + "score": 0.84, + "content": "r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )", + "type": "inline_equation" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 340, + 468, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 122, + 353 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 142, + 340, + 226, + 355 + ], + "score": 1.0, + "content": "Store the transitions", + "type": "text" + }, + { + "bbox": [ + 226, + 342, + 323, + 354 + ], + "score": 0.87, + "content": "\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 340, + 335, + 355 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 335, + 342, + 344, + 352 + ], + "score": 0.8, + "content": "R", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 340, + 468, + 355 + ], + "score": 1.0, + "content": "(Standard experience replay)", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 352, + 432, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 122, + 364 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 143, + 352, + 294, + 365 + ], + "score": 1.0, + "content": "Sample a set of the achieved goals of", + "type": "text" + }, + { + "bbox": [ + 294, + 354, + 305, + 364 + ], + "score": 0.88, + "content": "E _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 352, + 420, + 365 + ], + "score": 1.0, + "content": "as additional goals for reply", + "type": "text" + }, + { + "bbox": [ + 421, + 353, + 432, + 363 + ], + "score": 0.85, + "content": "G ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 362, + 207, + 377 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 123, + 376 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 142, + 363, + 158, + 376 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 158, + 363, + 191, + 377 + ], + "score": 0.92, + "content": "\\bar { g _ { p } ^ { \\prime } } \\in G ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 363, + 207, + 376 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 107, + 374, + 431, + 389 + ], + "spans": [ + { + "bbox": [ + 107, + 376, + 122, + 388 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 152, + 374, + 261, + 389 + ], + "score": 1.0, + "content": "Calculate a goal trajectory", + "type": "text" + }, + { + "bbox": [ + 261, + 376, + 314, + 389 + ], + "score": 0.93, + "content": "\\{ g _ { 0 } ^ { \\prime } , \\cdots , g _ { p } ^ { \\prime } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 374, + 342, + 389 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 343, + 376, + 392, + 388 + ], + "score": 0.56, + "content": "g _ { t } ^ { \\prime } = g ( t ; \\gamma )", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 376, + 431, + 388 + ], + "score": 0.4, + "content": "\\mathrm { ( H E R + ) }", + "type": "inline_equation" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 387, + 263, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 122, + 400 + ], + "score": 1.0, + "content": "15:", + "type": "text" + }, + { + "bbox": [ + 151, + 387, + 168, + 402 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 168, + 388, + 247, + 400 + ], + "score": 0.87, + "content": "t = \\{ 0 , \\cdots , p - 1 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 387, + 263, + 402 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 400, + 256, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 400, + 121, + 411 + ], + "score": 1.0, + "content": "16:", + "type": "text" + }, + { + "bbox": [ + 163, + 400, + 256, + 412 + ], + "score": 0.83, + "content": "r _ { t } ^ { \\prime } : = r ( s _ { i , t } , a _ { i , t } , g _ { t + 1 } ^ { \\prime } )", + "type": "inline_equation" + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 409, + 374, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 411, + 122, + 421 + ], + "score": 1.0, + "content": "17:", + "type": "text" + }, + { + "bbox": [ + 162, + 409, + 242, + 423 + ], + "score": 1.0, + "content": "Store the transition", + "type": "text" + }, + { + "bbox": [ + 242, + 410, + 374, + 423 + ], + "score": 0.82, + "content": "\\left( s _ { i , t } | g _ { t } ^ { \\prime } , a _ { i , t } , r _ { t } ^ { \\prime } , s _ { i , t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R", + "type": "inline_equation" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 421, + 187, + 433 + ], + "spans": [ + { + "bbox": [ + 106, + 421, + 122, + 433 + ], + "score": 1.0, + "content": "18:", + "type": "text" + }, + { + "bbox": [ + 152, + 421, + 187, + 433 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 432, + 178, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 122, + 444 + ], + "score": 1.0, + "content": "19:", + "type": "text" + }, + { + "bbox": [ + 142, + 432, + 178, + 444 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 442, + 167, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 122, + 455 + ], + "score": 1.0, + "content": "20:", + "type": "text" + }, + { + "bbox": [ + 132, + 442, + 167, + 455 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 454, + 157, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 454, + 157, + 466 + ], + "score": 1.0, + "content": "21: end for", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 221, + 468, + 466 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/Byf5-30qFX/Byf5-30qFX_model.json b/parse/train/Byf5-30qFX/Byf5-30qFX_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b8e05d88adcc0bad4a06c6c4f9872defb490fba --- /dev/null +++ b/parse/train/Byf5-30qFX/Byf5-30qFX_model.json @@ -0,0 +1,17985 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1176, + 1404, + 1176, + 1404, + 1512, + 298, + 1512 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1528, + 1405, + 1528, + 1405, + 1984, + 297, + 1984 + ], + "score": 0.979 + }, + { + "category_id": 0, + "poly": [ + 298, + 218, + 1154, + 218, + 1154, + 324, + 298, + 324 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 398, + 587, + 1301, + 587, + 1301, + 771, + 398, + 771 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 398, + 775, + 1302, + 775, + 1302, + 1049, + 398, + 1049 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 314, + 374, + 1131, + 374, + 1131, + 436, + 314, + 436 + ], + "score": 0.928 + }, + { + "category_id": 2, + "poly": [ + 332, + 2006, + 921, + 2006, + 921, + 2033, + 332, + 2033 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 302, + 1107, + 573, + 1107, + 573, + 1142, + 302, + 1142 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 816, + 75, + 816, + 104, + 300, + 104 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 773, + 518, + 926, + 518, + 926, + 553, + 773, + 553 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 857, + 2089, + 857, + 2112, + 841, + 2112 + ], + "score": 0.678 + }, + { + "category_id": 13, + "poly": [ + 826, + 1300, + 867, + 1300, + 867, + 1331, + 826, + 1331 + ], + "score": 0.43, + "latex": "\\mathrm { N g }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 218.0, + 1159.0, + 218.0, + 1159.0, + 270.0, + 295.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 273.0, + 757.0, + 273.0, + 757.0, + 328.0, + 294.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2001.0, + 925.0, + 2001.0, + 925.0, + 2040.0, + 330.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1104.0, + 579.0, + 1104.0, + 579.0, + 1151.0, + 294.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 817.0, + 72.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 518.0, + 932.0, + 518.0, + 932.0, + 556.0, + 769.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2088.0, + 861.0, + 2088.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1177.0, + 1404.0, + 1177.0, + 1404.0, + 1211.0, + 294.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1208.0, + 1405.0, + 1208.0, + 1405.0, + 1242.0, + 294.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1236.0, + 1405.0, + 1236.0, + 1405.0, + 1273.0, + 293.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1266.0, + 1404.0, + 1266.0, + 1404.0, + 1306.0, + 293.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1297.0, + 825.0, + 1297.0, + 825.0, + 1335.0, + 293.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1335.0, + 868.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1328.0, + 1406.0, + 1328.0, + 1406.0, + 1365.0, + 293.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1360.0, + 1404.0, + 1360.0, + 1404.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1391.0, + 1404.0, + 1391.0, + 1404.0, + 1422.0, + 296.0, + 1422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 1405.0, + 1421.0, + 1405.0, + 1455.0, + 294.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1452.0, + 1406.0, + 1452.0, + 1406.0, + 1486.0, + 294.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1484.0, + 632.0, + 1484.0, + 632.0, + 1515.0, + 296.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1529.0, + 1405.0, + 1529.0, + 1405.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1557.0, + 1405.0, + 1557.0, + 1405.0, + 1595.0, + 292.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1587.0, + 1405.0, + 1587.0, + 1405.0, + 1626.0, + 292.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1622.0, + 1402.0, + 1622.0, + 1402.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1647.0, + 1406.0, + 1647.0, + 1406.0, + 1688.0, + 292.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1679.0, + 1405.0, + 1679.0, + 1405.0, + 1716.0, + 294.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1403.0, + 1711.0, + 1403.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1739.0, + 1406.0, + 1739.0, + 1406.0, + 1780.0, + 291.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1771.0, + 1405.0, + 1771.0, + 1405.0, + 1806.0, + 294.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 1405.0, + 1802.0, + 1405.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 1406.0, + 1833.0, + 1406.0, + 1868.0, + 294.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1861.0, + 1406.0, + 1861.0, + 1406.0, + 1901.0, + 291.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1894.0, + 1405.0, + 1894.0, + 1405.0, + 1930.0, + 294.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1921.0, + 1406.0, + 1921.0, + 1406.0, + 1963.0, + 292.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1952.0, + 1406.0, + 1952.0, + 1406.0, + 1991.0, + 294.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 588.0, + 1304.0, + 588.0, + 1304.0, + 622.0, + 395.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 618.0, + 1305.0, + 618.0, + 1305.0, + 655.0, + 393.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 648.0, + 1306.0, + 648.0, + 1306.0, + 686.0, + 394.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 678.0, + 1306.0, + 678.0, + 1306.0, + 716.0, + 393.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 710.0, + 1306.0, + 710.0, + 1306.0, + 743.0, + 393.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 741.0, + 1189.0, + 741.0, + 1189.0, + 776.0, + 395.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 775.0, + 1304.0, + 775.0, + 1304.0, + 811.0, + 394.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 804.0, + 1306.0, + 804.0, + 1306.0, + 841.0, + 392.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 839.0, + 1304.0, + 839.0, + 1304.0, + 868.0, + 395.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 866.0, + 1304.0, + 866.0, + 1304.0, + 899.0, + 394.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 898.0, + 1305.0, + 898.0, + 1305.0, + 931.0, + 394.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 928.0, + 1306.0, + 928.0, + 1306.0, + 962.0, + 390.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 960.0, + 1305.0, + 960.0, + 1305.0, + 993.0, + 394.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 985.0, + 1305.0, + 985.0, + 1305.0, + 1026.0, + 393.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1019.0, + 994.0, + 1019.0, + 994.0, + 1052.0, + 393.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 371.0, + 1133.0, + 371.0, + 1133.0, + 416.0, + 310.0, + 416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 406.0, + 494.0, + 406.0, + 494.0, + 438.0, + 312.0, + 438.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 844, + 1403, + 844, + 1403, + 1241, + 298, + 1241 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1256, + 1403, + 1256, + 1403, + 1440, + 298, + 1440 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1623, + 1405, + 1623, + 1405, + 1900, + 298, + 1900 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1455, + 1402, + 1455, + 1402, + 1608, + 298, + 1608 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 306, + 229, + 1400, + 229, + 1400, + 577, + 306, + 577 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 300, + 765, + 1401, + 765, + 1401, + 829, + 300, + 829 + ], + "score": 0.945 + }, + { + "category_id": 4, + "poly": [ + 296, + 601, + 1403, + 601, + 1403, + 665, + 296, + 665 + ], + "score": 0.944 + }, + { + "category_id": 2, + "poly": [ + 329, + 2005, + 1236, + 2005, + 1236, + 2034, + 329, + 2034 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.886 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.723 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.153 + }, + { + "category_id": 15, + "poly": [ + 987.0, + 230.0, + 1102.0, + 230.0, + 1102.0, + 259.0, + 987.0, + 259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 246.0, + 877.0, + 246.0, + 877.0, + 274.0, + 740.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1003.0, + 246.0, + 1082.0, + 246.0, + 1082.0, + 279.0, + 1003.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 253.0, + 1241.0, + 253.0, + 1241.0, + 285.0, + 1108.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 486.0, + 278.0, + 536.0, + 278.0, + 536.0, + 305.0, + 486.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 280.0, + 814.0, + 280.0, + 814.0, + 297.0, + 782.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 278.0, + 1227.0, + 278.0, + 1227.0, + 304.0, + 1124.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 294.0, + 454.0, + 294.0, + 454.0, + 327.0, + 323.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 302.0, + 1203.0, + 302.0, + 1203.0, + 328.0, + 1144.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 301.0, + 1357.0, + 301.0, + 1357.0, + 333.0, + 1247.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 330.0, + 446.0, + 330.0, + 446.0, + 424.0, + 333.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 369.0, + 489.0, + 369.0, + 489.0, + 398.0, + 453.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 333.0, + 622.0, + 333.0, + 622.0, + 434.0, + 490.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 380.0, + 689.0, + 380.0, + 689.0, + 411.0, + 634.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 403.0, + 812.0, + 403.0, + 812.0, + 417.0, + 785.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 323.0, + 854.0, + 323.0, + 854.0, + 338.0, + 834.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 356.0, + 911.0, + 356.0, + 911.0, + 381.0, + 885.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 344.0, + 1070.0, + 344.0, + 1070.0, + 444.0, + 929.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 326.0, + 1357.0, + 326.0, + 1357.0, + 431.0, + 1249.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 434.0, + 337.0, + 434.0, + 337.0, + 454.0, + 317.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 435.0, + 465.0, + 435.0, + 465.0, + 453.0, + 449.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 436.0, + 1040.0, + 436.0, + 1040.0, + 460.0, + 956.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 470.0, + 451.0, + 470.0, + 451.0, + 504.0, + 324.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 453.0, + 469.0, + 469.0, + 469.0, + 469.0, + 483.0, + 453.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 460.0, + 530.0, + 460.0, + 530.0, + 488.0, + 492.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 460.0, + 1338.0, + 460.0, + 1338.0, + 490.0, + 1256.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 492.0, + 909.0, + 492.0, + 909.0, + 515.0, + 888.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 499.0, + 1238.0, + 499.0, + 1238.0, + 533.0, + 1105.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 525.0, + 812.0, + 525.0, + 812.0, + 539.0, + 785.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 522.0, + 1055.0, + 522.0, + 1055.0, + 555.0, + 937.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 523.0, + 1218.0, + 523.0, + 1218.0, + 553.0, + 1123.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 533.0, + 565.0, + 533.0, + 565.0, + 569.0, + 316.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 547.0, + 1032.0, + 547.0, + 1032.0, + 572.0, + 959.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 549.0, + 1200.0, + 549.0, + 1200.0, + 576.0, + 1141.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 452.5, + 1408.0, + 452.5, + 1408.0, + 483.5, + 1338.0, + 483.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 601.0, + 1404.0, + 601.0, + 1404.0, + 637.0, + 295.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 633.0, + 1195.0, + 633.0, + 1195.0, + 669.0, + 295.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 2000.0, + 1241.0, + 2000.0, + 1241.0, + 2039.0, + 333.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 842.0, + 1404.0, + 842.0, + 1404.0, + 879.0, + 293.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 874.0, + 1405.0, + 874.0, + 1405.0, + 910.0, + 293.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 899.0, + 1407.0, + 899.0, + 1407.0, + 944.0, + 292.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 934.0, + 1407.0, + 934.0, + 1407.0, + 973.0, + 292.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 964.0, + 1407.0, + 964.0, + 1407.0, + 1001.0, + 292.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 995.0, + 1405.0, + 995.0, + 1405.0, + 1033.0, + 293.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1027.0, + 1407.0, + 1027.0, + 1407.0, + 1061.0, + 293.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1057.0, + 1407.0, + 1057.0, + 1407.0, + 1093.0, + 294.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1088.0, + 1407.0, + 1088.0, + 1407.0, + 1124.0, + 293.0, + 1124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1115.0, + 1407.0, + 1115.0, + 1407.0, + 1156.0, + 292.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1147.0, + 1404.0, + 1147.0, + 1404.0, + 1184.0, + 294.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1215.0, + 293.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1076.0, + 1209.0, + 1076.0, + 1246.0, + 294.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1258.0, + 1404.0, + 1258.0, + 1404.0, + 1290.0, + 296.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1287.0, + 1407.0, + 1287.0, + 1407.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1318.0, + 1405.0, + 1318.0, + 1405.0, + 1351.0, + 292.0, + 1351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1348.0, + 1403.0, + 1348.0, + 1403.0, + 1384.0, + 293.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1378.0, + 1405.0, + 1378.0, + 1405.0, + 1414.0, + 294.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1409.0, + 1212.0, + 1409.0, + 1212.0, + 1444.0, + 294.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1625.0, + 1405.0, + 1625.0, + 1405.0, + 1658.0, + 297.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 1403.0, + 1655.0, + 1403.0, + 1688.0, + 295.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1684.0, + 1405.0, + 1684.0, + 1405.0, + 1721.0, + 295.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1406.0, + 1710.0, + 1406.0, + 1754.0, + 292.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1742.0, + 1406.0, + 1742.0, + 1406.0, + 1785.0, + 292.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1776.0, + 1405.0, + 1776.0, + 1405.0, + 1813.0, + 295.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1807.0, + 1405.0, + 1807.0, + 1405.0, + 1843.0, + 292.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1839.0, + 1405.0, + 1839.0, + 1405.0, + 1871.0, + 296.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1866.0, + 615.0, + 1866.0, + 615.0, + 1904.0, + 295.0, + 1904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1455.0, + 1403.0, + 1455.0, + 1403.0, + 1488.0, + 297.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1483.0, + 1406.0, + 1483.0, + 1406.0, + 1521.0, + 293.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1513.0, + 1404.0, + 1513.0, + 1404.0, + 1551.0, + 294.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1545.0, + 1404.0, + 1545.0, + 1404.0, + 1582.0, + 293.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1577.0, + 415.0, + 1577.0, + 415.0, + 1612.0, + 290.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 762.0, + 1406.0, + 762.0, + 1406.0, + 803.0, + 293.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 798.0, + 685.0, + 798.0, + 685.0, + 830.0, + 298.0, + 830.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1311, + 1404, + 1311, + 1404, + 1619, + 298, + 1619 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 846, + 1404, + 846, + 1404, + 1122, + 298, + 1122 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 294, + 1403, + 294, + 1403, + 571, + 298, + 571 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 586, + 1404, + 586, + 1404, + 831, + 298, + 831 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1711, + 1403, + 1711, + 1403, + 1866, + 298, + 1866 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1402, + 1881, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 1233, + 1400, + 1233, + 1400, + 1297, + 300, + 1297 + ], + "score": 0.955 + }, + { + "category_id": 0, + "poly": [ + 300, + 1165, + 581, + 1165, + 581, + 1201, + 300, + 1201 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 580, + 226, + 580, + 261, + 300, + 261 + ], + "score": 0.897 + }, + { + "category_id": 0, + "poly": [ + 300, + 1655, + 576, + 1655, + 576, + 1686, + 300, + 1686 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 816, + 76, + 816, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.588 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2112, + 842, + 2112 + ], + "score": 0.425 + }, + { + "category_id": 13, + "poly": [ + 298, + 1804, + 434, + 1804, + 434, + 1837, + 298, + 1837 + ], + "score": 0.93, + "latex": "g _ { t } = g ( t ; \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 906, + 1911, + 1097, + 1911, + 1097, + 1945, + 906, + 1945 + ], + "score": 0.93, + "latex": "f _ { g _ { t } } : S \\{ 0 , 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 687, + 1943, + 802, + 1943, + 802, + 1976, + 687, + 1976 + ], + "score": 0.92, + "latex": "f _ { g _ { t } } ( s ) = 1" + }, + { + "category_id": 13, + "poly": [ + 478, + 1912, + 558, + 1912, + 558, + 1944, + 478, + 1944 + ], + "score": 0.92, + "latex": "g _ { t } \\in \\mathcal G" + }, + { + "category_id": 13, + "poly": [ + 1051, + 2002, + 1146, + 2002, + 1146, + 2032, + 1051, + 2032 + ], + "score": 0.92, + "latex": "\\mathcal { G } = \\mathbb { R } ^ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1942, + 1318, + 1942, + 1318, + 1976, + 1111, + 1976 + ], + "score": 0.91, + "latex": "f _ { g _ { t } } ( s ) : = [ s = g _ { t } ]" + }, + { + "category_id": 13, + "poly": [ + 967, + 1774, + 1053, + 1774, + 1053, + 1805, + 967, + 1805 + ], + "score": 0.91, + "latex": "g _ { t } \\in \\mathcal G" + }, + { + "category_id": 13, + "poly": [ + 905, + 1945, + 983, + 1945, + 983, + 1973, + 905, + 1973 + ], + "score": 0.9, + "latex": "S = \\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 675, + 1407, + 795, + 1407, + 795, + 1436, + 675, + 1436 + ], + "score": 0.9, + "latex": "s _ { t } \\to s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1403, + 487, + 1403, + 487, + 1437, + 299, + 1437 + ], + "score": 0.9, + "latex": "\\bar { \\{ } s _ { 0 } , s _ { 1 } , \\dotsb , s _ { T } \\}" + }, + { + "category_id": 13, + "poly": [ + 945, + 1978, + 971, + 1978, + 971, + 2004, + 945, + 2004 + ], + "score": 0.84, + "latex": "g _ { t }" + }, + { + "category_id": 13, + "poly": [ + 520, + 1809, + 541, + 1809, + 541, + 1836, + 520, + 1836 + ], + "score": 0.81, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 1372, + 1883, + 1393, + 1883, + 1393, + 1911, + 1372, + 1911 + ], + "score": 0.79, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 1022, + 1979, + 1038, + 1979, + 1038, + 2000, + 1022, + 2000 + ], + "score": 0.71, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 522, + 1949, + 539, + 1949, + 539, + 1970, + 522, + 1970 + ], + "score": 0.71, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1160.0, + 587.0, + 1160.0, + 587.0, + 1210.0, + 290.0, + 1210.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 587.0, + 222.0, + 587.0, + 268.0, + 291.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1649.0, + 579.0, + 1649.0, + 579.0, + 1693.0, + 292.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1309.0, + 1405.0, + 1309.0, + 1405.0, + 1345.0, + 294.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1376.0, + 297.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1372.0, + 1407.0, + 1372.0, + 1407.0, + 1409.0, + 292.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 1404.0, + 674.0, + 1404.0, + 674.0, + 1439.0, + 488.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 1404.0, + 1405.0, + 1404.0, + 1405.0, + 1439.0, + 796.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1432.0, + 1408.0, + 1432.0, + 1408.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1462.0, + 1404.0, + 1462.0, + 1404.0, + 1501.0, + 292.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1493.0, + 1405.0, + 1493.0, + 1405.0, + 1530.0, + 292.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1525.0, + 1404.0, + 1525.0, + 1404.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1557.0, + 1402.0, + 1557.0, + 1402.0, + 1589.0, + 296.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1586.0, + 821.0, + 1586.0, + 821.0, + 1622.0, + 293.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 844.0, + 1405.0, + 844.0, + 1405.0, + 882.0, + 293.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 876.0, + 1406.0, + 876.0, + 1406.0, + 912.0, + 294.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 902.0, + 1406.0, + 902.0, + 1406.0, + 947.0, + 291.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 936.0, + 1406.0, + 936.0, + 1406.0, + 974.0, + 292.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 968.0, + 1404.0, + 968.0, + 1404.0, + 1005.0, + 292.0, + 1005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 998.0, + 1405.0, + 998.0, + 1405.0, + 1035.0, + 294.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1031.0, + 1404.0, + 1031.0, + 1404.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1060.0, + 1405.0, + 1060.0, + 1405.0, + 1095.0, + 293.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1090.0, + 518.0, + 1090.0, + 518.0, + 1128.0, + 296.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 296.0, + 1403.0, + 296.0, + 1403.0, + 329.0, + 296.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 324.0, + 1405.0, + 324.0, + 1405.0, + 363.0, + 292.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 358.0, + 1404.0, + 358.0, + 1404.0, + 391.0, + 297.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 384.0, + 1405.0, + 384.0, + 1405.0, + 425.0, + 292.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 456.0, + 291.0, + 456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 447.0, + 1404.0, + 447.0, + 1404.0, + 484.0, + 293.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1405.0, + 478.0, + 1405.0, + 515.0, + 294.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 504.0, + 1407.0, + 504.0, + 1407.0, + 548.0, + 292.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 540.0, + 1135.0, + 540.0, + 1135.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 585.0, + 1405.0, + 585.0, + 1405.0, + 622.0, + 294.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 616.0, + 1405.0, + 616.0, + 1405.0, + 653.0, + 294.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 649.0, + 1404.0, + 649.0, + 1404.0, + 682.0, + 294.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 679.0, + 1402.0, + 679.0, + 1402.0, + 712.0, + 293.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 709.0, + 1406.0, + 709.0, + 1406.0, + 744.0, + 292.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 738.0, + 1406.0, + 738.0, + 1406.0, + 774.0, + 291.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 768.0, + 1409.0, + 768.0, + 1409.0, + 804.0, + 293.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 891.0, + 798.0, + 891.0, + 837.0, + 295.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1713.0, + 1405.0, + 1713.0, + 1405.0, + 1746.0, + 296.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1739.0, + 1405.0, + 1739.0, + 1405.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1773.0, + 966.0, + 1773.0, + 966.0, + 1810.0, + 296.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1773.0, + 1405.0, + 1773.0, + 1405.0, + 1810.0, + 1054.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1803.0, + 297.0, + 1803.0, + 297.0, + 1842.0, + 292.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1803.0, + 519.0, + 1803.0, + 519.0, + 1842.0, + 435.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1803.0, + 1407.0, + 1803.0, + 1407.0, + 1842.0, + 542.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1836.0, + 871.0, + 1836.0, + 871.0, + 1870.0, + 296.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1880.0, + 1371.0, + 1880.0, + 1371.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1880.0, + 1404.0, + 1880.0, + 1404.0, + 1917.0, + 1394.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1909.0, + 477.0, + 1909.0, + 477.0, + 1950.0, + 295.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1909.0, + 905.0, + 1909.0, + 905.0, + 1950.0, + 559.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1950.0, + 1098.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1939.0, + 521.0, + 1939.0, + 521.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1939.0, + 686.0, + 1939.0, + 686.0, + 1979.0, + 540.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 1939.0, + 904.0, + 1939.0, + 904.0, + 1979.0, + 803.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1939.0, + 1110.0, + 1939.0, + 1110.0, + 1979.0, + 984.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1319.0, + 1939.0, + 1407.0, + 1939.0, + 1407.0, + 1979.0, + 1319.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 944.0, + 1972.0, + 944.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1972.0, + 1021.0, + 1972.0, + 1021.0, + 2009.0, + 972.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1972.0, + 1403.0, + 1972.0, + 1403.0, + 2009.0, + 1039.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1050.0, + 2001.0, + 1050.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 2001.0, + 1407.0, + 2001.0, + 1407.0, + 2038.0, + 1147.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1233.0, + 1404.0, + 1233.0, + 1404.0, + 1269.0, + 296.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1265.0, + 738.0, + 1265.0, + 738.0, + 1301.0, + 296.0, + 1301.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1444, + 1404, + 1444, + 1404, + 1630, + 298, + 1630 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1239, + 1404, + 1239, + 1404, + 1434, + 297, + 1434 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1734, + 1403, + 1734, + 1403, + 1866, + 299, + 1866 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1881, + 1403, + 1881, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 282, + 1359, + 282, + 1359, + 1174, + 297, + 1174 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 299, + 1677, + 890, + 1677, + 890, + 1709, + 299, + 1709 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.758 + }, + { + "category_id": 0, + "poly": [ + 296, + 227, + 1221, + 227, + 1221, + 263, + 296, + 263 + ], + "score": 0.522 + }, + { + "category_id": 1, + "poly": [ + 300, + 273, + 1157, + 273, + 1157, + 304, + 300, + 304 + ], + "score": 0.432 + }, + { + "category_id": 1, + "poly": [ + 296, + 227, + 1221, + 227, + 1221, + 263, + 296, + 263 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 1006, + 1368, + 1163, + 1368, + 1163, + 1402, + 1006, + 1402 + ], + "score": 0.93, + "latex": "f ( s _ { t } , g _ { t } ) = 1" + }, + { + "category_id": 13, + "poly": [ + 1077, + 795, + 1244, + 795, + 1244, + 835, + 1077, + 835 + ], + "score": 0.93, + "latex": "g _ { t } ^ { \\prime } = g _ { j , q - m + t } ^ { d e }" + }, + { + "category_id": 13, + "poly": [ + 960, + 1733, + 1317, + 1733, + 1317, + 1777, + 960, + 1777 + ], + "score": 0.92, + "latex": "g ^ { \\mathrm { d y n a m i c } } = \\{ s _ { t _ { 0 } } ^ { \\mathrm { o b j } } , s _ { t _ { 1 } } ^ { \\mathrm { o b j } } , \\cdot \\cdot \\cdot , s _ { t _ { T ^ { \\prime } } } ^ { \\mathrm { o b j } } \\}" + }, + { + "category_id": 13, + "poly": [ + 688, + 797, + 969, + 797, + 969, + 833, + 688, + 833 + ], + "score": 0.92, + "latex": "\\{ g _ { 0 } ^ { \\prime } , \\cdot \\cdot \\cdot , g _ { m } ^ { \\prime } \\} _ { m = \\operatorname* { m i n } \\{ p , q \\} }" + }, + { + "category_id": 13, + "poly": [ + 424, + 763, + 517, + 763, + 517, + 797, + 424, + 797 + ], + "score": 0.92, + "latex": "E _ { j } \\neq \\boldsymbol { \\mathcal { O } }" + }, + { + "category_id": 13, + "poly": [ + 680, + 1303, + 1112, + 1303, + 1112, + 1341, + 680, + 1341 + ], + "score": 0.91, + "latex": "f \\big ( s _ { t } , g _ { t } \\big ) = \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( \\| s _ { t } ^ { \\mathrm { o b j } } - g _ { t } \\| \\leq \\epsilon \\big )" + }, + { + "category_id": 13, + "poly": [ + 413, + 699, + 498, + 699, + 498, + 729, + 413, + 729 + ], + "score": 0.91, + "latex": "E _ { i } \\in \\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 572, + 729, + 739, + 729, + 739, + 764, + 572, + 764 + ], + "score": 0.91, + "latex": "E _ { j } ( i \\neq j ) \\in \\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 1197, + 1302, + 1239, + 1302, + 1239, + 1340, + 1197, + 1340 + ], + "score": 0.9, + "latex": "s _ { t } ^ { \\mathrm { o b j } }" + }, + { + "category_id": 13, + "poly": [ + 937, + 1476, + 1127, + 1476, + 1127, + 1510, + 937, + 1510 + ], + "score": 0.9, + "latex": "g _ { t } ^ { \\mathrm { s t a t i c } } = g ^ { \\mathrm { s t a t i c } } , \\forall t" + }, + { + "category_id": 13, + "poly": [ + 911, + 492, + 964, + 492, + 964, + 518, + 911, + 518 + ], + "score": 0.89, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 413, + 1015, + 569, + 1015, + 569, + 1046, + 413, + 1046 + ], + "score": 0.89, + "latex": "t = 1 , \\cdots , N" + }, + { + "category_id": 13, + "poly": [ + 413, + 395, + 611, + 395, + 611, + 426, + 413, + 426 + ], + "score": 0.89, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 552, + 1240, + 594, + 1240, + 594, + 1270, + 552, + 1270 + ], + "score": 0.88, + "latex": "s ^ { \\mathrm { o b j } }" + }, + { + "category_id": 13, + "poly": [ + 817, + 727, + 939, + 727, + 939, + 768, + 817, + 768 + ], + "score": 0.88, + "latex": "g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }" + }, + { + "category_id": 13, + "poly": [ + 469, + 831, + 698, + 831, + 698, + 864, + 469, + 864 + ], + "score": 0.87, + "latex": "t = \\{ 0 , \\cdots , m - 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 1308, + 798, + 1343, + 798, + 1343, + 832, + 1308, + 832 + ], + "score": 0.87, + "latex": "E _ { j }" + }, + { + "category_id": 13, + "poly": [ + 464, + 1778, + 627, + 1778, + 627, + 1805, + 464, + 1805 + ], + "score": 0.87, + "latex": "s _ { 0 } , s _ { 1 } , \\cdots , s _ { T }" + }, + { + "category_id": 13, + "poly": [ + 454, + 864, + 842, + 864, + 842, + 897, + 454, + 897 + ], + "score": 0.86, + "latex": "r _ { t } ^ { \\prime } : = r ( s _ { i , p - m + t } , a _ { i , p - m + t } , g _ { t + 1 } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 1125, + 1340, + 1235, + 1340, + 1235, + 1369, + 1125, + 1369 + ], + "score": 0.85, + "latex": "\\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } }" + }, + { + "category_id": 13, + "poly": [ + 413, + 547, + 612, + 547, + 612, + 577, + 413, + 577 + ], + "score": 0.85, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 573, + 1342, + 600, + 1342, + 600, + 1369, + 573, + 1369 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1141, + 491, + 1195, + 491, + 1195, + 519, + 1141, + 519 + ], + "score": 0.84, + "latex": "g _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 621, + 369, + 649, + 369, + 649, + 395, + 621, + 395 + ], + "score": 0.84, + "latex": "g _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 610, + 491, + 638, + 491, + 638, + 516, + 610, + 516 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 826, + 1373, + 852, + 1373, + 852, + 1399, + 826, + 1399 + ], + "score": 0.83, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1261, + 1405, + 1288, + 1405, + 1288, + 1429, + 1261, + 1429 + ], + "score": 0.83, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 597, + 430, + 624, + 430, + 624, + 455, + 597, + 455 + ], + "score": 0.83, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 630, + 608, + 897, + 608, + 897, + 641, + 630, + 641 + ], + "score": 0.83, + "latex": "\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 863, + 371, + 891, + 371, + 891, + 394, + 863, + 394 + ], + "score": 0.81, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 400, + 578, + 628, + 578, + 628, + 610, + 400, + 610 + ], + "score": 0.8, + "latex": "r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 1045, + 1803, + 1130, + 1803, + 1130, + 1836, + 1045, + 1836 + ], + "score": 0.8, + "latex": "g ^ { \\mathrm { d y n a m i c } }" + }, + { + "category_id": 13, + "poly": [ + 931, + 609, + 957, + 609, + 957, + 635, + 931, + 635 + ], + "score": 0.78, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 659, + 670, + 680, + 670, + 680, + 696, + 659, + 696 + ], + "score": 0.77, + "latex": "\\mathcal { E }" + }, + { + "category_id": 13, + "poly": [ + 627, + 1046, + 652, + 1046, + 652, + 1072, + 627, + 1072 + ], + "score": 0.77, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 476, + 334, + 639, + 334, + 639, + 365, + 476, + 365 + ], + "score": 0.76, + "latex": "\\mathbf { \\Omega } = 1 , 2 , \\cdots , M" + }, + { + "category_id": 13, + "poly": [ + 907, + 1046, + 932, + 1046, + 932, + 1072, + 907, + 1072 + ], + "score": 0.75, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 762, + 1077, + 788, + 1077, + 788, + 1103, + 762, + 1103 + ], + "score": 0.75, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 673, + 304, + 698, + 304, + 698, + 330, + 673, + 330 + ], + "score": 0.74, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 651, + 1343, + 668, + 1343, + 668, + 1365, + 651, + 1365 + ], + "score": 0.72, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 905, + 274, + 929, + 274, + 929, + 300, + 905, + 300 + ], + "score": 0.7, + "latex": "R" + }, + { + "category_id": 14, + "poly": [ + 399, + 455, + 562, + 455, + 562, + 489, + 399, + 489 + ], + "score": 0.6, + "latex": "a _ { t } \\pi ( s _ { t } | g _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 674, + 893, + 1225, + 893, + 1225, + 929, + 674, + 929 + ], + "score": 0.57, + "latex": "( \\stackrel { \\cdot } { s } _ { i , p - m + t } | \\bar { g } _ { t } ^ { \\prime } , a _ { i , p - m + t } , r _ { t } ^ { \\prime } , s _ { i , p - m + t + 1 } | g _ { t + 1 } ^ { \\prime } ) \\mathrm { i n } \\tilde { A }" + }, + { + "category_id": 13, + "poly": [ + 1215, + 893, + 1241, + 893, + 1241, + 921, + 1215, + 921 + ], + "score": 0.48, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 1141, + 280, + 1158, + 280, + 1158, + 300, + 1141, + 300 + ], + "score": 0.46, + "latex": "r" + }, + { + "category_id": 14, + "poly": [ + 675, + 889, + 1356, + 889, + 1356, + 929, + 675, + 929 + ], + "score": 0.44, + "latex": "\\left( s _ { i , p - m + t } | g _ { t } ^ { \\prime } , a _ { i , p - m + t } , r _ { t } ^ { \\prime } , s _ { i , p - m + t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R \\left( \\mathrm { D H E R } \\right)" + }, + { + "category_id": 13, + "poly": [ + 475, + 335, + 517, + 335, + 517, + 363, + 475, + 363 + ], + "score": 0.43, + "latex": "= 1" + }, + { + "category_id": 13, + "poly": [ + 380, + 1397, + 994, + 1397, + 994, + 1433, + 380, + 1433 + ], + "score": 0.41, + "latex": "r _ { t } : = r ( \\tilde { s } _ { t } , a _ { t } , g _ { t + 1 } ) = - \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } ( f ( s _ { t + 1 } , g _ { t + 1 } ) \\neq 1 )" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1677.0, + 893.0, + 1677.0, + 893.0, + 1713.0, + 295.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 218.0, + 1227.0, + 218.0, + 1227.0, + 275.0, + 292.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1444.0, + 1405.0, + 1444.0, + 1405.0, + 1480.0, + 293.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1471.0, + 936.0, + 1471.0, + 936.0, + 1515.0, + 287.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 1471.0, + 1409.0, + 1471.0, + 1409.0, + 1515.0, + 1128.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1506.0, + 1405.0, + 1506.0, + 1405.0, + 1542.0, + 293.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1573.0, + 294.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1568.0, + 1406.0, + 1568.0, + 1406.0, + 1603.0, + 293.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1600.0, + 675.0, + 1600.0, + 675.0, + 1631.0, + 294.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1235.0, + 551.0, + 1235.0, + 551.0, + 1276.0, + 294.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 595.0, + 1235.0, + 1405.0, + 1235.0, + 1405.0, + 1276.0, + 595.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1271.0, + 1401.0, + 1271.0, + 1401.0, + 1304.0, + 296.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1286.0, + 679.0, + 1286.0, + 679.0, + 1355.0, + 286.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1286.0, + 1196.0, + 1286.0, + 1196.0, + 1355.0, + 1113.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1286.0, + 1405.0, + 1286.0, + 1405.0, + 1355.0, + 1240.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1335.0, + 572.0, + 1335.0, + 572.0, + 1376.0, + 294.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1335.0, + 650.0, + 1335.0, + 650.0, + 1376.0, + 601.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1335.0, + 1124.0, + 1335.0, + 1124.0, + 1376.0, + 669.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1335.0, + 1409.0, + 1335.0, + 1409.0, + 1376.0, + 1236.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1365.0, + 825.0, + 1365.0, + 825.0, + 1404.0, + 295.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1365.0, + 1005.0, + 1365.0, + 1005.0, + 1404.0, + 853.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1365.0, + 1406.0, + 1365.0, + 1406.0, + 1404.0, + 1164.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1398.0, + 379.0, + 1398.0, + 379.0, + 1435.0, + 294.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1398.0, + 1260.0, + 1398.0, + 1260.0, + 1435.0, + 995.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1289.0, + 1398.0, + 1299.0, + 1398.0, + 1299.0, + 1435.0, + 1289.0, + 1435.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 278.0, + 1712.0, + 959.0, + 1712.0, + 959.0, + 1799.0, + 278.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 1738.0, + 1407.0, + 1738.0, + 1407.0, + 1773.0, + 1318.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1771.0, + 463.0, + 1771.0, + 463.0, + 1808.0, + 293.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 1771.0, + 1407.0, + 1771.0, + 1407.0, + 1808.0, + 628.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1795.0, + 1044.0, + 1795.0, + 1044.0, + 1842.0, + 289.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1795.0, + 1407.0, + 1795.0, + 1407.0, + 1842.0, + 1131.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1833.0, + 1047.0, + 1833.0, + 1047.0, + 1870.0, + 294.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.75, + 1744.5, + 1308.75, + 1744.5, + 1308.75, + 1778.0, + 1238.75, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1879.0, + 1405.0, + 1879.0, + 1405.0, + 1917.0, + 295.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1407.0, + 1942.0, + 1407.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1408.0, + 2000.0, + 1408.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 275.0, + 904.0, + 275.0, + 904.0, + 305.0, + 298.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 275.0, + 1140.0, + 275.0, + 1140.0, + 305.0, + 930.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 302.0, + 672.0, + 302.0, + 672.0, + 332.0, + 309.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 332.0, + 474.0, + 332.0, + 474.0, + 366.0, + 308.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 332.0, + 680.0, + 332.0, + 680.0, + 366.0, + 640.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 364.0, + 339.0, + 364.0, + 339.0, + 394.0, + 308.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 362.0, + 620.0, + 362.0, + 620.0, + 399.0, + 368.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 362.0, + 862.0, + 362.0, + 862.0, + 399.0, + 650.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 362.0, + 895.0, + 362.0, + 895.0, + 399.0, + 892.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 395.0, + 339.0, + 395.0, + 339.0, + 425.0, + 308.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 395.0, + 412.0, + 395.0, + 412.0, + 425.0, + 369.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 395.0, + 652.0, + 395.0, + 652.0, + 425.0, + 612.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 423.0, + 340.0, + 423.0, + 340.0, + 457.0, + 306.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 424.0, + 596.0, + 424.0, + 596.0, + 457.0, + 396.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 424.0, + 1029.0, + 424.0, + 1029.0, + 457.0, + 625.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 456.0, + 339.0, + 456.0, + 339.0, + 487.0, + 308.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 486.0, + 339.0, + 486.0, + 339.0, + 517.0, + 308.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 480.0, + 609.0, + 480.0, + 609.0, + 524.0, + 395.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 480.0, + 910.0, + 480.0, + 910.0, + 524.0, + 639.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 480.0, + 1140.0, + 480.0, + 1140.0, + 524.0, + 965.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 517.0, + 339.0, + 517.0, + 339.0, + 548.0, + 308.0, + 548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 516.0, + 465.0, + 516.0, + 465.0, + 547.0, + 370.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 546.0, + 340.0, + 546.0, + 340.0, + 579.0, + 305.0, + 579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 543.0, + 412.0, + 543.0, + 412.0, + 580.0, + 367.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 543.0, + 653.0, + 543.0, + 653.0, + 580.0, + 613.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 577.0, + 340.0, + 577.0, + 340.0, + 609.0, + 297.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 578.0, + 399.0, + 578.0, + 399.0, + 612.0, + 394.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 608.0, + 340.0, + 608.0, + 340.0, + 640.0, + 297.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 603.0, + 629.0, + 603.0, + 629.0, + 643.0, + 396.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 603.0, + 930.0, + 603.0, + 930.0, + 643.0, + 898.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 603.0, + 1299.0, + 603.0, + 1299.0, + 643.0, + 958.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 637.0, + 339.0, + 637.0, + 339.0, + 670.0, + 297.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 637.0, + 465.0, + 637.0, + 465.0, + 668.0, + 369.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 668.0, + 339.0, + 668.0, + 339.0, + 699.0, + 297.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 667.0, + 658.0, + 667.0, + 658.0, + 701.0, + 370.0, + 701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 698.0, + 340.0, + 698.0, + 340.0, + 764.0, + 295.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 698.0, + 412.0, + 698.0, + 412.0, + 782.0, + 369.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 499.0, + 698.0, + 571.0, + 698.0, + 571.0, + 782.0, + 499.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 698.0, + 816.0, + 698.0, + 816.0, + 782.0, + 740.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 698.0, + 954.0, + 698.0, + 954.0, + 782.0, + 940.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 764.0, + 339.0, + 764.0, + 339.0, + 796.0, + 297.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 760.0, + 423.0, + 760.0, + 423.0, + 797.0, + 396.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 760.0, + 580.0, + 760.0, + 580.0, + 797.0, + 518.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 796.0, + 341.0, + 796.0, + 341.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 783.0, + 687.0, + 783.0, + 687.0, + 846.0, + 416.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 783.0, + 1076.0, + 783.0, + 1076.0, + 846.0, + 970.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1245.0, + 783.0, + 1307.0, + 783.0, + 1307.0, + 846.0, + 1245.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 783.0, + 1355.0, + 783.0, + 1355.0, + 846.0, + 1344.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 832.0, + 340.0, + 832.0, + 340.0, + 864.0, + 297.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 831.0, + 468.0, + 831.0, + 468.0, + 865.0, + 425.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.0, + 831.0, + 741.0, + 831.0, + 741.0, + 865.0, + 699.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 863.0, + 339.0, + 863.0, + 339.0, + 894.0, + 297.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 858.0, + 453.0, + 858.0, + 453.0, + 903.0, + 449.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 858.0, + 847.0, + 858.0, + 847.0, + 903.0, + 843.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 889.0, + 340.0, + 889.0, + 340.0, + 926.0, + 296.0, + 926.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 885.0, + 673.0, + 885.0, + 673.0, + 932.0, + 448.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 924.0, + 339.0, + 924.0, + 339.0, + 956.0, + 295.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 922.0, + 521.0, + 922.0, + 521.0, + 955.0, + 425.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 953.0, + 338.0, + 953.0, + 338.0, + 986.0, + 295.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 953.0, + 476.0, + 953.0, + 476.0, + 983.0, + 394.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 983.0, + 338.0, + 983.0, + 338.0, + 1015.0, + 295.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 982.0, + 466.0, + 982.0, + 466.0, + 1013.0, + 369.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1014.0, + 338.0, + 1014.0, + 338.0, + 1046.0, + 295.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1013.0, + 412.0, + 1013.0, + 412.0, + 1046.0, + 368.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1013.0, + 611.0, + 1013.0, + 611.0, + 1046.0, + 570.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1042.0, + 339.0, + 1042.0, + 339.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1044.0, + 626.0, + 1044.0, + 626.0, + 1077.0, + 398.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1044.0, + 906.0, + 1044.0, + 906.0, + 1077.0, + 653.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1044.0, + 936.0, + 1044.0, + 936.0, + 1077.0, + 933.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1075.0, + 338.0, + 1075.0, + 338.0, + 1107.0, + 294.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1075.0, + 761.0, + 1075.0, + 761.0, + 1108.0, + 398.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 1075.0, + 792.0, + 1075.0, + 792.0, + 1108.0, + 789.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1105.0, + 339.0, + 1105.0, + 339.0, + 1137.0, + 295.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1104.0, + 465.0, + 1104.0, + 465.0, + 1136.0, + 368.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1135.0, + 438.0, + 1135.0, + 438.0, + 1169.0, + 295.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 270.0, + 904.0, + 270.0, + 904.0, + 309.0, + 296.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 270.0, + 1140.0, + 270.0, + 1140.0, + 309.0, + 930.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 218.0, + 1227.0, + 218.0, + 1227.0, + 275.0, + 292.0, + 275.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1183, + 1405, + 1183, + 1405, + 1581, + 296, + 1581 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 981, + 1404, + 981, + 1404, + 1168, + 297, + 1168 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1592, + 1402, + 1592, + 1402, + 1718, + 298, + 1718 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 363, + 237, + 1327, + 237, + 1327, + 728, + 363, + 728 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1403, + 1911, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 298, + 902, + 1399, + 902, + 1399, + 966, + 298, + 966 + ], + "score": 0.95 + }, + { + "category_id": 4, + "poly": [ + 295, + 737, + 1401, + 737, + 1401, + 800, + 295, + 800 + ], + "score": 0.94 + }, + { + "category_id": 0, + "poly": [ + 299, + 1811, + 540, + 1811, + 540, + 1847, + 299, + 1847 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.715 + }, + { + "category_id": 13, + "poly": [ + 702, + 1103, + 913, + 1103, + 913, + 1142, + 702, + 1142 + ], + "score": 0.93, + "latex": "\\dot { s } _ { t } = \\langle o _ { t } , g _ { t } ^ { a c } , g _ { t } ^ { d e } \\rangle" + }, + { + "category_id": 13, + "poly": [ + 298, + 1544, + 474, + 1544, + 474, + 1585, + 298, + 1585 + ], + "score": 0.93, + "latex": "\\{ g _ { j , 0 } ^ { d e } , \\cdot \\cdot \\cdot , g _ { j , t } ^ { d e } \\}" + }, + { + "category_id": 13, + "poly": [ + 552, + 1546, + 715, + 1546, + 715, + 1580, + 552, + 1580 + ], + "score": 0.93, + "latex": "t \\leq \\operatorname* { m i n } \\{ p , q \\}" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1477, + 1272, + 1477, + 1272, + 1512, + 1099, + 1512 + ], + "score": 0.93, + "latex": "t \\leq \\operatorname* { m i n } \\{ p , q \\}" + }, + { + "category_id": 13, + "poly": [ + 1104, + 1103, + 1204, + 1103, + 1204, + 1139, + 1104, + 1139 + ], + "score": 0.93, + "latex": "g _ { t } ^ { d e } = g _ { t }" + }, + { + "category_id": 13, + "poly": [ + 718, + 1475, + 786, + 1475, + 786, + 1515, + 718, + 1515 + ], + "score": 0.93, + "latex": "\\{ g _ { j , t } ^ { d e } \\}" + }, + { + "category_id": 13, + "poly": [ + 452, + 1441, + 578, + 1441, + 578, + 1479, + 452, + 1479 + ], + "score": 0.92, + "latex": "g _ { i , p } ^ { a c } = g _ { j , q } ^ { d e }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1072, + 727, + 1072, + 727, + 1106, + 297, + 1106 + ], + "score": 0.92, + "latex": "\\left\\{ \\left( s _ { 0 } , a _ { 0 } , r _ { 0 } , s _ { 1 } \\right) \\cdot \\cdot \\cdot , \\left( s _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } \\right) \\right\\}" + }, + { + "category_id": 13, + "poly": [ + 783, + 1337, + 826, + 1337, + 826, + 1373, + 783, + 1373 + ], + "score": 0.91, + "latex": "g _ { i , q } ^ { a c }" + }, + { + "category_id": 13, + "poly": [ + 750, + 1013, + 928, + 1013, + 928, + 1045, + 750, + 1045 + ], + "score": 0.91, + "latex": "\\left( { { s _ { t } } , { a _ { t } } , { r _ { t } } , { s _ { t + 1 } } } \\right)" + }, + { + "category_id": 13, + "poly": [ + 513, + 1369, + 557, + 1369, + 557, + 1410, + 513, + 1410 + ], + "score": 0.91, + "latex": "g _ { j , p } ^ { d e }" + }, + { + "category_id": 13, + "poly": [ + 427, + 1408, + 628, + 1408, + 628, + 1440, + 427, + 1440 + ], + "score": 0.9, + "latex": "\\{ E _ { 0 } , E _ { 1 } , \\stackrel { \\cdot } { E } _ { 2 } , \\cdot \\cdot \\cdot \\}" + }, + { + "category_id": 13, + "poly": [ + 1255, + 1106, + 1296, + 1106, + 1296, + 1139, + 1255, + 1139 + ], + "score": 0.9, + "latex": "g _ { t } ^ { a c }" + }, + { + "category_id": 13, + "poly": [ + 799, + 1514, + 832, + 1514, + 832, + 1547, + 799, + 1547 + ], + "score": 0.9, + "latex": "E _ { i } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1008, + 1479, + 1044, + 1479, + 1044, + 1513, + 1008, + 1513 + ], + "score": 0.89, + "latex": "E _ { j }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1106, + 338, + 1106, + 338, + 1139, + 297, + 1139 + ], + "score": 0.89, + "latex": "g _ { t } ^ { d e }" + }, + { + "category_id": 13, + "poly": [ + 546, + 1108, + 586, + 1108, + 586, + 1139, + 546, + 1139 + ], + "score": 0.89, + "latex": "g _ { t } ^ { a c }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1015, + 426, + 1015, + 426, + 1044, + 298, + 1044 + ], + "score": 0.89, + "latex": "s _ { t } \\ \\to \\ s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1408, + 1134, + 1408, + 1134, + 1438, + 1100, + 1438 + ], + "score": 0.88, + "latex": "E _ { i }" + }, + { + "category_id": 13, + "poly": [ + 940, + 1514, + 972, + 1514, + 972, + 1544, + 940, + 1544 + ], + "score": 0.88, + "latex": "E _ { i }" + }, + { + "category_id": 13, + "poly": [ + 690, + 989, + 852, + 989, + 852, + 1013, + 690, + 1013 + ], + "score": 0.87, + "latex": "s _ { 0 } , s _ { 1 } , \\cdots , s _ { T }" + }, + { + "category_id": 13, + "poly": [ + 641, + 1479, + 674, + 1479, + 674, + 1509, + 641, + 1509 + ], + "score": 0.86, + "latex": "E _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1224, + 1078, + 1250, + 1078, + 1250, + 1103, + 1224, + 1103 + ], + "score": 0.84, + "latex": "o _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1242, + 1016, + 1269, + 1016, + 1269, + 1042, + 1242, + 1042 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1015, + 1047, + 1015, + 1047, + 1042, + 1020, + 1042 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 371, + 1048, + 398, + 1048, + 398, + 1073, + 371, + 1073 + ], + "score": 0.83, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 667, + 1047, + 693, + 1047, + 693, + 1072, + 667, + 1072 + ], + "score": 0.82, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 877, + 1480, + 895, + 1480, + 895, + 1511, + 877, + 1511 + ], + "score": 0.81, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1187, + 1407, + 1223, + 1407, + 1223, + 1442, + 1187, + 1442 + ], + "score": 0.81, + "latex": "E _ { j }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1441, + 405, + 1441, + 405, + 1475, + 297, + 1475 + ], + "score": 0.78, + "latex": "\\exists i , j , p , q ," + }, + { + "category_id": 13, + "poly": [ + 981, + 1378, + 1000, + 1378, + 1000, + 1405, + 981, + 1405 + ], + "score": 0.77, + "latex": "p" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1374, + 1146, + 1374, + 1146, + 1405, + 1129, + 1405 + ], + "score": 0.76, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1045, + 1076, + 1045, + 1076, + 1071, + 1062, + 1071 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 297, + 1378, + 315, + 1378, + 315, + 1405, + 297, + 1405 + ], + "score": 0.74, + "latex": "q" + }, + { + "category_id": 13, + "poly": [ + 298, + 1046, + 310, + 1046, + 310, + 1071, + 298, + 1071 + ], + "score": 0.73, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 444, + 1374, + 459, + 1374, + 459, + 1400, + 444, + 1400 + ], + "score": 0.72, + "latex": "i" + }, + { + "category_id": 13, + "poly": [ + 1234, + 1407, + 1312, + 1407, + 1312, + 1441, + 1234, + 1441 + ], + "score": 0.68, + "latex": "( i \\neq j )" + }, + { + "category_id": 13, + "poly": [ + 616, + 1593, + 745, + 1593, + 745, + 1623, + 616, + 1623 + ], + "score": 0.67, + "latex": "\\mathrm { R L + D H E R }" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 595.0, + 1128.0, + 595.0, + 1128.0, + 620.0, + 1104.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 695.0, + 551.0, + 695.0, + 551.0, + 738.0, + 412.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 696.0, + 789.0, + 696.0, + 789.0, + 738.0, + 662.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 696.0, + 1030.0, + 696.0, + 1030.0, + 738.0, + 904.0, + 738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 701.0, + 1268.0, + 701.0, + 1268.0, + 733.0, + 1165.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 735.0, + 1406.0, + 735.0, + 1406.0, + 771.0, + 295.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 766.0, + 1185.0, + 766.0, + 1185.0, + 803.0, + 293.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1809.0, + 545.0, + 1809.0, + 545.0, + 1852.0, + 293.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1185.0, + 1403.0, + 1185.0, + 1403.0, + 1217.0, + 298.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1215.0, + 1403.0, + 1215.0, + 1403.0, + 1247.0, + 295.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1246.0, + 1402.0, + 1246.0, + 1402.0, + 1278.0, + 295.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1276.0, + 1403.0, + 1276.0, + 1403.0, + 1308.0, + 295.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1305.0, + 1403.0, + 1305.0, + 1403.0, + 1342.0, + 295.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1331.0, + 782.0, + 1331.0, + 782.0, + 1379.0, + 291.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1331.0, + 1408.0, + 1331.0, + 1408.0, + 1379.0, + 827.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1356.0, + 296.0, + 1356.0, + 296.0, + 1444.0, + 283.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1356.0, + 426.0, + 1356.0, + 426.0, + 1444.0, + 316.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 1356.0, + 980.0, + 1356.0, + 980.0, + 1444.0, + 629.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1356.0, + 1099.0, + 1356.0, + 1099.0, + 1444.0, + 1001.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1356.0, + 1186.0, + 1356.0, + 1186.0, + 1444.0, + 1147.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1224.0, + 1356.0, + 1233.0, + 1356.0, + 1233.0, + 1444.0, + 1224.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1313.0, + 1356.0, + 1417.0, + 1356.0, + 1417.0, + 1444.0, + 1313.0, + 1444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1430.0, + 296.0, + 1430.0, + 296.0, + 1491.0, + 286.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1430.0, + 451.0, + 1430.0, + 451.0, + 1491.0, + 406.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1430.0, + 1411.0, + 1430.0, + 1411.0, + 1491.0, + 579.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1465.0, + 640.0, + 1465.0, + 640.0, + 1530.0, + 285.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1465.0, + 717.0, + 1465.0, + 717.0, + 1530.0, + 675.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1465.0, + 876.0, + 1465.0, + 876.0, + 1530.0, + 787.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1465.0, + 1007.0, + 1465.0, + 1007.0, + 1530.0, + 896.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1465.0, + 1098.0, + 1465.0, + 1098.0, + 1530.0, + 1045.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 1465.0, + 1412.0, + 1465.0, + 1412.0, + 1530.0, + 1273.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1510.0, + 798.0, + 1510.0, + 798.0, + 1551.0, + 293.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 833.0, + 1510.0, + 939.0, + 1510.0, + 939.0, + 1551.0, + 833.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1510.0, + 1403.0, + 1510.0, + 1403.0, + 1551.0, + 973.0, + 1551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1532.0, + 297.0, + 1532.0, + 297.0, + 1595.0, + 290.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1532.0, + 551.0, + 1532.0, + 551.0, + 1595.0, + 475.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1532.0, + 735.0, + 1532.0, + 735.0, + 1595.0, + 716.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 978.0, + 689.0, + 978.0, + 689.0, + 1017.0, + 294.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 978.0, + 1405.0, + 978.0, + 1405.0, + 1017.0, + 853.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1009.0, + 297.0, + 1009.0, + 297.0, + 1049.0, + 292.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 1009.0, + 749.0, + 1009.0, + 749.0, + 1049.0, + 427.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1009.0, + 1019.0, + 1009.0, + 1019.0, + 1049.0, + 929.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1009.0, + 1241.0, + 1009.0, + 1241.0, + 1049.0, + 1048.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1049.0, + 1270.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1042.0, + 297.0, + 1042.0, + 297.0, + 1074.0, + 294.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 1042.0, + 370.0, + 1042.0, + 370.0, + 1074.0, + 311.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1042.0, + 666.0, + 1042.0, + 666.0, + 1074.0, + 399.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 1042.0, + 1061.0, + 1042.0, + 1061.0, + 1074.0, + 694.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1042.0, + 1404.0, + 1042.0, + 1404.0, + 1074.0, + 1077.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1072.0, + 1223.0, + 1072.0, + 1223.0, + 1108.0, + 728.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1072.0, + 1405.0, + 1072.0, + 1405.0, + 1108.0, + 1251.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1098.0, + 296.0, + 1098.0, + 296.0, + 1146.0, + 291.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1098.0, + 545.0, + 1098.0, + 545.0, + 1146.0, + 339.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1098.0, + 701.0, + 1098.0, + 701.0, + 1146.0, + 587.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1098.0, + 1103.0, + 1098.0, + 1103.0, + 1146.0, + 914.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1098.0, + 1254.0, + 1098.0, + 1254.0, + 1146.0, + 1205.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1098.0, + 1409.0, + 1098.0, + 1409.0, + 1146.0, + 1297.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1136.0, + 673.0, + 1136.0, + 673.0, + 1169.0, + 292.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1594.0, + 615.0, + 1594.0, + 615.0, + 1627.0, + 294.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 1594.0, + 1404.0, + 1594.0, + 1404.0, + 1627.0, + 746.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1623.0, + 1404.0, + 1623.0, + 1404.0, + 1661.0, + 292.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1656.0, + 1404.0, + 1656.0, + 1404.0, + 1689.0, + 294.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1686.0, + 978.0, + 1686.0, + 978.0, + 1722.0, + 293.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1911.0, + 1408.0, + 1911.0, + 1408.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1943.0, + 1405.0, + 1943.0, + 1405.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1408.0, + 1973.0, + 1408.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2003.0, + 876.0, + 2003.0, + 876.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 899.0, + 1405.0, + 899.0, + 1405.0, + 940.0, + 293.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 933.0, + 1326.0, + 933.0, + 1326.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1449, + 1404, + 1449, + 1404, + 1816, + 297, + 1816 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 769, + 1405, + 769, + 1405, + 1107, + 297, + 1107 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1122, + 1404, + 1122, + 1404, + 1435, + 297, + 1435 + ], + "score": 0.982 + }, + { + "category_id": 3, + "poly": [ + 311, + 222, + 1385, + 222, + 1385, + 581, + 311, + 581 + ], + "score": 0.968 + }, + { + "category_id": 4, + "poly": [ + 603, + 605, + 1095, + 605, + 1095, + 638, + 603, + 638 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 334, + 2005, + 700, + 2005, + 700, + 2033, + 334, + 2033 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 299, + 708, + 561, + 708, + 561, + 739, + 299, + 739 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 302, + 1928, + 1372, + 1928, + 1372, + 1960, + 302, + 1960 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 299, + 1866, + 501, + 1866, + 501, + 1896, + 299, + 1896 + ], + "score": 0.877 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.778 + }, + { + "category_id": 13, + "poly": [ + 1106, + 1666, + 1159, + 1666, + 1159, + 1695, + 1106, + 1695 + ], + "score": 0.92, + "latex": "g _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1235, + 1274, + 1404, + 1274, + 1404, + 1309, + 1235, + 1309 + ], + "score": 0.92, + "latex": "r ( s _ { t } , a _ { t } , g _ { t } ) =" + }, + { + "category_id": 13, + "poly": [ + 786, + 1602, + 848, + 1602, + 848, + 1635, + 786, + 1635 + ], + "score": 0.91, + "latex": "( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 376, + 1480, + 472, + 1480, + 472, + 1509, + 376, + 1509 + ], + "score": 0.9, + "latex": "3 0 \\times 4 0" + }, + { + "category_id": 13, + "poly": [ + 296, + 1305, + 659, + 1305, + 659, + 1348, + 296, + 1348 + ], + "score": 0.9, + "latex": "- \\mathbf { 1 } _ { \\mathrm { c o n d i t i o n } } \\big ( | s _ { t + 1 } ^ { \\mathrm { o b j } } - g _ { t + 1 } | \\geq \\epsilon \\big )" + }, + { + "category_id": 13, + "poly": [ + 1177, + 1481, + 1246, + 1481, + 1246, + 1509, + 1177, + 1509 + ], + "score": 0.9, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 1003, + 1666, + 1057, + 1666, + 1057, + 1695, + 1003, + 1695 + ], + "score": 0.89, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 742, + 1316, + 796, + 1316, + 796, + 1344, + 742, + 1344 + ], + "score": 0.89, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 850, + 1317, + 904, + 1317, + 904, + 1344, + 850, + 1344 + ], + "score": 0.89, + "latex": "g _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 381, + 1662, + 929, + 1662, + 929, + 1698, + 381, + 1698 + ], + "score": 0.88, + "latex": "r \\big ( s _ { t } , \\bar { a } _ { t } , g _ { t + 1 } \\big ) = - \\bar { \\mathbf { 1 } } _ { \\mathrm { c o n d i t i o n } } \\big ( \\big | s _ { t + 1 } ^ { \\mathrm { s n a k e } } - g _ { t + 1 } \\big | \\bar { \\neq } 0 \\big )" + }, + { + "category_id": 13, + "poly": [ + 859, + 1215, + 968, + 1215, + 968, + 1242, + 859, + 1242 + ], + "score": 0.87, + "latex": "\\epsilon = 0 . 0 1" + }, + { + "category_id": 13, + "poly": [ + 832, + 1348, + 857, + 1348, + 857, + 1372, + 832, + 1372 + ], + "score": 0.86, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 960, + 1698, + 986, + 1698, + 986, + 1724, + 960, + 1724 + ], + "score": 0.85, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 666, + 1347, + 694, + 1347, + 694, + 1372, + 666, + 1372 + ], + "score": 0.85, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 801, + 1699, + 829, + 1699, + 829, + 1724, + 801, + 1724 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 855, + 1245, + 974, + 1245, + 974, + 1273, + 855, + 1273 + ], + "score": 0.82, + "latex": "v = 0 . 0 1 1" + }, + { + "category_id": 13, + "poly": [ + 525, + 1377, + 541, + 1377, + 541, + 1399, + 525, + 1399 + ], + "score": 0.53, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 223.0, + 571.0, + 223.0, + 571.0, + 254.0, + 457.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 223.0, + 923.0, + 223.0, + 923.0, + 254.0, + 824.0, + 254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 226.0, + 1281.0, + 226.0, + 1281.0, + 253.0, + 1181.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 246.0, + 731.0, + 246.0, + 731.0, + 272.0, + 693.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 240.0, + 1089.0, + 240.0, + 1089.0, + 268.0, + 1051.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 251.0, + 1160.0, + 251.0, + 1160.0, + 271.0, + 1119.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 259.0, + 369.0, + 259.0, + 369.0, + 447.0, + 312.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 260.0, + 701.0, + 260.0, + 701.0, + 446.0, + 671.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 264.0, + 1060.0, + 264.0, + 1060.0, + 314.0, + 1034.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 265.0, + 1189.0, + 265.0, + 1189.0, + 286.0, + 1121.0, + 286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 302.0, + 371.0, + 302.0, + 371.0, + 329.0, + 333.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 281.0, + 731.0, + 281.0, + 731.0, + 313.0, + 691.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 301.0, + 1089.0, + 301.0, + 1089.0, + 447.0, + 1030.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1119.0, + 280.0, + 1203.0, + 280.0, + 1203.0, + 327.0, + 1119.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 318.0, + 638.0, + 318.0, + 638.0, + 350.0, + 569.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 319.0, + 731.0, + 319.0, + 731.0, + 351.0, + 691.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 318.0, + 997.0, + 318.0, + 997.0, + 350.0, + 929.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 348.0, + 371.0, + 348.0, + 371.0, + 374.0, + 333.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 347.0, + 654.0, + 347.0, + 654.0, + 393.0, + 568.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 361.0, + 729.0, + 361.0, + 729.0, + 388.0, + 691.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 345.0, + 1012.0, + 345.0, + 1012.0, + 392.0, + 927.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 378.0, + 1088.0, + 378.0, + 1088.0, + 404.0, + 1051.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 393.0, + 371.0, + 393.0, + 371.0, + 419.0, + 333.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 399.0, + 728.0, + 399.0, + 728.0, + 427.0, + 693.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 439.0, + 371.0, + 439.0, + 371.0, + 465.0, + 333.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 440.0, + 729.0, + 440.0, + 729.0, + 462.0, + 694.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 446.0, + 1089.0, + 446.0, + 1089.0, + 472.0, + 1051.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 464.0, + 391.0, + 464.0, + 391.0, + 486.0, + 370.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 461.0, + 451.0, + 461.0, + 451.0, + 489.0, + 415.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 459.0, + 557.0, + 459.0, + 557.0, + 490.0, + 464.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 574.0, + 464.0, + 607.0, + 464.0, + 607.0, + 486.0, + 574.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 461.0, + 660.0, + 461.0, + 660.0, + 489.0, + 625.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 465.0, + 750.0, + 465.0, + 750.0, + 485.0, + 729.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 461.0, + 809.0, + 461.0, + 809.0, + 488.0, + 775.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 461.0, + 864.0, + 461.0, + 864.0, + 488.0, + 827.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 461.0, + 914.0, + 461.0, + 914.0, + 488.0, + 880.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 935.0, + 464.0, + 968.0, + 464.0, + 968.0, + 486.0, + 935.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 464.0, + 1018.0, + 464.0, + 1018.0, + 486.0, + 984.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 464.0, + 1107.0, + 464.0, + 1107.0, + 485.0, + 1089.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 464.0, + 1167.0, + 464.0, + 1167.0, + 486.0, + 1135.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 461.0, + 1274.0, + 461.0, + 1274.0, + 488.0, + 1186.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 464.0, + 1325.0, + 464.0, + 1325.0, + 486.0, + 1293.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 464.0, + 1376.0, + 464.0, + 1376.0, + 486.0, + 1346.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 480.0, + 550.0, + 480.0, + 550.0, + 509.0, + 484.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 480.0, + 907.0, + 480.0, + 907.0, + 509.0, + 842.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 480.0, + 1266.0, + 480.0, + 1266.0, + 509.0, + 1199.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 521.0, + 1390.0, + 521.0, + 1390.0, + 552.0, + 310.0, + 552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 547.0, + 447.0, + 547.0, + 447.0, + 583.0, + 342.0, + 583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 551.0, + 888.0, + 551.0, + 888.0, + 578.0, + 704.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 549.0, + 1122.0, + 549.0, + 1122.0, + 582.0, + 1056.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 604.0, + 1099.0, + 604.0, + 1099.0, + 640.0, + 599.0, + 640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 2001.0, + 701.0, + 2001.0, + 701.0, + 2039.0, + 332.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 708.0, + 563.0, + 708.0, + 563.0, + 741.0, + 295.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 505.0, + 1865.0, + 505.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1445.0, + 1405.0, + 1445.0, + 1405.0, + 1485.0, + 291.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1478.0, + 375.0, + 1478.0, + 375.0, + 1513.0, + 292.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1478.0, + 1176.0, + 1478.0, + 1176.0, + 1513.0, + 473.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1478.0, + 1405.0, + 1478.0, + 1405.0, + 1513.0, + 1247.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1511.0, + 1404.0, + 1511.0, + 1404.0, + 1545.0, + 295.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1541.0, + 1406.0, + 1541.0, + 1406.0, + 1574.0, + 292.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1569.0, + 1405.0, + 1569.0, + 1405.0, + 1605.0, + 294.0, + 1605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1601.0, + 785.0, + 1601.0, + 785.0, + 1635.0, + 295.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1601.0, + 1402.0, + 1601.0, + 1402.0, + 1635.0, + 849.0, + 1635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1631.0, + 1405.0, + 1631.0, + 1405.0, + 1669.0, + 292.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1650.0, + 380.0, + 1650.0, + 380.0, + 1712.0, + 286.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1650.0, + 1002.0, + 1650.0, + 1002.0, + 1712.0, + 930.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1650.0, + 1105.0, + 1650.0, + 1105.0, + 1712.0, + 1058.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1160.0, + 1650.0, + 1406.0, + 1650.0, + 1406.0, + 1712.0, + 1160.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1693.0, + 800.0, + 1693.0, + 800.0, + 1727.0, + 294.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 1693.0, + 959.0, + 1693.0, + 959.0, + 1727.0, + 830.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1693.0, + 1404.0, + 1693.0, + 1404.0, + 1727.0, + 987.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1721.0, + 1405.0, + 1721.0, + 1405.0, + 1757.0, + 291.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1754.0, + 1405.0, + 1754.0, + 1405.0, + 1788.0, + 295.0, + 1788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1783.0, + 541.0, + 1783.0, + 541.0, + 1821.0, + 294.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 768.0, + 1405.0, + 768.0, + 1405.0, + 809.0, + 292.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 802.0, + 1406.0, + 802.0, + 1406.0, + 837.0, + 294.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 832.0, + 1405.0, + 832.0, + 1405.0, + 867.0, + 295.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 863.0, + 1405.0, + 863.0, + 1405.0, + 898.0, + 294.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 889.0, + 1405.0, + 889.0, + 1405.0, + 931.0, + 291.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 922.0, + 1405.0, + 922.0, + 1405.0, + 959.0, + 292.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 953.0, + 1403.0, + 953.0, + 1403.0, + 989.0, + 291.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 985.0, + 1403.0, + 985.0, + 1403.0, + 1016.0, + 296.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1011.0, + 1406.0, + 1011.0, + 1406.0, + 1053.0, + 292.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1045.0, + 1406.0, + 1045.0, + 1406.0, + 1080.0, + 295.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1076.0, + 1031.0, + 1076.0, + 1031.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1122.0, + 1405.0, + 1122.0, + 1405.0, + 1155.0, + 295.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1152.0, + 1404.0, + 1152.0, + 1404.0, + 1184.0, + 295.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1182.0, + 1407.0, + 1182.0, + 1407.0, + 1218.0, + 291.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1211.0, + 858.0, + 1211.0, + 858.0, + 1249.0, + 292.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1211.0, + 1406.0, + 1211.0, + 1406.0, + 1249.0, + 969.0, + 1249.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1242.0, + 854.0, + 1242.0, + 854.0, + 1278.0, + 294.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1242.0, + 1406.0, + 1242.0, + 1406.0, + 1278.0, + 975.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1272.0, + 1234.0, + 1272.0, + 1234.0, + 1311.0, + 292.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1303.0, + 295.0, + 1303.0, + 295.0, + 1357.0, + 290.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1303.0, + 741.0, + 1303.0, + 741.0, + 1357.0, + 660.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1303.0, + 849.0, + 1303.0, + 849.0, + 1357.0, + 797.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1303.0, + 1410.0, + 1303.0, + 1410.0, + 1357.0, + 905.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 665.0, + 1339.0, + 665.0, + 1377.0, + 294.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 1339.0, + 831.0, + 1339.0, + 831.0, + 1377.0, + 695.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 1339.0, + 1405.0, + 1339.0, + 1405.0, + 1377.0, + 858.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1369.0, + 524.0, + 1369.0, + 524.0, + 1406.0, + 294.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1369.0, + 1405.0, + 1369.0, + 1405.0, + 1406.0, + 542.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1400.0, + 885.0, + 1400.0, + 885.0, + 1440.0, + 294.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1922.0, + 1377.0, + 1922.0, + 1377.0, + 1965.0, + 293.0, + 1965.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1606, + 1404, + 1606, + 1404, + 1883, + 298, + 1883 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1346, + 1403, + 1346, + 1403, + 1593, + 297, + 1593 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 377, + 227, + 765, + 227, + 765, + 539, + 377, + 539 + ], + "score": 0.966 + }, + { + "category_id": 3, + "poly": [ + 928, + 227, + 1316, + 227, + 1316, + 539, + 928, + 539 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 368, + 1027, + 1404, + 1027, + 1404, + 1252, + 368, + 1252 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 298, + 1972, + 1397, + 1972, + 1397, + 2034, + 298, + 2034 + ], + "score": 0.942 + }, + { + "category_id": 1, + "poly": [ + 292, + 939, + 1401, + 939, + 1401, + 1003, + 292, + 1003 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 402, + 568, + 746, + 568, + 746, + 600, + 402, + 600 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 298, + 1920, + 1043, + 1920, + 1043, + 1949, + 298, + 1949 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 298, + 1291, + 1043, + 1291, + 1043, + 1321, + 298, + 1321 + ], + "score": 0.866 + }, + { + "category_id": 4, + "poly": [ + 950, + 568, + 1303, + 568, + 1303, + 600, + 950, + 600 + ], + "score": 0.828 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 369, + 734, + 1403, + 734, + 1403, + 913, + 369, + 913 + ], + "score": 0.611 + }, + { + "category_id": 1, + "poly": [ + 371, + 660, + 1402, + 660, + 1402, + 724, + 371, + 724 + ], + "score": 0.586 + }, + { + "category_id": 4, + "poly": [ + 950, + 567, + 1302, + 567, + 1302, + 600, + 950, + 600 + ], + "score": 0.326 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.217 + }, + { + "category_id": 13, + "poly": [ + 837, + 881, + 962, + 881, + 962, + 912, + 837, + 912 + ], + "score": 0.9, + "latex": "( - d + 1 . 0 )" + }, + { + "category_id": 13, + "poly": [ + 1319, + 851, + 1394, + 851, + 1394, + 882, + 1319, + 882 + ], + "score": 0.9, + "latex": "d \\geq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 726, + 1219, + 850, + 1219, + 850, + 1250, + 726, + 1250 + ], + "score": 0.9, + "latex": "( - d + 1 . 0 )" + }, + { + "category_id": 13, + "poly": [ + 533, + 882, + 600, + 882, + 600, + 909, + 533, + 909 + ], + "score": 0.9, + "latex": "d < \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 978, + 1189, + 1035, + 1189, + 1035, + 1220, + 978, + 1220 + ], + "score": 0.89, + "latex": "( - d )" + }, + { + "category_id": 13, + "poly": [ + 397, + 1219, + 488, + 1219, + 488, + 1248, + 397, + 1248 + ], + "score": 0.88, + "latex": "d = 0 . 0" + }, + { + "category_id": 13, + "poly": [ + 1029, + 851, + 1085, + 851, + 1085, + 881, + 1029, + 881 + ], + "score": 0.87, + "latex": "( - d )" + }, + { + "category_id": 13, + "poly": [ + 926, + 1147, + 983, + 1147, + 983, + 1177, + 926, + 1177 + ], + "score": 0.86, + "latex": "( - d )" + }, + { + "category_id": 13, + "poly": [ + 988, + 809, + 1044, + 809, + 1044, + 839, + 988, + 839 + ], + "score": 0.86, + "latex": "( - d )" + }, + { + "category_id": 13, + "poly": [ + 395, + 1102, + 544, + 1102, + 544, + 1135, + 395, + 1135 + ], + "score": 0.82, + "latex": "\\mathrm { D Q N + H E R }" + }, + { + "category_id": 13, + "poly": [ + 483, + 971, + 630, + 971, + 630, + 1001, + 483, + 1001 + ], + "score": 0.8, + "latex": "\\mathrm { D Q N + H E R }" + }, + { + "category_id": 13, + "poly": [ + 396, + 734, + 565, + 734, + 565, + 764, + 396, + 764 + ], + "score": 0.35, + "latex": "\\mathrm { D D P G } + \\mathrm { H E R }" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 223.0, + 664.0, + 223.0, + 664.0, + 257.0, + 538.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 247.0, + 442.0, + 247.0, + 442.0, + 280.0, + 396.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 265.0, + 411.0, + 265.0, + 411.0, + 475.0, + 373.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 291.0, + 440.0, + 291.0, + 440.0, + 322.0, + 398.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 332.0, + 441.0, + 332.0, + 441.0, + 366.0, + 397.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 342.0, + 722.0, + 342.0, + 722.0, + 366.0, + 678.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 358.0, + 756.0, + 358.0, + 756.0, + 381.0, + 678.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 377.0, + 440.0, + 377.0, + 440.0, + 407.0, + 397.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 374.0, + 749.0, + 374.0, + 749.0, + 397.0, + 678.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 419.0, + 441.0, + 419.0, + 441.0, + 452.0, + 397.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 463.0, + 440.0, + 463.0, + 440.0, + 493.0, + 398.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 492.0, + 460.0, + 492.0, + 460.0, + 516.0, + 440.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 489.0, + 528.0, + 489.0, + 528.0, + 520.0, + 491.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 489.0, + 588.0, + 489.0, + 588.0, + 519.0, + 549.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 489.0, + 646.0, + 489.0, + 646.0, + 519.0, + 608.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 489.0, + 705.0, + 489.0, + 705.0, + 519.0, + 666.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 489.0, + 762.0, + 489.0, + 762.0, + 520.0, + 726.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 510.0, + 636.0, + 510.0, + 636.0, + 542.0, + 563.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 223.0, + 1215.0, + 223.0, + 1215.0, + 257.0, + 1090.0, + 257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 243.0, + 992.0, + 243.0, + 992.0, + 273.0, + 950.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 266.0, + 963.0, + 266.0, + 963.0, + 472.0, + 927.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 296.0, + 992.0, + 296.0, + 992.0, + 329.0, + 951.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 343.0, + 1272.0, + 343.0, + 1272.0, + 364.0, + 1230.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 355.0, + 990.0, + 355.0, + 990.0, + 382.0, + 952.0, + 382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 359.0, + 1306.0, + 359.0, + 1306.0, + 379.0, + 1231.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 376.0, + 1299.0, + 376.0, + 1299.0, + 395.0, + 1231.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 408.0, + 991.0, + 408.0, + 991.0, + 438.0, + 951.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 463.0, + 991.0, + 463.0, + 991.0, + 492.0, + 950.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 492.0, + 1011.0, + 492.0, + 1011.0, + 516.0, + 992.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 489.0, + 1080.0, + 489.0, + 1080.0, + 519.0, + 1043.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 489.0, + 1143.0, + 489.0, + 1143.0, + 519.0, + 1100.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 489.0, + 1198.0, + 489.0, + 1198.0, + 519.0, + 1157.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 489.0, + 1258.0, + 489.0, + 1258.0, + 519.0, + 1216.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 491.0, + 1313.0, + 491.0, + 1313.0, + 518.0, + 1278.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 510.0, + 1188.0, + 510.0, + 1188.0, + 543.0, + 1117.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 564.0, + 748.0, + 564.0, + 748.0, + 602.0, + 399.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1916.0, + 1046.0, + 1916.0, + 1046.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1289.0, + 1045.0, + 1289.0, + 1045.0, + 1323.0, + 295.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 565.0, + 1302.0, + 565.0, + 1302.0, + 605.0, + 948.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 565.0, + 1301.0, + 565.0, + 1301.0, + 604.0, + 947.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1604.0, + 1404.0, + 1604.0, + 1404.0, + 1644.0, + 295.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1638.0, + 1405.0, + 1638.0, + 1405.0, + 1675.0, + 294.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1670.0, + 1404.0, + 1670.0, + 1404.0, + 1703.0, + 296.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1697.0, + 1404.0, + 1697.0, + 1404.0, + 1735.0, + 293.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1768.0, + 292.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 1405.0, + 1759.0, + 1405.0, + 1796.0, + 294.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1788.0, + 1404.0, + 1788.0, + 1404.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1857.0, + 293.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1852.0, + 551.0, + 1852.0, + 551.0, + 1888.0, + 293.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1345.0, + 1405.0, + 1345.0, + 1405.0, + 1382.0, + 294.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1377.0, + 1404.0, + 1377.0, + 1404.0, + 1410.0, + 295.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1408.0, + 1405.0, + 1408.0, + 1405.0, + 1441.0, + 292.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1438.0, + 1405.0, + 1438.0, + 1405.0, + 1472.0, + 295.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1466.0, + 1404.0, + 1466.0, + 1404.0, + 1507.0, + 291.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1500.0, + 1405.0, + 1500.0, + 1405.0, + 1533.0, + 293.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1405.0, + 1529.0, + 1405.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1560.0, + 1366.0, + 1560.0, + 1366.0, + 1597.0, + 293.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1024.0, + 1406.0, + 1024.0, + 1406.0, + 1066.0, + 364.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1056.0, + 478.0, + 1056.0, + 478.0, + 1097.0, + 390.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 1101.0, + 394.0, + 1101.0, + 394.0, + 1136.0, + 364.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1101.0, + 1092.0, + 1101.0, + 1092.0, + 1136.0, + 545.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1145.0, + 925.0, + 1145.0, + 925.0, + 1180.0, + 367.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1145.0, + 1187.0, + 1145.0, + 1187.0, + 1180.0, + 984.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 1187.0, + 977.0, + 1187.0, + 977.0, + 1222.0, + 369.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1187.0, + 1407.0, + 1187.0, + 1407.0, + 1222.0, + 1036.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1220.0, + 725.0, + 1220.0, + 725.0, + 1251.0, + 489.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1220.0, + 1237.0, + 1220.0, + 1237.0, + 1251.0, + 851.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1401.0, + 1971.0, + 1401.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1402.0, + 2000.0, + 1402.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 1403.0, + 936.0, + 1403.0, + 974.0, + 294.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 970.0, + 482.0, + 970.0, + 482.0, + 1006.0, + 294.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 970.0, + 1178.0, + 970.0, + 1178.0, + 1006.0, + 631.0, + 1006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 731.0, + 395.0, + 731.0, + 395.0, + 770.0, + 373.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.0, + 731.0, + 1403.0, + 731.0, + 1403.0, + 770.0, + 566.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 765.0, + 796.0, + 765.0, + 796.0, + 799.0, + 393.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 807.0, + 987.0, + 807.0, + 987.0, + 842.0, + 383.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 807.0, + 1247.0, + 807.0, + 1247.0, + 842.0, + 1045.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 848.0, + 1028.0, + 848.0, + 1028.0, + 886.0, + 388.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 848.0, + 1318.0, + 848.0, + 1318.0, + 886.0, + 1086.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 848.0, + 1406.0, + 848.0, + 1406.0, + 886.0, + 1395.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 881.0, + 532.0, + 881.0, + 532.0, + 916.0, + 396.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 881.0, + 836.0, + 881.0, + 836.0, + 916.0, + 601.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 881.0, + 1266.0, + 881.0, + 1266.0, + 916.0, + 963.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 657.0, + 1404.0, + 657.0, + 1404.0, + 696.0, + 371.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 688.0, + 1398.0, + 688.0, + 1398.0, + 729.0, + 392.0, + 729.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1043, + 1404, + 1043, + 1404, + 1380, + 297, + 1380 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1503, + 1404, + 1503, + 1404, + 1780, + 298, + 1780 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 675, + 1403, + 675, + 1403, + 919, + 298, + 919 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1795, + 1402, + 1795, + 1402, + 1917, + 298, + 1917 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 928, + 227, + 1319, + 227, + 1319, + 535, + 928, + 535 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 376, + 228, + 767, + 228, + 767, + 535, + 376, + 535 + ], + "score": 0.966 + }, + { + "category_id": 2, + "poly": [ + 298, + 1978, + 1403, + 1978, + 1403, + 2033, + 298, + 2033 + ], + "score": 0.933 + }, + { + "category_id": 4, + "poly": [ + 368, + 565, + 781, + 565, + 781, + 597, + 368, + 597 + ], + "score": 0.92 + }, + { + "category_id": 4, + "poly": [ + 856, + 565, + 1397, + 565, + 1397, + 597, + 856, + 597 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 814, + 76, + 814, + 104, + 299, + 104 + ], + "score": 0.89 + }, + { + "category_id": 0, + "poly": [ + 301, + 1440, + 680, + 1440, + 680, + 1469, + 301, + 1469 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.798 + }, + { + "category_id": 0, + "poly": [ + 298, + 981, + 1013, + 981, + 1013, + 1009, + 298, + 1009 + ], + "score": 0.793 + }, + { + "category_id": 1, + "poly": [ + 298, + 981, + 1013, + 981, + 1013, + 1009, + 298, + 1009 + ], + "score": 0.107 + }, + { + "category_id": 13, + "poly": [ + 858, + 798, + 976, + 798, + 976, + 827, + 858, + 827 + ], + "score": 0.9, + "latex": "v = 0 . 0 1 1" + }, + { + "category_id": 13, + "poly": [ + 352, + 799, + 470, + 799, + 470, + 827, + 352, + 827 + ], + "score": 0.89, + "latex": "v = 0 . 0 0 1" + }, + { + "category_id": 13, + "poly": [ + 1063, + 797, + 1117, + 797, + 1117, + 828, + 1063, + 828 + ], + "score": 0.89, + "latex": "1 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 692, + 828, + 811, + 828, + 811, + 857, + 692, + 857 + ], + "score": 0.89, + "latex": "v = 0 . 0 1 6" + }, + { + "category_id": 13, + "poly": [ + 1212, + 828, + 1330, + 828, + 1330, + 857, + 1212, + 857 + ], + "score": 0.89, + "latex": "v = 0 . 0 1 1" + }, + { + "category_id": 13, + "poly": [ + 725, + 859, + 850, + 859, + 850, + 888, + 725, + 888 + ], + "score": 0.88, + "latex": "v = 0 . 0 1 6" + }, + { + "category_id": 13, + "poly": [ + 1334, + 1795, + 1402, + 1795, + 1402, + 1825, + 1334, + 1825 + ], + "score": 0.85, + "latex": "100 \\%" + }, + { + "category_id": 13, + "poly": [ + 1055, + 565, + 1131, + 565, + 1131, + 594, + 1055, + 594 + ], + "score": 0.74, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 430, + 859, + 580, + 859, + 580, + 888, + 430, + 888 + ], + "score": 0.42, + "latex": "\\mathrm { \\Delta D D P G + H E R }" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 224.0, + 1200.0, + 224.0, + 1200.0, + 256.0, + 1102.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 247.0, + 994.0, + 247.0, + 994.0, + 279.0, + 949.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 267.0, + 960.0, + 267.0, + 960.0, + 472.0, + 927.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 298.0, + 993.0, + 298.0, + 993.0, + 331.0, + 949.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 343.0, + 1258.0, + 343.0, + 1258.0, + 363.0, + 1222.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 351.0, + 992.0, + 351.0, + 992.0, + 380.0, + 950.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 357.0, + 1305.0, + 357.0, + 1305.0, + 380.0, + 1222.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 373.0, + 1305.0, + 373.0, + 1305.0, + 396.0, + 1222.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 402.0, + 991.0, + 402.0, + 991.0, + 431.0, + 950.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 454.0, + 991.0, + 454.0, + 991.0, + 482.0, + 950.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 486.0, + 1322.0, + 486.0, + 1322.0, + 517.0, + 994.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 509.0, + 1193.0, + 509.0, + 1193.0, + 539.0, + 1109.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 225.0, + 648.0, + 225.0, + 648.0, + 256.0, + 549.0, + 256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 247.0, + 442.0, + 247.0, + 442.0, + 280.0, + 397.0, + 280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 267.0, + 409.0, + 267.0, + 409.0, + 470.0, + 376.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 293.0, + 440.0, + 293.0, + 440.0, + 323.0, + 399.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 328.0, + 701.0, + 328.0, + 701.0, + 348.0, + 666.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 339.0, + 440.0, + 339.0, + 440.0, + 369.0, + 399.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 341.0, + 749.0, + 341.0, + 749.0, + 364.0, + 665.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 357.0, + 741.0, + 357.0, + 741.0, + 380.0, + 665.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 375.0, + 741.0, + 375.0, + 741.0, + 394.0, + 666.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 386.0, + 437.0, + 386.0, + 437.0, + 412.0, + 401.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 389.0, + 755.0, + 389.0, + 755.0, + 412.0, + 665.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 431.0, + 437.0, + 431.0, + 437.0, + 458.0, + 401.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 493.0, + 458.0, + 493.0, + 458.0, + 513.0, + 442.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 486.0, + 652.0, + 486.0, + 652.0, + 517.0, + 484.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 486.0, + 772.0, + 486.0, + 772.0, + 517.0, + 654.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 510.0, + 641.0, + 510.0, + 641.0, + 539.0, + 558.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 494.5, + 660.0, + 494.5, + 660.0, + 507.5, + 642.0, + 507.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1968.0, + 1408.0, + 1968.0, + 1408.0, + 2015.0, + 326.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 464.0, + 2000.0, + 464.0, + 2040.0, + 294.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 364.0, + 559.0, + 781.0, + 559.0, + 781.0, + 604.0, + 364.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 562.0, + 1054.0, + 562.0, + 1054.0, + 602.0, + 850.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 562.0, + 1403.0, + 562.0, + 1403.0, + 602.0, + 1132.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1436.0, + 684.0, + 1436.0, + 684.0, + 1474.0, + 294.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 979.0, + 1018.0, + 979.0, + 1018.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1043.0, + 1405.0, + 1043.0, + 1405.0, + 1079.0, + 292.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1075.0, + 1405.0, + 1075.0, + 1405.0, + 1110.0, + 294.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1103.0, + 1405.0, + 1103.0, + 1405.0, + 1140.0, + 292.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1135.0, + 1406.0, + 1135.0, + 1406.0, + 1171.0, + 292.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1167.0, + 1406.0, + 1167.0, + 1406.0, + 1202.0, + 294.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1197.0, + 1406.0, + 1197.0, + 1406.0, + 1231.0, + 294.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1228.0, + 1404.0, + 1228.0, + 1404.0, + 1262.0, + 294.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1256.0, + 1405.0, + 1256.0, + 1405.0, + 1293.0, + 291.0, + 1293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1288.0, + 1404.0, + 1288.0, + 1404.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1316.0, + 1405.0, + 1316.0, + 1405.0, + 1357.0, + 291.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 759.0, + 1349.0, + 759.0, + 1384.0, + 294.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1503.0, + 1404.0, + 1503.0, + 1404.0, + 1539.0, + 296.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1537.0, + 1405.0, + 1537.0, + 1405.0, + 1570.0, + 296.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1566.0, + 1406.0, + 1566.0, + 1406.0, + 1602.0, + 293.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1596.0, + 1408.0, + 1596.0, + 1408.0, + 1632.0, + 293.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1626.0, + 1406.0, + 1626.0, + 1406.0, + 1661.0, + 292.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1656.0, + 1405.0, + 1656.0, + 1405.0, + 1691.0, + 294.0, + 1691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1687.0, + 1406.0, + 1687.0, + 1406.0, + 1721.0, + 293.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1716.0, + 1406.0, + 1716.0, + 1406.0, + 1753.0, + 291.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1750.0, + 529.0, + 1750.0, + 529.0, + 1783.0, + 294.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 675.0, + 1406.0, + 675.0, + 1406.0, + 709.0, + 292.0, + 709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 705.0, + 1405.0, + 705.0, + 1405.0, + 740.0, + 293.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 732.0, + 1404.0, + 732.0, + 1404.0, + 773.0, + 292.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 765.0, + 1407.0, + 765.0, + 1407.0, + 804.0, + 292.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 795.0, + 351.0, + 795.0, + 351.0, + 834.0, + 293.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 795.0, + 857.0, + 795.0, + 857.0, + 834.0, + 471.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 795.0, + 1062.0, + 795.0, + 1062.0, + 834.0, + 977.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 834.0, + 1118.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 828.0, + 691.0, + 828.0, + 691.0, + 861.0, + 293.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 828.0, + 1211.0, + 828.0, + 1211.0, + 861.0, + 812.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 828.0, + 1405.0, + 828.0, + 1405.0, + 861.0, + 1331.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 857.0, + 429.0, + 857.0, + 429.0, + 893.0, + 293.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 857.0, + 724.0, + 857.0, + 724.0, + 893.0, + 581.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 857.0, + 1405.0, + 857.0, + 1405.0, + 893.0, + 851.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 888.0, + 1181.0, + 888.0, + 1181.0, + 922.0, + 293.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1794.0, + 1333.0, + 1794.0, + 1333.0, + 1829.0, + 294.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1826.0, + 1406.0, + 1826.0, + 1406.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1858.0, + 1406.0, + 1858.0, + 1406.0, + 1891.0, + 296.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1888.0, + 1009.0, + 1888.0, + 1009.0, + 1921.0, + 294.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 979.0, + 1018.0, + 979.0, + 1018.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 954, + 1404, + 954, + 1404, + 1137, + 299, + 1137 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 580, + 1404, + 580, + 1404, + 733, + 299, + 733 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1251, + 1403, + 1251, + 1403, + 1343, + 299, + 1343 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 300, + 747, + 1402, + 747, + 1402, + 841, + 300, + 841 + ], + "score": 0.964 + }, + { + "category_id": 3, + "poly": [ + 307, + 234, + 1396, + 234, + 1396, + 425, + 307, + 425 + ], + "score": 0.953 + }, + { + "category_id": 0, + "poly": [ + 300, + 885, + 544, + 885, + 544, + 920, + 300, + 920 + ], + "score": 0.903 + }, + { + "category_id": 0, + "poly": [ + 301, + 1185, + 607, + 1185, + 607, + 1217, + 301, + 1217 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 815, + 76, + 815, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 0, + "poly": [ + 300, + 1390, + 487, + 1390, + 487, + 1423, + 300, + 1423 + ], + "score": 0.882 + }, + { + "category_id": 4, + "poly": [ + 299, + 433, + 1392, + 433, + 1392, + 466, + 299, + 466 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.772 + }, + { + "category_id": 1, + "poly": [ + 293, + 1439, + 1408, + 1439, + 1408, + 2038, + 293, + 2038 + ], + "score": 0.768 + }, + { + "category_id": 1, + "poly": [ + 294, + 523, + 1340, + 523, + 1340, + 555, + 294, + 555 + ], + "score": 0.428 + }, + { + "category_id": 0, + "poly": [ + 294, + 523, + 1340, + 523, + 1340, + 555, + 294, + 555 + ], + "score": 0.306 + }, + { + "category_id": 13, + "poly": [ + 1319, + 641, + 1394, + 641, + 1394, + 671, + 1319, + 671 + ], + "score": 0.8, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 823, + 750, + 899, + 750, + 899, + 778, + 823, + 778 + ], + "score": 0.72, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 729, + 703, + 804, + 703, + 804, + 730, + 729, + 730 + ], + "score": 0.63, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 634, + 780, + 709, + 780, + 709, + 810, + 634, + 810 + ], + "score": 0.5, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 294.0, + 527.0, + 294.0, + 527.0, + 336.0, + 506.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 386.0, + 516.0, + 386.0, + 516.0, + 423.0, + 366.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 395.0, + 627.0, + 395.0, + 627.0, + 431.0, + 547.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 386.0, + 802.0, + 386.0, + 802.0, + 423.0, + 648.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 385.0, + 1038.0, + 385.0, + 1038.0, + 420.0, + 922.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 398.0, + 1164.0, + 398.0, + 1164.0, + 424.0, + 1088.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 385.0, + 1324.0, + 385.0, + 1324.0, + 419.0, + 1204.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 881.0, + 548.0, + 881.0, + 548.0, + 929.0, + 291.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1182.0, + 612.0, + 1182.0, + 612.0, + 1225.0, + 297.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1392.0, + 489.0, + 1392.0, + 489.0, + 1425.0, + 297.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 429.0, + 1395.0, + 429.0, + 1395.0, + 470.0, + 302.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1343.0, + 520.0, + 1343.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 952.0, + 1405.0, + 952.0, + 1405.0, + 989.0, + 294.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 984.0, + 1404.0, + 984.0, + 1404.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1051.0, + 294.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1046.0, + 1405.0, + 1046.0, + 1405.0, + 1080.0, + 293.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1076.0, + 1406.0, + 1076.0, + 1406.0, + 1111.0, + 294.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1107.0, + 629.0, + 1107.0, + 629.0, + 1141.0, + 294.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 579.0, + 1405.0, + 579.0, + 1405.0, + 616.0, + 297.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 609.0, + 1405.0, + 609.0, + 1405.0, + 646.0, + 295.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 639.0, + 1318.0, + 639.0, + 1318.0, + 672.0, + 295.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 639.0, + 1404.0, + 639.0, + 1404.0, + 672.0, + 1395.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 672.0, + 1405.0, + 672.0, + 1405.0, + 705.0, + 297.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 701.0, + 728.0, + 701.0, + 728.0, + 737.0, + 293.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 701.0, + 1145.0, + 701.0, + 1145.0, + 737.0, + 805.0, + 737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1251.0, + 1404.0, + 1251.0, + 1404.0, + 1285.0, + 297.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1284.0, + 1404.0, + 1284.0, + 1404.0, + 1314.0, + 297.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1311.0, + 1059.0, + 1311.0, + 1059.0, + 1346.0, + 294.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 743.0, + 822.0, + 743.0, + 822.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 743.0, + 1403.0, + 743.0, + 1403.0, + 786.0, + 900.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 777.0, + 633.0, + 777.0, + 633.0, + 813.0, + 294.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 777.0, + 1404.0, + 777.0, + 1404.0, + 813.0, + 710.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 809.0, + 439.0, + 809.0, + 439.0, + 844.0, + 295.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1443.0, + 1404.0, + 1443.0, + 1404.0, + 1475.0, + 296.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1469.0, + 1405.0, + 1469.0, + 1405.0, + 1509.0, + 320.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1503.0, + 1282.0, + 1503.0, + 1282.0, + 1539.0, + 323.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 1404.0, + 1553.0, + 1404.0, + 1589.0, + 295.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1586.0, + 1404.0, + 1586.0, + 1404.0, + 1618.0, + 323.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1615.0, + 745.0, + 1615.0, + 745.0, + 1647.0, + 323.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1701.0, + 295.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1696.0, + 1404.0, + 1696.0, + 1404.0, + 1732.0, + 323.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1747.0, + 1405.0, + 1747.0, + 1405.0, + 1783.0, + 296.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1778.0, + 772.0, + 1778.0, + 772.0, + 1814.0, + 325.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1826.0, + 1405.0, + 1826.0, + 1405.0, + 1868.0, + 292.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1859.0, + 1405.0, + 1859.0, + 1405.0, + 1898.0, + 323.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1891.0, + 485.0, + 1891.0, + 485.0, + 1925.0, + 324.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1940.0, + 1401.0, + 1940.0, + 1401.0, + 1979.0, + 296.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2010.0, + 322.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 2004.0, + 826.0, + 2004.0, + 826.0, + 2037.0, + 325.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 520.0, + 1343.0, + 520.0, + 1343.0, + 557.0, + 294.0, + 557.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 816, + 75, + 816, + 105, + 299, + 105 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2112, + 836, + 2112 + ], + "score": 0.842 + }, + { + "category_id": 1, + "poly": [ + 294, + 1415, + 1402, + 1415, + 1402, + 1481, + 294, + 1481 + ], + "score": 0.728 + }, + { + "category_id": 1, + "poly": [ + 292, + 1332, + 1403, + 1332, + 1403, + 1398, + 292, + 1398 + ], + "score": 0.727 + }, + { + "category_id": 1, + "poly": [ + 301, + 1498, + 1402, + 1498, + 1402, + 1594, + 301, + 1594 + ], + "score": 0.721 + }, + { + "category_id": 1, + "poly": [ + 298, + 1726, + 1404, + 1726, + 1404, + 1821, + 298, + 1821 + ], + "score": 0.703 + }, + { + "category_id": 1, + "poly": [ + 293, + 1249, + 1402, + 1249, + 1402, + 1315, + 293, + 1315 + ], + "score": 0.695 + }, + { + "category_id": 1, + "poly": [ + 295, + 652, + 1397, + 652, + 1397, + 717, + 295, + 717 + ], + "score": 0.687 + }, + { + "category_id": 1, + "poly": [ + 298, + 1136, + 1403, + 1136, + 1403, + 1230, + 298, + 1230 + ], + "score": 0.634 + }, + { + "category_id": 1, + "poly": [ + 299, + 1022, + 1401, + 1022, + 1401, + 1119, + 299, + 1119 + ], + "score": 0.623 + }, + { + "category_id": 1, + "poly": [ + 301, + 229, + 1400, + 229, + 1400, + 324, + 301, + 324 + ], + "score": 0.618 + }, + { + "category_id": 1, + "poly": [ + 294, + 910, + 1403, + 910, + 1403, + 1006, + 294, + 1006 + ], + "score": 0.612 + }, + { + "category_id": 1, + "poly": [ + 297, + 1612, + 1404, + 1612, + 1404, + 1707, + 297, + 1707 + ], + "score": 0.603 + }, + { + "category_id": 1, + "poly": [ + 296, + 735, + 1403, + 735, + 1403, + 891, + 296, + 891 + ], + "score": 0.568 + }, + { + "category_id": 1, + "poly": [ + 296, + 456, + 1399, + 456, + 1399, + 520, + 296, + 520 + ], + "score": 0.564 + }, + { + "category_id": 1, + "poly": [ + 300, + 343, + 1397, + 343, + 1397, + 437, + 300, + 437 + ], + "score": 0.556 + }, + { + "category_id": 1, + "poly": [ + 295, + 539, + 1400, + 539, + 1400, + 633, + 295, + 633 + ], + "score": 0.543 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 831.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1410.0, + 1404.0, + 1410.0, + 1404.0, + 1456.0, + 292.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1448.0, + 946.0, + 1448.0, + 946.0, + 1482.0, + 324.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1372.0, + 291.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1365.0, + 897.0, + 1365.0, + 897.0, + 1399.0, + 321.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1498.0, + 1406.0, + 1498.0, + 1406.0, + 1536.0, + 294.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1528.0, + 1406.0, + 1528.0, + 1406.0, + 1569.0, + 321.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1562.0, + 1316.0, + 1562.0, + 1316.0, + 1596.0, + 324.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1725.0, + 1405.0, + 1725.0, + 1405.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1754.0, + 1406.0, + 1754.0, + 1406.0, + 1795.0, + 318.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1783.0, + 398.0, + 1783.0, + 398.0, + 1824.0, + 320.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1246.0, + 1405.0, + 1246.0, + 1405.0, + 1289.0, + 292.0, + 1289.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1282.0, + 720.0, + 1282.0, + 720.0, + 1316.0, + 320.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 648.0, + 1402.0, + 648.0, + 1402.0, + 692.0, + 293.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 683.0, + 800.0, + 683.0, + 800.0, + 719.0, + 321.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1134.0, + 1403.0, + 1134.0, + 1403.0, + 1175.0, + 295.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1168.0, + 1404.0, + 1168.0, + 1404.0, + 1202.0, + 323.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1193.0, + 409.0, + 1193.0, + 409.0, + 1236.0, + 321.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1024.0, + 1403.0, + 1024.0, + 1403.0, + 1058.0, + 295.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1056.0, + 1403.0, + 1056.0, + 1403.0, + 1090.0, + 325.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1085.0, + 929.0, + 1085.0, + 929.0, + 1118.0, + 323.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 229.0, + 1404.0, + 229.0, + 1404.0, + 267.0, + 296.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 296.0, + 322.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 291.0, + 821.0, + 291.0, + 821.0, + 325.0, + 323.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 911.0, + 1405.0, + 911.0, + 1405.0, + 946.0, + 296.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 941.0, + 1403.0, + 941.0, + 1403.0, + 975.0, + 322.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 972.0, + 1008.0, + 972.0, + 1008.0, + 1007.0, + 320.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1610.0, + 1404.0, + 1610.0, + 1404.0, + 1650.0, + 292.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1642.0, + 1404.0, + 1642.0, + 1404.0, + 1679.0, + 321.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1675.0, + 467.0, + 1675.0, + 467.0, + 1707.0, + 322.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 737.0, + 1404.0, + 737.0, + 1404.0, + 770.0, + 298.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 804.0, + 321.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 794.0, + 1405.0, + 794.0, + 1405.0, + 833.0, + 321.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 822.0, + 1405.0, + 822.0, + 1405.0, + 869.0, + 319.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 854.0, + 588.0, + 854.0, + 588.0, + 892.0, + 320.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 456.0, + 1405.0, + 456.0, + 1405.0, + 492.0, + 294.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 487.0, + 1094.0, + 487.0, + 1094.0, + 521.0, + 325.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 341.0, + 1401.0, + 341.0, + 1401.0, + 379.0, + 294.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 372.0, + 1405.0, + 372.0, + 1405.0, + 411.0, + 321.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 406.0, + 1169.0, + 406.0, + 1169.0, + 436.0, + 325.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 1405.0, + 536.0, + 1405.0, + 576.0, + 294.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 569.0, + 1405.0, + 569.0, + 1405.0, + 604.0, + 322.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 603.0, + 707.0, + 603.0, + 707.0, + 633.0, + 320.0, + 633.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 369, + 1332, + 1331, + 1332, + 1331, + 1763, + 369, + 1763 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 300, + 957, + 1403, + 957, + 1403, + 1050, + 300, + 1050 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 368, + 447, + 1404, + 447, + 1404, + 858, + 368, + 858 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 300, + 1150, + 1404, + 1150, + 1404, + 1241, + 300, + 1241 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 357, + 1404, + 357, + 1404, + 418, + 297, + 418 + ], + "score": 0.943 + }, + { + "category_id": 1, + "poly": [ + 297, + 1972, + 1402, + 1972, + 1402, + 2034, + 297, + 2034 + ], + "score": 0.938 + }, + { + "category_id": 4, + "poly": [ + 542, + 1801, + 1155, + 1801, + 1155, + 1833, + 542, + 1833 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 815, + 75, + 815, + 105, + 300, + 105 + ], + "score": 0.896 + }, + { + "category_id": 0, + "poly": [ + 301, + 228, + 449, + 228, + 449, + 261, + 301, + 261 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 862, + 2088, + 862, + 2112, + 836, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 1, + "poly": [ + 297, + 1258, + 1109, + 1258, + 1109, + 1291, + 297, + 1291 + ], + "score": 0.831 + }, + { + "category_id": 0, + "poly": [ + 299, + 898, + 440, + 898, + 440, + 928, + 299, + 928 + ], + "score": 0.762 + }, + { + "category_id": 0, + "poly": [ + 300, + 1092, + 452, + 1092, + 452, + 1122, + 300, + 1122 + ], + "score": 0.707 + }, + { + "category_id": 1, + "poly": [ + 300, + 1912, + 776, + 1912, + 776, + 1944, + 300, + 1944 + ], + "score": 0.631 + }, + { + "category_id": 0, + "poly": [ + 302, + 299, + 657, + 299, + 657, + 329, + 302, + 329 + ], + "score": 0.505 + }, + { + "category_id": 1, + "poly": [ + 302, + 299, + 657, + 299, + 657, + 329, + 302, + 329 + ], + "score": 0.458 + }, + { + "category_id": 0, + "poly": [ + 300, + 1912, + 776, + 1912, + 776, + 1944, + 300, + 1944 + ], + "score": 0.274 + }, + { + "category_id": 1, + "poly": [ + 300, + 1092, + 452, + 1092, + 452, + 1122, + 300, + 1122 + ], + "score": 0.169 + }, + { + "category_id": 1, + "poly": [ + 299, + 898, + 440, + 898, + 440, + 928, + 299, + 928 + ], + "score": 0.13 + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1323.0, + 1348.0, + 1323.0, + 1348.0, + 1480.0, + 1063.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1211.0, + 1399.0, + 1218.0, + 1399.0, + 1218.0, + 1408.0, + 1211.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1541.0, + 628.0, + 1541.0, + 628.0, + 1606.0, + 602.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1729.0, + 648.0, + 1729.0, + 648.0, + 1764.0, + 516.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1728.0, + 1249.0, + 1728.0, + 1249.0, + 1767.0, + 1012.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1797.0, + 1155.0, + 1797.0, + 1155.0, + 1838.0, + 541.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 225.0, + 454.0, + 225.0, + 454.0, + 268.0, + 296.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 899.0, + 443.0, + 899.0, + 443.0, + 931.0, + 296.0, + 931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1091.0, + 456.0, + 1091.0, + 456.0, + 1125.0, + 297.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 298.0, + 658.0, + 298.0, + 658.0, + 333.0, + 297.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 780.0, + 1912.0, + 780.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 955.0, + 1407.0, + 955.0, + 1407.0, + 993.0, + 295.0, + 993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 987.0, + 1404.0, + 987.0, + 1404.0, + 1022.0, + 294.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1019.0, + 1289.0, + 1019.0, + 1289.0, + 1053.0, + 296.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 446.0, + 1404.0, + 446.0, + 1404.0, + 483.0, + 371.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 479.0, + 1405.0, + 479.0, + 1405.0, + 514.0, + 395.0, + 514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 509.0, + 595.0, + 509.0, + 595.0, + 543.0, + 393.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 553.0, + 1403.0, + 553.0, + 1403.0, + 587.0, + 368.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 585.0, + 1226.0, + 585.0, + 1226.0, + 619.0, + 393.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 628.0, + 1405.0, + 628.0, + 1405.0, + 662.0, + 389.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 659.0, + 1404.0, + 659.0, + 1404.0, + 693.0, + 394.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 689.0, + 1403.0, + 689.0, + 1403.0, + 724.0, + 394.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 716.0, + 1406.0, + 716.0, + 1406.0, + 761.0, + 391.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 763.0, + 1403.0, + 763.0, + 1403.0, + 797.0, + 385.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 796.0, + 1404.0, + 796.0, + 1404.0, + 831.0, + 395.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 822.0, + 635.0, + 822.0, + 635.0, + 862.0, + 395.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1149.0, + 1405.0, + 1149.0, + 1405.0, + 1186.0, + 295.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1218.0, + 293.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 591.0, + 1211.0, + 591.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 355.0, + 1405.0, + 355.0, + 1405.0, + 391.0, + 295.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 385.0, + 1308.0, + 385.0, + 1308.0, + 421.0, + 294.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1971.0, + 1402.0, + 1971.0, + 1402.0, + 2007.0, + 296.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1404.0, + 2000.0, + 1404.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1255.0, + 1112.0, + 1255.0, + 1112.0, + 1295.0, + 294.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 780.0, + 1912.0, + 780.0, + 1948.0, + 296.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 298.0, + 658.0, + 298.0, + 658.0, + 333.0, + 297.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1091.0, + 456.0, + 1091.0, + 456.0, + 1125.0, + 297.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 899.0, + 443.0, + 899.0, + 443.0, + 931.0, + 296.0, + 931.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 308, + 1404, + 308, + 1404, + 431, + 298, + 431 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 301, + 617, + 1303, + 617, + 1303, + 1299, + 301, + 1299 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 292, + 229, + 1401, + 229, + 1401, + 293, + 292, + 293 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 302, + 445, + 1398, + 445, + 1398, + 539, + 302, + 539 + ], + "score": 0.946 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 815, + 75, + 815, + 105, + 299, + 105 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.84 + }, + { + "category_id": 0, + "poly": [ + 301, + 573, + 836, + 573, + 836, + 606, + 301, + 606 + ], + "score": 0.816 + }, + { + "category_id": 13, + "poly": [ + 727, + 1045, + 874, + 1045, + 874, + 1083, + 727, + 1083 + ], + "score": 0.93, + "latex": "\\{ g _ { 0 } ^ { \\prime } , \\cdots , g _ { p } ^ { \\prime } \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 397, + 421, + 397, + 421, + 434, + 297, + 434 + ], + "score": 0.93, + "latex": "g _ { t } ^ { d e } = g ( { \\dot { t } } )" + }, + { + "category_id": 13, + "poly": [ + 748, + 230, + 825, + 230, + 825, + 264, + 748, + 264 + ], + "score": 0.93, + "latex": "g ( t ; \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 413, + 739, + 611, + 739, + 611, + 769, + 413, + 769 + ], + "score": 0.92, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 441, + 1011, + 533, + 1011, + 533, + 1049, + 441, + 1049 + ], + "score": 0.92, + "latex": "\\bar { g _ { p } ^ { \\prime } } \\in G ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1006, + 371, + 1045, + 371, + 1045, + 402, + 1006, + 402 + ], + "score": 0.9, + "latex": "g _ { t } ^ { a c }" + }, + { + "category_id": 13, + "poly": [ + 911, + 836, + 964, + 836, + 964, + 862, + 911, + 862 + ], + "score": 0.89, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 818, + 985, + 849, + 985, + 849, + 1012, + 818, + 1012 + ], + "score": 0.88, + "latex": "E _ { i }" + }, + { + "category_id": 13, + "poly": [ + 630, + 951, + 898, + 951, + 898, + 984, + 630, + 984 + ], + "score": 0.87, + "latex": "\\left( s _ { t } | g _ { t } , a _ { t } , r _ { t } , s _ { t + 1 } | g _ { t + 1 } \\right)" + }, + { + "category_id": 13, + "poly": [ + 469, + 1079, + 687, + 1079, + 687, + 1113, + 469, + 1113 + ], + "score": 0.87, + "latex": "t = \\{ 0 , \\cdots , p - 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 1142, + 836, + 1195, + 836, + 1195, + 863, + 1142, + 863 + ], + "score": 0.85, + "latex": "g _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 413, + 892, + 612, + 892, + 612, + 921, + 413, + 921 + ], + "score": 0.85, + "latex": "t = 0 , \\cdots , T - 1" + }, + { + "category_id": 13, + "poly": [ + 1170, + 982, + 1201, + 982, + 1201, + 1009, + 1170, + 1009 + ], + "score": 0.85, + "latex": "G ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 399, + 922, + 627, + 922, + 627, + 954, + 399, + 954 + ], + "score": 0.84, + "latex": "r _ { t } : = r ( s _ { t } , a _ { t } , g _ { t + 1 } )" + }, + { + "category_id": 13, + "poly": [ + 611, + 836, + 638, + 836, + 638, + 860, + 611, + 860 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 466, + 447, + 542, + 447, + 542, + 475, + 466, + 475 + ], + "score": 0.84, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 621, + 712, + 650, + 712, + 650, + 739, + 621, + 739 + ], + "score": 0.84, + "latex": "g _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 454, + 1113, + 713, + 1113, + 713, + 1145, + 454, + 1145 + ], + "score": 0.83, + "latex": "r _ { t } ^ { \\prime } : = r ( s _ { i , t } , a _ { i , t } , g _ { t + 1 } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 597, + 774, + 624, + 774, + 624, + 799, + 597, + 799 + ], + "score": 0.83, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 903, + 447, + 980, + 447, + 980, + 475, + 903, + 475 + ], + "score": 0.83, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 476, + 677, + 639, + 677, + 639, + 708, + 476, + 708 + ], + "score": 0.82, + "latex": "\\mathbf { \\Omega } = 1 , 2 , \\cdots , M" + }, + { + "category_id": 13, + "poly": [ + 674, + 1140, + 1040, + 1140, + 1040, + 1176, + 674, + 1176 + ], + "score": 0.82, + "latex": "\\left( s _ { i , t } | g _ { t } ^ { \\prime } , a _ { i , t } , r _ { t } ^ { \\prime } , s _ { i , t + 1 } | g _ { t + 1 } ^ { \\prime } \\right) \\mathrm { i n } R" + }, + { + "category_id": 13, + "poly": [ + 863, + 715, + 891, + 715, + 891, + 738, + 863, + 738 + ], + "score": 0.81, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 932, + 952, + 957, + 952, + 957, + 979, + 932, + 979 + ], + "score": 0.8, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 475, + 478, + 550, + 478, + 550, + 506, + 475, + 506 + ], + "score": 0.79, + "latex": "\\mathrm { H E R + }" + }, + { + "category_id": 13, + "poly": [ + 673, + 648, + 698, + 648, + 698, + 674, + 673, + 674 + ], + "score": 0.75, + "latex": "R" + }, + { + "category_id": 14, + "poly": [ + 400, + 798, + 562, + 798, + 562, + 832, + 400, + 832 + ], + "score": 0.74, + "latex": "a _ { t } \\doteq \\pi ( s _ { t } | g _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 905, + 618, + 929, + 618, + 929, + 644, + 905, + 644 + ], + "score": 0.74, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 679, + 372, + 693, + 372, + 693, + 396, + 679, + 396 + ], + "score": 0.59, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1142, + 1045, + 1142, + 1045, + 1170, + 1019, + 1170 + ], + "score": 0.57, + "latex": "R" + }, + { + "category_id": 13, + "poly": [ + 953, + 1045, + 1090, + 1045, + 1090, + 1080, + 953, + 1080 + ], + "score": 0.56, + "latex": "g _ { t } ^ { \\prime } = g ( t ; \\gamma )" + }, + { + "category_id": 13, + "poly": [ + 1141, + 624, + 1158, + 624, + 1158, + 644, + 1141, + 644 + ], + "score": 0.41, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 475, + 678, + 516, + 678, + 516, + 706, + 475, + 706 + ], + "score": 0.41, + "latex": "= 1" + }, + { + "category_id": 13, + "poly": [ + 1109, + 1046, + 1198, + 1046, + 1198, + 1079, + 1109, + 1079 + ], + "score": 0.4, + "latex": "\\mathrm { ( H E R + ) }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 817.0, + 73.0, + 817.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 568.0, + 842.0, + 568.0, + 842.0, + 612.0, + 295.0, + 612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 309.0, + 1406.0, + 309.0, + 1406.0, + 342.0, + 294.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 337.0, + 1405.0, + 337.0, + 1405.0, + 374.0, + 292.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 365.0, + 678.0, + 365.0, + 678.0, + 407.0, + 291.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 365.0, + 1005.0, + 365.0, + 1005.0, + 407.0, + 694.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 365.0, + 1406.0, + 365.0, + 1406.0, + 407.0, + 1046.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 395.0, + 296.0, + 395.0, + 296.0, + 439.0, + 292.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 395.0, + 1316.0, + 395.0, + 1316.0, + 439.0, + 422.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 614.0, + 904.0, + 614.0, + 904.0, + 650.0, + 297.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 614.0, + 1140.0, + 614.0, + 1140.0, + 650.0, + 930.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 647.0, + 672.0, + 647.0, + 672.0, + 676.0, + 310.0, + 676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 674.0, + 474.0, + 674.0, + 474.0, + 710.0, + 307.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 674.0, + 680.0, + 674.0, + 680.0, + 710.0, + 640.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 707.0, + 341.0, + 707.0, + 341.0, + 740.0, + 307.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 703.0, + 620.0, + 703.0, + 620.0, + 744.0, + 367.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 703.0, + 862.0, + 703.0, + 862.0, + 744.0, + 651.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 703.0, + 895.0, + 703.0, + 895.0, + 744.0, + 892.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 737.0, + 341.0, + 737.0, + 341.0, + 771.0, + 307.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 736.0, + 412.0, + 736.0, + 412.0, + 768.0, + 368.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 736.0, + 653.0, + 736.0, + 653.0, + 768.0, + 612.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 767.0, + 339.0, + 767.0, + 339.0, + 801.0, + 307.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 767.0, + 596.0, + 767.0, + 596.0, + 803.0, + 397.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 625.0, + 767.0, + 1030.0, + 767.0, + 1030.0, + 803.0, + 625.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 800.0, + 338.0, + 800.0, + 338.0, + 830.0, + 308.0, + 830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 800.0, + 567.0, + 800.0, + 567.0, + 833.0, + 563.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 828.0, + 339.0, + 828.0, + 339.0, + 863.0, + 307.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 824.0, + 610.0, + 824.0, + 610.0, + 867.0, + 394.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 824.0, + 910.0, + 824.0, + 910.0, + 867.0, + 639.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 824.0, + 1141.0, + 824.0, + 1141.0, + 867.0, + 965.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 860.0, + 338.0, + 860.0, + 338.0, + 891.0, + 308.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 857.0, + 467.0, + 857.0, + 467.0, + 892.0, + 369.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 891.0, + 339.0, + 891.0, + 339.0, + 923.0, + 306.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 369.0, + 889.0, + 412.0, + 889.0, + 412.0, + 923.0, + 369.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 889.0, + 652.0, + 889.0, + 652.0, + 923.0, + 613.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 921.0, + 339.0, + 921.0, + 339.0, + 952.0, + 298.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 922.0, + 631.0, + 922.0, + 631.0, + 956.0, + 628.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 951.0, + 339.0, + 951.0, + 339.0, + 982.0, + 297.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 945.0, + 629.0, + 945.0, + 629.0, + 987.0, + 395.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 945.0, + 931.0, + 945.0, + 931.0, + 987.0, + 899.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 945.0, + 1300.0, + 945.0, + 1300.0, + 987.0, + 958.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 981.0, + 339.0, + 981.0, + 339.0, + 1013.0, + 297.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 980.0, + 817.0, + 980.0, + 817.0, + 1016.0, + 398.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 980.0, + 1169.0, + 980.0, + 1169.0, + 1016.0, + 850.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1008.0, + 342.0, + 1008.0, + 342.0, + 1046.0, + 297.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1009.0, + 440.0, + 1009.0, + 440.0, + 1046.0, + 396.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 1009.0, + 576.0, + 1009.0, + 576.0, + 1046.0, + 534.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1046.0, + 339.0, + 1046.0, + 339.0, + 1078.0, + 298.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1041.0, + 726.0, + 1041.0, + 726.0, + 1083.0, + 423.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 1041.0, + 952.0, + 1041.0, + 952.0, + 1083.0, + 875.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1041.0, + 1108.0, + 1041.0, + 1108.0, + 1083.0, + 1091.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1041.0, + 1204.0, + 1041.0, + 1204.0, + 1083.0, + 1199.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1080.0, + 339.0, + 1080.0, + 339.0, + 1112.0, + 297.0, + 1112.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1075.0, + 468.0, + 1075.0, + 468.0, + 1119.0, + 422.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1075.0, + 731.0, + 1075.0, + 731.0, + 1119.0, + 688.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1112.0, + 338.0, + 1112.0, + 338.0, + 1142.0, + 297.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1106.0, + 718.0, + 1106.0, + 718.0, + 1150.0, + 714.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1142.0, + 339.0, + 1142.0, + 339.0, + 1172.0, + 297.0, + 1172.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1137.0, + 673.0, + 1137.0, + 673.0, + 1177.0, + 450.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 1137.0, + 1049.0, + 1137.0, + 1049.0, + 1177.0, + 1046.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1172.0, + 339.0, + 1172.0, + 339.0, + 1204.0, + 297.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1170.0, + 522.0, + 1170.0, + 522.0, + 1204.0, + 423.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1200.0, + 341.0, + 1200.0, + 341.0, + 1234.0, + 295.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1200.0, + 496.0, + 1200.0, + 496.0, + 1234.0, + 396.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1231.0, + 341.0, + 1231.0, + 341.0, + 1265.0, + 293.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1230.0, + 466.0, + 1230.0, + 466.0, + 1264.0, + 368.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1262.0, + 438.0, + 1262.0, + 438.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 747.0, + 228.0, + 747.0, + 264.0, + 294.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 228.0, + 1402.0, + 228.0, + 1402.0, + 264.0, + 826.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1004.0, + 261.0, + 1004.0, + 297.0, + 296.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 446.0, + 465.0, + 446.0, + 465.0, + 480.0, + 298.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 446.0, + 902.0, + 446.0, + 902.0, + 480.0, + 543.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 446.0, + 1404.0, + 446.0, + 1404.0, + 480.0, + 981.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 475.0, + 474.0, + 475.0, + 474.0, + 513.0, + 296.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 475.0, + 1404.0, + 475.0, + 1404.0, + 513.0, + 551.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 507.0, + 839.0, + 507.0, + 839.0, + 542.0, + 296.0, + 542.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/HygrdpVKvr/HygrdpVKvr_content_list.json b/parse/train/HygrdpVKvr/HygrdpVKvr_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9582c9460615cdeb89714fcfb0d9a5742c8f112d --- /dev/null +++ b/parse/train/HygrdpVKvr/HygrdpVKvr_content_list.json @@ -0,0 +1,1647 @@ +[ + { + "type": "text", + "text": "NAS EVALUATION IS FRUSTRATINGLY HARD ", + "text_level": 1, + "bbox": [ + 174, + 101, + 705, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Antoine Yang∗ Ecole Polytechnique ´ † France ", + "bbox": [ + 184, + 145, + 325, + 188 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pedro M Esperanc¸a Huawei Noah’s Ark Lab London, UK ", + "bbox": [ + 382, + 145, + 545, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fabio Maria Carlucci Huawei Noah’s Ark Lab London, UK ", + "bbox": [ + 606, + 145, + 769, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 226, + 544, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Architecture Search (NAS) is an exciting new field which promises to be as much as a game-changer as Convolutional Neural Networks were in 2012. Despite many great works leading to substantial improvements on a variety of tasks, comparison between different methods is still very much an open issue. While most algorithms are tested on the same datasets, there is no shared experimental protocol followed by all. As such, and due to the under-use of ablation studies, there is a lack of clarity regarding why certain methods are more effective than others. Our first contribution is a benchmark of 8 NAS methods on 5 datasets. To overcome the hurdle of comparing methods with different search spaces, we propose using a method’s relative improvement over the randomly sampled average architecture, which effectively removes advantages arising from expertly engineered search spaces or training protocols. Surprisingly, we find that many NAS techniques struggle to significantly beat the average architecture baseline. We perform further experiments with the commonly used DARTS search space in order to understand the contribution of each component in the NAS pipeline. These experiments highlight that: (i) the use of tricks in the evaluation protocol has a predominant impact on the reported performance of architectures; (ii) the cell-based search space has a very narrow accuracy range, such that the seed has a considerable impact on architecture rankings; (iii) the hand-designed macrostructure (cells) is more important than the searched micro-structure (operations); and (iv) the depth-gap is a real phenomenon, evidenced by the change in rankings between 8 and 20 cell architectures. To conclude, we suggest best practices, that we hope will prove useful for the community and help mitigate current NAS pitfalls, e.g. difficulties in reproducibility and comparison of search methods. The code used is available at https://github.com/antoyang/NAS-Benchmark. ", + "bbox": [ + 232, + 256, + 764, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 628, + 336, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the deep learning revolution helped us move away from hand crafted features (Krizhevsky et al., 2012) and reach new heights (He et al., 2016; Szegedy et al., 2017), so does Neural Architecture Search (NAS) hold the promise of freeing us from hand-crafted architectures, which requires tedious and expensive tuning for each new task or dataset. Identifying the optimal architecture is indeed a key pillar of any Automated Machine Learning (AutoML) pipeline. Research in the last two years has proceeded at a rapid pace and many search strategies have been proposed, from Reinforcement Learning (Zoph & Le, 2017; Pham et al., 2018), to Evolutionary Algorithms (Real et al., 2017), to Gradient-based methods (Liu et al., 2019; Liang et al., 2019). Still, it remains unclear which approach and search algorithm is preferable. Typically, methods have been evaluated on accuracy alone, even though accuracy is influenced by many other factors besides the search algorithm. Comparison between published search algorithms for NAS is therefore either very difficult (complex training protocols with no code available) or simply impossible (different search spaces), as previously pointed out (Li & Talwalkar, 2019; Sciuto et al., 2019; Lindauer & Hutter, 2019). ", + "bbox": [ + 174, + 660, + 825, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NAS methods have been typically decomposed into three components (Elsken et al., 2019; Li & Talwalkar, 2019): search space, search strategy and model evaluation strategy. This division is important to keep in mind, as an improvement in any of these elements will lead to a better final performance. But is a method with a more (manually) tuned search space a better AutoML algorithm? If the key idea behind NAS is to find the optimal architecture, without human intervention, why are we devoting so much energy to infuse expert knowledge into the pipeline? Furthermore, the lack of ablation studies in most works makes it harder to pinpoint which components are instrumental to the final performance, which can easily lead to Hypothesizing After the Results are Known (HARKing; Gencoglu et al., 2019). ", + "bbox": [ + 178, + 847, + 825, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 825, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Paradoxically, the huge effort invested in finding better search spaces and training protocols, has led to a situation in which any randomly sampled architecture performs almost as well as those obtained by the search strategies. Our findings suggest that most of the gains in accuracy in recent contributions to NAS have come from manual improvements in the training protocol, not in the search algorithms. ", + "bbox": [ + 174, + 208, + 825, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a step towards understanding which methods are more effective, we have collected code for 8 reasonably fast (search time of less than 4 days) NAS algorithms, and benchmarked them on 5 well known CV datasets. Using a simple metric—the relative improvement over the average architecture of the search space—we find that most NAS methods perform very similarly and rarely substantially above this baseline. The methods used are DARTS, StacNAS, PDARTS, MANAS, CNAS, NSGANET, ENAS and NAO. The datasets used are CIFAR10, CIFAR100, SPORT8, MIT67 and FLOWERS102. ", + "bbox": [ + 174, + 285, + 825, + 382 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through a number of additional experiments on the widely used DARTS search space (Liu et al., 2019), we will show that: (a) how you train your model has a much bigger impact than the actual architecture chosen; (b) different architectures from the same search space perform very similarly, so much so that (c) hyperparameters, like the number of cells, or the seed itself have a very significant effect on the ranking; and (d) the specific operations themselves have less impact on the final accuracy than the hand-designed macro-structure of the network. Notably, we find that the $2 0 0 +$ architectures sampled from this search space (available from the link in the abstract) are all within a range of one percentage point (top-1 accuracy) after a standard full training on CIFAR10. Finally, we include some observations on how to foster reproducibility and a discussion on how to potentially avoid some of the encountered pitfalls. ", + "bbox": [ + 174, + 388, + 825, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 563, + 339, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As mentioned, NAS methods have the potential to truly revolutionize the field, but to do so it is crucial that future research avoids common mistakes. Some of these concerns have been recently raised by the community. ", + "bbox": [ + 178, + 603, + 821, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For example, Li & Talwalkar (2019) highlight that most NAS methods a) fail to compare against an adequate baseline, such as a properly implemented random search strategy, b) are overly complex, with no ablation to properly assign credit to the important components, and c) fail to provide all details needed for successfully reproducing their results. In our paper we go one step further and argue that the relative improvement over the average (randomly sampled) architecture is an useful tool to quantify the effectiveness of a proposed solution and compare it with competing methods. To partly answer their second point, and understand how much the final accuracy depends on the specific architecture, we implement an in-depth study of the widely employed DARTS (Liu et al., 2019) search space and perform an ablation on the commonly used training techniques (e.g. Cutout, DropPath, AutoAugment). ", + "bbox": [ + 174, + 652, + 825, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, Sciuto et al. (2019) also took the important step of systematically using fair baselines, and suggest random search with early stopping, averaged over multiple seeds, as an extremely competitive baseline. They find that the search spaces of three methods investigated (DARTS, ENAS, NAO) have been expertly engineered to the extent that any randomly selected architecture performs very well. In contrast, we show that even random sampling (without search) provides an incredibly competitive baseline. Our relative improvement metric allows us to isolate the contribution of the search strategy from the effects of the search space and training pipeline. Thus, we further confirm the authors’ claim, showing that indeed the average architecture performs extremely well and that how you train a model has more impact than any specific architecture. ", + "bbox": [ + 174, + 797, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 NAS BENCHMARK ", + "text_level": 1, + "bbox": [ + 176, + 102, + 362, + 118 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we present a systematic evaluation of 8 methods on 5 datasets using a strategy that is designed to reveal the quality of each method’s search strategy, removing the effect of the manuallyengineered training protocol and search space. The goal is to find general trends and highlight common features rather than just pin-pointing the most accurate algorithm. ", + "bbox": [ + 174, + 136, + 823, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Understanding why methods are effective is not an easy task: most introduce variations to previous search spaces, search strategies, and training protocols—with ablations disentangling the contribution of each component often incomplete or missing. In other words, how can we be sure that a new state-of-the-art method is not so simply due to a better engineered search space or training protocol? To address this issue we compare a set of 8 methods with randomly sampled architectures from their respective search spaces, and trained with the same protocol as the searched architectures. ", + "bbox": [ + 174, + 199, + 825, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ultimate goal behind NAS should be to return the optimal model for any dataset given, at least within the limits of a certain task, and we feel that the current practices of searching almost exclusively on CIFAR10 go against this principle. Indeed, to avoid the very concrete risk of overfitting to this set of data, NAS methods should be tested on a variety of tasks. For this reason we run experiments on 5 different datasets. ", + "bbox": [ + 174, + 290, + 825, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1 METHODOLOGY ", + "text_level": 1, + "bbox": [ + 176, + 381, + 326, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Criteria for dataset selection. We selected datasets to cover a variety of subtasks within image classification. In addition to the standard CIFAR10 we select CIFAR100 for a more challenging object classification problem (Krizhevsky, 2009); SPORT8 for action classification (Li & Fei-Fei, 2007); MIT67 for scene classification (Quattoni & Torralba, 2009); and FLOWERS102 for finegrained object classification (Nilsback & Zisserman, 2008). More details are given in the Appendix. ", + "bbox": [ + 174, + 410, + 825, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Criteria for method selection. We selected methods which (a) have open-source code, or provided it upon request, and (b) have a reasonable running time, specifically a search time under 4 GPU-days on CIFAR10. The selected methods are: DARTS (Liu et al., 2019), StacNAS (Li et al., 2019), PDARTS (Xin Chen, 2019), MANAS (Carlucci et al., 2019), CNAS (Weng et al., 2019), NSGANET (Lu et al., 2018), ENAS (Pham et al., 2018), and NAO (Luo et al., 2018). With the exception of NAO and NSGANET, all methods are DARTS variants and use weight sharing. ", + "bbox": [ + 174, + 500, + 825, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation protocol. NAS algorithms usually consist of two phases: (i) search, producing the best architecture according to the search algorithm used; (ii) augmentation, consisting in training from scratch the best model found in the search phase. We evaluate methods as follows: 1. Sample 8 architectures from the search space, uniformly at random, and use the method’s code to augment these architectures (same augment seed for all); 2. Use the method’s code to search for 8 architectures and augment them (different search seed, same augment seed); 3. Report mean and standard deviation of the top-1 test accuracy, obtained at the end of the augmentation, for both the randomly sampled and the searched architectures; ", + "bbox": [ + 174, + 603, + 825, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since both learned and randomly sampled architectures share the same search space and training protocol, calculating a relative improvement over this random baseline as $R I = 1 0 0 \\times ( A c c _ { m } -$ $A c c _ { r } ) / A c c _ { r }$ can offer insights into the quality of the search strategy alone. $A c c _ { m }$ and $A c c _ { r }$ represent the top-1 accuracy of the search method and random sampling strategies, respectively. A good, general-purpose NAS method is expected to yield $R I > 0$ consistently over different searches and across different subtasks. We emphasize that the comparison is not against random search, but rather against random sampling, i.e., the average architecture of the search space. For example, in the DARTS search space, for each edge in the graph that defines a cell we select one out of eight possible operations (e.g. pooling or convolutions) with uniform probability $1 / 8$ . ", + "bbox": [ + 174, + 722, + 825, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperparameters are optimized on CIFAR10, according to the values reported by the corresponding authors. Since most methods do not include their optimization as part of the search routine, we assumed them to be robust and generalizable to other tasks. As such, aside from scaling down the architecture depending on dataset size, experiments on other datasets use the same hyperparameters. Other training details and references are given in the Appendix. ", + "bbox": [ + 174, + 854, + 823, + 924 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/edf5f8a33f5bb147c60a4fc8d6e36c6306bb7276930131c944ec9914b79e4870.jpg", + "image_caption": [ + "Figure 1: Comparison of search methods and random sampling from their respective search spaces. Methods lying in the diagonal perform the same as the average architecture, while methods above the diagonal outperform it. See also Table 1. " + ], + "image_footnote": [], + "bbox": [ + 176, + 107, + 818, + 441 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3203460b279ee612939f03c89b01e0c7ce0f8773d113c8c7e58f41bc3d8a1d1d.jpg", + "image_caption": [ + "Figure 2: Performance and computational cost of the search phase on CIFAR10. " + ], + "image_footnote": [], + "bbox": [ + 173, + 513, + 459, + 699 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c01388e0b766ac927995cde360d93dfc9a5de1b355216979bcf6a2dd53aecde9.jpg", + "table_caption": [ + "Table 1: Relative improvement metric, $\\begin{array} { r l } { R I } & { { } = } \\end{array}$ $1 0 0 \\times ( A c c _ { m } - A c c _ { r } ) / A c c _ { r }$ (in $\\%$ ), where $A c c _ { m }$ and $A c c _ { r }$ are the accuracies of the search method and random sampling baseline, respectively. " + ], + "table_footnote": [], + "table_body": "
C10C100S8M67F102
DARTS0.320.23-0.130.100.25
PDARTS0.521.200.511.190.20
NSGANET-0.481.370.432.001.47
ENAS0.01-3.440.670.130.47
CNAS0.74-0.89-1.06-0.66-2.48
MANAS0.18-0.200.331.480.70
StacNAS NAO0.43 0.442.87 -0.010.38 -2.050.05 -1.53-0.16 -0.13
", + "bbox": [ + 498, + 584, + 820, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 751, + 279, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1 shows the evaluation results on the 5 datasets, from which we draw two main conclusions. First, the improvements over random sampling tend to be small. In some cases the average performance of a method is even below the average randomly sampled architecture, which suggests that the search methods are not converging to desirable architectures. Second, the small range of accuracies obtained hints at narrow search spaces, where even the worst architectures perform reasonably well. See Section 5 for more experiments corroborating this conclusion. ", + "bbox": [ + 173, + 776, + 823, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe also that, on CIFAR10, the top half of best-performing methods (PDARTS, MANAS, DARTS, StacNAS) all perform similarly and positively in relation to their respective search spaces, but more variance is seen on the other datasets. This could be explained by the fact that most methods’ hyperparameters have been optimized on CIFAR10 and might not generalize as well on different datasets. As a matter of fact, we found that all NAS methods neglect to report the time needed to optimize hyperparameters. In addition, Table 1 shows the relative improvement metric $R I$ (see intro to Section 3) for each method and dataset. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 823, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The computational cost of searching for architectures is a limiting factor in their applicability and, therefore, an important variable in the evaluation of NAS algorithms. Figure 2 shows the performance as well and the computational cost of the search phase on CIFAR10. ", + "bbox": [ + 174, + 152, + 821, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 COMPARISON OF TRAINING PROTOCOLS ", + "text_level": 1, + "bbox": [ + 174, + 226, + 537, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section we attempt to shed some light on the surprising results of the previous section. We noticed that there was a much larger differences between the random baselines of different methods than the actual increase in performance of each approach. We hypothesized that how a network is trained (the training protocol) has a larger impact on the final accuracy than which architecture is trained, for each search space. To test this, we performed sensitivity analysis using the most common performanceboosting training protocols. ", + "bbox": [ + 174, + 262, + 549, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 METHODOLOGY ", + "text_level": 1, + "bbox": [ + 176, + 429, + 326, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We decided to evaluate architectures from the commonly used DARTS search space (Liu et al., 2019) on the CIFAR10 dataset. We use the following process: 1) sample 8 random architectures, 2) train them with different training protocols (details below) and 3) report mean, standard deviation and maximum of the top-1 test accuracy at the end of the training process. ", + "bbox": [ + 174, + 458, + 549, + 556 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/caa55a31af78adab243a490ab57323412ef1ea1347232a5cd9f8350973a2321e.jpg", + "image_caption": [ + "Figure 3: Comparison of different augmentation protocols for the DARTS search space on CIFAR10. Same colored dots represent minimum and maximum accuracies in the 8 runs. " + ], + "image_footnote": [], + "bbox": [ + 562, + 262, + 820, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training protocols. The simplest training protocol, which we will call Base is similar to the one used in DARTS, but with all tricks disabled: the model is simply trained for 600 epochs. On the other extreme, our full protocol uses several tricks which have been used in recent works (Xie et al., 2019b; Nayman et al., 2019): Auxiliary Towers (A), DropPath (D; Larsson & Shakhnarovich, 2017), Cutout (C; DeVries & Taylor, 2017), AutoAugment (AA; Cubuk et al., 2018), extended training for 1500 epochs (1500E), and increased number of channels (50C). In between these two extremes, by selectively enabling and disabling each component, we evaluated a further 8 intermediate training protocols. When active, DropPath probability is 0.2, cutout length is 16, auxiliary tower weight is 0.4, and AutoAugment combined with Cutout are used after standard data pre-processing techniques previously described, as in Popien (2019). ", + "bbox": [ + 174, + 582, + 547, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 590, + 825, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2 RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 747, + 277, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3, a large difference of over 3 percentage points (p.p.) exists between the simplest and the most advanced training protocols. Indeed, this is much higher than any improvement over random sampling observed in the previous section: for example, on CIFAR10, the best improvement observed was 0.69 p.p. In other words, the training protocol is often far more important than the architecture used. Note that the best accuracy of the 8 random architectures training with the best protocol is $9 8 . 1 5 \\%$ , which is only 0.25 p.p. below state-of-the-art (Nayman et al., 2019). ", + "bbox": [ + 174, + 776, + 823, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To summarize, it seems that most recent state-of-the-art results, though impressive, cannot always be attributed to superior search strategies. Rather, they are often the result of expert knowledge applied to the evaluation protocol. In Figure 10 (Appendix A.3.1) we show similar results when training a ResNet-50 (He et al., 2016) with the same protocols. ", + "bbox": [ + 174, + 868, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f9ef12213677fc9a4669a42a87f74fd716606dd30dbac4f7517b2b24659054a0.jpg", + "image_caption": [ + "Figure 4: Training curves for the 214 randomly sampled architectures. Inset plot shows the histogram of accuracies at different epochs. " + ], + "image_footnote": [], + "bbox": [ + 178, + 106, + 480, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ab1d0af1522db09538fa5918d8022741f2b541c8dfb7d338233ba12f9eb1f4f.jpg", + "image_caption": [ + "Figure 5: Correlation between accuracies at different epochs and final accuracy, using raw and smoothed accuracies over a window $w$ . " + ], + "image_footnote": [], + "bbox": [ + 514, + 106, + 816, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 STUDY OF DARTS’ SEARCH SPACE ", + "text_level": 1, + "bbox": [ + 174, + 356, + 500, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 DISTRIBUTION OF THE RANDOM SAMPLING WITHIN DARTS SEARCH SPACE ", + "text_level": 1, + "bbox": [ + 174, + 398, + 741, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To better understand the results from the previous section, we sampled a considerable number of architectures (214) from the most commonly used search space (Liu et al., 2019) and fully trained them with the matching training protocol (Cutout+DropPath+Auxiliary Towers). This allows us to get a sense of how much variance exists between the different models (training statistics are made available at the link in the abstract). ", + "bbox": [ + 174, + 433, + 825, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As we can observe from Figure 4, architectures sampled from this search space all perform similarly, with a mean of $9 7 . 0 3 \\pm 0 . 2 3$ . The worst architecture we found had an accuracy of 96.18, while the best achieved 97.56. ", + "bbox": [ + 176, + 510, + 825, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To put this into perspective, many methods using the same training protocol, fall within (or very close to) the standard deviation of the average architecture. Furthermore, as we can observe in Figure 7, the number of cells (a human-picked hyperparameter) has a much larger impact on the final accuracy. ", + "bbox": [ + 174, + 559, + 823, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Figure 5 we used the training statistics of the 214 models to plot the correlation between test accuracies at different epochs: it grows slowly in an almost linear fashion. We note that using the moving average of the accuracies yields a stronger correlation, which could be useful for methods using early stopping. ", + "bbox": [ + 174, + 622, + 825, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2 OPERATIONS", + "text_level": 1, + "bbox": [ + 174, + 715, + 305, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test whether the results from the previous section were due to the choice of available operations, we developed an intentionally sub-optimal search space containing 4 plain convolutions $( 1 \\times 1 , 3 \\times 3$ , $7 \\times 7 .$ , $1 1 \\times 1 1$ ), 2 max pooling operators $( 3 \\times 3 , 5 \\times 5 )$ plus the none and skip connect operations. This proposed search space is clearly more parameter inefficient compared to the commonly used DARTS one (which uses both dilated and separable ones), and we expect it to perform worse. ", + "bbox": [ + 174, + 750, + 825, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We sampled 56 architectures from this new search space and trained them with the DARTS training protocol (Cutout $^ +$ DropPath $^ +$ Auxiliary Towers), for fair comparison with the results from the previous section. Figure 6 shows the resulting histogram, together with the one obtained from the classical DARTS space of operations. The two distributions are only shifted by 0.18 accuracy points. Given the minor difference in performance, the specific operations are not a key ingredient in the success of this search space. Very likely, it’s the well engineered cell structure that allows the model to perform as well as it does. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e91248ab0a2f0aabd9ab36f7d3f7c4ee456017c62591d2c5a5bc19fc63e0e6f.jpg", + "image_caption": [ + "Figure 6: Histograms of the final accuracies (600 epochs) for architectures sampled from the DARTS search space (214 models) and our modified version (56 models). " + ], + "image_footnote": [], + "bbox": [ + 178, + 106, + 480, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbe7e776ae2cf3008835cf33d721f2209476268f702144b2c37653630acca5b7.jpg", + "image_caption": [ + "Figure 7: Performance of 16 randomly sampled architectures with different numbers of cells. Error bars represent standard deviation. " + ], + "image_footnote": [], + "bbox": [ + 514, + 107, + 816, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0587d112298e61ac522ef6c9052d3064e3454621c12eee6f15cad59d794e6af4.jpg", + "image_caption": [ + "Figure 8: Changes in the ranking of different architectures when trained with two different seeds (A and B). " + ], + "image_footnote": [], + "bbox": [ + 179, + 348, + 482, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5f3c33ff2e3ac55ede03bb632eba1c2d1e204c46b832377bf24ea864a9fd54c4.jpg", + "image_caption": [ + "Figure 9: Changes in the ranking of different architectures when trained with different numbers of cells (same seed). " + ], + "image_footnote": [], + "bbox": [ + 514, + 348, + 816, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3 DOES CHANGING SEED AND NUMBER OF CELLS AFFECT RANKING? ", + "text_level": 1, + "bbox": [ + 176, + 597, + 678, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A necessary practice for many weight-sharing methods (Liu et al., 2019; Pham et al., 2018) is to restart the training from scratch after the search phase, with a different number of cells. Recent works have warned that this procedure might negatively affect ranking; similarly, the role of the seed has been previously recognized as as a fundamental element in reproducibility (Li & Talwalkar, 2019; Sciuto et al., 2019). ", + "bbox": [ + 174, + 623, + 825, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test the impact of seed, we randomly sampled 32 architectures and trained them with two different seeds (Figure 8). Ranking is heavily influenced, as the Kendall tau correlation between the two sets of training is 0.48. On average, the test accuracy changes by $0 . 1 3 \\% \\pm 0 . 0 8$ (max change is $0 . 3 9 \\%$ , which is substantial considering the small gap between random architectures and NAS methods. ", + "bbox": [ + 173, + 700, + 825, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test the depth-gap we trained another 32 with different number of cells (Figure 9). The correlation between the two different depths is not very strong as measured by Kendall Tau (0.54), with architectures shifting up and down the rankings by up to 18 positions (out of 32). Methods employing weight sharing (WS) would see an even more pronounced effect as the architectures normally chosen at 8 cells would have been training sub-optimally due to the WS itself (Sciuto et al., 2019). ", + "bbox": [ + 174, + 762, + 823, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These findings point towards two issues. The first is that since the seed has such a large effect on ranking, it stands to reason that the final accuracy reported should be averaged over multiple seeds. The second is that, if the lottery ticket hypothesis holds—so that specific sub-networks are better mainly due to their lucky initialization; Frankle & Carbin., 2018—together with our findings, this could be an additional reason why methods searching on a different number of cells than the final model, struggle to significantly improve on the average randomly sampled architecture. ", + "bbox": [ + 173, + 840, + 825, + 924 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6 DISCUSSION AND BEST PRACTICES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 495, + 117 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section we offer some suggestions on how to mitigate the issues in NAS research. ", + "bbox": [ + 171, + 132, + 751, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Augmention tricks: while achieving higher accuracies is clearly a desirable goal, we have shown in section 4, that using well engineered training protocols can hide the contribution of the search algorithm. We therefore suggest that both results, with and without training tricks, should be reported. An example of best practice is found in Hundt et al. (2019). ", + "bbox": [ + 174, + 154, + 823, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Search Space: it is difficult to evaluate the effectiveness of any given proposed method without a measure of how good randomly sampled architectures are. This is not the same thing as performing a random search which is a search strategy in itself; random sampling is simply used to establish how good the average model is. A simple approach to measure the variability of any new given search space could be to randomly sample $k$ architectures and report mean and standard deviation. We hope that future works will attempt to develop more expressive search spaces, capable of producing both good and bad network designs. Restricted search spaces, while guaranteeing good performance and quick results, will inevitably be constrained by the bounds of expert knowledge (local optima) and will be incapable of reaching more truly innovative solutions (closer to the global optima). As our findings in section 5.2 suggest, the overall wiring (the macro-structure) is an extremely influential component in the final performance. As such, future research could investigate the optimal wiring at a global level: an interesting work in this direction is Xie et al. (2019a). ", + "bbox": [ + 173, + 217, + 825, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple datasets: as the true goal of AutoML is to minimize the need for human experts, focusing the research efforts on a single dataset will inevitably lead to algorithmic overfitting and/or methods heavily dependent on hyperparameter tuning. The best solution for this is likely to test NAS algorithms on a battery of datasets, with different characteristics: image sizes, number of samples, class granularity and learning task. ", + "bbox": [ + 174, + 391, + 823, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Investigating hidden components: as our experiments in Sections 4 and 5.2 show, the DARTS search space is not only effective due to specific operations that are being chosen, but in greater part due to the overall macro-structure and the training protocol used. We suggest that proper ablation studies can lead to better understanding of the contributions of each element of the pipeline. ", + "bbox": [ + 174, + 468, + 825, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The importance of reproducibility: reproducibility is of extreme relevance in all sciences. To this end, it is very important that authors release not only their best found architecture but also the corresponding seed (if they did not average over multiple ones), as well as the code and the detailed training protocol (including hyperparameters). To this end, NAS-Bench-101 (Ying et al., 2019), a dataset mapping architectures to their accuracy, can be extremely useful, as it allows the quality of search strategies to be assessed in isolation from other NAS components (e.g. search space, training protocol) in a quick and reproducible fashion. The code for this paper is open-source (link in the abstract). We also open-source the 270 trained architectures used in Section 5. ", + "bbox": [ + 174, + 530, + 825, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hyperparameter tuning cost: tuning hyperparameters in NAS is an extremely costly component. Therefore, we argue that either (i) hyperparameters are general enough so that they do not require tuning for further tasks, or (2) the cost is included in the search budget. ", + "bbox": [ + 174, + 648, + 825, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7 CONCLUSIONS ", + "text_level": 1, + "bbox": [ + 176, + 710, + 328, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "AutoML, and NAS in particular, have the potential to truly democratize the use of machine learning for all, and could bring forth very notable improvements on a variety of tasks. To truly step forward, a principled approach, with a focus on fairness and reproducibility is needed. ", + "bbox": [ + 176, + 742, + 823, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we have shown that, for many NAS methods, the search space has been engineered such that all architectures perform similarly well and that their relative ranking can easily shift. We have furthermore showed that the training protocol itself has a higher impact on the final accuracy than the actual network. Finally, we have provided some suggestions on how to make future research more robust to these issues. ", + "bbox": [ + 174, + 791, + 823, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We hope that our findings will help the community focus their efforts towards a more general approach to automated neural architecture design. Only then can we expect to learn from NASgenerated architectures as opposed to the current paradigm where search spaces are heavily influenced by our current (human) expert knowledge. ", + "bbox": [ + 174, + 867, + 823, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 103, + 287, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Fabio Carlucci, Pedro M. Esperanc¸a, Marco Singh, Antoine Yang, Victor Gabillon, Hang Xu, Zewei Chen, and Jun Wang. MANAS: Multi-Agent Neural Architecture Search. arxiv:1909.01051, 2019. ", + "bbox": [ + 174, + 126, + 823, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv:1805.09501, 2018. ", + "bbox": [ + 173, + 178, + 823, + 208 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout. arXiv:1708.04552, 2017. ", + "bbox": [ + 171, + 217, + 825, + 246 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey. Journal of Machine Learning Research, 20(55):1–21, 2019. ", + "bbox": [ + 173, + 256, + 821, + 286 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural networks. In arXiv:1803.03635, 2018. ", + "bbox": [ + 171, + 295, + 821, + 324 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Oguzhan Gencoglu, Mark van Gils, Esin Guldogan, Chamin Morikawa, Mehmet Suzen, Mathias ¨ Gruber, Jussi Leinonen, and Heikki Huttunen. Hark side of deep learning–from grad student descent to automated machine learning. arXiv:1904.07633, 2019. ", + "bbox": [ + 176, + 333, + 823, + 377 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016. ", + "bbox": [ + 173, + 386, + 821, + 416 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Zhang Xiangyu Rein Shaoqing He, Kaiming and Jian Sun. Delving deep into rectifiers: Surpassing human level performance on imagenet classification. In Computer Vision and Pattern Recognition (CVPR), 2015. ", + "bbox": [ + 174, + 425, + 823, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016. ", + "bbox": [ + 173, + 478, + 823, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Andrew Hundt, Varun Jain, and Gregory D. Hager. sharpDARTS: Faster and more accurate differentiable architecture search. arXiv:1903.09900, 2019. ", + "bbox": [ + 173, + 517, + 823, + 546 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. ", + "bbox": [ + 174, + 556, + 821, + 585 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105, 2012. ", + "bbox": [ + 174, + 594, + 823, + 637 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Michael Maire Larsson, Gustav and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks without residuals. In International Conference on Learning Representations (ICLR), 2017. ", + "bbox": [ + 173, + 647, + 823, + 678 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Guilin Li, Xing Zhang, Zitong Wang, Zhenguo Li, and Tong Zhang. StacNAS: Towards stable and consistent optimization for differentiable Neural Architecture Search. arXiv:1909.11926, 2019. ", + "bbox": [ + 173, + 686, + 823, + 715 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Li-Jia Li and Li Fei-Fei. What, where and who? classifying events by scene and object recognition. In International Conference on Computer Vision (ICCV), pp. 1–8, 2007. ", + "bbox": [ + 174, + 726, + 823, + 756 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search. arXiv:1902.07638, 2019. ", + "bbox": [ + 169, + 765, + 823, + 794 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hanwen Liang, Shifeng Zhang, Jiacheng Sun, Xingqiu He, Weiran Huang, Kechen Zhuang, and Zhenguo Li. DARTS $^ +$ : Improved differentiable architecture search with early stopping. arXiv:1909.06035, 2019. ", + "bbox": [ + 173, + 803, + 823, + 845 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Marius Lindauer and Frank Hutter. Best practices for scientific research on neural architecture search. arXiv:1909.02453, 2019. ", + "bbox": [ + 169, + 856, + 823, + 885 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. In International Conference on Learning Representations (ICLR), 2019. ", + "bbox": [ + 173, + 895, + 821, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations (ICLR), 2017. ", + "bbox": [ + 169, + 103, + 823, + 132 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Zhichao Lu, Ian Whalen, Vishnu Boddeti, Yashesh Dhebar, Kalyanmoy Deb, Erik Goodman, and Wolfgang Banzhaf. Nsga-net: a multi-objective genetic algorithm for neural architecture search. In GECCO-2019, 2018. ", + "bbox": [ + 176, + 141, + 823, + 184 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization. In Advances in Neural Information Processing Systems (NIPS), pp. 7816–7827, 2018. ", + "bbox": [ + 169, + 193, + 823, + 223 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, and Lihi Zelnik-Manor. Xnas: Neural architecture search with expert advice. arXiv:1906.08031, 2019. ", + "bbox": [ + 173, + 231, + 823, + 261 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yurii E Nesterov. A method for solving the convex programming problem with convergence rate o (1/kˆ 2). In Dokl. akad. nauk Sssr, volume 269, pp. 543–547, 1983. ", + "bbox": [ + 176, + 270, + 821, + 299 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pp. 722–729, 2008. ", + "bbox": [ + 173, + 308, + 825, + 352 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture search via parameter sharing. In International Conference on Machine Learning (ICML), pp. 4092–4101, 2018. ", + "bbox": [ + 173, + 359, + 823, + 404 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Philip Popien. AutoAugment. GitHub repository, https://github.com/DeepVoltaire/ AutoAugment, 2019. ", + "bbox": [ + 173, + 411, + 821, + 441 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In Computer Vision and Pattern Recognition (CVPR), pp. 413–420, 2009. ", + "bbox": [ + 173, + 450, + 823, + 481 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan, Quoc V Le, and Alexey Kurakin. Large-scale evolution of image classifiers. In International Conference on Machine Learning (ICML), pp. 2902–2911, 2017. ", + "bbox": [ + 176, + 488, + 823, + 532 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Christian Sciuto, Kaicheng Yu, Martin Jaggi, Claudiu Musat, and Mathieu Salzmann. Evaluating the search phase of neural architecture search. arXiv:1902.08142, 2019. ", + "bbox": [ + 171, + 540, + 823, + 570 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014. ", + "bbox": [ + 178, + 579, + 823, + 622 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4, Inception-ResNet and the impact of residual connections on learning. In AAAI Conference on Artificial Intelligence, 2017. ", + "bbox": [ + 173, + 631, + 823, + 674 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yu Weng, Tianbao Zhou, Lei Liu, and Chunlei Xia. Automatic convolutional neural architecture search for image classification under different scenes. IEEE Access, 7:38495–38506, 2019. ", + "bbox": [ + 173, + 683, + 821, + 713 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Saining Xie, Alexander Kirillov, Ross Girshick, and Kaiming He. Exploring randomly wired neural networks for image recognition. arXiv preprint arXiv:1904.01569, 2019a. ", + "bbox": [ + 173, + 720, + 825, + 751 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. SNAS: Stochastic neural architecture search. In International Conference on Learning Representations (ICLR), 2019b. ", + "bbox": [ + 173, + 760, + 823, + 790 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jun Wu Qi Tian Xin Chen, Lingxi Xie. Progressive differentiable architecture search. In International Conference on Computer Vision (ICCV), 2019. ", + "bbox": [ + 171, + 797, + 823, + 827 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. NASBench-101: Towards reproducible neural architecture search. In International Conference on Machine Learning (ICML 2019), pp. 7105–7114, 2019. ", + "bbox": [ + 176, + 837, + 823, + 880 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Barret Zoph and Quoc Le. Neural architecture search with reinforcement learning. In International Conference on Learning Representations (ICLR), 2017. ", + "bbox": [ + 174, + 888, + 823, + 917 + ], + "page_idx": 9 + }, + { + "type": "table", + "img_path": "images/320203340cda9ff05fec56f6573c4eaf1ef276b1e36acceee02cf27a14c0f25e.jpg", + "table_caption": [ + "Table 2: Hyperparameters for different NAS methods. “S” denotes search stage and “A” denotes augmentation stage. For learning rates, $a \\downarrow b$ means cosine annealing from $a$ to $b$ . " + ], + "table_footnote": [], + "table_body": "
Method
DARTSStacNASPDARTSMANASCNAS
batch size S6464966464
batch size A969612812864
init channels S1616161616
init channels A3636363636
epochs S50100+10025+25+255060
epochs A600600600600200
optimizer S/ASGDSGDSGDSGDAdam
learning rates S.025↓.001.025↓.001.025↓.0.025↓.001.025-.003
learning rates A.025↓.0.025↓.0.025↓.0.025←.0.025-.001
weight decay S/A3 ×10-43×10-43×10-43×10-43×10-4
optimizer archAdamAdamAdamAdam
learning rates arch3×10-43×10-46×10-43×10-4
weight decay arch10-310-310-310-3
nb cells S CIFAR814/205/11/1786
nb cells A CIFAR2020202020
nb cells S other datasets88/88/8/886
nb cellsA other datasets88888
nb intermediate nodes44446
", + "bbox": [ + 176, + 145, + 820, + 428 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 174, + 460, + 299, + 476 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "This section details the datasets and the hyperparameters used for each method on each dataset. Search spaces were naturally left unchanged. Hyperparameters were chosen as close as possible to the original paper and occasionally updated to more recent implementations. The network size was tuned similarly for all methods for SPORT8, MIT67 and FLOWERS102. All experiments were run on NVIDIA Tesla V100 GPUs. ", + "bbox": [ + 174, + 493, + 825, + 563 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A.1 METHODS AND HYPERPARAMETERS ", + "text_level": 1, + "bbox": [ + 176, + 582, + 467, + 595 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "During search, models are trained/validated on the training/validation subsets, respectively During the final evaluation, the model is trained on the training $^ +$ validation subsets and tested on the test subset. ", + "bbox": [ + 174, + 608, + 825, + 650 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Common hyperparameters. All 8 methods share a common number of hyperparameters precised here. When SGD optimizer is used, momentum is .9 while when Adam is used, momentum is $\\beta = ( 0 . 5 , 0 . 9 9 9 )$ . Gradient clipping is set at 5. ", + "bbox": [ + 174, + 657, + 825, + 700 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "DARTS, StacNAS, PDARTS, MANAS, CNAS common hyperparameters. These methods are inspired by DARTS code-wise and consequently share a common number of hyperparameters, which we precise in table 1. ", + "bbox": [ + 174, + 707, + 825, + 748 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "DARTS. We used the following repository : https://github.com/khanrc/pt.darts. It notably updates the official implementation to a pytorch version posterior to 0.4. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), path dropout of probability 0.2 (Larsson & Shakhnarovich, 2017), and auxiliary tower with weight 0.4. ", + "bbox": [ + 174, + 756, + 825, + 811 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "StacNAS. We used an unofficial implementation provided by the authors. The search process consists of 2 stages, of which the details are given in table 1. Additional enhancements are the same as DARTS. ", + "bbox": [ + 174, + 819, + 825, + 861 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "PDARTS. We used the official implementation $:$ https://github.com/chenxin061/pdarts. The search process consists of 3 stages, of which general details are given in table 1. At stage 1, 2 and 3 respectively, the number of operations decreases from 8 to 5 to 3, and the dropout probability on skip-connect increases from 0.0 to 0.4 to 0.7 for CIFAR10, SPORT8, MIT67 and FLOWERS102 (0.1 to 0.2 to 0.3 for CIFAR100). Discovered cells are restricted to keep at most 2 skip-connect operations. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), DropPath of probability 0.3 (Larsson & Shakhnarovich, 2017) and auxiliary tower with weight 0.4. ", + "bbox": [ + 174, + 867, + 823, + 922 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 823, + 146 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "MANAS. We used an unofficial implementation provided by the authors. The reward baseline is 5, gamma is 0.1 (0.07 for SPORT8, 0.05 for FLOWERS102, 0.01 for MIT67) and the Boltzmann temperature decays from 1000 to 200 (300 to 100 for SPORT8, 5000 to 2000 for MIT67). Additional enhancements are the same as DARTS. ", + "bbox": [ + 174, + 152, + 823, + 208 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "CNAS. We used the official implementation : https://github.com/tianbaochou/CNAS. Label Smoothing is used with epsilon 0.1. Other additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), DropPath of probability 0.25 (Larsson & Shakhnarovich, 2017). ", + "bbox": [ + 174, + 215, + 821, + 257 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "NSGANET. We used the official implementation: https://github.com/ianwhale/nsga-net. The search is done in the micro search space (with 2 cells to search, 9 operations in the search space, 5 blocks in each cell) on 8 layers networks. The population size is 40, the number of generations is 30 and the number of offsprings created by generation is 20. Networks are trained for 20 epochs, with batch size 128, and the initial number of channels is 16. For architecture evaluation, the network is composed of 20 cells for CIFAR10 and CIFAR100, and 8 cells for SPORT8, MIT67 and FLOWERS102. The final selected models are trained for 600 epochs with batch size 96 and the initial number of channels 34. Momentum SGD is used with initial learning rate $\\eta _ { w } = 0 . 0 2 5$ (annealed down to zero following a cosine schedule), and weight decay $3 \\times 1 0 ^ { - \\hat { 4 } }$ . The filter increment is set to 4, and squeeze and excitation is used. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), DropPath of probability 0.2 (Larsson & Shakhnarovich, 2017) and auxiliary tower with weight 0.4. ", + "bbox": [ + 173, + 263, + 825, + 417 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "NAO. We used the official Pytorch implementation: https://github.com/renqianluo/NAO pytorch. The LSTM model used to encode architecture has a token embedding size of 48 and a hidden state size of 96. The LSTM model used to decode architecture has a hidden state size of 96. The encoder and decoder are trained using Adam for 1000 epochs with a learning rate of 0.001. The trade-off parameters is $\\lambda = 0 . 9$ . The step size to perform continuous optimization is $\\eta = 1 0$ . The number of nodes is fixed to 5, normal cell is stacked 3 (2 for SPORT8, MIT67 and FLOWERS102) times to form the CNN architecture, which corresponds to a 11 (8 for SPORT8, MIT67 and FLOWERS102) cells network and the initial number of channels is 20. Networks are trained for 100 epochs with batch size 64 for training, and validated for 20 epochs with batch size 500. For architecture evaluation, the final CNN architecture is a 20 cells network (8 cells network). This network is trained for 600 epochs with batch size 128 for the training set (96 for both for SPORT8, MIT67 and FLOWERS102), 500 for the validation set (128 for both for SPORT8, MIT67 and FLOWERS102) and the initial number of channels is 36. Momentum SGD is used with initial learning rate $\\eta _ { w } = 0 . 0 2 5$ (annealed down to zero following a cosine schedule) and weight decay $3 \\times 1 0 ^ { - \\overline { { 4 } } }$ . Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), DropPath of probability 0.2 (Larsson & Shakhnarovich, 2017), dropout of probability 0.4 (Srivastava et al., 2014), and auxiliary tower with weight 0.4. ", + "bbox": [ + 173, + 424, + 825, + 660 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "ENAS. We used the official tensorflow implementation for experiments on CIFAR10 and CIFAR100: https://github.com/melodyguan/enas Experiments on SPORT8, MIT67 and FLOWERS102 are done using a more recent version in Pytorch for the search : https://github.com/MengTianjian/enas-pytorch. Because only the search process is implemented there, the evaluation code used is the same as DARTS. During the search, networks are composed of 8 cells. The shared parameters $w$ are trained with Nesterov momentum (Nesterov, 1983), weight decay $1 0 ^ { - 4 }$ , gradient clipping 5, batch size 160, 20 output filters, and a cosine learning rate schedule with $l _ { m } a x = 0 . 0 5$ , $l _ { m } i n = 0 . 0 0 1$ , $T _ { 0 } = 1 0$ , $T _ { m } u l = 2$ (Loshchilov & Hutter, 2017). Each architecture search is run for 150 epochs. $w$ are initialized with He initialization (He & Sun, 2015). The policy parameters $\\theta$ are initialized uniformly in [-0.1, 0.1], and trained with Adam at a learning rate of 0.00035. A tanh constant of 1.10 and a temperature of 2.5 is applied to the controllers logits, and the controller entropy is added to the reward with weight 0.1. For the evaluation, the architecture searched is extended to 17 cells (8 for SPORT8, MIT67 and FLOWERS102), trained for 630 epochs, with batch size 144, and a cosine learning rate schedule with $l _ { \\mathrm { m a x } } = 0 . 0 5$ , $l _ { \\mathrm { m i n } } = 0 . 0 0 1$ , $T _ { 0 } = 1 0$ , $T _ { \\mathrm { m u l } } = 2$ . ", + "bbox": [ + 173, + 667, + 825, + 876 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.2 DATASETS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 290, + 117 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We present here the datasets used and how they are pre-processed. ", + "bbox": [ + 174, + 130, + 609, + 143 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "CIFAR10. The CIFAR10 dataset (Krizhevsky, 2009) is a dataset of 10 classes and consists of 50, 000 training images and 10, 000 test images of size $3 2 \\times 3 2$ . ", + "bbox": [ + 174, + 150, + 823, + 179 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "CIFAR100. The CIFAR100 dataset (Krizhevsky, 2009) is a dataset of 100 classes and consists of 50, 000 training images and 10, 000 test images of size $3 2 \\times 3 2$ . ", + "bbox": [ + 174, + 185, + 821, + 214 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Each of these datasets is split into a training, validation and testing subsets of size 25, 000, 25, 000 and 10, 000 respectively. For both these datasets, we use standard data pre-processing and augmentation techniques, i.e. subtracting the channel mean and dividing by the channel standard deviation; centrally padding the training images to $4 0 \\times 4 0$ and randomly cropping them back to $3 2 \\times 3 2$ ; and randomly clipping them horizontally. ", + "bbox": [ + 174, + 220, + 825, + 291 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SPORT8. This is an action recognition dataset containing 8 sport event categories and a total of 1579 images (Li & Fei-Fei, 2007). The tiny size of this dataset stresses the generalization capabilities of any NAS method applied to it. ", + "bbox": [ + 174, + 297, + 825, + 340 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "MIT67. This is a dataset of 67 classes representing different indoor scenes and consists of 15, 620 images of different sizes (Quattoni & Torralba, 2009). ", + "bbox": [ + 176, + 347, + 823, + 376 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "FLOWERS102. This is a dataset of 102 classes representing different species of flowers and consists of 8, 189 images of different sizes (Nilsback & Zisserman, 2008). ", + "bbox": [ + 176, + 382, + 820, + 411 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Each of these datasets is split into a training, validation and testing subsets with proportions $4 0 / 4 0 / 2 0 \\ ( \\% )$ . For each one, we use use standard data pre-processing and augmentation techniques, i.e. subtracting the channel mean and dividing the channel standard deviation, cropping the training images to random size and aspect ratio, resizing them to $2 2 4 \\times 2 2 4$ , and randomly changing their brightness, contrast, and saturation, while resizing test images to $2 5 6 \\times 2 5 6$ and cropping them at the center. ", + "bbox": [ + 173, + 417, + 825, + 501 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.3 ADDITIONAL RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 518, + 374, + 532 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/ce230d7557a836be95b1c534a22b98a18804cb6ce095b27f9043206e3ab14ff8.jpg", + "image_caption": [ + "A.3.1 DIFFERENT TRAINING PROTOCOLS FOR RESNET-50 ", + "Figure 10: Extension of Figure 3, including results obtained by training a ResNet-50 on CIFAR10. Bars with darker shade are for ResNet-50 and bars with lighter shade are for DARTS (same as Figure 3). Result are for 8 runs of each training protocol. For ResNet-50, the auxiliary tower was added after layer 2. As DropPath (Larsson & Shakhnarovich, 2017) would not have been straightforward to apply, we instead implemented Stocastic Depth (Huang et al., 2016), to a similar effect. " + ], + "image_footnote": [], + "bbox": [ + 269, + 573, + 723, + 712 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/parse/train/HygrdpVKvr/HygrdpVKvr_middle.json b/parse/train/HygrdpVKvr/HygrdpVKvr_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..076b134ebd14e011016b0f2557d79e91febde869 --- /dev/null +++ b/parse/train/HygrdpVKvr/HygrdpVKvr_middle.json @@ -0,0 +1,34763 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 432, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 434, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 434, + 101 + ], + "score": 1.0, + "content": "NAS EVALUATION IS FRUSTRATINGLY HARD", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 113, + 115, + 199, + 149 + ], + "lines": [ + { + "bbox": [ + 111, + 113, + 177, + 128 + ], + "spans": [ + { + "bbox": [ + 111, + 113, + 177, + 128 + ], + "score": 1.0, + "content": "Antoine Yang∗", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 125, + 201, + 141 + ], + "spans": [ + { + "bbox": [ + 111, + 125, + 201, + 141 + ], + "score": 1.0, + "content": "Ecole Polytechnique ´ †", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 137, + 143, + 151 + ], + "spans": [ + { + "bbox": [ + 111, + 137, + 143, + 151 + ], + "score": 1.0, + "content": "France", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 234, + 115, + 334, + 148 + ], + "lines": [ + { + "bbox": [ + 233, + 114, + 322, + 128 + ], + "spans": [ + { + "bbox": [ + 233, + 114, + 322, + 128 + ], + "score": 1.0, + "content": "Pedro M Esperanc¸a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 126, + 334, + 137 + ], + "spans": [ + { + "bbox": [ + 234, + 126, + 334, + 137 + ], + "score": 1.0, + "content": "Huawei Noah’s Ark Lab", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 137, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 233, + 137, + 288, + 149 + ], + "score": 1.0, + "content": "London, UK", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 371, + 115, + 471, + 148 + ], + "lines": [ + { + "bbox": [ + 371, + 115, + 466, + 126 + ], + "spans": [ + { + "bbox": [ + 371, + 115, + 466, + 126 + ], + "score": 1.0, + "content": "Fabio Maria Carlucci", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 371, + 126, + 471, + 137 + ], + "spans": [ + { + "bbox": [ + 371, + 126, + 471, + 137 + ], + "score": 1.0, + "content": "Huawei Noah’s Ark Lab", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 137, + 425, + 149 + ], + "spans": [ + { + "bbox": [ + 370, + 137, + 425, + 149 + ], + "score": 1.0, + "content": "London, UK", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 278, + 179, + 333, + 191 + ], + "lines": [ + { + "bbox": [ + 276, + 177, + 336, + 193 + ], + "spans": [ + { + "bbox": [ + 276, + 177, + 336, + 193 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 142, + 203, + 468, + 478 + ], + "lines": [ + { + "bbox": [ + 141, + 203, + 469, + 216 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 469, + 216 + ], + "score": 1.0, + "content": "Neural Architecture Search (NAS) is an exciting new field which promises to be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 214, + 469, + 226 + ], + "spans": [ + { + "bbox": [ + 141, + 214, + 469, + 226 + ], + "score": 1.0, + "content": "as much as a game-changer as Convolutional Neural Networks were in 2012. De-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 226, + 469, + 238 + ], + "spans": [ + { + "bbox": [ + 141, + 226, + 469, + 238 + ], + "score": 1.0, + "content": "spite many great works leading to substantial improvements on a variety of tasks,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 236, + 470, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 470, + 249 + ], + "score": 1.0, + "content": "comparison between different methods is still very much an open issue. While", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 248, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 469, + 259 + ], + "score": 1.0, + "content": "most algorithms are tested on the same datasets, there is no shared experimental", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 259, + 469, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 469, + 270 + ], + "score": 1.0, + "content": "protocol followed by all. As such, and due to the under-use of ablation studies,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 270, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 469, + 281 + ], + "score": 1.0, + "content": "there is a lack of clarity regarding why certain methods are more effective than", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 280, + 469, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 469, + 292 + ], + "score": 1.0, + "content": "others. Our first contribution is a benchmark of 8 NAS methods on 5 datasets.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 291, + 469, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 291, + 469, + 303 + ], + "score": 1.0, + "content": "To overcome the hurdle of comparing methods with different search spaces, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "score": 1.0, + "content": "propose using a method’s relative improvement over the randomly sampled av-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 313, + 469, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 326 + ], + "score": 1.0, + "content": "erage architecture, which effectively removes advantages arising from expertly", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "score": 1.0, + "content": "engineered search spaces or training protocols. Surprisingly, we find that many", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 334, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 348 + ], + "score": 1.0, + "content": "NAS techniques struggle to significantly beat the average architecture baseline.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 345, + 469, + 359 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 469, + 359 + ], + "score": 1.0, + "content": "We perform further experiments with the commonly used DARTS search space", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 357, + 469, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 469, + 369 + ], + "score": 1.0, + "content": "in order to understand the contribution of each component in the NAS pipeline.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "score": 1.0, + "content": "These experiments highlight that: (i) the use of tricks in the evaluation protocol", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "score": 1.0, + "content": "has a predominant impact on the reported performance of architectures; (ii) the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "score": 1.0, + "content": "cell-based search space has a very narrow accuracy range, such that the seed has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 402, + 469, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 469, + 412 + ], + "score": 1.0, + "content": "a considerable impact on architecture rankings; (iii) the hand-designed macro-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 412, + 469, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 469, + 424 + ], + "score": 1.0, + "content": "structure (cells) is more important than the searched micro-structure (operations);", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 422, + 470, + 436 + ], + "spans": [ + { + "bbox": [ + 141, + 422, + 470, + 436 + ], + "score": 1.0, + "content": "and (iv) the depth-gap is a real phenomenon, evidenced by the change in rankings", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 434, + 469, + 446 + ], + "spans": [ + { + "bbox": [ + 142, + 434, + 469, + 446 + ], + "score": 1.0, + "content": "between 8 and 20 cell architectures. To conclude, we suggest best practices, that", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 445, + 469, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 445, + 469, + 456 + ], + "score": 1.0, + "content": "we hope will prove useful for the community and help mitigate current NAS pit-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 456, + 469, + 467 + ], + "spans": [ + { + "bbox": [ + 141, + 456, + 469, + 467 + ], + "score": 1.0, + "content": "falls, e.g. difficulties in reproducibility and comparison of search methods. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 142, + 467, + 429, + 478 + ], + "spans": [ + { + "bbox": [ + 142, + 467, + 429, + 478 + ], + "score": 1.0, + "content": "code used is available at https://github.com/antoyang/NAS-Benchmark.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 206, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 208, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 208, + 514 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "As the deep learning revolution helped us move away from hand crafted features (Krizhevsky et al.,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "2012) and reach new heights (He et al., 2016; Szegedy et al., 2017), so does Neural Architecture", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "Search (NAS) hold the promise of freeing us from hand-crafted architectures, which requires te-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "dious and expensive tuning for each new task or dataset. Identifying the optimal architecture is", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "indeed a key pillar of any Automated Machine Learning (AutoML) pipeline. Research in the last", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "two years has proceeded at a rapid pace and many search strategies have been proposed, from Rein-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "forcement Learning (Zoph & Le, 2017; Pham et al., 2018), to Evolutionary Algorithms (Real et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "2017), to Gradient-based methods (Liu et al., 2019; Liang et al., 2019). Still, it remains unclear", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "which approach and search algorithm is preferable. Typically, methods have been evaluated on", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 621, + 504, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 504, + 634 + ], + "score": 1.0, + "content": "accuracy alone, even though accuracy is influenced by many other factors besides the search algo-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "rithm. Comparison between published search algorithms for NAS is therefore either very difficult", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "(complex training protocols with no code available) or simply impossible (different search spaces),", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 654, + 488, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 488, + 667 + ], + "score": 1.0, + "content": "as previously pointed out (Li & Talwalkar, 2019; Sciuto et al., 2019; Lindauer & Hutter, 2019).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 109, + 671, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 107, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 107, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "NAS methods have been typically decomposed into three components (Elsken et al., 2019; Li &", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 107, + 682, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 107, + 682, + 506, + 695 + ], + "score": 1.0, + "content": "Talwalkar, 2019): search space, search strategy and model evaluation strategy. This division is", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 50.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 700, + 515, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 699, + 517, + 713 + ], + "spans": [ + { + "bbox": [ + 117, + 699, + 517, + 713 + ], + "score": 1.0, + "content": "∗antoineyang3@gmail.com, {pedro.esperanca,fabio.maria.carlucci}@huawei.com", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 709, + 470, + 724 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 470, + 724 + ], + "score": 1.0, + "content": "†This work was done when the first author was an intern at Huawei", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 317, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 317, + 732 + ], + "score": 1.0, + "content": "Noah’s Ark Lab, London, United Kingdom.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 80, + 432, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 434, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 434, + 101 + ], + "score": 1.0, + "content": "NAS EVALUATION IS FRUSTRATINGLY HARD", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 113, + 115, + 199, + 149 + ], + "lines": [ + { + "bbox": [ + 111, + 113, + 177, + 128 + ], + "spans": [ + { + "bbox": [ + 111, + 113, + 177, + 128 + ], + "score": 1.0, + "content": "Antoine Yang∗", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 125, + 201, + 141 + ], + "spans": [ + { + "bbox": [ + 111, + 125, + 201, + 141 + ], + "score": 1.0, + "content": "Ecole Polytechnique ´ †", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 137, + 143, + 151 + ], + "spans": [ + { + "bbox": [ + 111, + 137, + 143, + 151 + ], + "score": 1.0, + "content": "France", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 111, + 113, + 201, + 151 + ] + }, + { + "type": "text", + "bbox": [ + 234, + 115, + 334, + 148 + ], + "lines": [ + { + "bbox": [ + 233, + 114, + 322, + 128 + ], + "spans": [ + { + "bbox": [ + 233, + 114, + 322, + 128 + ], + "score": 1.0, + "content": "Pedro M Esperanc¸a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 126, + 334, + 137 + ], + "spans": [ + { + "bbox": [ + 234, + 126, + 334, + 137 + ], + "score": 1.0, + "content": "Huawei Noah’s Ark Lab", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 233, + 137, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 233, + 137, + 288, + 149 + ], + "score": 1.0, + "content": "London, UK", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 233, + 114, + 334, + 149 + ] + }, + { + "type": "text", + "bbox": [ + 371, + 115, + 471, + 148 + ], + "lines": [ + { + "bbox": [ + 371, + 115, + 466, + 126 + ], + "spans": [ + { + "bbox": [ + 371, + 115, + 466, + 126 + ], + "score": 1.0, + "content": "Fabio Maria Carlucci", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 371, + 126, + 471, + 137 + ], + "spans": [ + { + "bbox": [ + 371, + 126, + 471, + 137 + ], + "score": 1.0, + "content": "Huawei Noah’s Ark Lab", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 370, + 137, + 425, + 149 + ], + "spans": [ + { + "bbox": [ + 370, + 137, + 425, + 149 + ], + "score": 1.0, + "content": "London, UK", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 370, + 115, + 471, + 149 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 179, + 333, + 191 + ], + "lines": [ + { + "bbox": [ + 276, + 177, + 336, + 193 + ], + "spans": [ + { + "bbox": [ + 276, + 177, + 336, + 193 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 142, + 203, + 468, + 478 + ], + "lines": [ + { + "bbox": [ + 141, + 203, + 469, + 216 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 469, + 216 + ], + "score": 1.0, + "content": "Neural Architecture Search (NAS) is an exciting new field which promises to be", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 214, + 469, + 226 + ], + "spans": [ + { + "bbox": [ + 141, + 214, + 469, + 226 + ], + "score": 1.0, + "content": "as much as a game-changer as Convolutional Neural Networks were in 2012. De-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 226, + 469, + 238 + ], + "spans": [ + { + "bbox": [ + 141, + 226, + 469, + 238 + ], + "score": 1.0, + "content": "spite many great works leading to substantial improvements on a variety of tasks,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 236, + 470, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 470, + 249 + ], + "score": 1.0, + "content": "comparison between different methods is still very much an open issue. While", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 248, + 469, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 248, + 469, + 259 + ], + "score": 1.0, + "content": "most algorithms are tested on the same datasets, there is no shared experimental", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 259, + 469, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 259, + 469, + 270 + ], + "score": 1.0, + "content": "protocol followed by all. As such, and due to the under-use of ablation studies,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 270, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 270, + 469, + 281 + ], + "score": 1.0, + "content": "there is a lack of clarity regarding why certain methods are more effective than", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 280, + 469, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 469, + 292 + ], + "score": 1.0, + "content": "others. Our first contribution is a benchmark of 8 NAS methods on 5 datasets.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 291, + 469, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 291, + 469, + 303 + ], + "score": 1.0, + "content": "To overcome the hurdle of comparing methods with different search spaces, we", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "spans": [ + { + "bbox": [ + 141, + 302, + 470, + 314 + ], + "score": 1.0, + "content": "propose using a method’s relative improvement over the randomly sampled av-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 313, + 469, + 326 + ], + "spans": [ + { + "bbox": [ + 141, + 313, + 469, + 326 + ], + "score": 1.0, + "content": "erage architecture, which effectively removes advantages arising from expertly", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 469, + 337 + ], + "score": 1.0, + "content": "engineered search spaces or training protocols. Surprisingly, we find that many", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 334, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 348 + ], + "score": 1.0, + "content": "NAS techniques struggle to significantly beat the average architecture baseline.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 345, + 469, + 359 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 469, + 359 + ], + "score": 1.0, + "content": "We perform further experiments with the commonly used DARTS search space", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 141, + 357, + 469, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 357, + 469, + 369 + ], + "score": 1.0, + "content": "in order to understand the contribution of each component in the NAS pipeline.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "score": 1.0, + "content": "These experiments highlight that: (i) the use of tricks in the evaluation protocol", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "score": 1.0, + "content": "has a predominant impact on the reported performance of architectures; (ii) the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "score": 1.0, + "content": "cell-based search space has a very narrow accuracy range, such that the seed has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 402, + 469, + 412 + ], + "spans": [ + { + "bbox": [ + 141, + 402, + 469, + 412 + ], + "score": 1.0, + "content": "a considerable impact on architecture rankings; (iii) the hand-designed macro-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 141, + 412, + 469, + 424 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 469, + 424 + ], + "score": 1.0, + "content": "structure (cells) is more important than the searched micro-structure (operations);", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 422, + 470, + 436 + ], + "spans": [ + { + "bbox": [ + 141, + 422, + 470, + 436 + ], + "score": 1.0, + "content": "and (iv) the depth-gap is a real phenomenon, evidenced by the change in rankings", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 142, + 434, + 469, + 446 + ], + "spans": [ + { + "bbox": [ + 142, + 434, + 469, + 446 + ], + "score": 1.0, + "content": "between 8 and 20 cell architectures. To conclude, we suggest best practices, that", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 141, + 445, + 469, + 456 + ], + "spans": [ + { + "bbox": [ + 141, + 445, + 469, + 456 + ], + "score": 1.0, + "content": "we hope will prove useful for the community and help mitigate current NAS pit-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 141, + 456, + 469, + 467 + ], + "spans": [ + { + "bbox": [ + 141, + 456, + 469, + 467 + ], + "score": 1.0, + "content": "falls, e.g. difficulties in reproducibility and comparison of search methods. The", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 142, + 467, + 429, + 478 + ], + "spans": [ + { + "bbox": [ + 142, + 467, + 429, + 478 + ], + "score": 1.0, + "content": "code used is available at https://github.com/antoyang/NAS-Benchmark.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 23, + "bbox_fs": [ + 141, + 203, + 470, + 478 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 498, + 206, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 208, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 208, + 514 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "text", + "bbox": [ + 107, + 523, + 505, + 666 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "As the deep learning revolution helped us move away from hand crafted features (Krizhevsky et al.,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 547 + ], + "score": 1.0, + "content": "2012) and reach new heights (He et al., 2016; Szegedy et al., 2017), so does Neural Architecture", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 506, + 558 + ], + "score": 1.0, + "content": "Search (NAS) hold the promise of freeing us from hand-crafted architectures, which requires te-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "dious and expensive tuning for each new task or dataset. Identifying the optimal architecture is", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "indeed a key pillar of any Automated Machine Learning (AutoML) pipeline. Research in the last", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "two years has proceeded at a rapid pace and many search strategies have been proposed, from Rein-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "forcement Learning (Zoph & Le, 2017; Pham et al., 2018), to Evolutionary Algorithms (Real et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "2017), to Gradient-based methods (Liu et al., 2019; Liang et al., 2019). Still, it remains unclear", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "which approach and search algorithm is preferable. Typically, methods have been evaluated on", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 621, + 504, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 504, + 634 + ], + "score": 1.0, + "content": "accuracy alone, even though accuracy is influenced by many other factors besides the search algo-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "rithm. Comparison between published search algorithms for NAS is therefore either very difficult", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 656 + ], + "score": 1.0, + "content": "(complex training protocols with no code available) or simply impossible (different search spaces),", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 654, + 488, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 488, + 667 + ], + "score": 1.0, + "content": "as previously pointed out (Li & Talwalkar, 2019; Sciuto et al., 2019; Lindauer & Hutter, 2019).", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 523, + 506, + 667 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 671, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 107, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 107, + 671, + 505, + 684 + ], + "score": 1.0, + "content": "NAS methods have been typically decomposed into three components (Elsken et al., 2019; Li &", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 107, + 682, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 107, + 682, + 506, + 695 + ], + "score": 1.0, + "content": "Talwalkar, 2019): search space, search strategy and model evaluation strategy. This division is", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "important to keep in mind, as an improvement in any of these elements will lead to a better final per-", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "formance. But is a method with a more (manually) tuned search space a better AutoML algorithm?", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "If the key idea behind NAS is to find the optimal architecture, without human intervention, why are", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "we devoting so much energy to infuse expert knowledge into the pipeline? Furthermore, the lack of", + "type": "text", + "cross_page": true + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "ablation studies in most works makes it harder to pinpoint which components are instrumental to the", + "type": "text", + "cross_page": true + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "score": 1.0, + "content": "final performance, which can easily lead to Hypothesizing After the Results are Known (HARKing;", + "type": "text", + "cross_page": true + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 200, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 200, + 161 + ], + "score": 1.0, + "content": "Gencoglu et al., 2019).", + "type": "text", + "cross_page": true + } + ], + "index": 6 + } + ], + "index": 50.5, + "bbox_fs": [ + 107, + 671, + 506, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 159 + ], + "lines": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 505, + 95 + ], + "score": 1.0, + "content": "important to keep in mind, as an improvement in any of these elements will lead to a better final per-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "formance. But is a method with a more (manually) tuned search space a better AutoML algorithm?", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "If the key idea behind NAS is to find the optimal architecture, without human intervention, why are", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 127 + ], + "score": 1.0, + "content": "we devoting so much energy to infuse expert knowledge into the pipeline? Furthermore, the lack of", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 505, + 139 + ], + "score": 1.0, + "content": "ablation studies in most works makes it harder to pinpoint which components are instrumental to the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 151 + ], + "score": 1.0, + "content": "final performance, which can easily lead to Hypothesizing After the Results are Known (HARKing;", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 149, + 200, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 200, + 161 + ], + "score": 1.0, + "content": "Gencoglu et al., 2019).", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "Paradoxically, the huge effort invested in finding better search spaces and training protocols, has", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "led to a situation in which any randomly sampled architecture performs almost as well as those", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "obtained by the search strategies. Our findings suggest that most of the gains in accuracy in recent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "contributions to NAS have come from manual improvements in the training protocol, not in the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 181, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 181, + 222 + ], + "score": 1.0, + "content": "search algorithms.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "score": 1.0, + "content": "As a step towards understanding which methods are more effective, we have collected code for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "score": 1.0, + "content": "8 reasonably fast (search time of less than 4 days) NAS algorithms, and benchmarked them on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 246, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 104, + 246, + 506, + 262 + ], + "score": 1.0, + "content": "5 well known CV datasets. Using a simple metric—the relative improvement over the average", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "architecture of the search space—we find that most NAS methods perform very similarly and rarely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "substantially above this baseline. The methods used are DARTS, StacNAS, PDARTS, MANAS,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "score": 1.0, + "content": "CNAS, NSGANET, ENAS and NAO. The datasets used are CIFAR10, CIFAR100, SPORT8, MIT67", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 189, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 189, + 303 + ], + "score": 1.0, + "content": "and FLOWERS102.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "Through a number of additional experiments on the widely used DARTS search space (Liu et al.,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 319, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 506, + 333 + ], + "score": 1.0, + "content": "2019), we will show that: (a) how you train your model has a much bigger impact than the actual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "architecture chosen; (b) different architectures from the same search space perform very similarly,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 340, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 356 + ], + "score": 1.0, + "content": "so much so that (c) hyperparameters, like the number of cells, or the seed itself have a very sig-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "nificant effect on the ranking; and (d) the specific operations themselves have less impact on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "final accuracy than the hand-designed macro-structure of the network. Notably, we find that the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 131, + 385 + ], + "score": 0.83, + "content": "2 0 0 +", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "architectures sampled from this search space (available from the link in the abstract) are all", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "within a range of one percentage point (top-1 accuracy) after a standard full training on CIFAR10.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 396, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 409 + ], + "score": 1.0, + "content": "Finally, we include some observations on how to foster reproducibility and a discussion on how to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 308, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 308, + 419 + ], + "score": 1.0, + "content": "potentially avoid some of the encountered pitfalls.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "title", + "bbox": [ + 108, + 446, + 208, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 445, + 211, + 461 + ], + "spans": [ + { + "bbox": [ + 104, + 445, + 211, + 461 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 109, + 478, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 107, + 479, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 107, + 479, + 505, + 489 + ], + "score": 1.0, + "content": "As mentioned, NAS methods have the potential to truly revolutionize the field, but to do so it is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "crucial that future research avoids common mistakes. Some of these concerns have been recently", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 501, + 209, + 513 + ], + "spans": [ + { + "bbox": [ + 107, + 501, + 209, + 513 + ], + "score": 1.0, + "content": "raised by the community.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "For example, Li & Talwalkar (2019) highlight that most NAS methods a) fail to compare against an", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "adequate baseline, such as a properly implemented random search strategy, b) are overly complex,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "with no ablation to properly assign credit to the important components, and c) fail to provide all", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "details needed for successfully reproducing their results. In our paper we go one step further and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "argue that the relative improvement over the average (randomly sampled) architecture is an useful", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 573, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 504, + 584 + ], + "score": 1.0, + "content": "tool to quantify the effectiveness of a proposed solution and compare it with competing methods.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "To partly answer their second point, and understand how much the final accuracy depends on the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "specific architecture, we implement an in-depth study of the widely employed DARTS (Liu et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "2019) search space and perform an ablation on the commonly used training techniques (e.g. Cutout,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 214, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 214, + 629 + ], + "score": 1.0, + "content": "DropPath, AutoAugment).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "In addition, Sciuto et al. (2019) also took the important step of systematically using fair baselines,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "and suggest random search with early stopping, averaged over multiple seeds, as an extremely com-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "petitive baseline. They find that the search spaces of three methods investigated (DARTS, ENAS,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "NAO) have been expertly engineered to the extent that any randomly selected architecture performs", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "score": 1.0, + "content": "very well. In contrast, we show that even random sampling (without search) provides an incredibly", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "competitive baseline. Our relative improvement metric allows us to isolate the contribution of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "search strategy from the effects of the search space and training pipeline. Thus, we further confirm", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "the authors’ claim, showing that indeed the average architecture performs extremely well and that", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 721, + 388, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 388, + 734 + ], + "score": 1.0, + "content": "how you train a model has more impact than any specific architecture.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 13, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 159 + ], + "lines": [], + "index": 3, + "bbox_fs": [ + 105, + 83, + 506, + 161 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 165, + 505, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 164, + 505, + 178 + ], + "score": 1.0, + "content": "Paradoxically, the huge effort invested in finding better search spaces and training protocols, has", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 505, + 189 + ], + "score": 1.0, + "content": "led to a situation in which any randomly sampled architecture performs almost as well as those", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "obtained by the search strategies. Our findings suggest that most of the gains in accuracy in recent", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "contributions to NAS have come from manual improvements in the training protocol, not in the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 210, + 181, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 181, + 222 + ], + "score": 1.0, + "content": "search algorithms.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 164, + 505, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 226, + 505, + 303 + ], + "lines": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 506, + 238 + ], + "score": 1.0, + "content": "As a step towards understanding which methods are more effective, we have collected code for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 506, + 249 + ], + "score": 1.0, + "content": "8 reasonably fast (search time of less than 4 days) NAS algorithms, and benchmarked them on", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 246, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 104, + 246, + 506, + 262 + ], + "score": 1.0, + "content": "5 well known CV datasets. Using a simple metric—the relative improvement over the average", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "architecture of the search space—we find that most NAS methods perform very similarly and rarely", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "substantially above this baseline. The methods used are DARTS, StacNAS, PDARTS, MANAS,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "score": 1.0, + "content": "CNAS, NSGANET, ENAS and NAO. The datasets used are CIFAR10, CIFAR100, SPORT8, MIT67", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 189, + 303 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 189, + 303 + ], + "score": 1.0, + "content": "and FLOWERS102.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 226, + 506, + 303 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 308, + 505, + 418 + ], + "lines": [ + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "Through a number of additional experiments on the widely used DARTS search space (Liu et al.,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 319, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 506, + 333 + ], + "score": 1.0, + "content": "2019), we will show that: (a) how you train your model has a much bigger impact than the actual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 343 + ], + "score": 1.0, + "content": "architecture chosen; (b) different architectures from the same search space perform very similarly,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 340, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 356 + ], + "score": 1.0, + "content": "so much so that (c) hyperparameters, like the number of cells, or the seed itself have a very sig-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "nificant effect on the ranking; and (d) the specific operations themselves have less impact on the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "final accuracy than the hand-designed macro-structure of the network. Notably, we find that the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 374, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 131, + 385 + ], + "score": 0.83, + "content": "2 0 0 +", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 374, + 505, + 387 + ], + "score": 1.0, + "content": "architectures sampled from this search space (available from the link in the abstract) are all", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "within a range of one percentage point (top-1 accuracy) after a standard full training on CIFAR10.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 396, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 506, + 409 + ], + "score": 1.0, + "content": "Finally, we include some observations on how to foster reproducibility and a discussion on how to", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 407, + 308, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 308, + 419 + ], + "score": 1.0, + "content": "potentially avoid some of the encountered pitfalls.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 308, + 506, + 419 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 446, + 208, + 459 + ], + "lines": [ + { + "bbox": [ + 104, + 445, + 211, + 461 + ], + "spans": [ + { + "bbox": [ + 104, + 445, + 211, + 461 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 109, + 478, + 503, + 511 + ], + "lines": [ + { + "bbox": [ + 107, + 479, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 107, + 479, + 505, + 489 + ], + "score": 1.0, + "content": "As mentioned, NAS methods have the potential to truly revolutionize the field, but to do so it is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "crucial that future research avoids common mistakes. Some of these concerns have been recently", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 107, + 501, + 209, + 513 + ], + "spans": [ + { + "bbox": [ + 107, + 501, + 209, + 513 + ], + "score": 1.0, + "content": "raised by the community.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 479, + 505, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "For example, Li & Talwalkar (2019) highlight that most NAS methods a) fail to compare against an", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 505, + 541 + ], + "score": 1.0, + "content": "adequate baseline, such as a properly implemented random search strategy, b) are overly complex,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "with no ablation to properly assign credit to the important components, and c) fail to provide all", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 506, + 564 + ], + "score": 1.0, + "content": "details needed for successfully reproducing their results. In our paper we go one step further and", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "argue that the relative improvement over the average (randomly sampled) architecture is an useful", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 573, + 504, + 584 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 504, + 584 + ], + "score": 1.0, + "content": "tool to quantify the effectiveness of a proposed solution and compare it with competing methods.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "To partly answer their second point, and understand how much the final accuracy depends on the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "specific architecture, we implement an in-depth study of the widely employed DARTS (Liu et al.,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "2019) search space and perform an ablation on the commonly used training techniques (e.g. Cutout,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 214, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 214, + 629 + ], + "score": 1.0, + "content": "DropPath, AutoAugment).", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 516, + 506, + 629 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "In addition, Sciuto et al. (2019) also took the important step of systematically using fair baselines,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 505, + 657 + ], + "score": 1.0, + "content": "and suggest random search with early stopping, averaged over multiple seeds, as an extremely com-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "petitive baseline. They find that the search spaces of three methods investigated (DARTS, ENAS,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "NAO) have been expertly engineered to the extent that any randomly selected architecture performs", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 504, + 689 + ], + "score": 1.0, + "content": "very well. In contrast, we show that even random sampling (without search) provides an incredibly", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "competitive baseline. Our relative improvement metric allows us to isolate the contribution of the", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "search strategy from the effects of the search space and training pipeline. Thus, we further confirm", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "the authors’ claim, showing that indeed the average architecture performs extremely well and that", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 721, + 388, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 388, + 734 + ], + "score": 1.0, + "content": "how you train a model has more impact than any specific architecture.", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 633, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 222, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 224, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 224, + 96 + ], + "score": 1.0, + "content": "3 NAS BENCHMARK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 108, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 120 + ], + "score": 1.0, + "content": "In this section we present a systematic evaluation of 8 methods on 5 datasets using a strategy that is", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 120, + 504, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 504, + 132 + ], + "score": 1.0, + "content": "designed to reveal the quality of each method’s search strategy, removing the effect of the manually-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 143 + ], + "score": 1.0, + "content": "engineered training protocol and search space. The goal is to find general trends and highlight", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 142, + 407, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 407, + 154 + ], + "score": 1.0, + "content": "common features rather than just pin-pointing the most accurate algorithm.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 158, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "Understanding why methods are effective is not an easy task: most introduce variations to previous", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 170, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 181 + ], + "score": 1.0, + "content": "search spaces, search strategies, and training protocols—with ablations disentangling the contribu-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "tion of each component often incomplete or missing. In other words, how can we be sure that a new", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "state-of-the-art method is not so simply due to a better engineered search space or training protocol?", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 202, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 506, + 215 + ], + "score": 1.0, + "content": "To address this issue we compare a set of 8 methods with randomly sampled architectures from their", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 468, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 468, + 226 + ], + "score": 1.0, + "content": "respective search spaces, and trained with the same protocol as the searched architectures.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 505, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "The ultimate goal behind NAS should be to return the optimal model for any dataset given, at least", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "within the limits of a certain task, and we feel that the current practices of searching almost exclu-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 266 + ], + "score": 1.0, + "content": "sively on CIFAR10 go against this principle. Indeed, to avoid the very concrete risk of overfitting", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 506, + 276 + ], + "score": 1.0, + "content": "to this set of data, NAS methods should be tested on a variety of tasks. For this reason we run", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 274, + 250, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 250, + 286 + ], + "score": 1.0, + "content": "experiments on 5 different datasets.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 200, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 202, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 202, + 315 + ], + "score": 1.0, + "content": "3.1 METHODOLOGY", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "Criteria for dataset selection. We selected datasets to cover a variety of subtasks within image", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "score": 1.0, + "content": "classification. In addition to the standard CIFAR10 we select CIFAR100 for a more challenging", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "object classification problem (Krizhevsky, 2009); SPORT8 for action classification (Li & Fei-Fei,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "score": 1.0, + "content": "2007); MIT67 for scene classification (Quattoni & Torralba, 2009); and FLOWERS102 for fine-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "grained object classification (Nilsback & Zisserman, 2008). More details are given in the Appendix.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 396, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "Criteria for method selection. We selected methods which (a) have open-source code, or pro-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "vided it upon request, and (b) have a reasonable running time, specifically a search time under 4", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "score": 1.0, + "content": "GPU-days on CIFAR10. The selected methods are: DARTS (Liu et al., 2019), StacNAS (Li et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "2019), PDARTS (Xin Chen, 2019), MANAS (Carlucci et al., 2019), CNAS (Weng et al., 2019),", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "NSGANET (Lu et al., 2018), ENAS (Pham et al., 2018), and NAO (Luo et al., 2018). With the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 450, + 477, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 477, + 465 + ], + "score": 1.0, + "content": "exception of NAO and NSGANET, all methods are DARTS variants and use weight sharing.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "Evaluation protocol. NAS algorithms usually consist of two phases: (i) search, producing the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "score": 1.0, + "content": "best architecture according to the search algorithm used; (ii) augmentation, consisting in training", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "from scratch the best model found in the search phase. We evaluate methods as follows: 1. Sample", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "8 architectures from the search space, uniformly at random, and use the method’s code to augment", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "these architectures (same augment seed for all); 2. Use the method’s code to search for 8 architec-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "tures and augment them (different search seed, same augment seed); 3. Report mean and standard", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "deviation of the top-1 test accuracy, obtained at the end of the augmentation, for both the randomly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 268, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 268, + 567 + ], + "score": 1.0, + "content": "sampled and the searched architectures;", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Since both learned and randomly sampled architectures share the same search space and training", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 581, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 408, + 597 + ], + "score": 1.0, + "content": "protocol, calculating a relative improvement over this random baseline as", + "type": "text" + }, + { + "bbox": [ + 408, + 583, + 505, + 594 + ], + "score": 0.9, + "content": "R I = 1 0 0 \\times ( A c c _ { m } -", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 157, + 606 + ], + "score": 0.9, + "content": "A c c _ { r } ) / A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 594, + 401, + 607 + ], + "score": 1.0, + "content": "can offer insights into the quality of the search strategy alone.", + "type": "text" + }, + { + "bbox": [ + 402, + 594, + 426, + 605 + ], + "score": 0.91, + "content": "A c c _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 594, + 443, + 607 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 444, + 595, + 465, + 605 + ], + "score": 0.9, + "content": "A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "represent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "the top-1 accuracy of the search method and random sampling strategies, respectively. A good,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 309, + 629 + ], + "score": 1.0, + "content": "general-purpose NAS method is expected to yield", + "type": "text" + }, + { + "bbox": [ + 309, + 616, + 343, + 627 + ], + "score": 0.89, + "content": "R I > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "consistently over different searches and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "across different subtasks. We emphasize that the comparison is not against random search, but", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "rather against random sampling, i.e., the average architecture of the search space. For example, in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "the DARTS search space, for each edge in the graph that defines a cell we select one out of eight", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 659, + 428, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 408, + 673 + ], + "score": 1.0, + "content": "possible operations (e.g. pooling or convolutions) with uniform probability", + "type": "text" + }, + { + "bbox": [ + 408, + 660, + 424, + 672 + ], + "score": 0.74, + "content": "1 / 8", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 659, + 428, + 673 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "Hyperparameters are optimized on CIFAR10, according to the values reported by the corresponding", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "authors. Since most methods do not include their optimization as part of the search routine, we", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "assumed them to be robust and generalizable to other tasks. As such, aside from scaling down the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "architecture depending on dataset size, experiments on other datasets use the same hyperparameters.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 720, + 362, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 362, + 734 + ], + "score": 1.0, + "content": "Other training details and references are given in the Appendix.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 222, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 224, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 224, + 96 + ], + "score": 1.0, + "content": "3 NAS BENCHMARK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 108, + 504, + 153 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 506, + 120 + ], + "score": 1.0, + "content": "In this section we present a systematic evaluation of 8 methods on 5 datasets using a strategy that is", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 120, + 504, + 132 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 504, + 132 + ], + "score": 1.0, + "content": "designed to reveal the quality of each method’s search strategy, removing the effect of the manually-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 131, + 506, + 143 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 506, + 143 + ], + "score": 1.0, + "content": "engineered training protocol and search space. The goal is to find general trends and highlight", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 142, + 407, + 154 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 407, + 154 + ], + "score": 1.0, + "content": "common features rather than just pin-pointing the most accurate algorithm.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 109, + 506, + 154 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 158, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "Understanding why methods are effective is not an easy task: most introduce variations to previous", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 170, + 505, + 181 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 505, + 181 + ], + "score": 1.0, + "content": "search spaces, search strategies, and training protocols—with ablations disentangling the contribu-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 193 + ], + "score": 1.0, + "content": "tion of each component often incomplete or missing. In other words, how can we be sure that a new", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 204 + ], + "score": 1.0, + "content": "state-of-the-art method is not so simply due to a better engineered search space or training protocol?", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 202, + 506, + 215 + ], + "spans": [ + { + "bbox": [ + 106, + 202, + 506, + 215 + ], + "score": 1.0, + "content": "To address this issue we compare a set of 8 methods with randomly sampled architectures from their", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 213, + 468, + 226 + ], + "spans": [ + { + "bbox": [ + 105, + 213, + 468, + 226 + ], + "score": 1.0, + "content": "respective search spaces, and trained with the same protocol as the searched architectures.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 158, + 506, + 226 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 230, + 505, + 285 + ], + "lines": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 242 + ], + "score": 1.0, + "content": "The ultimate goal behind NAS should be to return the optimal model for any dataset given, at least", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 106, + 240, + 505, + 252 + ], + "score": 1.0, + "content": "within the limits of a certain task, and we feel that the current practices of searching almost exclu-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 506, + 266 + ], + "score": 1.0, + "content": "sively on CIFAR10 go against this principle. Indeed, to avoid the very concrete risk of overfitting", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 262, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 506, + 276 + ], + "score": 1.0, + "content": "to this set of data, NAS methods should be tested on a variety of tasks. For this reason we run", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 274, + 250, + 286 + ], + "spans": [ + { + "bbox": [ + 106, + 274, + 250, + 286 + ], + "score": 1.0, + "content": "experiments on 5 different datasets.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 230, + 506, + 286 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 302, + 200, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 202, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 202, + 315 + ], + "score": 1.0, + "content": "3.1 METHODOLOGY", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 325, + 505, + 380 + ], + "lines": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 505, + 338 + ], + "score": 1.0, + "content": "Criteria for dataset selection. We selected datasets to cover a variety of subtasks within image", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 505, + 349 + ], + "score": 1.0, + "content": "classification. In addition to the standard CIFAR10 we select CIFAR100 for a more challenging", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 506, + 360 + ], + "score": 1.0, + "content": "object classification problem (Krizhevsky, 2009); SPORT8 for action classification (Li & Fei-Fei,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 371 + ], + "score": 1.0, + "content": "2007); MIT67 for scene classification (Quattoni & Torralba, 2009); and FLOWERS102 for fine-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "grained object classification (Nilsback & Zisserman, 2008). More details are given in the Appendix.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 324, + 506, + 382 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 396, + 505, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "Criteria for method selection. We selected methods which (a) have open-source code, or pro-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 506, + 420 + ], + "score": 1.0, + "content": "vided it upon request, and (b) have a reasonable running time, specifically a search time under 4", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 505, + 431 + ], + "score": 1.0, + "content": "GPU-days on CIFAR10. The selected methods are: DARTS (Liu et al., 2019), StacNAS (Li et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 428, + 505, + 442 + ], + "score": 1.0, + "content": "2019), PDARTS (Xin Chen, 2019), MANAS (Carlucci et al., 2019), CNAS (Weng et al., 2019),", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "NSGANET (Lu et al., 2018), ENAS (Pham et al., 2018), and NAO (Luo et al., 2018). With the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 450, + 477, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 450, + 477, + 465 + ], + "score": 1.0, + "content": "exception of NAO and NSGANET, all methods are DARTS variants and use weight sharing.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 395, + 506, + 465 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 478, + 505, + 566 + ], + "lines": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "Evaluation protocol. NAS algorithms usually consist of two phases: (i) search, producing the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 505, + 503 + ], + "score": 1.0, + "content": "best architecture according to the search algorithm used; (ii) augmentation, consisting in training", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 505, + 513 + ], + "score": 1.0, + "content": "from scratch the best model found in the search phase. We evaluate methods as follows: 1. Sample", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 505, + 523 + ], + "score": 1.0, + "content": "8 architectures from the search space, uniformly at random, and use the method’s code to augment", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 505, + 534 + ], + "score": 1.0, + "content": "these architectures (same augment seed for all); 2. Use the method’s code to search for 8 architec-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "tures and augment them (different search seed, same augment seed); 3. Report mean and standard", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "deviation of the top-1 test accuracy, obtained at the end of the augmentation, for both the randomly", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 555, + 268, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 268, + 567 + ], + "score": 1.0, + "content": "sampled and the searched architectures;", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 478, + 506, + 567 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 572, + 505, + 672 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Since both learned and randomly sampled architectures share the same search space and training", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 581, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 408, + 597 + ], + "score": 1.0, + "content": "protocol, calculating a relative improvement over this random baseline as", + "type": "text" + }, + { + "bbox": [ + 408, + 583, + 505, + 594 + ], + "score": 0.9, + "content": "R I = 1 0 0 \\times ( A c c _ { m } -", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 594, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 157, + 606 + ], + "score": 0.9, + "content": "A c c _ { r } ) / A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 594, + 401, + 607 + ], + "score": 1.0, + "content": "can offer insights into the quality of the search strategy alone.", + "type": "text" + }, + { + "bbox": [ + 402, + 594, + 426, + 605 + ], + "score": 0.91, + "content": "A c c _ { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 594, + 443, + 607 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 444, + 595, + 465, + 605 + ], + "score": 0.9, + "content": "A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 594, + 505, + 607 + ], + "score": 1.0, + "content": "represent", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "the top-1 accuracy of the search method and random sampling strategies, respectively. A good,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 309, + 629 + ], + "score": 1.0, + "content": "general-purpose NAS method is expected to yield", + "type": "text" + }, + { + "bbox": [ + 309, + 616, + 343, + 627 + ], + "score": 0.89, + "content": "R I > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "consistently over different searches and", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "across different subtasks. We emphasize that the comparison is not against random search, but", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "rather against random sampling, i.e., the average architecture of the search space. For example, in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 661 + ], + "score": 1.0, + "content": "the DARTS search space, for each edge in the graph that defines a cell we select one out of eight", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 659, + 428, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 408, + 673 + ], + "score": 1.0, + "content": "possible operations (e.g. pooling or convolutions) with uniform probability", + "type": "text" + }, + { + "bbox": [ + 408, + 660, + 424, + 672 + ], + "score": 0.74, + "content": "1 / 8", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 659, + 428, + 673 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 572, + 506, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "Hyperparameters are optimized on CIFAR10, according to the values reported by the corresponding", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "authors. Since most methods do not include their optimization as part of the search routine, we", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "assumed them to be robust and generalizable to other tasks. As such, aside from scaling down the", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "architecture depending on dataset size, experiments on other datasets use the same hyperparameters.", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 720, + 362, + 734 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 362, + 734 + ], + "score": 1.0, + "content": "Other training details and references are given in the Appendix.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 47, + "bbox_fs": [ + 105, + 676, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 85, + 501, + 350 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 85, + 501, + 350 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 85, + 501, + 350 + ], + "spans": [ + { + "bbox": [ + 108, + 85, + 501, + 350 + ], + "score": 0.977, + "type": "image", + "image_path": "edf5f8a33f5bb147c60a4fc8d6e36c6306bb7276930131c944ec9914b79e4870.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 85, + 501, + 173.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 173.33333333333331, + 501, + 261.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 261.66666666666663, + 501, + 349.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 357, + 505, + 391 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "Figure 1: Comparison of search methods and random sampling from their respective search spaces.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "Methods lying in the diagonal perform the same as the average architecture, while methods above", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 379, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 286, + 392 + ], + "score": 1.0, + "content": "the diagonal outperform it. See also Table 1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 106, + 407, + 281, + 554 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 407, + 281, + 554 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 407, + 281, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 281, + 554 + ], + "score": 0.968, + "type": "image", + "image_path": "3203460b279ee612939f03c89b01e0c7ce0f8773d113c8c7e58f41bc3d8a1d1d.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 106, + 407, + 281, + 420.3636363636364 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 420.3636363636364, + 281, + 433.72727272727275 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 433.72727272727275, + 281, + 447.0909090909091 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 447.0909090909091, + 281, + 460.4545454545455 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 460.4545454545455, + 281, + 473.81818181818187 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 473.81818181818187, + 281, + 487.18181818181824 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 487.18181818181824, + 281, + 500.5454545454546 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 500.5454545454546, + 281, + 513.909090909091 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 513.909090909091, + 281, + 527.2727272727274 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 106, + 527.2727272727274, + 281, + 540.6363636363637 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 540.6363636363637, + 281, + 554.0000000000001 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 560, + 286, + 583 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 560, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 287, + 572 + ], + "score": 1.0, + "content": "Figure 2: Performance and computational", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 571, + 259, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 259, + 583 + ], + "score": 1.0, + "content": "cost of the search phase on CIFAR10.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 14.25 + }, + { + "type": "table", + "bbox": [ + 305, + 463, + 502, + 557 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 304, + 410, + 504, + 455 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 303, + 410, + 475, + 422 + ], + "score": 1.0, + "content": "Table 1: Relative improvement metric,", + "type": "text" + }, + { + "bbox": [ + 475, + 411, + 505, + 421 + ], + "score": 0.85, + "content": "\\begin{array} { r l } { R I } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 420, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 421, + 434 + ], + "score": 0.91, + "content": "1 0 0 \\times ( A c c _ { m } - A c c _ { r } ) / A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 420, + 435, + 435 + ], + "score": 1.0, + "content": "(in", + "type": "text" + }, + { + "bbox": [ + 435, + 422, + 445, + 432 + ], + "score": 0.65, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 420, + 479, + 435 + ], + "score": 1.0, + "content": "), where", + "type": "text" + }, + { + "bbox": [ + 479, + 422, + 504, + 433 + ], + "score": 0.88, + "content": "A c c _ { m }", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 303, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 303, + 432, + 321, + 444 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 322, + 433, + 343, + 443 + ], + "score": 0.88, + "content": "A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "are the accuracies of the search method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 303, + 443, + 483, + 456 + ], + "spans": [ + { + "bbox": [ + 303, + 443, + 483, + 456 + ], + "score": 1.0, + "content": "and random sampling baseline, respectively.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "table_body", + "bbox": [ + 305, + 463, + 502, + 557 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 305, + 463, + 502, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 502, + 557 + ], + "score": 0.975, + "html": "
C10C100S8M67F102
DARTS0.320.23-0.130.100.25
PDARTS0.521.200.511.190.20
NSGANET-0.481.370.432.001.47
ENAS0.01-3.440.670.130.47
CNAS0.74-0.89-1.06-0.66-2.48
MANAS0.18-0.200.331.480.70
StacNAS NAO0.43 0.442.87 -0.010.38 -2.050.05 -1.53-0.16 -0.13
", + "type": "table", + "image_path": "c01388e0b766ac927995cde360d93dfc9a5de1b355216979bcf6a2dd53aecde9.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 305, + 463, + 502, + 476.42857142857144 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 305, + 476.42857142857144, + 502, + 489.8571428571429 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 305, + 489.8571428571429, + 502, + 503.28571428571433 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 305, + 503.28571428571433, + 502, + 516.7142857142858 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 305, + 516.7142857142858, + 502, + 530.1428571428572 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 305, + 530.1428571428572, + 502, + 543.5714285714287 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 305, + 543.5714285714287, + 502, + 557.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 23.25 + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 171, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 171, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 171, + 608 + ], + "score": 1.0, + "content": "3.2 RESULTS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 615, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Figure 1 shows the evaluation results on the 5 datasets, from which we draw two main conclusions.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "First, the improvements over random sampling tend to be small. In some cases the average perfor-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "mance of a method is even below the average randomly sampled architecture, which suggests that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "the search methods are not converging to desirable architectures. Second, the small range of accura-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "score": 1.0, + "content": "cies obtained hints at narrow search spaces, where even the worst architectures perform reasonably", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 670, + 396, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 396, + 684 + ], + "score": 1.0, + "content": "well. See Section 5 for more experiments corroborating this conclusion.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "We observe also that, on CIFAR10, the top half of best-performing methods (PDARTS, MANAS,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "DARTS, StacNAS) all perform similarly and positively in relation to their respective search spaces,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "but more variance is seen on the other datasets. This could be explained by the fact that most", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "methods’ hyperparameters have been optimized on CIFAR10 and might not generalize as well on", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 85, + 501, + 350 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 85, + 501, + 350 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 85, + 501, + 350 + ], + "spans": [ + { + "bbox": [ + 108, + 85, + 501, + 350 + ], + "score": 0.977, + "type": "image", + "image_path": "edf5f8a33f5bb147c60a4fc8d6e36c6306bb7276930131c944ec9914b79e4870.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 108, + 85, + 501, + 173.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 108, + 173.33333333333331, + 501, + 261.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 108, + 261.66666666666663, + 501, + 349.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 357, + 505, + 391 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 505, + 370 + ], + "score": 1.0, + "content": "Figure 1: Comparison of search methods and random sampling from their respective search spaces.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "Methods lying in the diagonal perform the same as the average architecture, while methods above", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 379, + 286, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 286, + 392 + ], + "score": 1.0, + "content": "the diagonal outperform it. See also Table 1.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "image", + "bbox": [ + 106, + 407, + 281, + 554 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 407, + 281, + 554 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 407, + 281, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 281, + 554 + ], + "score": 0.968, + "type": "image", + "image_path": "3203460b279ee612939f03c89b01e0c7ce0f8773d113c8c7e58f41bc3d8a1d1d.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 106, + 407, + 281, + 420.3636363636364 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 420.3636363636364, + 281, + 433.72727272727275 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 106, + 433.72727272727275, + 281, + 447.0909090909091 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 447.0909090909091, + 281, + 460.4545454545455 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 460.4545454545455, + 281, + 473.81818181818187 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 106, + 473.81818181818187, + 281, + 487.18181818181824 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 106, + 487.18181818181824, + 281, + 500.5454545454546 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 106, + 500.5454545454546, + 281, + 513.909090909091 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 106, + 513.909090909091, + 281, + 527.2727272727274 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 106, + 527.2727272727274, + 281, + 540.6363636363637 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 106, + 540.6363636363637, + 281, + 554.0000000000001 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 560, + 286, + 583 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 560, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 287, + 572 + ], + "score": 1.0, + "content": "Figure 2: Performance and computational", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 571, + 259, + 583 + ], + "spans": [ + { + "bbox": [ + 106, + 571, + 259, + 583 + ], + "score": 1.0, + "content": "cost of the search phase on CIFAR10.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 14.25 + }, + { + "type": "table", + "bbox": [ + 305, + 463, + 502, + 557 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 304, + 410, + 504, + 455 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 303, + 410, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 303, + 410, + 475, + 422 + ], + "score": 1.0, + "content": "Table 1: Relative improvement metric,", + "type": "text" + }, + { + "bbox": [ + 475, + 411, + 505, + 421 + ], + "score": 0.85, + "content": "\\begin{array} { r l } { R I } & { { } = } \\end{array}", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 420, + 504, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 421, + 421, + 434 + ], + "score": 0.91, + "content": "1 0 0 \\times ( A c c _ { m } - A c c _ { r } ) / A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 420, + 435, + 435 + ], + "score": 1.0, + "content": "(in", + "type": "text" + }, + { + "bbox": [ + 435, + 422, + 445, + 432 + ], + "score": 0.65, + "content": "\\%", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 420, + 479, + 435 + ], + "score": 1.0, + "content": "), where", + "type": "text" + }, + { + "bbox": [ + 479, + 422, + 504, + 433 + ], + "score": 0.88, + "content": "A c c _ { m }", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 303, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 303, + 432, + 321, + 444 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 322, + 433, + 343, + 443 + ], + "score": 0.88, + "content": "A c c _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "are the accuracies of the search method", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 303, + 443, + 483, + 456 + ], + "spans": [ + { + "bbox": [ + 303, + 443, + 483, + 456 + ], + "score": 1.0, + "content": "and random sampling baseline, respectively.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "table_body", + "bbox": [ + 305, + 463, + 502, + 557 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 305, + 463, + 502, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 463, + 502, + 557 + ], + "score": 0.975, + "html": "
C10C100S8M67F102
DARTS0.320.23-0.130.100.25
PDARTS0.521.200.511.190.20
NSGANET-0.481.370.432.001.47
ENAS0.01-3.440.670.130.47
CNAS0.74-0.89-1.06-0.66-2.48
MANAS0.18-0.200.331.480.70
StacNAS NAO0.43 0.442.87 -0.010.38 -2.050.05 -1.53-0.16 -0.13
", + "type": "table", + "image_path": "c01388e0b766ac927995cde360d93dfc9a5de1b355216979bcf6a2dd53aecde9.jpg" + } + ] + } + ], + "index": 26, + "virtual_lines": [ + { + "bbox": [ + 305, + 463, + 502, + 476.42857142857144 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 305, + 476.42857142857144, + 502, + 489.8571428571429 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 305, + 489.8571428571429, + 502, + 503.28571428571433 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 305, + 503.28571428571433, + 502, + 516.7142857142858 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 305, + 516.7142857142858, + 502, + 530.1428571428572 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 305, + 530.1428571428572, + 502, + 543.5714285714287 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 305, + 543.5714285714287, + 502, + 557.0000000000001 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 23.25 + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 171, + 606 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 171, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 171, + 608 + ], + "score": 1.0, + "content": "3.2 RESULTS", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 615, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Figure 1 shows the evaluation results on the 5 datasets, from which we draw two main conclusions.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "First, the improvements over random sampling tend to be small. In some cases the average perfor-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 506, + 651 + ], + "score": 1.0, + "content": "mance of a method is even below the average randomly sampled architecture, which suggests that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "the search methods are not converging to desirable architectures. Second, the small range of accura-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 505, + 672 + ], + "score": 1.0, + "content": "cies obtained hints at narrow search spaces, where even the worst architectures perform reasonably", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 670, + 396, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 396, + 684 + ], + "score": 1.0, + "content": "well. See Section 5 for more experiments corroborating this conclusion.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 616, + 506, + 684 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "We observe also that, on CIFAR10, the top half of best-performing methods (PDARTS, MANAS,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "DARTS, StacNAS) all perform similarly and positively in relation to their respective search spaces,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "but more variance is seen on the other datasets. This could be explained by the fact that most", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "methods’ hyperparameters have been optimized on CIFAR10 and might not generalize as well on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "different datasets. As a matter of fact, we found that all NAS methods neglect to report the time", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "needed to optimize hyperparameters. In addition, Table 1 shows the relative improvement metric", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 104, + 332, + 117 + ], + "spans": [ + { + "bbox": [ + 107, + 105, + 121, + 115 + ], + "score": 0.39, + "content": "R I", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 121, + 104, + 332, + 117 + ], + "score": 1.0, + "content": "(see intro to Section 3) for each method and dataset.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "different datasets. As a matter of fact, we found that all NAS methods neglect to report the time", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "needed to optimize hyperparameters. In addition, Table 1 shows the relative improvement metric", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 104, + 332, + 117 + ], + "spans": [ + { + "bbox": [ + 107, + 105, + 121, + 115 + ], + "score": 0.39, + "content": "R I", + "type": "inline_equation" + }, + { + "bbox": [ + 121, + 104, + 332, + 117 + ], + "score": 1.0, + "content": "(see intro to Section 3) for each method and dataset.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 503, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 504, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 504, + 133 + ], + "score": 1.0, + "content": "The computational cost of searching for architectures is a limiting factor in their applicability and,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "therefore, an important variable in the evaluation of NAS algorithms. Figure 2 shows the perfor-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 410, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 410, + 156 + ], + "score": 1.0, + "content": "mance as well and the computational cost of the search phase on CIFAR10.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 107, + 179, + 329, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 178, + 330, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 330, + 193 + ], + "score": 1.0, + "content": "4 COMPARISON OF TRAINING PROTOCOLS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 336, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 336, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 336, + 220 + ], + "score": 1.0, + "content": "In this section we attempt to shed some light on the sur-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 219, + 337, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 337, + 231 + ], + "score": 1.0, + "content": "prising results of the previous section. We noticed that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 231, + 336, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 336, + 242 + ], + "score": 1.0, + "content": "there was a much larger differences between the random", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 241, + 336, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 336, + 253 + ], + "score": 1.0, + "content": "baselines of different methods than the actual increase in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 252, + 336, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 336, + 264 + ], + "score": 1.0, + "content": "performance of each approach. We hypothesized that how", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 263, + 337, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 337, + 275 + ], + "score": 1.0, + "content": "a network is trained (the training protocol) has a larger", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 274, + 336, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 336, + 285 + ], + "score": 1.0, + "content": "impact on the final accuracy than which architecture is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 336, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 336, + 297 + ], + "score": 1.0, + "content": "trained, for each search space. To test this, we performed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 295, + 336, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 336, + 308 + ], + "score": 1.0, + "content": "sensitivity analysis using the most common performance-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 307, + 219, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 219, + 320 + ], + "score": 1.0, + "content": "boosting training protocols.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 340, + 200, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 202, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 202, + 353 + ], + "score": 1.0, + "content": "4.1 METHODOLOGY", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 336, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 336, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 336, + 376 + ], + "score": 1.0, + "content": "We decided to evaluate architectures from the commonly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 374, + 336, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 336, + 386 + ], + "score": 1.0, + "content": "used DARTS search space (Liu et al., 2019) on the CI-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 385, + 336, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 336, + 398 + ], + "score": 1.0, + "content": "FAR10 dataset. We use the following process: 1) sample", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 396, + 336, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 336, + 408 + ], + "score": 1.0, + "content": "8 random architectures, 2) train them with different train-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 407, + 337, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 337, + 419 + ], + "score": 1.0, + "content": "ing protocols (details below) and 3) report mean, standard", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 418, + 337, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 337, + 431 + ], + "score": 1.0, + "content": "deviation and maximum of the top-1 test accuracy at the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 429, + 217, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 217, + 442 + ], + "score": 1.0, + "content": "end of the training process.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 344, + 208, + 502, + 387 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 344, + 208, + 502, + 387 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 344, + 208, + 502, + 387 + ], + "spans": [ + { + "bbox": [ + 344, + 208, + 502, + 387 + ], + "score": 0.97, + "type": "image", + "image_path": "caa55a31af78adab243a490ab57323412ef1ea1347232a5cd9f8350973a2321e.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 344, + 208, + 502, + 221.76923076923077 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 344, + 221.76923076923077, + 502, + 235.53846153846155 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 344, + 235.53846153846155, + 502, + 249.30769230769232 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 344, + 249.30769230769232, + 502, + 263.0769230769231 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 344, + 263.0769230769231, + 502, + 276.84615384615387 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 344, + 276.84615384615387, + 502, + 290.61538461538464 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 344, + 290.61538461538464, + 502, + 304.3846153846154 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 344, + 304.3846153846154, + 502, + 318.1538461538462 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 344, + 318.1538461538462, + 502, + 331.92307692307696 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 344, + 331.92307692307696, + 502, + 345.69230769230774 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 344, + 345.69230769230774, + 502, + 359.4615384615385 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 344, + 359.4615384615385, + 502, + 373.2307692307693 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 344, + 373.2307692307693, + 502, + 387.00000000000006 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 344, + 392, + 504, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 343, + 390, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 343, + 390, + 505, + 405 + ], + "score": 1.0, + "content": "Figure 3: Comparison of different aug-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 343, + 404, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 343, + 404, + 505, + 414 + ], + "score": 1.0, + "content": "mentation protocols for the DARTS", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 343, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 343, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "search space on CIFAR10. Same col-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 344, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 344, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "ored dots represent minimum and max-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 343, + 437, + 469, + 448 + ], + "spans": [ + { + "bbox": [ + 343, + 437, + 469, + 448 + ], + "score": 1.0, + "content": "imum accuracies in the 8 runs.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 335, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 337, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 337, + 475 + ], + "score": 1.0, + "content": "Training protocols. The simplest training protocol,", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 43 + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 505, + 571 + ], + "lines": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "which we will call Base is similar to the one used in DARTS, but with all tricks disabled: the model", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "is simply trained for 600 epochs. On the other extreme, our full protocol uses several tricks which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "have been used in recent works (Xie et al., 2019b; Nayman et al., 2019): Auxiliary Towers (A),", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 504, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 518 + ], + "score": 1.0, + "content": "DropPath (D; Larsson & Shakhnarovich, 2017), Cutout (C; DeVries & Taylor, 2017), AutoAug-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "ment (AA; Cubuk et al., 2018), extended training for 1500 epochs (1500E), and increased number", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "of channels (50C). In between these two extremes, by selectively enabling and disabling each com-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "ponent, we evaluated a further 8 intermediate training protocols. When active, DropPath probability", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "is 0.2, cutout length is 16, auxiliary tower weight is 0.4, and AutoAugment combined with Cutout", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 560, + 495, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 495, + 572 + ], + "score": 1.0, + "content": "are used after standard data pre-processing techniques previously described, as in Popien (2019).", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 48 + }, + { + "type": "title", + "bbox": [ + 107, + 592, + 170, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 172, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 172, + 605 + ], + "score": 1.0, + "content": "4.2 RESULTS", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "As shown in Figure 3, a large difference of over 3 percentage points (p.p.) exists between the sim-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 627, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 638 + ], + "score": 1.0, + "content": "plest and the most advanced training protocols. Indeed, this is much higher than any improvement", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "over random sampling observed in the previous section: for example, on CIFAR10, the best im-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "provement observed was 0.69 p.p. In other words, the training protocol is often far more important", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "than the architecture used. Note that the best accuracy of the 8 random architectures training with", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 671, + 495, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 184, + 683 + ], + "score": 1.0, + "content": "the best protocol is", + "type": "text" + }, + { + "bbox": [ + 185, + 671, + 217, + 681 + ], + "score": 0.89, + "content": "9 8 . 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 671, + 495, + 683 + ], + "score": 1.0, + "content": ", which is only 0.25 p.p. below state-of-the-art (Nayman et al., 2019).", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 56.5 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "To summarize, it seems that most recent state-of-the-art results, though impressive, cannot always be", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "attributed to superior search strategies. Rather, they are often the result of expert knowledge applied", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "to the evaluation protocol. In Figure 10 (Appendix A.3.1) we show similar results when training a", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 720, + 319, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 319, + 733 + ], + "score": 1.0, + "content": "ResNet-50 (He et al., 2016) with the same protocols.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 61.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 117 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 503, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 504, + 133 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 504, + 133 + ], + "score": 1.0, + "content": "The computational cost of searching for architectures is a limiting factor in their applicability and,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "therefore, an important variable in the evaluation of NAS algorithms. Figure 2 shows the perfor-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 410, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 410, + 156 + ], + "score": 1.0, + "content": "mance as well and the computational cost of the search phase on CIFAR10.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 122, + 505, + 156 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 179, + 329, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 178, + 330, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 330, + 193 + ], + "score": 1.0, + "content": "4 COMPARISON OF TRAINING PROTOCOLS", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 208, + 336, + 318 + ], + "lines": [ + { + "bbox": [ + 105, + 208, + 336, + 220 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 336, + 220 + ], + "score": 1.0, + "content": "In this section we attempt to shed some light on the sur-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 219, + 337, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 337, + 231 + ], + "score": 1.0, + "content": "prising results of the previous section. We noticed that", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 231, + 336, + 242 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 336, + 242 + ], + "score": 1.0, + "content": "there was a much larger differences between the random", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 241, + 336, + 253 + ], + "spans": [ + { + "bbox": [ + 106, + 241, + 336, + 253 + ], + "score": 1.0, + "content": "baselines of different methods than the actual increase in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 252, + 336, + 264 + ], + "spans": [ + { + "bbox": [ + 106, + 252, + 336, + 264 + ], + "score": 1.0, + "content": "performance of each approach. We hypothesized that how", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 263, + 337, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 337, + 275 + ], + "score": 1.0, + "content": "a network is trained (the training protocol) has a larger", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 274, + 336, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 336, + 285 + ], + "score": 1.0, + "content": "impact on the final accuracy than which architecture is", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 285, + 336, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 336, + 297 + ], + "score": 1.0, + "content": "trained, for each search space. To test this, we performed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 295, + 336, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 336, + 308 + ], + "score": 1.0, + "content": "sensitivity analysis using the most common performance-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 307, + 219, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 219, + 320 + ], + "score": 1.0, + "content": "boosting training protocols.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 208, + 337, + 320 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 340, + 200, + 351 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 202, + 353 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 202, + 353 + ], + "score": 1.0, + "content": "4.1 METHODOLOGY", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 363, + 336, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 336, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 336, + 376 + ], + "score": 1.0, + "content": "We decided to evaluate architectures from the commonly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 374, + 336, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 336, + 386 + ], + "score": 1.0, + "content": "used DARTS search space (Liu et al., 2019) on the CI-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 385, + 336, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 336, + 398 + ], + "score": 1.0, + "content": "FAR10 dataset. We use the following process: 1) sample", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 396, + 336, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 336, + 408 + ], + "score": 1.0, + "content": "8 random architectures, 2) train them with different train-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 407, + 337, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 407, + 337, + 419 + ], + "score": 1.0, + "content": "ing protocols (details below) and 3) report mean, standard", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 418, + 337, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 337, + 431 + ], + "score": 1.0, + "content": "deviation and maximum of the top-1 test accuracy at the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 429, + 217, + 442 + ], + "spans": [ + { + "bbox": [ + 106, + 429, + 217, + 442 + ], + "score": 1.0, + "content": "end of the training process.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 362, + 337, + 442 + ] + }, + { + "type": "image", + "bbox": [ + 344, + 208, + 502, + 387 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 344, + 208, + 502, + 387 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 344, + 208, + 502, + 387 + ], + "spans": [ + { + "bbox": [ + 344, + 208, + 502, + 387 + ], + "score": 0.97, + "type": "image", + "image_path": "caa55a31af78adab243a490ab57323412ef1ea1347232a5cd9f8350973a2321e.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 344, + 208, + 502, + 221.76923076923077 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 344, + 221.76923076923077, + 502, + 235.53846153846155 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 344, + 235.53846153846155, + 502, + 249.30769230769232 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 344, + 249.30769230769232, + 502, + 263.0769230769231 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 344, + 263.0769230769231, + 502, + 276.84615384615387 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 344, + 276.84615384615387, + 502, + 290.61538461538464 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 344, + 290.61538461538464, + 502, + 304.3846153846154 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 344, + 304.3846153846154, + 502, + 318.1538461538462 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 344, + 318.1538461538462, + 502, + 331.92307692307696 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 344, + 331.92307692307696, + 502, + 345.69230769230774 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 344, + 345.69230769230774, + 502, + 359.4615384615385 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 344, + 359.4615384615385, + 502, + 373.2307692307693 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 344, + 373.2307692307693, + 502, + 387.00000000000006 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 344, + 392, + 504, + 447 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 343, + 390, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 343, + 390, + 505, + 405 + ], + "score": 1.0, + "content": "Figure 3: Comparison of different aug-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 343, + 404, + 505, + 414 + ], + "spans": [ + { + "bbox": [ + 343, + 404, + 505, + 414 + ], + "score": 1.0, + "content": "mentation protocols for the DARTS", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 343, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 343, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "search space on CIFAR10. Same col-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 344, + 425, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 344, + 425, + 505, + 437 + ], + "score": 1.0, + "content": "ored dots represent minimum and max-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 343, + 437, + 469, + 448 + ], + "spans": [ + { + "bbox": [ + 343, + 437, + 469, + 448 + ], + "score": 1.0, + "content": "imum accuracies in the 8 runs.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 461, + 335, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 337, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 337, + 475 + ], + "score": 1.0, + "content": "Training protocols. The simplest training protocol,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 505, + 484 + ], + "score": 1.0, + "content": "which we will call Base is similar to the one used in DARTS, but with all tricks disabled: the model", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 104, + 483, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 483, + 505, + 495 + ], + "score": 1.0, + "content": "is simply trained for 600 epochs. On the other extreme, our full protocol uses several tricks which", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 505, + 506 + ], + "score": 1.0, + "content": "have been used in recent works (Xie et al., 2019b; Nayman et al., 2019): Auxiliary Towers (A),", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 504, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 505, + 518 + ], + "score": 1.0, + "content": "DropPath (D; Larsson & Shakhnarovich, 2017), Cutout (C; DeVries & Taylor, 2017), AutoAug-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 528 + ], + "score": 1.0, + "content": "ment (AA; Cubuk et al., 2018), extended training for 1500 epochs (1500E), and increased number", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 505, + 540 + ], + "score": 1.0, + "content": "of channels (50C). In between these two extremes, by selectively enabling and disabling each com-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 537, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 505, + 551 + ], + "score": 1.0, + "content": "ponent, we evaluated a further 8 intermediate training protocols. When active, DropPath probability", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 506, + 561 + ], + "score": 1.0, + "content": "is 0.2, cutout length is 16, auxiliary tower weight is 0.4, and AutoAugment combined with Cutout", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 560, + 495, + 572 + ], + "spans": [ + { + "bbox": [ + 106, + 560, + 495, + 572 + ], + "score": 1.0, + "content": "are used after standard data pre-processing techniques previously described, as in Popien (2019).", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 43, + "bbox_fs": [ + 105, + 459, + 337, + 475 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 468, + 505, + 571 + ], + "lines": [], + "index": 48, + "bbox_fs": [ + 104, + 471, + 506, + 572 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 107, + 592, + 170, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 172, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 172, + 605 + ], + "score": 1.0, + "content": "4.2 RESULTS", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 107, + 615, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 506, + 629 + ], + "score": 1.0, + "content": "As shown in Figure 3, a large difference of over 3 percentage points (p.p.) exists between the sim-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 627, + 505, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 638 + ], + "score": 1.0, + "content": "plest and the most advanced training protocols. Indeed, this is much higher than any improvement", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "over random sampling observed in the previous section: for example, on CIFAR10, the best im-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 506, + 662 + ], + "score": 1.0, + "content": "provement observed was 0.69 p.p. In other words, the training protocol is often far more important", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "than the architecture used. Note that the best accuracy of the 8 random architectures training with", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 106, + 671, + 495, + 683 + ], + "spans": [ + { + "bbox": [ + 106, + 671, + 184, + 683 + ], + "score": 1.0, + "content": "the best protocol is", + "type": "text" + }, + { + "bbox": [ + 185, + 671, + 217, + 681 + ], + "score": 0.89, + "content": "9 8 . 1 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 671, + 495, + 683 + ], + "score": 1.0, + "content": ", which is only 0.25 p.p. below state-of-the-art (Nayman et al., 2019).", + "type": "text" + } + ], + "index": 59 + } + ], + "index": 56.5, + "bbox_fs": [ + 105, + 616, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 699 + ], + "score": 1.0, + "content": "To summarize, it seems that most recent state-of-the-art results, though impressive, cannot always be", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 711 + ], + "score": 1.0, + "content": "attributed to superior search strategies. Rather, they are often the result of expert knowledge applied", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 723 + ], + "score": 1.0, + "content": "to the evaluation protocol. In Figure 10 (Appendix A.3.1) we show similar results when training a", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 720, + 319, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 319, + 733 + ], + "score": 1.0, + "content": "ResNet-50 (He et al., 2016) with the same protocols.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 61.5, + "bbox_fs": [ + 105, + 688, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 84, + 294, + 208 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 84, + 294, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 84, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 109, + 84, + 294, + 208 + ], + "score": 0.972, + "type": "image", + "image_path": "f9ef12213677fc9a4669a42a87f74fd716606dd30dbac4f7517b2b24659054a0.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 109, + 84, + 294, + 97.77777777777777 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 97.77777777777777, + 294, + 111.55555555555554 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 111.55555555555554, + 294, + 125.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 125.33333333333331, + 294, + 139.1111111111111 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 139.1111111111111, + 294, + 152.88888888888886 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 152.88888888888886, + 294, + 166.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 166.66666666666663, + 294, + 180.4444444444444 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 180.4444444444444, + 294, + 194.22222222222217 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 194.22222222222217, + 294, + 207.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 298, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 214, + 299, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 299, + 229 + ], + "score": 1.0, + "content": "Figure 4: Training curves for the 214 randomly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 226, + 299, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 299, + 239 + ], + "score": 1.0, + "content": "sampled architectures. Inset plot shows the his-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 237, + 270, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 270, + 250 + ], + "score": 1.0, + "content": "togram of accuracies at different epochs.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 315, + 84, + 500, + 208 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 84, + 500, + 208 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 315, + 84, + 500, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 500, + 208 + ], + "score": 0.972, + "type": "image", + "image_path": "3ab1d0af1522db09538fa5918d8022741f2b541c8dfb7d338233ba12f9eb1f4f.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 315, + 84, + 500, + 97.77777777777777 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 315, + 97.77777777777777, + 500, + 111.55555555555554 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 315, + 111.55555555555554, + 500, + 125.33333333333331 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 315, + 125.33333333333331, + 500, + 139.1111111111111 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 315, + 139.1111111111111, + 500, + 152.88888888888886 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 315, + 152.88888888888886, + 500, + 166.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 315, + 166.66666666666663, + 500, + 180.4444444444444 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 315, + 180.4444444444444, + 500, + 194.22222222222217 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 315, + 194.22222222222217, + 500, + 207.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 216, + 505, + 249 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 312, + 216, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 216, + 504, + 227 + ], + "score": 1.0, + "content": "Figure 5: Correlation between accuracies at dif-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 311, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "ferent epochs and final accuracy, using raw and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 238, + 471, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 459, + 249 + ], + "score": 1.0, + "content": "smoothed accuracies over a window", + "type": "text" + }, + { + "bbox": [ + 459, + 240, + 467, + 248 + ], + "score": 0.69, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 238, + 471, + 249 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 282, + 306, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 307, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 307, + 297 + ], + "score": 1.0, + "content": "5 STUDY OF DARTS’ SEARCH SPACE", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 316, + 454, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 456, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 456, + 329 + ], + "score": 1.0, + "content": "5.1 DISTRIBUTION OF THE RANDOM SAMPLING WITHIN DARTS SEARCH SPACE", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "To better understand the results from the previous section, we sampled a considerable number of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "architectures (214) from the most commonly used search space (Liu et al., 2019) and fully trained", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "them with the matching training protocol (Cutout+DropPath+Auxiliary Towers). This allows us to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "get a sense of how much variance exists between the different models (training statistics are made", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 388, + 250, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 250, + 400 + ], + "score": 1.0, + "content": "available at the link in the abstract).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 108, + 404, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 105, + 403, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 418 + ], + "score": 1.0, + "content": "As we can observe from Figure 4, architectures sampled from this search space all perform similarly,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 169, + 428 + ], + "score": 1.0, + "content": "with a mean of", + "type": "text" + }, + { + "bbox": [ + 169, + 416, + 223, + 426 + ], + "score": 0.91, + "content": "9 7 . 0 3 \\pm 0 . 2 3", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 415, + 505, + 428 + ], + "score": 1.0, + "content": ". The worst architecture we found had an accuracy of 96.18, while the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 426, + 190, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 190, + 438 + ], + "score": 1.0, + "content": "best achieved 97.56.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "To put this into perspective, many methods using the same training protocol, fall within (or very", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "close to) the standard deviation of the average architecture. Furthermore, as we can observe in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 506, + 479 + ], + "score": 1.0, + "content": "Figure 7, the number of cells (a human-picked hyperparameter) has a much larger impact on the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 476, + 167, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 167, + 489 + ], + "score": 1.0, + "content": "final accuracy.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "In Figure 5 we used the training statistics of the 214 models to plot the correlation between test", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "accuracies at different epochs: it grows slowly in an almost linear fashion. We note that using the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "moving average of the accuracies yields a stronger correlation, which could be useful for methods", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 525, + 192, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 192, + 541 + ], + "score": 1.0, + "content": "using early stopping.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5 + }, + { + "type": "title", + "bbox": [ + 107, + 567, + 187, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 189, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 189, + 580 + ], + "score": 1.0, + "content": "5.2 OPERATIONS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "To test whether the results from the previous section were due to the choice of available operations,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 455, + 618 + ], + "score": 1.0, + "content": "we developed an intentionally sub-optimal search space containing 4 plain convolutions", + "type": "text" + }, + { + "bbox": [ + 456, + 605, + 501, + 616 + ], + "score": 0.65, + "content": "( 1 \\times 1 , 3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 605, + 505, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 131, + 627 + ], + "score": 0.69, + "content": "7 \\times 7 .", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 615, + 133, + 630 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 133, + 616, + 168, + 627 + ], + "score": 0.75, + "content": "1 1 \\times 1 1", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 615, + 275, + 630 + ], + "score": 1.0, + "content": "), 2 max pooling operators", + "type": "text" + }, + { + "bbox": [ + 276, + 616, + 331, + 627 + ], + "score": 0.6, + "content": "( 3 \\times 3 , 5 \\times 5 )", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "plus the none and skip connect operations.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "This proposed search space is clearly more parameter inefficient compared to the commonly used", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 637, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 482, + 652 + ], + "score": 1.0, + "content": "DARTS one (which uses both dilated and separable ones), and we expect it to perform worse.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45 + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "We sampled 56 architectures from this new search space and trained them with the DARTS train-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 190, + 677 + ], + "score": 1.0, + "content": "ing protocol (Cutout", + "type": "text" + }, + { + "bbox": [ + 190, + 667, + 196, + 675 + ], + "score": 0.32, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 666, + 234, + 677 + ], + "score": 1.0, + "content": "DropPath", + "type": "text" + }, + { + "bbox": [ + 234, + 667, + 240, + 675 + ], + "score": 0.34, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Auxiliary Towers), for fair comparison with the results from the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "previous section. Figure 6 shows the resulting histogram, together with the one obtained from the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "classical DARTS space of operations. The two distributions are only shifted by 0.18 accuracy points.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Given the minor difference in performance, the specific operations are not a key ingredient in the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "success of this search space. Very likely, it’s the well engineered cell structure that allows the model", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 721, + 224, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 224, + 733 + ], + "score": 1.0, + "content": "to perform as well as it does.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 84, + 294, + 208 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 84, + 294, + 208 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 84, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 109, + 84, + 294, + 208 + ], + "score": 0.972, + "type": "image", + "image_path": "f9ef12213677fc9a4669a42a87f74fd716606dd30dbac4f7517b2b24659054a0.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 109, + 84, + 294, + 97.77777777777777 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 97.77777777777777, + 294, + 111.55555555555554 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 111.55555555555554, + 294, + 125.33333333333331 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 125.33333333333331, + 294, + 139.1111111111111 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 139.1111111111111, + 294, + 152.88888888888886 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 152.88888888888886, + 294, + 166.66666666666663 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 166.66666666666663, + 294, + 180.4444444444444 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 180.4444444444444, + 294, + 194.22222222222217 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 194.22222222222217, + 294, + 207.99999999999994 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 298, + 249 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 214, + 299, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 299, + 229 + ], + "score": 1.0, + "content": "Figure 4: Training curves for the 214 randomly", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 226, + 299, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 299, + 239 + ], + "score": 1.0, + "content": "sampled architectures. Inset plot shows the his-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 237, + 270, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 237, + 270, + 250 + ], + "score": 1.0, + "content": "togram of accuracies at different epochs.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + } + ], + "index": 11.5 + }, + { + "type": "image", + "bbox": [ + 315, + 84, + 500, + 208 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 84, + 500, + 208 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 315, + 84, + 500, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 84, + 500, + 208 + ], + "score": 0.972, + "type": "image", + "image_path": "3ab1d0af1522db09538fa5918d8022741f2b541c8dfb7d338233ba12f9eb1f4f.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 315, + 84, + 500, + 97.77777777777777 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 315, + 97.77777777777777, + 500, + 111.55555555555554 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 315, + 111.55555555555554, + 500, + 125.33333333333331 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 315, + 125.33333333333331, + 500, + 139.1111111111111 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 315, + 139.1111111111111, + 500, + 152.88888888888886 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 315, + 152.88888888888886, + 500, + 166.66666666666663 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 315, + 166.66666666666663, + 500, + 180.4444444444444 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 315, + 180.4444444444444, + 500, + 194.22222222222217 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 315, + 194.22222222222217, + 500, + 207.99999999999994 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 216, + 505, + 249 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 312, + 216, + 504, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 216, + 504, + 227 + ], + "score": 1.0, + "content": "Figure 5: Correlation between accuracies at dif-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 311, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 311, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "ferent epochs and final accuracy, using raw and", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 238, + 471, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 459, + 249 + ], + "score": 1.0, + "content": "smoothed accuracies over a window", + "type": "text" + }, + { + "bbox": [ + 459, + 240, + 467, + 248 + ], + "score": 0.69, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 238, + 471, + 249 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 282, + 306, + 295 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 307, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 307, + 297 + ], + "score": 1.0, + "content": "5 STUDY OF DARTS’ SEARCH SPACE", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 316, + 454, + 328 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 456, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 456, + 329 + ], + "score": 1.0, + "content": "5.1 DISTRIBUTION OF THE RANDOM SAMPLING WITHIN DARTS SEARCH SPACE", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 343, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "spans": [ + { + "bbox": [ + 106, + 343, + 505, + 356 + ], + "score": 1.0, + "content": "To better understand the results from the previous section, we sampled a considerable number of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 505, + 366 + ], + "score": 1.0, + "content": "architectures (214) from the most commonly used search space (Liu et al., 2019) and fully trained", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 506, + 379 + ], + "score": 1.0, + "content": "them with the matching training protocol (Cutout+DropPath+Auxiliary Towers). This allows us to", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 506, + 390 + ], + "score": 1.0, + "content": "get a sense of how much variance exists between the different models (training statistics are made", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 388, + 250, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 250, + 400 + ], + "score": 1.0, + "content": "available at the link in the abstract).", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 343, + 506, + 400 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 404, + 505, + 438 + ], + "lines": [ + { + "bbox": [ + 105, + 403, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 418 + ], + "score": 1.0, + "content": "As we can observe from Figure 4, architectures sampled from this search space all perform similarly,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 169, + 428 + ], + "score": 1.0, + "content": "with a mean of", + "type": "text" + }, + { + "bbox": [ + 169, + 416, + 223, + 426 + ], + "score": 0.91, + "content": "9 7 . 0 3 \\pm 0 . 2 3", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 415, + 505, + 428 + ], + "score": 1.0, + "content": ". The worst architecture we found had an accuracy of 96.18, while the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 426, + 190, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 190, + 438 + ], + "score": 1.0, + "content": "best achieved 97.56.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 403, + 505, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 443, + 504, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 456 + ], + "score": 1.0, + "content": "To put this into perspective, many methods using the same training protocol, fall within (or very", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "close to) the standard deviation of the average architecture. Furthermore, as we can observe in", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 104, + 465, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 506, + 479 + ], + "score": 1.0, + "content": "Figure 7, the number of cells (a human-picked hyperparameter) has a much larger impact on the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 476, + 167, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 167, + 489 + ], + "score": 1.0, + "content": "final accuracy.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 443, + 506, + 489 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 493, + 505, + 538 + ], + "lines": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 493, + 505, + 505 + ], + "score": 1.0, + "content": "In Figure 5 we used the training statistics of the 214 models to plot the correlation between test", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 505, + 516 + ], + "score": 1.0, + "content": "accuracies at different epochs: it grows slowly in an almost linear fashion. We note that using the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 527 + ], + "score": 1.0, + "content": "moving average of the accuracies yields a stronger correlation, which could be useful for methods", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 525, + 192, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 192, + 541 + ], + "score": 1.0, + "content": "using early stopping.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 493, + 505, + 541 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 567, + 187, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 189, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 189, + 580 + ], + "score": 1.0, + "content": "5.2 OPERATIONS", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 606 + ], + "score": 1.0, + "content": "To test whether the results from the previous section were due to the choice of available operations,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 455, + 618 + ], + "score": 1.0, + "content": "we developed an intentionally sub-optimal search space containing 4 plain convolutions", + "type": "text" + }, + { + "bbox": [ + 456, + 605, + 501, + 616 + ], + "score": 0.65, + "content": "( 1 \\times 1 , 3 \\times 3", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 605, + 505, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 131, + 627 + ], + "score": 0.69, + "content": "7 \\times 7 .", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 615, + 133, + 630 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 133, + 616, + 168, + 627 + ], + "score": 0.75, + "content": "1 1 \\times 1 1", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 615, + 275, + 630 + ], + "score": 1.0, + "content": "), 2 max pooling operators", + "type": "text" + }, + { + "bbox": [ + 276, + 616, + 331, + 627 + ], + "score": 0.6, + "content": "( 3 \\times 3 , 5 \\times 5 )", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "plus the none and skip connect operations.", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "This proposed search space is clearly more parameter inefficient compared to the commonly used", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 637, + 482, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 482, + 652 + ], + "score": 1.0, + "content": "DARTS one (which uses both dilated and separable ones), and we expect it to perform worse.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 594, + 505, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 654, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "We sampled 56 architectures from this new search space and trained them with the DARTS train-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 190, + 677 + ], + "score": 1.0, + "content": "ing protocol (Cutout", + "type": "text" + }, + { + "bbox": [ + 190, + 667, + 196, + 675 + ], + "score": 0.32, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 197, + 666, + 234, + 677 + ], + "score": 1.0, + "content": "DropPath", + "type": "text" + }, + { + "bbox": [ + 234, + 667, + 240, + 675 + ], + "score": 0.34, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "Auxiliary Towers), for fair comparison with the results from the", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "previous section. Figure 6 shows the resulting histogram, together with the one obtained from the", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 506, + 701 + ], + "score": 1.0, + "content": "classical DARTS space of operations. The two distributions are only shifted by 0.18 accuracy points.", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "Given the minor difference in performance, the specific operations are not a key ingredient in the", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "success of this search space. Very likely, it’s the well engineered cell structure that allows the model", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 721, + 224, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 224, + 733 + ], + "score": 1.0, + "content": "to perform as well as it does.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 51, + "bbox_fs": [ + 105, + 654, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 84, + 294, + 209 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 84, + 294, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 84, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 109, + 84, + 294, + 209 + ], + "score": 0.97, + "type": "image", + "image_path": "5e91248ab0a2f0aabd9ab36f7d3f7c4ee456017c62591d2c5a5bc19fc63e0e6f.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 109, + 84, + 294, + 97.88888888888889 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 97.88888888888889, + 294, + 111.77777777777777 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 111.77777777777777, + 294, + 125.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 125.66666666666666, + 294, + 139.55555555555554 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 139.55555555555554, + 294, + 153.44444444444443 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 153.44444444444443, + 294, + 167.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 167.33333333333331, + 294, + 181.2222222222222 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 181.2222222222222, + 294, + 195.1111111111111 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 195.1111111111111, + 294, + 208.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 299, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 299, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 299, + 228 + ], + "score": 1.0, + "content": "Figure 6: Histograms of the final accuracies", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 226, + 299, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 299, + 238 + ], + "score": 1.0, + "content": "(600 epochs) for architectures sampled from", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 236, + 299, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 299, + 250 + ], + "score": 1.0, + "content": "the DARTS search space (214 models) and our", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 249, + 228, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 228, + 260 + ], + "score": 1.0, + "content": "modified version (56 models).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 11.75 + }, + { + "type": "image", + "bbox": [ + 315, + 85, + 500, + 208 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 85, + 500, + 208 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 315, + 85, + 500, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 85, + 500, + 208 + ], + "score": 0.972, + "type": "image", + "image_path": "dbe7e776ae2cf3008835cf33d721f2209476268f702144b2c37653630acca5b7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 315, + 85, + 500, + 98.66666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 315, + 98.66666666666667, + 500, + 112.33333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 315, + 112.33333333333334, + 500, + 126.00000000000001 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 315, + 126.00000000000001, + 500, + 139.66666666666669 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 315, + 139.66666666666669, + 500, + 153.33333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 315, + 153.33333333333334, + 500, + 167.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 315, + 167.0, + 500, + 180.66666666666666 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 315, + 180.66666666666666, + 500, + 194.33333333333331 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 315, + 194.33333333333331, + 500, + 207.99999999999997 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 216, + 505, + 249 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 312, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "Figure 7: Performance of 16 randomly sampled", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 312, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "architectures with different numbers of cells.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 238, + 472, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 472, + 249 + ], + "score": 1.0, + "content": "Error bars represent standard deviation.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + } + ], + "index": 18.0 + }, + { + "type": "image", + "bbox": [ + 110, + 276, + 295, + 408 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 276, + 295, + 408 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 276, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 110, + 276, + 295, + 408 + ], + "score": 0.969, + "type": "image", + "image_path": "0587d112298e61ac522ef6c9052d3064e3454621c12eee6f15cad59d794e6af4.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 110, + 276, + 295, + 289.2 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 110, + 289.2, + 295, + 302.4 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 110, + 302.4, + 295, + 315.59999999999997 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 110, + 315.59999999999997, + 295, + 328.79999999999995 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 110, + 328.79999999999995, + 295, + 341.99999999999994 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 110, + 341.99999999999994, + 295, + 355.19999999999993 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 110, + 355.19999999999993, + 295, + 368.3999999999999 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 110, + 368.3999999999999, + 295, + 381.5999999999999 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 110, + 381.5999999999999, + 295, + 394.7999999999999 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 110, + 394.7999999999999, + 295, + 407.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 416, + 298, + 449 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 415, + 299, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 299, + 428 + ], + "score": 1.0, + "content": "Figure 8: Changes in the ranking of differ-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 426, + 299, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 299, + 439 + ], + "score": 1.0, + "content": "ent architectures when trained with two differ-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 438, + 190, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 190, + 450 + ], + "score": 1.0, + "content": "ent seeds (A and B).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "index": 37.75 + }, + { + "type": "image", + "bbox": [ + 315, + 276, + 500, + 408 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 276, + 500, + 408 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 315, + 276, + 500, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 500, + 408 + ], + "score": 0.97, + "type": "image", + "image_path": "5f3c33ff2e3ac55ede03bb632eba1c2d1e204c46b832377bf24ea864a9fd54c4.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 315, + 276, + 500, + 289.2 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 315, + 289.2, + 500, + 302.4 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 315, + 302.4, + 500, + 315.59999999999997 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 315, + 315.59999999999997, + 500, + 328.79999999999995 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 315, + 328.79999999999995, + 500, + 341.99999999999994 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 315, + 341.99999999999994, + 500, + 355.19999999999993 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 315, + 355.19999999999993, + 500, + 368.3999999999999 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 315, + 368.3999999999999, + 500, + 381.5999999999999 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 315, + 381.5999999999999, + 500, + 394.7999999999999 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 315, + 394.7999999999999, + 500, + 407.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 416, + 504, + 449 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 312, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "Figure 9: Changes in the ranking of different", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 312, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "architectures when trained with different num-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 312, + 437, + 415, + 450 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 415, + 450 + ], + "score": 1.0, + "content": "bers of cells (same seed).", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + } + ], + "index": 44.25 + }, + { + "type": "title", + "bbox": [ + 108, + 473, + 415, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 416, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 416, + 485 + ], + "score": 1.0, + "content": "5.3 DOES CHANGING SEED AND NUMBER OF CELLS AFFECT RANKING?", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "A necessary practice for many weight-sharing methods (Liu et al., 2019; Pham et al., 2018) is to", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "restart the training from scratch after the search phase, with a different number of cells. Recent", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "works have warned that this procedure might negatively affect ranking; similarly, the role of the", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 525, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 542 + ], + "score": 1.0, + "content": "seed has been previously recognized as as a fundamental element in reproducibility (Li & Talwalkar,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 538, + 212, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 212, + 551 + ], + "score": 1.0, + "content": "2019; Sciuto et al., 2019).", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54 + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "To test the impact of seed, we randomly sampled 32 architectures and trained them with two different", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "seeds (Figure 8). Ranking is heavily influenced, as the Kendall tau correlation between the two sets", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 349, + 589 + ], + "score": 1.0, + "content": "of training is 0.48. On average, the test accuracy changes by", + "type": "text" + }, + { + "bbox": [ + 349, + 577, + 407, + 588 + ], + "score": 0.9, + "content": "0 . 1 3 \\% \\pm 0 . 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 577, + 470, + 589 + ], + "score": 1.0, + "content": "(max change is", + "type": "text" + }, + { + "bbox": [ + 471, + 577, + 499, + 588 + ], + "score": 0.86, + "content": "0 . 3 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 577, + 505, + 589 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 588, + 491, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 491, + 600 + ], + "score": 1.0, + "content": "which is substantial considering the small gap between random architectures and NAS methods.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 58.5 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "To test the depth-gap we trained another 32 with different number of cells (Figure 9). The corre-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "lation between the two different depths is not very strong as measured by Kendall Tau (0.54), with", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 104, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "architectures shifting up and down the rankings by up to 18 positions (out of 32). Methods employ-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "ing weight sharing (WS) would see an even more pronounced effect as the architectures normally", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 649, + 501, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 501, + 662 + ], + "score": 1.0, + "content": "chosen at 8 cells would have been training sub-optimally due to the WS itself (Sciuto et al., 2019).", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 63 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "These findings point towards two issues. The first is that since the seed has such a large effect on", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "ranking, it stands to reason that the final accuracy reported should be averaged over multiple seeds.", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "The second is that, if the lottery ticket hypothesis holds—so that specific sub-networks are better", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "mainly due to their lucky initialization; Frankle & Carbin., 2018—together with our findings, this", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "could be an additional reason why methods searching on a different number of cells than the final", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 721, + 457, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 457, + 733 + ], + "score": 1.0, + "content": "model, struggle to significantly improve on the average randomly sampled architecture.", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 68.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 84, + 294, + 209 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 109, + 84, + 294, + 209 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 84, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 109, + 84, + 294, + 209 + ], + "score": 0.97, + "type": "image", + "image_path": "5e91248ab0a2f0aabd9ab36f7d3f7c4ee456017c62591d2c5a5bc19fc63e0e6f.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 109, + 84, + 294, + 97.88888888888889 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 109, + 97.88888888888889, + 294, + 111.77777777777777 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 109, + 111.77777777777777, + 294, + 125.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 109, + 125.66666666666666, + 294, + 139.55555555555554 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 109, + 139.55555555555554, + 294, + 153.44444444444443 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 109, + 153.44444444444443, + 294, + 167.33333333333331 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 167.33333333333331, + 294, + 181.2222222222222 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 181.2222222222222, + 294, + 195.1111111111111 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 109, + 195.1111111111111, + 294, + 208.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 216, + 299, + 260 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 216, + 299, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 299, + 228 + ], + "score": 1.0, + "content": "Figure 6: Histograms of the final accuracies", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 226, + 299, + 238 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 299, + 238 + ], + "score": 1.0, + "content": "(600 epochs) for architectures sampled from", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 236, + 299, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 299, + 250 + ], + "score": 1.0, + "content": "the DARTS search space (214 models) and our", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 249, + 228, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 228, + 260 + ], + "score": 1.0, + "content": "modified version (56 models).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 19.5 + } + ], + "index": 11.75 + }, + { + "type": "image", + "bbox": [ + 315, + 85, + 500, + 208 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 85, + 500, + 208 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 315, + 85, + 500, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 85, + 500, + 208 + ], + "score": 0.972, + "type": "image", + "image_path": "dbe7e776ae2cf3008835cf33d721f2209476268f702144b2c37653630acca5b7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 315, + 85, + 500, + 98.66666666666667 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 315, + 98.66666666666667, + 500, + 112.33333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 315, + 112.33333333333334, + 500, + 126.00000000000001 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 315, + 126.00000000000001, + 500, + 139.66666666666669 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 315, + 139.66666666666669, + 500, + 153.33333333333334 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 315, + 153.33333333333334, + 500, + 167.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 315, + 167.0, + 500, + 180.66666666666666 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 315, + 180.66666666666666, + 500, + 194.33333333333331 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 315, + 194.33333333333331, + 500, + 207.99999999999997 + ], + "spans": [], + "index": 17 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 216, + 505, + 249 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 312, + 215, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 312, + 215, + 505, + 227 + ], + "score": 1.0, + "content": "Figure 7: Performance of 16 randomly sampled", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 226, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 312, + 226, + 505, + 238 + ], + "score": 1.0, + "content": "architectures with different numbers of cells.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 238, + 472, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 472, + 249 + ], + "score": 1.0, + "content": "Error bars represent standard deviation.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + } + ], + "index": 18.0 + }, + { + "type": "image", + "bbox": [ + 110, + 276, + 295, + 408 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 276, + 295, + 408 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 276, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 110, + 276, + 295, + 408 + ], + "score": 0.969, + "type": "image", + "image_path": "0587d112298e61ac522ef6c9052d3064e3454621c12eee6f15cad59d794e6af4.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 110, + 276, + 295, + 289.2 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 110, + 289.2, + 295, + 302.4 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 110, + 302.4, + 295, + 315.59999999999997 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 110, + 315.59999999999997, + 295, + 328.79999999999995 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 110, + 328.79999999999995, + 295, + 341.99999999999994 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 110, + 341.99999999999994, + 295, + 355.19999999999993 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 110, + 355.19999999999993, + 295, + 368.3999999999999 + ], + "spans": [], + "index": 31 + }, + { + "bbox": [ + 110, + 368.3999999999999, + 295, + 381.5999999999999 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 110, + 381.5999999999999, + 295, + 394.7999999999999 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 110, + 394.7999999999999, + 295, + 407.9999999999999 + ], + "spans": [], + "index": 34 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 416, + 298, + 449 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 415, + 299, + 428 + ], + "spans": [ + { + "bbox": [ + 106, + 415, + 299, + 428 + ], + "score": 1.0, + "content": "Figure 8: Changes in the ranking of differ-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 426, + 299, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 299, + 439 + ], + "score": 1.0, + "content": "ent architectures when trained with two differ-", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 438, + 190, + 450 + ], + "spans": [ + { + "bbox": [ + 106, + 438, + 190, + 450 + ], + "score": 1.0, + "content": "ent seeds (A and B).", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "index": 37.75 + }, + { + "type": "image", + "bbox": [ + 315, + 276, + 500, + 408 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 315, + 276, + 500, + 408 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 315, + 276, + 500, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 500, + 408 + ], + "score": 0.97, + "type": "image", + "image_path": "5f3c33ff2e3ac55ede03bb632eba1c2d1e204c46b832377bf24ea864a9fd54c4.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 315, + 276, + 500, + 289.2 + ], + "spans": [], + "index": 35 + }, + { + "bbox": [ + 315, + 289.2, + 500, + 302.4 + ], + "spans": [], + "index": 36 + }, + { + "bbox": [ + 315, + 302.4, + 500, + 315.59999999999997 + ], + "spans": [], + "index": 37 + }, + { + "bbox": [ + 315, + 315.59999999999997, + 500, + 328.79999999999995 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 315, + 328.79999999999995, + 500, + 341.99999999999994 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 315, + 341.99999999999994, + 500, + 355.19999999999993 + ], + "spans": [], + "index": 40 + }, + { + "bbox": [ + 315, + 355.19999999999993, + 500, + 368.3999999999999 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 315, + 368.3999999999999, + 500, + 381.5999999999999 + ], + "spans": [], + "index": 42 + }, + { + "bbox": [ + 315, + 381.5999999999999, + 500, + 394.7999999999999 + ], + "spans": [], + "index": 43 + }, + { + "bbox": [ + 315, + 394.7999999999999, + 500, + 407.9999999999999 + ], + "spans": [], + "index": 44 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 312, + 416, + 504, + 449 + ], + "group_id": 3, + "lines": [ + { + "bbox": [ + 312, + 415, + 505, + 428 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 505, + 428 + ], + "score": 1.0, + "content": "Figure 9: Changes in the ranking of different", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 312, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "architectures when trained with different num-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 312, + 437, + 415, + 450 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 415, + 450 + ], + "score": 1.0, + "content": "bers of cells (same seed).", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 49 + } + ], + "index": 44.25 + }, + { + "type": "title", + "bbox": [ + 108, + 473, + 415, + 483 + ], + "lines": [ + { + "bbox": [ + 105, + 472, + 416, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 416, + 485 + ], + "score": 1.0, + "content": "5.3 DOES CHANGING SEED AND NUMBER OF CELLS AFFECT RANKING?", + "type": "text" + } + ], + "index": 51 + } + ], + "index": 51 + }, + { + "type": "text", + "bbox": [ + 107, + 494, + 505, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 506 + ], + "score": 1.0, + "content": "A necessary practice for many weight-sharing methods (Liu et al., 2019; Pham et al., 2018) is to", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 506, + 518 + ], + "score": 1.0, + "content": "restart the training from scratch after the search phase, with a different number of cells. Recent", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 529 + ], + "score": 1.0, + "content": "works have warned that this procedure might negatively affect ranking; similarly, the role of the", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 525, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 542 + ], + "score": 1.0, + "content": "seed has been previously recognized as as a fundamental element in reproducibility (Li & Talwalkar,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 538, + 212, + 551 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 212, + 551 + ], + "score": 1.0, + "content": "2019; Sciuto et al., 2019).", + "type": "text" + } + ], + "index": 56 + } + ], + "index": 54, + "bbox_fs": [ + 105, + 495, + 506, + 551 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 555, + 505, + 599 + ], + "lines": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 106, + 555, + 505, + 567 + ], + "score": 1.0, + "content": "To test the impact of seed, we randomly sampled 32 architectures and trained them with two different", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 505, + 578 + ], + "score": 1.0, + "content": "seeds (Figure 8). Ranking is heavily influenced, as the Kendall tau correlation between the two sets", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 577, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 349, + 589 + ], + "score": 1.0, + "content": "of training is 0.48. On average, the test accuracy changes by", + "type": "text" + }, + { + "bbox": [ + 349, + 577, + 407, + 588 + ], + "score": 0.9, + "content": "0 . 1 3 \\% \\pm 0 . 0 8", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 577, + 470, + 589 + ], + "score": 1.0, + "content": "(max change is", + "type": "text" + }, + { + "bbox": [ + 471, + 577, + 499, + 588 + ], + "score": 0.86, + "content": "0 . 3 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 577, + 505, + 589 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 588, + 491, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 491, + 600 + ], + "score": 1.0, + "content": "which is substantial considering the small gap between random architectures and NAS methods.", + "type": "text" + } + ], + "index": 60 + } + ], + "index": 58.5, + "bbox_fs": [ + 105, + 555, + 505, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "To test the depth-gap we trained another 32 with different number of cells (Figure 9). The corre-", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "lation between the two different depths is not very strong as measured by Kendall Tau (0.54), with", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 104, + 626, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 626, + 506, + 641 + ], + "score": 1.0, + "content": "architectures shifting up and down the rankings by up to 18 positions (out of 32). Methods employ-", + "type": "text" + } + ], + "index": 63 + }, + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 651 + ], + "score": 1.0, + "content": "ing weight sharing (WS) would see an even more pronounced effect as the architectures normally", + "type": "text" + } + ], + "index": 64 + }, + { + "bbox": [ + 105, + 649, + 501, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 501, + 662 + ], + "score": 1.0, + "content": "chosen at 8 cells would have been training sub-optimally due to the WS itself (Sciuto et al., 2019).", + "type": "text" + } + ], + "index": 65 + } + ], + "index": 63, + "bbox_fs": [ + 104, + 605, + 506, + 662 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "These findings point towards two issues. The first is that since the seed has such a large effect on", + "type": "text" + } + ], + "index": 66 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "ranking, it stands to reason that the final accuracy reported should be averaged over multiple seeds.", + "type": "text" + } + ], + "index": 67 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "The second is that, if the lottery ticket hypothesis holds—so that specific sub-networks are better", + "type": "text" + } + ], + "index": 68 + }, + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 504, + 711 + ], + "score": 1.0, + "content": "mainly due to their lucky initialization; Frankle & Carbin., 2018—together with our findings, this", + "type": "text" + } + ], + "index": 69 + }, + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 721 + ], + "score": 1.0, + "content": "could be an additional reason why methods searching on a different number of cells than the final", + "type": "text" + } + ], + "index": 70 + }, + { + "bbox": [ + 105, + 721, + 457, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 457, + 733 + ], + "score": 1.0, + "content": "model, struggle to significantly improve on the average randomly sampled architecture.", + "type": "text" + } + ], + "index": 71 + } + ], + "index": 68.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 303, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 305, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 305, + 95 + ], + "score": 1.0, + "content": "6 DISCUSSION AND BEST PRACTICES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 105, + 460, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 462, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 462, + 119 + ], + "score": 1.0, + "content": "In this section we offer some suggestions on how to mitigate the issues in NAS research.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Augmention tricks: while achieving higher accuracies is clearly a desirable goal, we have shown", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "in section 4, that using well engineered training protocols can hide the contribution of the search al-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 145, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 157 + ], + "score": 1.0, + "content": "gorithm. We therefore suggest that both results, with and without training tricks, should be reported.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 156, + 346, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 346, + 167 + ], + "score": 1.0, + "content": "An example of best practice is found in Hundt et al. (2019).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 506, + 185 + ], + "score": 1.0, + "content": "Search Space: it is difficult to evaluate the effectiveness of any given proposed method without a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "score": 1.0, + "content": "measure of how good randomly sampled architectures are. This is not the same thing as performing", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "a random search which is a search strategy in itself; random sampling is simply used to establish how", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "good the average model is. A simple approach to measure the variability of any new given search", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 246, + 229 + ], + "score": 1.0, + "content": "space could be to randomly sample", + "type": "text" + }, + { + "bbox": [ + 247, + 217, + 254, + 226 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "architectures and report mean and standard deviation. We hope", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "that future works will attempt to develop more expressive search spaces, capable of producing both", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "good and bad network designs. Restricted search spaces, while guaranteeing good performance and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "score": 1.0, + "content": "quick results, will inevitably be constrained by the bounds of expert knowledge (local optima) and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "will be incapable of reaching more truly innovative solutions (closer to the global optima). As our", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "findings in section 5.2 suggest, the overall wiring (the macro-structure) is an extremely influential", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "component in the final performance. As such, future research could investigate the optimal wiring", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 293, + 403, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 403, + 306 + ], + "score": 1.0, + "content": "at a global level: an interesting work in this direction is Xie et al. (2019a).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 504, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 324 + ], + "score": 1.0, + "content": "Multiple datasets: as the true goal of AutoML is to minimize the need for human experts, focusing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "the research efforts on a single dataset will inevitably lead to algorithmic overfitting and/or methods", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "heavily dependent on hyperparameter tuning. The best solution for this is likely to test NAS algo-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "rithms on a battery of datasets, with different characteristics: image sizes, number of samples, class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 354, + 226, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 226, + 367 + ], + "score": 1.0, + "content": "granularity and learning task.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "Investigating hidden components: as our experiments in Sections 4 and 5.2 show, the DARTS", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "search space is not only effective due to specific operations that are being chosen, but in greater part", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "score": 1.0, + "content": "due to the overall macro-structure and the training protocol used. We suggest that proper ablation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 475, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 475, + 416 + ], + "score": 1.0, + "content": "studies can lead to better understanding of the contributions of each element of the pipeline.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 420, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "score": 1.0, + "content": "The importance of reproducibility: reproducibility is of extreme relevance in all sciences. To", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "this end, it is very important that authors release not only their best found architecture but also the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "corresponding seed (if they did not average over multiple ones), as well as the code and the detailed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "score": 1.0, + "content": "training protocol (including hyperparameters). To this end, NAS-Bench-101 (Ying et al., 2019), a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "dataset mapping architectures to their accuracy, can be extremely useful, as it allows the quality of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "score": 1.0, + "content": "search strategies to be assessed in isolation from other NAS components (e.g. search space, training", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "protocol) in a quick and reproducible fashion. The code for this paper is open-source (link in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 422, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 422, + 509 + ], + "score": 1.0, + "content": "abstract). We also open-source the 270 trained architectures used in Section 5.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "Hyperparameter tuning cost: tuning hyperparameters in NAS is an extremely costly component.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "score": 1.0, + "content": "Therefore, we argue that either (i) hyperparameters are general enough so that they do not require", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 536, + 392, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 392, + 549 + ], + "score": 1.0, + "content": "tuning for further tasks, or (2) the cost is included in the search budget.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "title", + "bbox": [ + 108, + 563, + 201, + 576 + ], + "lines": [ + { + "bbox": [ + 104, + 561, + 203, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 561, + 203, + 579 + ], + "score": 1.0, + "content": "7 CONCLUSIONS", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 588, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "AutoML, and NAS in particular, have the potential to truly democratize the use of machine learning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "for all, and could bring forth very notable improvements on a variety of tasks. To truly step forward,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 610, + 416, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 416, + 622 + ], + "score": 1.0, + "content": "a principled approach, with a focus on fairness and reproducibility is needed.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "In this paper we have shown that, for many NAS methods, the search space has been engineered", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "such that all architectures perform similarly well and that their relative ranking can easily shift. We", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "have furthermore showed that the training protocol itself has a higher impact on the final accuracy", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "than the actual network. Finally, we have provided some suggestions on how to make future research", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 671, + 219, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 219, + 683 + ], + "score": 1.0, + "content": "more robust to these issues.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "We hope that our findings will help the community focus their efforts towards a more general", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "approach to automated neural architecture design. Only then can we expect to learn from NAS-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "generated architectures as opposed to the current paradigm where search spaces are heavily influ-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 303, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 303, + 734 + ], + "score": 1.0, + "content": "enced by our current (human) expert knowledge.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 303, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 305, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 305, + 95 + ], + "score": 1.0, + "content": "6 DISCUSSION AND BEST PRACTICES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 105, + 105, + 460, + 117 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 462, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 462, + 119 + ], + "score": 1.0, + "content": "In this section we offer some suggestions on how to mitigate the issues in NAS research.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 105, + 462, + 119 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 122, + 504, + 167 + ], + "lines": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Augmention tricks: while achieving higher accuracies is clearly a desirable goal, we have shown", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 134, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 134, + 505, + 146 + ], + "score": 1.0, + "content": "in section 4, that using well engineered training protocols can hide the contribution of the search al-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 145, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 505, + 157 + ], + "score": 1.0, + "content": "gorithm. We therefore suggest that both results, with and without training tricks, should be reported.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 156, + 346, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 156, + 346, + 167 + ], + "score": 1.0, + "content": "An example of best practice is found in Hundt et al. (2019).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 122, + 505, + 167 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 172, + 505, + 304 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 506, + 185 + ], + "score": 1.0, + "content": "Search Space: it is difficult to evaluate the effectiveness of any given proposed method without a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 505, + 197 + ], + "score": 1.0, + "content": "measure of how good randomly sampled architectures are. This is not the same thing as performing", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 105, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "a random search which is a search strategy in itself; random sampling is simply used to establish how", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 205, + 505, + 218 + ], + "score": 1.0, + "content": "good the average model is. A simple approach to measure the variability of any new given search", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 217, + 505, + 229 + ], + "spans": [ + { + "bbox": [ + 105, + 217, + 246, + 229 + ], + "score": 1.0, + "content": "space could be to randomly sample", + "type": "text" + }, + { + "bbox": [ + 247, + 217, + 254, + 226 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 217, + 505, + 229 + ], + "score": 1.0, + "content": "architectures and report mean and standard deviation. We hope", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 506, + 240 + ], + "score": 1.0, + "content": "that future works will attempt to develop more expressive search spaces, capable of producing both", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 238, + 506, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 506, + 251 + ], + "score": 1.0, + "content": "good and bad network designs. Restricted search spaces, while guaranteeing good performance and", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 262 + ], + "score": 1.0, + "content": "quick results, will inevitably be constrained by the bounds of expert knowledge (local optima) and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 273 + ], + "score": 1.0, + "content": "will be incapable of reaching more truly innovative solutions (closer to the global optima). As our", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "findings in section 5.2 suggest, the overall wiring (the macro-structure) is an extremely influential", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "component in the final performance. As such, future research could investigate the optimal wiring", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 293, + 403, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 403, + 306 + ], + "score": 1.0, + "content": "at a global level: an interesting work in this direction is Xie et al. (2019a).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 173, + 506, + 306 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 310, + 504, + 365 + ], + "lines": [ + { + "bbox": [ + 105, + 308, + 506, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 506, + 324 + ], + "score": 1.0, + "content": "Multiple datasets: as the true goal of AutoML is to minimize the need for human experts, focusing", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "the research efforts on a single dataset will inevitably lead to algorithmic overfitting and/or methods", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "spans": [ + { + "bbox": [ + 105, + 331, + 505, + 345 + ], + "score": 1.0, + "content": "heavily dependent on hyperparameter tuning. The best solution for this is likely to test NAS algo-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 506, + 356 + ], + "score": 1.0, + "content": "rithms on a battery of datasets, with different characteristics: image sizes, number of samples, class", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 354, + 226, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 226, + 367 + ], + "score": 1.0, + "content": "granularity and learning task.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 308, + 506, + 367 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 505, + 383 + ], + "score": 1.0, + "content": "Investigating hidden components: as our experiments in Sections 4 and 5.2 show, the DARTS", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 506, + 395 + ], + "score": 1.0, + "content": "search space is not only effective due to specific operations that are being chosen, but in greater part", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 406 + ], + "score": 1.0, + "content": "due to the overall macro-structure and the training protocol used. We suggest that proper ablation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 475, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 475, + 416 + ], + "score": 1.0, + "content": "studies can lead to better understanding of the contributions of each element of the pipeline.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 371, + 506, + 416 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 420, + 505, + 508 + ], + "lines": [ + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 506, + 433 + ], + "score": 1.0, + "content": "The importance of reproducibility: reproducibility is of extreme relevance in all sciences. To", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "this end, it is very important that authors release not only their best found architecture but also the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "corresponding seed (if they did not average over multiple ones), as well as the code and the detailed", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "score": 1.0, + "content": "training protocol (including hyperparameters). To this end, NAS-Bench-101 (Ying et al., 2019), a", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "dataset mapping architectures to their accuracy, can be extremely useful, as it allows the quality of", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 505, + 489 + ], + "score": 1.0, + "content": "search strategies to be assessed in isolation from other NAS components (e.g. search space, training", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 505, + 499 + ], + "score": 1.0, + "content": "protocol) in a quick and reproducible fashion. The code for this paper is open-source (link in the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 497, + 422, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 422, + 509 + ], + "score": 1.0, + "content": "abstract). We also open-source the 270 trained architectures used in Section 5.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 420, + 506, + 509 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 514, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 106, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "Hyperparameter tuning cost: tuning hyperparameters in NAS is an extremely costly component.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 539 + ], + "score": 1.0, + "content": "Therefore, we argue that either (i) hyperparameters are general enough so that they do not require", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 536, + 392, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 392, + 549 + ], + "score": 1.0, + "content": "tuning for further tasks, or (2) the cost is included in the search budget.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 105, + 514, + 505, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 563, + 201, + 576 + ], + "lines": [ + { + "bbox": [ + 104, + 561, + 203, + 579 + ], + "spans": [ + { + "bbox": [ + 104, + 561, + 203, + 579 + ], + "score": 1.0, + "content": "7 CONCLUSIONS", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 588, + 504, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 587, + 505, + 601 + ], + "score": 1.0, + "content": "AutoML, and NAS in particular, have the potential to truly democratize the use of machine learning", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 612 + ], + "score": 1.0, + "content": "for all, and could bring forth very notable improvements on a variety of tasks. To truly step forward,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 610, + 416, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 416, + 622 + ], + "score": 1.0, + "content": "a principled approach, with a focus on fairness and reproducibility is needed.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 587, + 505, + 622 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 627, + 504, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "In this paper we have shown that, for many NAS methods, the search space has been engineered", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "such that all architectures perform similarly well and that their relative ranking can easily shift. We", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "have furthermore showed that the training protocol itself has a higher impact on the final accuracy", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "than the actual network. Finally, we have provided some suggestions on how to make future research", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 671, + 219, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 219, + 683 + ], + "score": 1.0, + "content": "more robust to these issues.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 626, + 506, + 683 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "We hope that our findings will help the community focus their efforts towards a more general", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "approach to automated neural architecture design. Only then can we expect to learn from NAS-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "generated architectures as opposed to the current paradigm where search spaces are heavily influ-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 303, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 303, + 734 + ], + "score": 1.0, + "content": "enced by our current (human) expert knowledge.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 48.5, + "bbox_fs": [ + 105, + 687, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Fabio Carlucci, Pedro M. Esperanc¸a, Marco Singh, Antoine Yang, Victor Gabillon, Hang Xu, Zewei", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Chen, and Jun Wang. MANAS: Multi-Agent Neural Architecture Search. arxiv:1909.01051,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 143, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 143, + 135 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 141, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 504, + 155 + ], + "score": 1.0, + "content": "Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment:", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 393, + 165 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 393, + 165 + ], + "score": 1.0, + "content": "Learning augmentation policies from data. arXiv:1805.09501, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 105, + 172, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 184, + 271, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 184, + 271, + 195 + ], + "score": 1.0, + "content": "with cutout. arXiv:1708.04552, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 503, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 218 + ], + "score": 1.0, + "content": "Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 214, + 357, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 357, + 227 + ], + "score": 1.0, + "content": "Journal of Machine Learning Research, 20(55):1–21, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 105, + 234, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 246, + 273, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 273, + 257 + ], + "score": 1.0, + "content": "networks. In arXiv:1803.03635, 2018.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 108, + 264, + 504, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "Oguzhan Gencoglu, Mark van Gils, Esin Guldogan, Chamin Morikawa, Mehmet Suzen, Mathias ¨", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "Gruber, Jussi Leinonen, and Heikki Huttunen. Hark side of deep learning–from grad student", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 287, + 381, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 287, + 381, + 299 + ], + "score": 1.0, + "content": "descent to automated machine learning. arXiv:1904.07633, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 503, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 322 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 318, + 442, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 442, + 330 + ], + "score": 1.0, + "content": "nition. In Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 352 + ], + "score": 1.0, + "content": "Zhang Xiangyu Rein Shaoqing He, Kaiming and Jian Sun. Delving deep into rectifiers: Surpassing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "human level performance on imagenet classification. In Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 360, + 178, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 178, + 372 + ], + "score": 1.0, + "content": "(CVPR), 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 390, + 485, + 403 + ], + "spans": [ + { + "bbox": [ + 115, + 390, + 485, + 403 + ], + "score": 1.0, + "content": "stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "Andrew Hundt, Varun Jain, and Gregory D. Hager. sharpDARTS: Faster and more accurate differ-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 421, + 334, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 334, + 433 + ], + "score": 1.0, + "content": "entiable architecture search. arXiv:1903.09900, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 503, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 455 + ], + "score": 1.0, + "content": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 452, + 188, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 188, + 464 + ], + "score": 1.0, + "content": "of Toronto, 2009.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 504, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 493, + 142, + 505 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 142, + 505 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 504, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "Michael Maire Larsson, Gustav and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 524, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 480, + 537 + ], + "score": 1.0, + "content": "without residuals. In International Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 106, + 544, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Guilin Li, Xing Zhang, Zitong Wang, Zhenguo Li, and Tong Zhang. StacNAS: Towards stable and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 555, + 499, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 499, + 568 + ], + "score": 1.0, + "content": "consistent optimization for differentiable Neural Architecture Search. arXiv:1909.11926, 2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 504, + 588 + ], + "score": 1.0, + "content": "Li-Jia Li and Li Fei-Fei. What, where and who? classifying events by scene and object recognition.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 586, + 407, + 599 + ], + "spans": [ + { + "bbox": [ + 116, + 586, + 407, + 599 + ], + "score": 1.0, + "content": "In International Conference on Computer Vision (ICCV), pp. 1–8, 2007.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 104, + 606, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 504, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 504, + 619 + ], + "score": 1.0, + "content": "Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 617, + 219, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 219, + 628 + ], + "score": 1.0, + "content": "arXiv:1902.07638, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "score": 1.0, + "content": "Hanwen Liang, Shifeng Zhang, Jiacheng Sun, Xingqiu He, Weiran Huang, Kechen Zhuang,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 647, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 226, + 662 + ], + "score": 1.0, + "content": "and Zhenguo Li. DARTS", + "type": "text" + }, + { + "bbox": [ + 226, + 649, + 232, + 658 + ], + "score": 0.26, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 647, + 505, + 662 + ], + "score": 1.0, + "content": ": Improved differentiable architecture search with early stopping.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 659, + 219, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 659, + 219, + 669 + ], + "score": 1.0, + "content": "arXiv:1909.06035, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 104, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Marius Lindauer and Frank Hutter. Best practices for scientific research on neural architecture", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 690, + 251, + 701 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 251, + 701 + ], + "score": 1.0, + "content": "search. arXiv:1909.02453, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. In", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 396, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 396, + 732 + ], + "score": 1.0, + "content": "International Conference on Learning Representations (ICLR), 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 176, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 100, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "Fabio Carlucci, Pedro M. Esperanc¸a, Marco Singh, Antoine Yang, Victor Gabillon, Hang Xu, Zewei", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 116, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Chen, and Jun Wang. MANAS: Multi-Agent Neural Architecture Search. arxiv:1909.01051,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 121, + 143, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 121, + 143, + 135 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 100, + 505, + 135 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 141, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 106, + 142, + 504, + 155 + ], + "spans": [ + { + "bbox": [ + 106, + 142, + 504, + 155 + ], + "score": 1.0, + "content": "Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment:", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 153, + 393, + 165 + ], + "spans": [ + { + "bbox": [ + 115, + 153, + 393, + 165 + ], + "score": 1.0, + "content": "Learning augmentation policies from data. arXiv:1805.09501, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 142, + 504, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 172, + 505, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 184, + 271, + 195 + ], + "spans": [ + { + "bbox": [ + 116, + 184, + 271, + 195 + ], + "score": 1.0, + "content": "with cutout. arXiv:1708.04552, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 172, + 505, + 195 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 203, + 503, + 227 + ], + "lines": [ + { + "bbox": [ + 105, + 201, + 505, + 218 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 505, + 218 + ], + "score": 1.0, + "content": "Thomas Elsken, Jan Hendrik Metzen, and Frank Hutter. Neural architecture search: A survey.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 214, + 357, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 214, + 357, + 227 + ], + "score": 1.0, + "content": "Journal of Machine Learning Research, 20(55):1–21, 2019.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 201, + 505, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 234, + 503, + 257 + ], + "lines": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 505, + 248 + ], + "score": 1.0, + "content": "Jonathan Frankle and Michael Carbin. The lottery ticket hypothesis: Finding sparse, trainable neural", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 246, + 273, + 257 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 273, + 257 + ], + "score": 1.0, + "content": "networks. In arXiv:1803.03635, 2018.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 235, + 505, + 257 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 264, + 504, + 299 + ], + "lines": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 264, + 505, + 278 + ], + "score": 1.0, + "content": "Oguzhan Gencoglu, Mark van Gils, Esin Guldogan, Chamin Morikawa, Mehmet Suzen, Mathias ¨", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 276, + 505, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 505, + 289 + ], + "score": 1.0, + "content": "Gruber, Jussi Leinonen, and Heikki Huttunen. Hark side of deep learning–from grad student", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 116, + 287, + 381, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 287, + 381, + 299 + ], + "score": 1.0, + "content": "descent to automated machine learning. arXiv:1904.07633, 2019.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 106, + 264, + 505, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 306, + 503, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 305, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 505, + 322 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 318, + 442, + 330 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 442, + 330 + ], + "score": 1.0, + "content": "nition. In Computer Vision and Pattern Recognition (CVPR), pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 305, + 505, + 330 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 337, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 105, + 337, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 505, + 352 + ], + "score": 1.0, + "content": "Zhang Xiangyu Rein Shaoqing He, Kaiming and Jian Sun. Delving deep into rectifiers: Surpassing", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 349, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 115, + 349, + 505, + 361 + ], + "score": 1.0, + "content": "human level performance on imagenet classification. In Computer Vision and Pattern Recognition", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 360, + 178, + 372 + ], + "spans": [ + { + "bbox": [ + 116, + 360, + 178, + 372 + ], + "score": 1.0, + "content": "(CVPR), 2015.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 337, + 505, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 379, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 390, + 485, + 403 + ], + "spans": [ + { + "bbox": [ + 115, + 390, + 485, + 403 + ], + "score": 1.0, + "content": "stochastic depth. In European conference on computer vision, pp. 646–661. Springer, 2016.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 106, + 379, + 505, + 403 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 410, + 504, + 433 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 505, + 423 + ], + "score": 1.0, + "content": "Andrew Hundt, Varun Jain, and Gregory D. Hager. sharpDARTS: Faster and more accurate differ-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 421, + 334, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 334, + 433 + ], + "score": 1.0, + "content": "entiable architecture search. arXiv:1903.09900, 2019.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 106, + 410, + 505, + 433 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 503, + 464 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 455 + ], + "score": 1.0, + "content": "Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 452, + 188, + 464 + ], + "spans": [ + { + "bbox": [ + 116, + 452, + 188, + 464 + ], + "score": 1.0, + "content": "of Toronto, 2009.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 439, + 505, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 471, + 504, + 505 + ], + "lines": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "spans": [ + { + "bbox": [ + 105, + 471, + 505, + 485 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 115, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "lutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 493, + 142, + 505 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 142, + 505 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 471, + 505, + 505 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 513, + 504, + 537 + ], + "lines": [ + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 106, + 513, + 505, + 526 + ], + "score": 1.0, + "content": "Michael Maire Larsson, Gustav and Gregory Shakhnarovich. Fractalnet: Ultra-deep neural networks", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 524, + 480, + 537 + ], + "spans": [ + { + "bbox": [ + 116, + 524, + 480, + 537 + ], + "score": 1.0, + "content": "without residuals. In International Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5, + "bbox_fs": [ + 106, + 513, + 505, + 537 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 544, + 504, + 567 + ], + "lines": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 545, + 505, + 558 + ], + "score": 1.0, + "content": "Guilin Li, Xing Zhang, Zitong Wang, Zhenguo Li, and Tong Zhang. StacNAS: Towards stable and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 555, + 499, + 568 + ], + "spans": [ + { + "bbox": [ + 115, + 555, + 499, + 568 + ], + "score": 1.0, + "content": "consistent optimization for differentiable Neural Architecture Search. arXiv:1909.11926, 2019.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 106, + 545, + 505, + 568 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 504, + 599 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 504, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 504, + 588 + ], + "score": 1.0, + "content": "Li-Jia Li and Li Fei-Fei. What, where and who? classifying events by scene and object recognition.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 586, + 407, + 599 + ], + "spans": [ + { + "bbox": [ + 116, + 586, + 407, + 599 + ], + "score": 1.0, + "content": "In International Conference on Computer Vision (ICCV), pp. 1–8, 2007.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 574, + 504, + 599 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 606, + 504, + 629 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 504, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 504, + 619 + ], + "score": 1.0, + "content": "Liam Li and Ameet Talwalkar. Random search and reproducibility for neural architecture search.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 617, + 219, + 628 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 219, + 628 + ], + "score": 1.0, + "content": "arXiv:1902.07638, 2019.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 606, + 504, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 636, + 504, + 670 + ], + "lines": [ + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 635, + 506, + 651 + ], + "score": 1.0, + "content": "Hanwen Liang, Shifeng Zhang, Jiacheng Sun, Xingqiu He, Weiran Huang, Kechen Zhuang,", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 647, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 115, + 647, + 226, + 662 + ], + "score": 1.0, + "content": "and Zhenguo Li. DARTS", + "type": "text" + }, + { + "bbox": [ + 226, + 649, + 232, + 658 + ], + "score": 0.26, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 647, + 505, + 662 + ], + "score": 1.0, + "content": ": Improved differentiable architecture search with early stopping.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 659, + 219, + 669 + ], + "spans": [ + { + "bbox": [ + 115, + 659, + 219, + 669 + ], + "score": 1.0, + "content": "arXiv:1909.06035, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 635, + 506, + 669 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 678, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 106, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "Marius Lindauer and Frank Hutter. Best practices for scientific research on neural architecture", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 116, + 690, + 251, + 701 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 251, + 701 + ], + "score": 1.0, + "content": "search. arXiv:1909.02453, 2019.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 678, + 505, + 701 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable architecture search. In", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 115, + 720, + 396, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 396, + 732 + ], + "score": 1.0, + "content": "International Conference on Learning Representations (ICLR), 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 709, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Interna-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 367, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 367, + 106 + ], + "score": 1.0, + "content": "tional Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 108, + 112, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Zhichao Lu, Ian Whalen, Vishnu Boddeti, Yashesh Dhebar, Kalyanmoy Deb, Erik Goodman, and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "Wolfgang Banzhaf. Nsga-net: a multi-objective genetic algorithm for neural architecture search.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 214, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 214, + 146 + ], + "score": 1.0, + "content": "In GECCO-2019, 2018.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 104, + 153, + 504, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 461, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 461, + 177 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems (NIPS), pp. 7816–7827, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 106, + 183, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 505, + 196 + ], + "score": 1.0, + "content": "Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, and Lihi Zelnik-Manor. Xnas:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 405, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 405, + 207 + ], + "score": 1.0, + "content": "Neural architecture search with expert advice. arXiv:1906.08031, 2019.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 108, + 214, + 503, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 228 + ], + "score": 1.0, + "content": "Yurii E Nesterov. A method for solving the convex programming problem with convergence rate o", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 226, + 387, + 237 + ], + "spans": [ + { + "bbox": [ + 117, + 226, + 387, + 237 + ], + "score": 1.0, + "content": "(1/kˆ 2). In Dokl. akad. nauk Sssr, volume 269, pp. 543–547, 1983.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 254, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 505, + 270 + ], + "score": 1.0, + "content": "of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pp.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 267, + 182, + 278 + ], + "spans": [ + { + "bbox": [ + 116, + 267, + 182, + 278 + ], + "score": 1.0, + "content": "722–729, 2008.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 285, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "search via parameter sharing. In International Conference on Machine Learning (ICML), pp.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 308, + 191, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 308, + 191, + 320 + ], + "score": 1.0, + "content": "4092–4101, 2018.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 503, + 350 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "score": 1.0, + "content": "Philip Popien. AutoAugment. GitHub repository, https://github.com/DeepVoltaire/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 339, + 212, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 212, + 350 + ], + "score": 1.0, + "content": "AutoAugment, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In Computer Vision and Pattern", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 368, + 284, + 381 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 284, + 381 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 413–420, 2009.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 108, + 387, + 504, + 422 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 117, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "Quoc V Le, and Alexey Kurakin. Large-scale evolution of image classifiers. In International", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 410, + 379, + 423 + ], + "spans": [ + { + "bbox": [ + 117, + 410, + 379, + 423 + ], + "score": 1.0, + "content": "Conference on Machine Learning (ICML), pp. 2902–2911, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 105, + 428, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 443 + ], + "score": 1.0, + "content": "Christian Sciuto, Kaicheng Yu, Martin Jaggi, Claudiu Musat, and Mathieu Salzmann. Evaluating", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 440, + 407, + 452 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 407, + 452 + ], + "score": 1.0, + "content": "the search phase of neural architecture search. arXiv:1902.08142, 2019.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 109, + 459, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 117, + 471, + 504, + 482 + ], + "spans": [ + { + "bbox": [ + 117, + 471, + 504, + 482 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 482, + 291, + 494 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 291, + 494 + ], + "score": 1.0, + "content": "learning research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 504, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "Inception-ResNet and the impact of residual connections on learning. In AAAI Conference on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 523, + 231, + 535 + ], + "spans": [ + { + "bbox": [ + 116, + 523, + 231, + 535 + ], + "score": 1.0, + "content": "Artificial Intelligence, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 541, + 503, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "Yu Weng, Tianbao Zhou, Lei Liu, and Chunlei Xia. Automatic convolutional neural architecture", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 552, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 115, + 552, + 482, + 565 + ], + "score": 1.0, + "content": "search for image classification under different scenes. IEEE Access, 7:38495–38506, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 106, + 571, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Saining Xie, Alexander Kirillov, Ross Girshick, and Kaiming He. Exploring randomly wired neural", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 583, + 415, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 415, + 596 + ], + "score": 1.0, + "content": "networks for image recognition. arXiv preprint arXiv:1904.01569, 2019a.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 504, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 504, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 504, + 615 + ], + "score": 1.0, + "content": "Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. SNAS: Stochastic neural architecture search.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 613, + 411, + 626 + ], + "spans": [ + { + "bbox": [ + 116, + 613, + 411, + 626 + ], + "score": 1.0, + "content": "In International Conference on Learning Representations (ICLR), 2019b.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 105, + 632, + 504, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "Jun Wu Qi Tian Xin Chen, Lingxi Xie. Progressive differentiable architecture search. In Interna-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 643, + 331, + 656 + ], + "spans": [ + { + "bbox": [ + 116, + 643, + 331, + 656 + ], + "score": 1.0, + "content": "tional Conference on Computer Vision (ICCV), 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 108, + 663, + 504, + 697 + ], + "lines": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. NAS-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 673, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 115, + 673, + 505, + 687 + ], + "score": 1.0, + "content": "Bench-101: Towards reproducible neural architecture search. In International Conference on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 685, + 340, + 698 + ], + "spans": [ + { + "bbox": [ + 115, + 685, + 340, + 698 + ], + "score": 1.0, + "content": "Machine Learning (ICML 2019), pp. 7105–7114, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 107, + 704, + 504, + 727 + ], + "lines": [ + { + "bbox": [ + 105, + 704, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 505, + 717 + ], + "score": 1.0, + "content": "Barret Zoph and Quoc Le. Neural architecture search with reinforcement learning. In International", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 715, + 341, + 727 + ], + "spans": [ + { + "bbox": [ + 116, + 715, + 341, + 727 + ], + "score": 1.0, + "content": "Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 294, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 82, + 504, + 105 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In Interna-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 367, + 106 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 367, + 106 + ], + "score": 1.0, + "content": "tional Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 106, + 82, + 505, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 112, + 504, + 146 + ], + "lines": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "Zhichao Lu, Ian Whalen, Vishnu Boddeti, Yashesh Dhebar, Kalyanmoy Deb, Erik Goodman, and", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 123, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 116, + 123, + 505, + 137 + ], + "score": 1.0, + "content": "Wolfgang Banzhaf. Nsga-net: a multi-objective genetic algorithm for neural architecture search.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 134, + 214, + 146 + ], + "spans": [ + { + "bbox": [ + 116, + 134, + 214, + 146 + ], + "score": 1.0, + "content": "In GECCO-2019, 2018.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 112, + 505, + 146 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 153, + 504, + 177 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "Renqian Luo, Fei Tian, Tao Qin, Enhong Chen, and Tie-Yan Liu. Neural architecture optimization.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 164, + 461, + 177 + ], + "spans": [ + { + "bbox": [ + 115, + 164, + 461, + 177 + ], + "score": 1.0, + "content": "In Advances in Neural Information Processing Systems (NIPS), pp. 7816–7827, 2018.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 154, + 505, + 177 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 183, + 504, + 207 + ], + "lines": [ + { + "bbox": [ + 106, + 185, + 505, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 185, + 505, + 196 + ], + "score": 1.0, + "content": "Niv Nayman, Asaf Noy, Tal Ridnik, Itamar Friedman, Rong Jin, and Lihi Zelnik-Manor. Xnas:", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 195, + 405, + 207 + ], + "spans": [ + { + "bbox": [ + 116, + 195, + 405, + 207 + ], + "score": 1.0, + "content": "Neural architecture search with expert advice. arXiv:1906.08031, 2019.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 106, + 185, + 505, + 207 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 214, + 503, + 237 + ], + "lines": [ + { + "bbox": [ + 106, + 213, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 505, + 228 + ], + "score": 1.0, + "content": "Yurii E Nesterov. A method for solving the convex programming problem with convergence rate o", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 117, + 226, + 387, + 237 + ], + "spans": [ + { + "bbox": [ + 117, + 226, + 387, + 237 + ], + "score": 1.0, + "content": "(1/kˆ 2). In Dokl. akad. nauk Sssr, volume 269, pp. 543–547, 1983.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 213, + 505, + 237 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 244, + 505, + 279 + ], + "lines": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 505, + 257 + ], + "score": 1.0, + "content": "Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 254, + 505, + 270 + ], + "spans": [ + { + "bbox": [ + 115, + 254, + 505, + 270 + ], + "score": 1.0, + "content": "of classes. In Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pp.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 267, + 182, + 278 + ], + "spans": [ + { + "bbox": [ + 116, + 267, + 182, + 278 + ], + "score": 1.0, + "content": "722–729, 2008.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 244, + 505, + 278 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 285, + 504, + 320 + ], + "lines": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 506, + 299 + ], + "score": 1.0, + "content": "Hieu Pham, Melody Guan, Barret Zoph, Quoc Le, and Jeff Dean. Efficient neural architecture", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 115, + 296, + 506, + 311 + ], + "score": 1.0, + "content": "search via parameter sharing. In International Conference on Machine Learning (ICML), pp.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 308, + 191, + 320 + ], + "spans": [ + { + "bbox": [ + 116, + 308, + 191, + 320 + ], + "score": 1.0, + "content": "4092–4101, 2018.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 285, + 506, + 320 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 503, + 350 + ], + "lines": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 505, + 341 + ], + "score": 1.0, + "content": "Philip Popien. AutoAugment. GitHub repository, https://github.com/DeepVoltaire/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 339, + 212, + 350 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 212, + 350 + ], + "score": 1.0, + "content": "AutoAugment, 2019.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 326, + 505, + 350 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 357, + 504, + 381 + ], + "lines": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "Ariadna Quattoni and Antonio Torralba. Recognizing indoor scenes. In Computer Vision and Pattern", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 116, + 368, + 284, + 381 + ], + "spans": [ + { + "bbox": [ + 116, + 368, + 284, + 381 + ], + "score": 1.0, + "content": "Recognition (CVPR), pp. 413–420, 2009.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 356, + 506, + 381 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 387, + 504, + 422 + ], + "lines": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 505, + 400 + ], + "score": 1.0, + "content": "Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Jie Tan,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 117, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 117, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "Quoc V Le, and Alexey Kurakin. Large-scale evolution of image classifiers. In International", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 410, + 379, + 423 + ], + "spans": [ + { + "bbox": [ + 117, + 410, + 379, + 423 + ], + "score": 1.0, + "content": "Conference on Machine Learning (ICML), pp. 2902–2911, 2017.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 106, + 387, + 505, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 428, + 504, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 427, + 506, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 427, + 506, + 443 + ], + "score": 1.0, + "content": "Christian Sciuto, Kaicheng Yu, Martin Jaggi, Claudiu Musat, and Mathieu Salzmann. Evaluating", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 116, + 440, + 407, + 452 + ], + "spans": [ + { + "bbox": [ + 116, + 440, + 407, + 452 + ], + "score": 1.0, + "content": "the search phase of neural architecture search. arXiv:1902.08142, 2019.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 427, + 506, + 452 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 459, + 504, + 493 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 505, + 472 + ], + "score": 1.0, + "content": "Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 117, + 471, + 504, + 482 + ], + "spans": [ + { + "bbox": [ + 117, + 471, + 504, + 482 + ], + "score": 1.0, + "content": "Dropout: a simple way to prevent neural networks from overfitting. The journal of machine", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 116, + 482, + 291, + 494 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 291, + 494 + ], + "score": 1.0, + "content": "learning research, 15(1):1929–1958, 2014.", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 106, + 459, + 505, + 494 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 500, + 504, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander A Alemi. Inception-v4,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 505, + 525 + ], + "score": 1.0, + "content": "Inception-ResNet and the impact of residual connections on learning. In AAAI Conference on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 116, + 523, + 231, + 535 + ], + "spans": [ + { + "bbox": [ + 116, + 523, + 231, + 535 + ], + "score": 1.0, + "content": "Artificial Intelligence, 2017.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30, + "bbox_fs": [ + 106, + 500, + 505, + 535 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 541, + 503, + 565 + ], + "lines": [ + { + "bbox": [ + 106, + 541, + 505, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 555 + ], + "score": 1.0, + "content": "Yu Weng, Tianbao Zhou, Lei Liu, and Chunlei Xia. Automatic convolutional neural architecture", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 552, + 482, + 565 + ], + "spans": [ + { + "bbox": [ + 115, + 552, + 482, + 565 + ], + "score": 1.0, + "content": "search for image classification under different scenes. IEEE Access, 7:38495–38506, 2019.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 541, + 505, + 565 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 571, + 505, + 595 + ], + "lines": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 505, + 585 + ], + "score": 1.0, + "content": "Saining Xie, Alexander Kirillov, Ross Girshick, and Kaiming He. Exploring randomly wired neural", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 583, + 415, + 596 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 415, + 596 + ], + "score": 1.0, + "content": "networks for image recognition. arXiv preprint arXiv:1904.01569, 2019a.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 572, + 505, + 596 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 504, + 626 + ], + "lines": [ + { + "bbox": [ + 106, + 601, + 504, + 615 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 504, + 615 + ], + "score": 1.0, + "content": "Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. SNAS: Stochastic neural architecture search.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 116, + 613, + 411, + 626 + ], + "spans": [ + { + "bbox": [ + 116, + 613, + 411, + 626 + ], + "score": 1.0, + "content": "In International Conference on Learning Representations (ICLR), 2019b.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 106, + 601, + 504, + 626 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 632, + 504, + 655 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 505, + 646 + ], + "score": 1.0, + "content": "Jun Wu Qi Tian Xin Chen, Lingxi Xie. Progressive differentiable architecture search. In Interna-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 643, + 331, + 656 + ], + "spans": [ + { + "bbox": [ + 116, + 643, + 331, + 656 + ], + "score": 1.0, + "content": "tional Conference on Computer Vision (ICCV), 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 632, + 505, + 656 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 663, + 504, + 697 + ], + "lines": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 106, + 663, + 505, + 675 + ], + "score": 1.0, + "content": "Chris Ying, Aaron Klein, Eric Christiansen, Esteban Real, Kevin Murphy, and Frank Hutter. NAS-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 673, + 505, + 687 + ], + "spans": [ + { + "bbox": [ + 115, + 673, + 505, + 687 + ], + "score": 1.0, + "content": "Bench-101: Towards reproducible neural architecture search. In International Conference on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 685, + 340, + 698 + ], + "spans": [ + { + "bbox": [ + 115, + 685, + 340, + 698 + ], + "score": 1.0, + "content": "Machine Learning (ICML 2019), pp. 7105–7114, 2019.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 663, + 505, + 698 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 704, + 504, + 727 + ], + "lines": [ + { + "bbox": [ + 105, + 704, + 505, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 704, + 505, + 717 + ], + "score": 1.0, + "content": "Barret Zoph and Quoc Le. Neural architecture search with reinforcement learning. In International", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 715, + 341, + 727 + ], + "spans": [ + { + "bbox": [ + 116, + 715, + 341, + 727 + ], + "score": 1.0, + "content": "Conference on Learning Representations (ICLR), 2017.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 704, + 505, + 727 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 115, + 502, + 339 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 89, + 502, + 112 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "Table 2: Hyperparameters for different NAS methods. “S” denotes search stage and “A” denotes", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 434, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 264, + 112 + ], + "score": 1.0, + "content": "augmentation stage. For learning rates,", + "type": "text" + }, + { + "bbox": [ + 264, + 100, + 286, + 111 + ], + "score": 0.87, + "content": "a \\downarrow b", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 100, + 406, + 112 + ], + "score": 1.0, + "content": "means cosine annealing from", + "type": "text" + }, + { + "bbox": [ + 406, + 102, + 412, + 110 + ], + "score": 0.64, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 100, + 424, + 112 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 424, + 101, + 429, + 110 + ], + "score": 0.64, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 100, + 434, + 112 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 115, + 502, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 115, + 502, + 339 + ], + "spans": [ + { + "bbox": [ + 108, + 115, + 502, + 339 + ], + "score": 0.983, + "html": "
Method
DARTSStacNASPDARTSMANASCNAS
batch size S6464966464
batch size A969612812864
init channels S1616161616
init channels A3636363636
epochs S50100+10025+25+255060
epochs A600600600600200
optimizer S/ASGDSGDSGDSGDAdam
learning rates S.025↓.001.025↓.001.025↓.0.025↓.001.025-.003
learning rates A.025↓.0.025↓.0.025↓.0.025←.0.025-.001
weight decay S/A3 ×10-43×10-43×10-43×10-43×10-4
optimizer archAdamAdamAdamAdam
learning rates arch3×10-43×10-46×10-43×10-4
weight decay arch10-310-310-310-3
nb cells S CIFAR814/205/11/1786
nb cells A CIFAR2020202020
nb cells S other datasets88/88/8/886
nb cellsA other datasets88888
nb intermediate nodes44446
", + "type": "table", + "image_path": "320203340cda9ff05fec56f6573c4eaf1ef276b1e36acceee02cf27a14c0f25e.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 115, + 502, + 189.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 189.66666666666669, + 502, + 264.33333333333337 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 264.33333333333337, + 502, + 339.00000000000006 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "title", + "bbox": [ + 107, + 365, + 183, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 185, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 185, + 380 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "This section details the datasets and the hyperparameters used for each method on each dataset.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "Search spaces were naturally left unchanged. Hyperparameters were chosen as close as possible to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "the original paper and occasionally updated to more recent implementations. The network size was", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "score": 1.0, + "content": "tuned similarly for all methods for SPORT8, MIT67 and FLOWERS102. All experiments were run", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 435, + 234, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 234, + 447 + ], + "score": 1.0, + "content": "on NVIDIA Tesla V100 GPUs.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "title", + "bbox": [ + 108, + 461, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 288, + 474 + ], + "score": 1.0, + "content": "A.1 METHODS AND HYPERPARAMETERS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "score": 1.0, + "content": "During search, models are trained/validated on the training/validation subsets, respectively During", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 335, + 506 + ], + "score": 1.0, + "content": "the final evaluation, the model is trained on the training", + "type": "text" + }, + { + "bbox": [ + 335, + 495, + 342, + 503 + ], + "score": 0.38, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "validation subsets and tested on the test", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 504, + 137, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 137, + 516 + ], + "score": 1.0, + "content": "subset.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "score": 1.0, + "content": "Common hyperparameters. All 8 methods share a common number of hyperparameters precised", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "here. When SGD optimizer is used, momentum is .9 while when Adam is used, momentum is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 543, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 107, + 543, + 174, + 555 + ], + "score": 0.87, + "content": "\\beta = ( 0 . 5 , 0 . 9 9 9 )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 543, + 294, + 556 + ], + "score": 1.0, + "content": ". Gradient clipping is set at 5.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "DARTS, StacNAS, PDARTS, MANAS, CNAS common hyperparameters. These methods are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 570, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 585 + ], + "score": 1.0, + "content": "inspired by DARTS code-wise and consequently share a common number of hyperparameters,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 582, + 221, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 221, + 595 + ], + "score": 1.0, + "content": "which we precise in table 1.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 104, + 597, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 597, + 505, + 613 + ], + "score": 1.0, + "content": "DARTS. We used the following repository : https://github.com/khanrc/pt.darts. It notably up-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "dates the official implementation to a pytorch version posterior to 0.4. Additional enhancements", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "include cutout of size 16 (DeVries & Taylor, 2017), path dropout of probability 0.2 (Larsson &", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 632, + 346, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 346, + 644 + ], + "score": 1.0, + "content": "Shakhnarovich, 2017), and auxiliary tower with weight 0.4.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "StacNAS. We used an unofficial implementation provided by the authors. The search process con-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "sists of 2 stages, of which the details are given in table 1. Additional enhancements are the same as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 670, + 144, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 144, + 682 + ], + "score": 1.0, + "content": "DARTS.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 297, + 700 + ], + "score": 1.0, + "content": "PDARTS. We used the official implementation", + "type": "text" + }, + { + "bbox": [ + 297, + 690, + 303, + 697 + ], + "score": 0.53, + "content": ":", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "https://github.com/chenxin061/pdarts. The search", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "process consists of 3 stages, of which general details are given in table 1. At stage 1, 2 and 3", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "respectively, the number of operations decreases from 8 to 5 to 3, and the dropout probability on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "skip-connect increases from 0.0 to 0.4 to 0.7 for CIFAR10, SPORT8, MIT67 and FLOWERS102", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 115, + 502, + 339 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 89, + 502, + 112 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 102 + ], + "score": 1.0, + "content": "Table 2: Hyperparameters for different NAS methods. “S” denotes search stage and “A” denotes", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 434, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 264, + 112 + ], + "score": 1.0, + "content": "augmentation stage. For learning rates,", + "type": "text" + }, + { + "bbox": [ + 264, + 100, + 286, + 111 + ], + "score": 0.87, + "content": "a \\downarrow b", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 100, + 406, + 112 + ], + "score": 1.0, + "content": "means cosine annealing from", + "type": "text" + }, + { + "bbox": [ + 406, + 102, + 412, + 110 + ], + "score": 0.64, + "content": "a", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 100, + 424, + 112 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 424, + 101, + 429, + 110 + ], + "score": 0.64, + "content": "b", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 100, + 434, + 112 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 115, + 502, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 115, + 502, + 339 + ], + "spans": [ + { + "bbox": [ + 108, + 115, + 502, + 339 + ], + "score": 0.983, + "html": "
Method
DARTSStacNASPDARTSMANASCNAS
batch size S6464966464
batch size A969612812864
init channels S1616161616
init channels A3636363636
epochs S50100+10025+25+255060
epochs A600600600600200
optimizer S/ASGDSGDSGDSGDAdam
learning rates S.025↓.001.025↓.001.025↓.0.025↓.001.025-.003
learning rates A.025↓.0.025↓.0.025↓.0.025←.0.025-.001
weight decay S/A3 ×10-43×10-43×10-43×10-43×10-4
optimizer archAdamAdamAdamAdam
learning rates arch3×10-43×10-46×10-43×10-4
weight decay arch10-310-310-310-3
nb cells S CIFAR814/205/11/1786
nb cells A CIFAR2020202020
nb cells S other datasets88/88/8/886
nb cellsA other datasets88888
nb intermediate nodes44446
", + "type": "table", + "image_path": "320203340cda9ff05fec56f6573c4eaf1ef276b1e36acceee02cf27a14c0f25e.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 108, + 115, + 502, + 189.66666666666669 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 189.66666666666669, + 502, + 264.33333333333337 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 264.33333333333337, + 502, + 339.00000000000006 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "title", + "bbox": [ + 107, + 365, + 183, + 377 + ], + "lines": [ + { + "bbox": [ + 105, + 363, + 185, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 185, + 380 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 391, + 505, + 446 + ], + "lines": [ + { + "bbox": [ + 106, + 390, + 505, + 404 + ], + "spans": [ + { + "bbox": [ + 106, + 390, + 505, + 404 + ], + "score": 1.0, + "content": "This section details the datasets and the hyperparameters used for each method on each dataset.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 506, + 415 + ], + "score": 1.0, + "content": "Search spaces were naturally left unchanged. Hyperparameters were chosen as close as possible to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 506, + 426 + ], + "score": 1.0, + "content": "the original paper and occasionally updated to more recent implementations. The network size was", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 422, + 506, + 438 + ], + "score": 1.0, + "content": "tuned similarly for all methods for SPORT8, MIT67 and FLOWERS102. All experiments were run", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 435, + 234, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 234, + 447 + ], + "score": 1.0, + "content": "on NVIDIA Tesla V100 GPUs.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 390, + 506, + 447 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 461, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 106, + 461, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 288, + 474 + ], + "score": 1.0, + "content": "A.1 METHODS AND HYPERPARAMETERS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 482, + 505, + 515 + ], + "lines": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 496 + ], + "score": 1.0, + "content": "During search, models are trained/validated on the training/validation subsets, respectively During", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 492, + 505, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 492, + 335, + 506 + ], + "score": 1.0, + "content": "the final evaluation, the model is trained on the training", + "type": "text" + }, + { + "bbox": [ + 335, + 495, + 342, + 503 + ], + "score": 0.38, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 492, + 505, + 506 + ], + "score": 1.0, + "content": "validation subsets and tested on the test", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 504, + 137, + 516 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 137, + 516 + ], + "score": 1.0, + "content": "subset.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 481, + 505, + 516 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 521, + 505, + 555 + ], + "lines": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 522, + 505, + 533 + ], + "score": 1.0, + "content": "Common hyperparameters. All 8 methods share a common number of hyperparameters precised", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 532, + 505, + 545 + ], + "score": 1.0, + "content": "here. When SGD optimizer is used, momentum is .9 while when Adam is used, momentum is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 107, + 543, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 107, + 543, + 174, + 555 + ], + "score": 0.87, + "content": "\\beta = ( 0 . 5 , 0 . 9 9 9 )", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 543, + 294, + 556 + ], + "score": 1.0, + "content": ". Gradient clipping is set at 5.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 522, + 505, + 556 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 560, + 505, + 593 + ], + "lines": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 559, + 505, + 573 + ], + "score": 1.0, + "content": "DARTS, StacNAS, PDARTS, MANAS, CNAS common hyperparameters. These methods are", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 570, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 506, + 585 + ], + "score": 1.0, + "content": "inspired by DARTS code-wise and consequently share a common number of hyperparameters,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 582, + 221, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 582, + 221, + 595 + ], + "score": 1.0, + "content": "which we precise in table 1.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 559, + 506, + 595 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 599, + 505, + 643 + ], + "lines": [ + { + "bbox": [ + 104, + 597, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 597, + 505, + 613 + ], + "score": 1.0, + "content": "DARTS. We used the following repository : https://github.com/khanrc/pt.darts. It notably up-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 622 + ], + "score": 1.0, + "content": "dates the official implementation to a pytorch version posterior to 0.4. Additional enhancements", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 505, + 633 + ], + "score": 1.0, + "content": "include cutout of size 16 (DeVries & Taylor, 2017), path dropout of probability 0.2 (Larsson &", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 632, + 346, + 644 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 346, + 644 + ], + "score": 1.0, + "content": "Shakhnarovich, 2017), and auxiliary tower with weight 0.4.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 104, + 597, + 505, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 682 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 505, + 662 + ], + "score": 1.0, + "content": "StacNAS. We used an unofficial implementation provided by the authors. The search process con-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 672 + ], + "score": 1.0, + "content": "sists of 2 stages, of which the details are given in table 1. Additional enhancements are the same as", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 670, + 144, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 144, + 682 + ], + "score": 1.0, + "content": "DARTS.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 648, + 505, + 682 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 687, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 686, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 297, + 700 + ], + "score": 1.0, + "content": "PDARTS. We used the official implementation", + "type": "text" + }, + { + "bbox": [ + 297, + 690, + 303, + 697 + ], + "score": 0.53, + "content": ":", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 686, + 506, + 700 + ], + "score": 1.0, + "content": "https://github.com/chenxin061/pdarts. The search", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 712 + ], + "score": 1.0, + "content": "process consists of 3 stages, of which general details are given in table 1. At stage 1, 2 and 3", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 506, + 722 + ], + "score": 1.0, + "content": "respectively, the number of operations decreases from 8 to 5 to 3, and the dropout probability on", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "skip-connect increases from 0.0 to 0.4 to 0.7 for CIFAR10, SPORT8, MIT67 and FLOWERS102", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(0.1 to 0.2 to 0.3 for CIFAR100). Discovered cells are restricted to keep at most 2 skip-connect", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "operations. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), DropPath", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 464, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 464, + 117 + ], + "score": 1.0, + "content": "of probability 0.3 (Larsson & Shakhnarovich, 2017) and auxiliary tower with weight 0.4.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 686, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "(0.1 to 0.2 to 0.3 for CIFAR100). Discovered cells are restricted to keep at most 2 skip-connect", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "operations. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017), DropPath", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 464, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 464, + 117 + ], + "score": 1.0, + "content": "of probability 0.3 (Larsson & Shakhnarovich, 2017) and auxiliary tower with weight 0.4.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 133 + ], + "score": 1.0, + "content": "MANAS. We used an unofficial implementation provided by the authors. The reward baseline is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "5, gamma is 0.1 (0.07 for SPORT8, 0.05 for FLOWERS102, 0.01 for MIT67) and the Boltzmann", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "temperature decays from 1000 to 200 (300 to 100 for SPORT8, 5000 to 2000 for MIT67). Additional", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 265, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 265, + 165 + ], + "score": 1.0, + "content": "enhancements are the same as DARTS.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 503, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 170, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 185 + ], + "score": 1.0, + "content": "CNAS. We used the official implementation : https://github.com/tianbaochou/CNAS. Label", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Smoothing is used with epsilon 0.1. Other additional enhancements include cutout of size 16 (De-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 193, + 457, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 457, + 205 + ], + "score": 1.0, + "content": "Vries & Taylor, 2017), DropPath of probability 0.25 (Larsson & Shakhnarovich, 2017).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 209, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "NSGANET. We used the official implementation: https://github.com/ianwhale/nsga-net. The search", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "is done in the micro search space (with 2 cells to search, 9 operations in the search space, 5 blocks in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 232, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 245 + ], + "score": 1.0, + "content": "each cell) on 8 layers networks. The population size is 40, the number of generations is 30 and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "number of offsprings created by generation is 20. Networks are trained for 20 epochs, with batch size", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "128, and the initial number of channels is 16. For architecture evaluation, the network is composed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "score": 1.0, + "content": "of 20 cells for CIFAR10 and CIFAR100, and 8 cells for SPORT8, MIT67 and FLOWERS102. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "final selected models are trained for 600 epochs with batch size 96 and the initial number of channels", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 319, + 300 + ], + "score": 1.0, + "content": "34. Momentum SGD is used with initial learning rate", + "type": "text" + }, + { + "bbox": [ + 320, + 287, + 369, + 298 + ], + "score": 0.9, + "content": "\\eta _ { w } = 0 . 0 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 285, + 506, + 300 + ], + "score": 1.0, + "content": "(annealed down to zero following", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 260, + 311 + ], + "score": 1.0, + "content": "a cosine schedule), and weight decay", + "type": "text" + }, + { + "bbox": [ + 261, + 297, + 300, + 308 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - \\hat { 4 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 297, + 506, + 311 + ], + "score": 1.0, + "content": ". The filter increment is set to 4, and squeeze and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "excitation is used. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "DropPath of probability 0.2 (Larsson & Shakhnarovich, 2017) and auxiliary tower with weight 0.4.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 505, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "score": 1.0, + "content": "NAO. We used the official Pytorch implementation: https://github.com/renqianluo/NAO pytorch.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "The LSTM model used to encode architecture has a token embedding size of 48 and a hidden state", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "score": 1.0, + "content": "size of 96. The LSTM model used to decode architecture has a hidden state size of 96. The encoder", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "and decoder are trained using Adam for 1000 epochs with a learning rate of 0.001. The trade-off", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 163, + 393 + ], + "score": 1.0, + "content": "parameters is", + "type": "text" + }, + { + "bbox": [ + 163, + 381, + 198, + 391 + ], + "score": 0.9, + "content": "\\lambda = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 381, + 415, + 393 + ], + "score": 1.0, + "content": ". The step size to perform continuous optimization is", + "type": "text" + }, + { + "bbox": [ + 416, + 381, + 447, + 392 + ], + "score": 0.9, + "content": "\\eta = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 381, + 506, + 393 + ], + "score": 1.0, + "content": ". The number", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "of nodes is fixed to 5, normal cell is stacked 3 (2 for SPORT8, MIT67 and FLOWERS102) times to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "score": 1.0, + "content": "form the CNN architecture, which corresponds to a 11 (8 for SPORT8, MIT67 and FLOWERS102)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "cells network and the initial number of channels is 20. Networks are trained for 100 epochs with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "batch size 64 for training, and validated for 20 epochs with batch size 500. For architecture evalua-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 435, + 504, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 504, + 446 + ], + "score": 1.0, + "content": "tion, the final CNN architecture is a 20 cells network (8 cells network). This network is trained for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "600 epochs with batch size 128 for the training set (96 for both for SPORT8, MIT67 and FLOW-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "ERS102), 500 for the validation set (128 for both for SPORT8, MIT67 and FLOWERS102) and the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 436, + 481 + ], + "score": 1.0, + "content": "initial number of channels is 36. Momentum SGD is used with initial learning rate", + "type": "text" + }, + { + "bbox": [ + 437, + 469, + 486, + 479 + ], + "score": 0.9, + "content": "\\eta _ { w } = 0 . 0 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "(an-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 378, + 492 + ], + "score": 1.0, + "content": "nealed down to zero following a cosine schedule) and weight decay", + "type": "text" + }, + { + "bbox": [ + 378, + 478, + 417, + 489 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - \\overline { { 4 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 478, + 505, + 492 + ], + "score": 1.0, + "content": ". Additional enhance-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "ments include cutout of size 16 (DeVries & Taylor, 2017), DropPath of probability 0.2 (Larsson &", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Shakhnarovich, 2017), dropout of probability 0.4 (Srivastava et al., 2014), and auxiliary tower with", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 512, + 154, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 154, + 525 + ], + "score": 1.0, + "content": "weight 0.4.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 541 + ], + "score": 1.0, + "content": "ENAS. We used the official tensorflow implementation for experiments on CIFAR10", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "and CIFAR100: https://github.com/melodyguan/enas Experiments on SPORT8, MIT67 and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "FLOWERS102 are done using a more recent version in Pytorch for the search :", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "https://github.com/MengTianjian/enas-pytorch. Because only the search process is implemented", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "there, the evaluation code used is the same as DARTS. During the search, networks are composed", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 241, + 596 + ], + "score": 1.0, + "content": "of 8 cells. The shared parameters", + "type": "text" + }, + { + "bbox": [ + 241, + 586, + 250, + 594 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "are trained with Nesterov momentum (Nesterov, 1983), weight", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 132, + 608 + ], + "score": 1.0, + "content": "decay", + "type": "text" + }, + { + "bbox": [ + 132, + 594, + 153, + 605 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 594, + 506, + 608 + ], + "score": 1.0, + "content": ", gradient clipping 5, batch size 160, 20 output filters, and a cosine learning rate schedule", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 127, + 618 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 606, + 182, + 617 + ], + "score": 0.86, + "content": "l _ { m } a x = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 605, + 186, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 186, + 606, + 245, + 617 + ], + "score": 0.88, + "content": "l _ { m } i n = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 605, + 249, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 249, + 606, + 285, + 617 + ], + "score": 0.85, + "content": "T _ { 0 } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 605, + 289, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 289, + 606, + 333, + 617 + ], + "score": 0.81, + "content": "T _ { m } u l = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "(Loshchilov & Hutter, 2017). Each archi-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 254, + 629 + ], + "score": 1.0, + "content": "tecture search is run for 150 epochs.", + "type": "text" + }, + { + "bbox": [ + 255, + 618, + 263, + 627 + ], + "score": 0.64, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "are initialized with He initialization (He & Sun, 2015). The", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 180, + 640 + ], + "score": 1.0, + "content": "policy parameters", + "type": "text" + }, + { + "bbox": [ + 180, + 628, + 186, + 637 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "are initialized uniformly in [-0.1, 0.1], and trained with Adam at a learning rate", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "of 0.00035. A tanh constant of 1.10 and a temperature of 2.5 is applied to the controllers logits, and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "the controller entropy is added to the reward with weight 0.1. For the evaluation, the architecture", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "searched is extended to 17 cells (8 for SPORT8, MIT67 and FLOWERS102), trained for 630 epochs,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 354, + 684 + ], + "score": 1.0, + "content": "with batch size 144, and a cosine learning rate schedule with", + "type": "text" + }, + { + "bbox": [ + 355, + 672, + 404, + 683 + ], + "score": 0.88, + "content": "l _ { \\mathrm { m a x } } = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 671, + 408, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 408, + 672, + 460, + 683 + ], + "score": 0.84, + "content": "l _ { \\mathrm { m i n } } = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 671, + 465, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 465, + 672, + 501, + 683 + ], + "score": 0.84, + "content": "T _ { 0 } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 671, + 505, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 682, + 148, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 143, + 694 + ], + "score": 0.9, + "content": "T _ { \\mathrm { m u l } } = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 682, + 148, + 695 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 45 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 116 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 504, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 122, + 506, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 506, + 133 + ], + "score": 1.0, + "content": "MANAS. We used an unofficial implementation provided by the authors. The reward baseline is", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "5, gamma is 0.1 (0.07 for SPORT8, 0.05 for FLOWERS102, 0.01 for MIT67) and the Boltzmann", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 506, + 156 + ], + "score": 1.0, + "content": "temperature decays from 1000 to 200 (300 to 100 for SPORT8, 5000 to 2000 for MIT67). Additional", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 154, + 265, + 165 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 265, + 165 + ], + "score": 1.0, + "content": "enhancements are the same as DARTS.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 122, + 506, + 165 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 171, + 503, + 204 + ], + "lines": [ + { + "bbox": [ + 105, + 170, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 185 + ], + "score": 1.0, + "content": "CNAS. We used the official implementation : https://github.com/tianbaochou/CNAS. Label", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 195 + ], + "score": 1.0, + "content": "Smoothing is used with epsilon 0.1. Other additional enhancements include cutout of size 16 (De-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 193, + 457, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 193, + 457, + 205 + ], + "score": 1.0, + "content": "Vries & Taylor, 2017), DropPath of probability 0.25 (Larsson & Shakhnarovich, 2017).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 170, + 506, + 205 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 209, + 505, + 331 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "NSGANET. We used the official implementation: https://github.com/ianwhale/nsga-net. The search", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "is done in the micro search space (with 2 cells to search, 9 operations in the search space, 5 blocks in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 232, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 506, + 245 + ], + "score": 1.0, + "content": "each cell) on 8 layers networks. The population size is 40, the number of generations is 30 and the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 255 + ], + "score": 1.0, + "content": "number of offsprings created by generation is 20. Networks are trained for 20 epochs, with batch size", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 505, + 266 + ], + "score": 1.0, + "content": "128, and the initial number of channels is 16. For architecture evaluation, the network is composed", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 276 + ], + "score": 1.0, + "content": "of 20 cells for CIFAR10 and CIFAR100, and 8 cells for SPORT8, MIT67 and FLOWERS102. The", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "final selected models are trained for 600 epochs with batch size 96 and the initial number of channels", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 285, + 506, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 285, + 319, + 300 + ], + "score": 1.0, + "content": "34. Momentum SGD is used with initial learning rate", + "type": "text" + }, + { + "bbox": [ + 320, + 287, + 369, + 298 + ], + "score": 0.9, + "content": "\\eta _ { w } = 0 . 0 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 285, + 506, + 300 + ], + "score": 1.0, + "content": "(annealed down to zero following", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 297, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 260, + 311 + ], + "score": 1.0, + "content": "a cosine schedule), and weight decay", + "type": "text" + }, + { + "bbox": [ + 261, + 297, + 300, + 308 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - \\hat { 4 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 297, + 506, + 311 + ], + "score": 1.0, + "content": ". The filter increment is set to 4, and squeeze and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 105, + 308, + 505, + 321 + ], + "score": 1.0, + "content": "excitation is used. Additional enhancements include cutout of size 16 (DeVries & Taylor, 2017),", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 320, + 505, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 320, + 505, + 333 + ], + "score": 1.0, + "content": "DropPath of probability 0.2 (Larsson & Shakhnarovich, 2017) and auxiliary tower with weight 0.4.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 15, + "bbox_fs": [ + 104, + 210, + 506, + 333 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 505, + 523 + ], + "lines": [ + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 504, + 349 + ], + "score": 1.0, + "content": "NAO. We used the official Pytorch implementation: https://github.com/renqianluo/NAO pytorch.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 359 + ], + "score": 1.0, + "content": "The LSTM model used to encode architecture has a token embedding size of 48 and a hidden state", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 359, + 505, + 370 + ], + "score": 1.0, + "content": "size of 96. The LSTM model used to decode architecture has a hidden state size of 96. The encoder", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 382 + ], + "score": 1.0, + "content": "and decoder are trained using Adam for 1000 epochs with a learning rate of 0.001. The trade-off", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 381, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 163, + 393 + ], + "score": 1.0, + "content": "parameters is", + "type": "text" + }, + { + "bbox": [ + 163, + 381, + 198, + 391 + ], + "score": 0.9, + "content": "\\lambda = 0 . 9", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 381, + 415, + 393 + ], + "score": 1.0, + "content": ". The step size to perform continuous optimization is", + "type": "text" + }, + { + "bbox": [ + 416, + 381, + 447, + 392 + ], + "score": 0.9, + "content": "\\eta = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 381, + 506, + 393 + ], + "score": 1.0, + "content": ". The number", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "of nodes is fixed to 5, normal cell is stacked 3 (2 for SPORT8, MIT67 and FLOWERS102) times to", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 402, + 504, + 414 + ], + "score": 1.0, + "content": "form the CNN architecture, which corresponds to a 11 (8 for SPORT8, MIT67 and FLOWERS102)", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "cells network and the initial number of channels is 20. Networks are trained for 100 epochs with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "batch size 64 for training, and validated for 20 epochs with batch size 500. For architecture evalua-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 435, + 504, + 446 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 504, + 446 + ], + "score": 1.0, + "content": "tion, the final CNN architecture is a 20 cells network (8 cells network). This network is trained for", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "600 epochs with batch size 128 for the training set (96 for both for SPORT8, MIT67 and FLOW-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 470 + ], + "score": 1.0, + "content": "ERS102), 500 for the validation set (128 for both for SPORT8, MIT67 and FLOWERS102) and the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 468, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 436, + 481 + ], + "score": 1.0, + "content": "initial number of channels is 36. Momentum SGD is used with initial learning rate", + "type": "text" + }, + { + "bbox": [ + 437, + 469, + 486, + 479 + ], + "score": 0.9, + "content": "\\eta _ { w } = 0 . 0 2 5", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 468, + 505, + 481 + ], + "score": 1.0, + "content": "(an-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 478, + 505, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 478, + 378, + 492 + ], + "score": 1.0, + "content": "nealed down to zero following a cosine schedule) and weight decay", + "type": "text" + }, + { + "bbox": [ + 378, + 478, + 417, + 489 + ], + "score": 0.91, + "content": "3 \\times 1 0 ^ { - \\overline { { 4 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 478, + 505, + 492 + ], + "score": 1.0, + "content": ". Additional enhance-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 490, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 505, + 502 + ], + "score": 1.0, + "content": "ments include cutout of size 16 (DeVries & Taylor, 2017), DropPath of probability 0.2 (Larsson &", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Shakhnarovich, 2017), dropout of probability 0.4 (Srivastava et al., 2014), and auxiliary tower with", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 512, + 154, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 154, + 525 + ], + "score": 1.0, + "content": "weight 0.4.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 336, + 506, + 525 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 529, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 106, + 529, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 506, + 541 + ], + "score": 1.0, + "content": "ENAS. We used the official tensorflow implementation for experiments on CIFAR10", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "and CIFAR100: https://github.com/melodyguan/enas Experiments on SPORT8, MIT67 and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 564 + ], + "score": 1.0, + "content": "FLOWERS102 are done using a more recent version in Pytorch for the search :", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 505, + 574 + ], + "score": 1.0, + "content": "https://github.com/MengTianjian/enas-pytorch. Because only the search process is implemented", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 505, + 585 + ], + "score": 1.0, + "content": "there, the evaluation code used is the same as DARTS. During the search, networks are composed", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 241, + 596 + ], + "score": 1.0, + "content": "of 8 cells. The shared parameters", + "type": "text" + }, + { + "bbox": [ + 241, + 586, + 250, + 594 + ], + "score": 0.73, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "are trained with Nesterov momentum (Nesterov, 1983), weight", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 594, + 506, + 608 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 132, + 608 + ], + "score": 1.0, + "content": "decay", + "type": "text" + }, + { + "bbox": [ + 132, + 594, + 153, + 605 + ], + "score": 0.9, + "content": "1 0 ^ { - 4 }", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 594, + 506, + 608 + ], + "score": 1.0, + "content": ", gradient clipping 5, batch size 160, 20 output filters, and a cosine learning rate schedule", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 605, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 127, + 618 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 606, + 182, + 617 + ], + "score": 0.86, + "content": "l _ { m } a x = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 605, + 186, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 186, + 606, + 245, + 617 + ], + "score": 0.88, + "content": "l _ { m } i n = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 605, + 249, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 249, + 606, + 285, + 617 + ], + "score": 0.85, + "content": "T _ { 0 } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 605, + 289, + 618 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 289, + 606, + 333, + 617 + ], + "score": 0.81, + "content": "T _ { m } u l = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 605, + 505, + 618 + ], + "score": 1.0, + "content": "(Loshchilov & Hutter, 2017). Each archi-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 616, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 254, + 629 + ], + "score": 1.0, + "content": "tecture search is run for 150 epochs.", + "type": "text" + }, + { + "bbox": [ + 255, + 618, + 263, + 627 + ], + "score": 0.64, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 616, + 505, + 629 + ], + "score": 1.0, + "content": "are initialized with He initialization (He & Sun, 2015). The", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 180, + 640 + ], + "score": 1.0, + "content": "policy parameters", + "type": "text" + }, + { + "bbox": [ + 180, + 628, + 186, + 637 + ], + "score": 0.75, + "content": "\\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "are initialized uniformly in [-0.1, 0.1], and trained with Adam at a learning rate", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 505, + 650 + ], + "score": 1.0, + "content": "of 0.00035. A tanh constant of 1.10 and a temperature of 2.5 is applied to the controllers logits, and", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 662 + ], + "score": 1.0, + "content": "the controller entropy is added to the reward with weight 0.1. For the evaluation, the architecture", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 660, + 505, + 673 + ], + "score": 1.0, + "content": "searched is extended to 17 cells (8 for SPORT8, MIT67 and FLOWERS102), trained for 630 epochs,", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 671, + 505, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 354, + 684 + ], + "score": 1.0, + "content": "with batch size 144, and a cosine learning rate schedule with", + "type": "text" + }, + { + "bbox": [ + 355, + 672, + 404, + 683 + ], + "score": 0.88, + "content": "l _ { \\mathrm { m a x } } = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 404, + 671, + 408, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 408, + 672, + 460, + 683 + ], + "score": 0.84, + "content": "l _ { \\mathrm { m i n } } = 0 . 0 0 1", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 671, + 465, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 465, + 672, + 501, + 683 + ], + "score": 0.84, + "content": "T _ { 0 } = 1 0", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 671, + 505, + 684 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 106, + 682, + 148, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 682, + 143, + 694 + ], + "score": 0.9, + "content": "T _ { \\mathrm { m u l } } = 2", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 682, + 148, + 695 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 529, + 506, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 178, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "score": 1.0, + "content": "A.2 DATASETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 373, + 114 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 374, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 374, + 118 + ], + "score": 1.0, + "content": "We present here the datasets used and how they are pre-processed.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 119, + 504, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 119, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 506, + 132 + ], + "score": 1.0, + "content": "CIFAR10. The CIFAR10 dataset (Krizhevsky, 2009) is a dataset of 10 classes and consists of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 131, + 359, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 327, + 143 + ], + "score": 1.0, + "content": "50, 000 training images and 10, 000 test images of size", + "type": "text" + }, + { + "bbox": [ + 327, + 131, + 356, + 142 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 131, + 359, + 143 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 147, + 503, + 170 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "CIFAR100. The CIFAR100 dataset (Krizhevsky, 2009) is a dataset of 100 classes and consists of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 159, + 360, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 327, + 171 + ], + "score": 1.0, + "content": "50, 000 training images and 10, 000 test images of size", + "type": "text" + }, + { + "bbox": [ + 327, + 159, + 356, + 169 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 159, + 360, + 171 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 107, + 175, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "Each of these datasets is split into a training, validation and testing subsets of size 25, 000, 25, 000", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "and 10, 000 respectively. For both these datasets, we use standard data pre-processing and augmen-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "tation techniques, i.e. subtracting the channel mean and dividing by the channel standard deviation;", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 271, + 221 + ], + "score": 1.0, + "content": "centrally padding the training images to", + "type": "text" + }, + { + "bbox": [ + 271, + 209, + 300, + 219 + ], + "score": 0.88, + "content": "4 0 \\times 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 208, + 454, + 221 + ], + "score": 1.0, + "content": "and randomly cropping them back to", + "type": "text" + }, + { + "bbox": [ + 454, + 209, + 483, + 219 + ], + "score": 0.88, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "; and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 220, + 257, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 257, + 232 + ], + "score": 1.0, + "content": "randomly clipping them horizontally.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 250 + ], + "score": 1.0, + "content": "SPORT8. This is an action recognition dataset containing 8 sport event categories and a total of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "1579 images (Li & Fei-Fei, 2007). The tiny size of this dataset stresses the generalization capabilities", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 259, + 241, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 241, + 271 + ], + "score": 1.0, + "content": "of any NAS method applied to it.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 108, + 275, + 504, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "score": 1.0, + "content": "MIT67. This is a dataset of 67 classes representing different indoor scenes and consists of 15, 620", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 286, + 324, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 324, + 298 + ], + "score": 1.0, + "content": "images of different sizes (Quattoni & Torralba, 2009).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 108, + 303, + 502, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "FLOWERS102. This is a dataset of 102 classes representing different species of flowers and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 314, + 405, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 405, + 326 + ], + "score": 1.0, + "content": "consists of 8, 189 images of different sizes (Nilsback & Zisserman, 2008).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "Each of these datasets is split into a training, validation and testing subsets with proportions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 167, + 354 + ], + "score": 0.64, + "content": "4 0 / 4 0 / 2 0 \\ ( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 342, + 505, + 355 + ], + "score": 1.0, + "content": ". For each one, we use use standard data pre-processing and augmentation tech-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "niques, i.e. subtracting the channel mean and dividing the channel standard deviation, cropping the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 366, + 378 + ], + "score": 1.0, + "content": "training images to random size and aspect ratio, resizing them to", + "type": "text" + }, + { + "bbox": [ + 366, + 364, + 405, + 375 + ], + "score": 0.9, + "content": "2 2 4 \\times 2 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 363, + 505, + 378 + ], + "score": 1.0, + "content": ", and randomly changing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 387, + 389 + ], + "score": 1.0, + "content": "their brightness, contrast, and saturation, while resizing test images to", + "type": "text" + }, + { + "bbox": [ + 388, + 375, + 427, + 386 + ], + "score": 0.88, + "content": "2 5 6 \\times 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 374, + 505, + 389 + ], + "score": 1.0, + "content": "and cropping them", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 160, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 160, + 398 + ], + "score": 1.0, + "content": "at the center.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "title", + "bbox": [ + 108, + 411, + 229, + 422 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 230, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 230, + 423 + ], + "score": 1.0, + "content": "A.3 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 165, + 454, + 443, + 564 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 107, + 431, + 361, + 443 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 431, + 361, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 361, + 444 + ], + "score": 1.0, + "content": "A.3.1 DIFFERENT TRAINING PROTOCOLS FOR RESNET-50", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "image_body", + "bbox": [ + 165, + 454, + 443, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 454, + 443, + 564 + ], + "spans": [ + { + "bbox": [ + 165, + 454, + 443, + 564 + ], + "score": 0.969, + "type": "image", + "image_path": "ce230d7557a836be95b1c534a22b98a18804cb6ce095b27f9043206e3ab14ff8.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 165, + 454, + 443, + 490.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 165, + 490.6666666666667, + 443, + 527.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 165, + 527.3333333333334, + 443, + 564.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 579, + 505, + 646 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 580, + 504, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 504, + 592 + ], + "score": 1.0, + "content": "Figure 10: Extension of Figure 3, including results obtained by training a ResNet-50 on CIFAR10.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "Bars with darker shade are for ResNet-50 and bars with lighter shade are for DARTS (same as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "Figure 3). Result are for 8 runs of each training protocol. For ResNet-50, the auxiliary tower", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "was added after layer 2. As DropPath (Larsson & Shakhnarovich, 2017) would not have been", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "straightforward to apply, we instead implemented Stocastic Depth (Huang et al., 2016), to a similar", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 634, + 134, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 134, + 647 + ], + "score": 1.0, + "content": "effect.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + } + ], + "index": 27 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 293, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 294, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 178, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 180, + 95 + ], + "score": 1.0, + "content": "A.2 DATASETS", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 103, + 373, + 114 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 374, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 374, + 118 + ], + "score": 1.0, + "content": "We present here the datasets used and how they are pre-processed.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 100, + 374, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 119, + 504, + 142 + ], + "lines": [ + { + "bbox": [ + 106, + 119, + 506, + 132 + ], + "spans": [ + { + "bbox": [ + 106, + 119, + 506, + 132 + ], + "score": 1.0, + "content": "CIFAR10. The CIFAR10 dataset (Krizhevsky, 2009) is a dataset of 10 classes and consists of", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 131, + 359, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 131, + 327, + 143 + ], + "score": 1.0, + "content": "50, 000 training images and 10, 000 test images of size", + "type": "text" + }, + { + "bbox": [ + 327, + 131, + 356, + 142 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 131, + 359, + 143 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 119, + 506, + 143 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 147, + 503, + 170 + ], + "lines": [ + { + "bbox": [ + 106, + 147, + 505, + 160 + ], + "spans": [ + { + "bbox": [ + 106, + 147, + 505, + 160 + ], + "score": 1.0, + "content": "CIFAR100. The CIFAR100 dataset (Krizhevsky, 2009) is a dataset of 100 classes and consists of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 159, + 360, + 171 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 327, + 171 + ], + "score": 1.0, + "content": "50, 000 training images and 10, 000 test images of size", + "type": "text" + }, + { + "bbox": [ + 327, + 159, + 356, + 169 + ], + "score": 0.89, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 159, + 360, + 171 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 147, + 505, + 171 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 175, + 505, + 231 + ], + "lines": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "spans": [ + { + "bbox": [ + 106, + 176, + 505, + 188 + ], + "score": 1.0, + "content": "Each of these datasets is split into a training, validation and testing subsets of size 25, 000, 25, 000", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "and 10, 000 respectively. For both these datasets, we use standard data pre-processing and augmen-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 210 + ], + "score": 1.0, + "content": "tation techniques, i.e. subtracting the channel mean and dividing by the channel standard deviation;", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 208, + 505, + 221 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 271, + 221 + ], + "score": 1.0, + "content": "centrally padding the training images to", + "type": "text" + }, + { + "bbox": [ + 271, + 209, + 300, + 219 + ], + "score": 0.88, + "content": "4 0 \\times 4 0", + "type": "inline_equation" + }, + { + "bbox": [ + 300, + 208, + 454, + 221 + ], + "score": 1.0, + "content": "and randomly cropping them back to", + "type": "text" + }, + { + "bbox": [ + 454, + 209, + 483, + 219 + ], + "score": 0.88, + "content": "3 2 \\times 3 2", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 208, + 505, + 221 + ], + "score": 1.0, + "content": "; and", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 220, + 257, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 257, + 232 + ], + "score": 1.0, + "content": "randomly clipping them horizontally.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 176, + 505, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 236, + 505, + 270 + ], + "lines": [ + { + "bbox": [ + 105, + 235, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 506, + 250 + ], + "score": 1.0, + "content": "SPORT8. This is an action recognition dataset containing 8 sport event categories and a total of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 505, + 260 + ], + "score": 1.0, + "content": "1579 images (Li & Fei-Fei, 2007). The tiny size of this dataset stresses the generalization capabilities", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 259, + 241, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 241, + 271 + ], + "score": 1.0, + "content": "of any NAS method applied to it.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 235, + 506, + 271 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 275, + 504, + 298 + ], + "lines": [ + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 274, + 505, + 288 + ], + "score": 1.0, + "content": "MIT67. This is a dataset of 67 classes representing different indoor scenes and consists of 15, 620", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 286, + 324, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 324, + 298 + ], + "score": 1.0, + "content": "images of different sizes (Quattoni & Torralba, 2009).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 274, + 505, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 303, + 502, + 326 + ], + "lines": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "FLOWERS102. This is a dataset of 102 classes representing different species of flowers and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 314, + 405, + 326 + ], + "spans": [ + { + "bbox": [ + 106, + 314, + 405, + 326 + ], + "score": 1.0, + "content": "consists of 8, 189 images of different sizes (Nilsback & Zisserman, 2008).", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 302, + 505, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 331, + 505, + 397 + ], + "lines": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 331, + 505, + 344 + ], + "score": 1.0, + "content": "Each of these datasets is split into a training, validation and testing subsets with proportions", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 167, + 354 + ], + "score": 0.64, + "content": "4 0 / 4 0 / 2 0 \\ ( \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 342, + 505, + 355 + ], + "score": 1.0, + "content": ". For each one, we use use standard data pre-processing and augmentation tech-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 366 + ], + "score": 1.0, + "content": "niques, i.e. subtracting the channel mean and dividing the channel standard deviation, cropping the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 366, + 378 + ], + "score": 1.0, + "content": "training images to random size and aspect ratio, resizing them to", + "type": "text" + }, + { + "bbox": [ + 366, + 364, + 405, + 375 + ], + "score": 0.9, + "content": "2 2 4 \\times 2 2 4", + "type": "inline_equation" + }, + { + "bbox": [ + 405, + 363, + 505, + 378 + ], + "score": 1.0, + "content": ", and randomly changing", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 374, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 387, + 389 + ], + "score": 1.0, + "content": "their brightness, contrast, and saturation, while resizing test images to", + "type": "text" + }, + { + "bbox": [ + 388, + 375, + 427, + 386 + ], + "score": 0.88, + "content": "2 5 6 \\times 2 5 6", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 374, + 505, + 389 + ], + "score": 1.0, + "content": "and cropping them", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 385, + 160, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 160, + 398 + ], + "score": 1.0, + "content": "at the center.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 331, + 505, + 398 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 411, + 229, + 422 + ], + "lines": [ + { + "bbox": [ + 106, + 410, + 230, + 423 + ], + "spans": [ + { + "bbox": [ + 106, + 410, + 230, + 423 + ], + "score": 1.0, + "content": "A.3 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 165, + 454, + 443, + 564 + ], + "blocks": [ + { + "type": "image_caption", + "bbox": [ + 107, + 431, + 361, + 443 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 431, + 361, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 431, + 361, + 444 + ], + "score": 1.0, + "content": "A.3.1 DIFFERENT TRAINING PROTOCOLS FOR RESNET-50", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "image_body", + "bbox": [ + 165, + 454, + 443, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 165, + 454, + 443, + 564 + ], + "spans": [ + { + "bbox": [ + 165, + 454, + 443, + 564 + ], + "score": 0.969, + "type": "image", + "image_path": "ce230d7557a836be95b1c534a22b98a18804cb6ce095b27f9043206e3ab14ff8.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 165, + 454, + 443, + 490.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 165, + 490.6666666666667, + 443, + 527.3333333333334 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 165, + 527.3333333333334, + 443, + 564.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 579, + 505, + 646 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 580, + 504, + 592 + ], + "spans": [ + { + "bbox": [ + 106, + 580, + 504, + 592 + ], + "score": 1.0, + "content": "Figure 10: Extension of Figure 3, including results obtained by training a ResNet-50 on CIFAR10.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "Bars with darker shade are for ResNet-50 and bars with lighter shade are for DARTS (same as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 506, + 615 + ], + "score": 1.0, + "content": "Figure 3). Result are for 8 runs of each training protocol. For ResNet-50, the auxiliary tower", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "was added after layer 2. As DropPath (Larsson & Shakhnarovich, 2017) would not have been", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 506, + 637 + ], + "score": 1.0, + "content": "straightforward to apply, we instead implemented Stocastic Depth (Huang et al., 2016), to a similar", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 634, + 134, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 134, + 647 + ], + "score": 1.0, + "content": "effect.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + } + ], + "index": 27 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/KLH36ELmwIB/KLH36ELmwIB_middle.json b/parse/train/KLH36ELmwIB/KLH36ELmwIB_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..bd6741622440caeae8c0b5c510b86e2196c5ed2b --- /dev/null +++ b/parse/train/KLH36ELmwIB/KLH36ELmwIB_middle.json @@ -0,0 +1,47000 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "DARTS-: ROBUSTLY STEPPING OUT OF PERFOR-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 98, + 420, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 98, + 420, + 118 + ], + "score": 1.0, + "content": "MANCE COLLAPSE WITHOUT INDICATORS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 133, + 498, + 190 + ], + "lines": [ + { + "bbox": [ + 110, + 134, + 498, + 147 + ], + "spans": [ + { + "bbox": [ + 110, + 134, + 163, + 147 + ], + "score": 1.0, + "content": "Xiangxiang", + "type": "text" + }, + { + "bbox": [ + 164, + 135, + 187, + 146 + ], + "score": 0.73, + "content": "\\mathbf { C h u ^ { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 134, + 348, + 147 + ], + "score": 1.0, + "content": ", Xiaoxing Wang1,2∗, Bo Zhang1, Shun", + "type": "text" + }, + { + "bbox": [ + 348, + 135, + 376, + 146 + ], + "score": 0.71, + "content": "\\mathbf { L u ^ { 1 , 3 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 134, + 471, + 147 + ], + "score": 1.0, + "content": ", Xiaolin Wei1, Junchi", + "type": "text" + }, + { + "bbox": [ + 471, + 134, + 498, + 146 + ], + "score": 0.36, + "content": "\\mathbf { Y a n } ^ { 2 \\dagger }", + "type": "inline_equation" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 144, + 470, + 160 + ], + "spans": [ + { + "bbox": [ + 111, + 144, + 470, + 160 + ], + "score": 1.0, + "content": "1Meituan, 2Shanghai Jiao Tong University, 3University of Chinese Academy of Sciences", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 157, + 413, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 413, + 171 + ], + "score": 1.0, + "content": "{chuxiangxiang,zhangbo97,weixiaolin02}@meituan.com", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 168, + 320, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 320, + 181 + ], + "score": 1.0, + "content": "{figure1 wxx,yanjunchi}@sjtu.edu.cn", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 180, + 264, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 180, + 264, + 191 + ], + "score": 1.0, + "content": "lushun19@mails.ucas.ac.cn", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 278, + 220, + 333, + 232 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 335, + 233 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 335, + 233 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 246, + 468, + 422 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "score": 1.0, + "content": "Despite the fast development of differentiable architecture search (DARTS), it", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 469, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 469, + 271 + ], + "score": 1.0, + "content": "suffers from long-standing performance instability, which extremely limits its ap-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "score": 1.0, + "content": "plication. Existing robustifying methods draw clues from the resulting deterio-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "score": 1.0, + "content": "rated behavior instead of finding out its causing factor. Various indicators such", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 469, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 469, + 303 + ], + "score": 1.0, + "content": "as Hessian eigenvalues are proposed as a signal to stop searching before the per-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "score": 1.0, + "content": "formance collapses. However, these indicator-based methods tend to easily reject", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "score": 1.0, + "content": "good architectures if the thresholds are inappropriately set, let alone the search-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 324, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 470, + 335 + ], + "score": 1.0, + "content": "ing is intrinsically noisy. In this paper, we undertake a more subtle and direct", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 335, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 470, + 346 + ], + "score": 1.0, + "content": "approach to resolve the collapse. We first demonstrate that skip connections have", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 469, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 469, + 358 + ], + "score": 1.0, + "content": "a clear advantage over other candidate operations, where it can easily recover", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "score": 1.0, + "content": "from a disadvantageous state and become dominant. We conjecture that this priv-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "score": 1.0, + "content": "ilege is causing degenerated performance. Therefore, we propose to factor out", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "score": 1.0, + "content": "this benefit with an auxiliary skip connection, ensuring a fairer competition for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "score": 1.0, + "content": "all operations. We call this approach DARTS-. Extensive experiments on various", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 401, + 469, + 412 + ], + "spans": [ + { + "bbox": [ + 142, + 401, + 469, + 412 + ], + "score": 1.0, + "content": "datasets verify that it can substantially improve robustness. Our code is available", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 412, + 397, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 397, + 423 + ], + "score": 1.0, + "content": "at https://github.com/Meituan-AutoML/DARTS-.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 15.5 + }, + { + "type": "title", + "bbox": [ + 108, + 448, + 206, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 208, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 208, + 464 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "Recent studies (Zela et al., 2020; Liang et al., 2019; Chu et al., 2020b) have shown that one critical", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "issue for differentiable architecture search (Liu et al., 2019b) regarding the performance collapse due", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "score": 1.0, + "content": "to superfluous skip connections. Accordingly, some empirical indicators for detecting the occurrence", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "of collapse have been produced. R-DARTS (Zela et al., 2020) shows that the loss landscape has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "more curvatures (characterized by higher Hessian eigenvalues w.r.t. architectural weights) when", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 529, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 543 + ], + "score": 1.0, + "content": "the derived architecture generalizes poorly. By regularizing for a lower Hessian eigenvalue, Zela", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "et al. (2020); Chen & Hsieh (2020) attempt to stabilize the search process. Meanwhile, by directly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "constraining the number of skip connections to a fixed number (typically 2), the collapse issue", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "becomes less pronounced (Chen et al., 2019b; Liang et al., 2019). These indicator-based approaches", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "have several main drawbacks. Firstly, robustness relies heavily on the quality of the indicator. An", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "imprecise indicator either inevitably accepts poor models or mistakenly reject good ones. Secondly,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "indicators impose strong priors by directly manipulating the inferred model, which is somewhat", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 606, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 620 + ], + "score": 1.0, + "content": "suspicious, akin to touching the test set. Thirdly, extra computing cost (Zela et al., 2020) or careful", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "tuning of hyper-parameters (Chen et al., 2019b; Liang et al., 2019) are required. Therefore, it’s", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 628, + 262, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 262, + 641 + ], + "score": 1.0, + "content": "natural to ask the following questions:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 133, + 652, + 504, + 674 + ], + "lines": [ + { + "bbox": [ + 134, + 651, + 504, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 651, + 504, + 664 + ], + "score": 1.0, + "content": "• Can we resolve the collapse without handcrafted indicators and restrictions to interfere with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 662, + 328, + 675 + ], + "spans": [ + { + "bbox": [ + 142, + 662, + 328, + 675 + ], + "score": 1.0, + "content": "the searching and/or discretization procedure?", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 133, + 680, + 485, + 691 + ], + "lines": [ + { + "bbox": [ + 133, + 678, + 487, + 694 + ], + "spans": [ + { + "bbox": [ + 133, + 678, + 487, + 694 + ], + "score": 1.0, + "content": "• Is it possible to achieve robustness in DARTS without tuning extra hyper-parameters?", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 711, + 263, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 264, + 722 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 264, + 722 + ], + "score": 1.0, + "content": "∗Work done as an intern at Meituan Inc.", + "type": "text" + } + ] + }, + { + "bbox": [ + 118, + 720, + 208, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 208, + 734 + ], + "score": 1.0, + "content": "†Correspondent Author.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "DARTS-: ROBUSTLY STEPPING OUT OF PERFOR-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 98, + 420, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 98, + 420, + 118 + ], + "score": 1.0, + "content": "MANCE COLLAPSE WITHOUT INDICATORS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "list", + "bbox": [ + 112, + 133, + 498, + 190 + ], + "lines": [ + { + "bbox": [ + 110, + 134, + 498, + 147 + ], + "spans": [ + { + "bbox": [ + 110, + 134, + 163, + 147 + ], + "score": 1.0, + "content": "Xiangxiang", + "type": "text" + }, + { + "bbox": [ + 164, + 135, + 187, + 146 + ], + "score": 0.73, + "content": "\\mathbf { C h u ^ { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 134, + 348, + 147 + ], + "score": 1.0, + "content": ", Xiaoxing Wang1,2∗, Bo Zhang1, Shun", + "type": "text" + }, + { + "bbox": [ + 348, + 135, + 376, + 146 + ], + "score": 0.71, + "content": "\\mathbf { L u ^ { 1 , 3 * } }", + "type": "inline_equation" + }, + { + "bbox": [ + 376, + 134, + 471, + 147 + ], + "score": 1.0, + "content": ", Xiaolin Wei1, Junchi", + "type": "text" + }, + { + "bbox": [ + 471, + 134, + 498, + 146 + ], + "score": 0.36, + "content": "\\mathbf { Y a n } ^ { 2 \\dagger }", + "type": "inline_equation" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 111, + 144, + 470, + 160 + ], + "spans": [ + { + "bbox": [ + 111, + 144, + 470, + 160 + ], + "score": 1.0, + "content": "1Meituan, 2Shanghai Jiao Tong University, 3University of Chinese Academy of Sciences", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 157, + 413, + 171 + ], + "spans": [ + { + "bbox": [ + 112, + 157, + 413, + 171 + ], + "score": 1.0, + "content": "{chuxiangxiang,zhangbo97,weixiaolin02}@meituan.com", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 168, + 320, + 181 + ], + "spans": [ + { + "bbox": [ + 112, + 168, + 320, + 181 + ], + "score": 1.0, + "content": "{figure1 wxx,yanjunchi}@sjtu.edu.cn", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 112, + 180, + 264, + 191 + ], + "spans": [ + { + "bbox": [ + 112, + 180, + 264, + 191 + ], + "score": 1.0, + "content": "lushun19@mails.ucas.ac.cn", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + } + ], + "index": 4, + "bbox_fs": [ + 110, + 134, + 498, + 191 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 220, + 333, + 232 + ], + "lines": [ + { + "bbox": [ + 276, + 219, + 335, + 233 + ], + "spans": [ + { + "bbox": [ + 276, + 219, + 335, + 233 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 143, + 246, + 468, + 422 + ], + "lines": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 470, + 259 + ], + "score": 1.0, + "content": "Despite the fast development of differentiable architecture search (DARTS), it", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 469, + 271 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 469, + 271 + ], + "score": 1.0, + "content": "suffers from long-standing performance instability, which extremely limits its ap-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 269, + 469, + 281 + ], + "score": 1.0, + "content": "plication. Existing robustifying methods draw clues from the resulting deterio-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "spans": [ + { + "bbox": [ + 141, + 280, + 470, + 292 + ], + "score": 1.0, + "content": "rated behavior instead of finding out its causing factor. Various indicators such", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 469, + 303 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 469, + 303 + ], + "score": 1.0, + "content": "as Hessian eigenvalues are proposed as a signal to stop searching before the per-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 142, + 302, + 470, + 313 + ], + "score": 1.0, + "content": "formance collapses. However, these indicator-based methods tend to easily reject", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 312, + 470, + 325 + ], + "score": 1.0, + "content": "good architectures if the thresholds are inappropriately set, let alone the search-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 324, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 324, + 470, + 335 + ], + "score": 1.0, + "content": "ing is intrinsically noisy. In this paper, we undertake a more subtle and direct", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 335, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 335, + 470, + 346 + ], + "score": 1.0, + "content": "approach to resolve the collapse. We first demonstrate that skip connections have", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 345, + 469, + 358 + ], + "spans": [ + { + "bbox": [ + 141, + 345, + 469, + 358 + ], + "score": 1.0, + "content": "a clear advantage over other candidate operations, where it can easily recover", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 470, + 369 + ], + "score": 1.0, + "content": "from a disadvantageous state and become dominant. We conjecture that this priv-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 469, + 380 + ], + "score": 1.0, + "content": "ilege is causing degenerated performance. Therefore, we propose to factor out", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "spans": [ + { + "bbox": [ + 141, + 378, + 470, + 392 + ], + "score": 1.0, + "content": "this benefit with an auxiliary skip connection, ensuring a fairer competition for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "spans": [ + { + "bbox": [ + 141, + 389, + 470, + 402 + ], + "score": 1.0, + "content": "all operations. We call this approach DARTS-. Extensive experiments on various", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 142, + 401, + 469, + 412 + ], + "spans": [ + { + "bbox": [ + 142, + 401, + 469, + 412 + ], + "score": 1.0, + "content": "datasets verify that it can substantially improve robustness. Our code is available", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 141, + 412, + 397, + 423 + ], + "spans": [ + { + "bbox": [ + 141, + 412, + 397, + 423 + ], + "score": 1.0, + "content": "at https://github.com/Meituan-AutoML/DARTS-.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 15.5, + "bbox_fs": [ + 141, + 246, + 470, + 423 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 448, + 206, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 447, + 208, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 208, + 464 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 107, + 474, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 487 + ], + "score": 1.0, + "content": "Recent studies (Zela et al., 2020; Liang et al., 2019; Chu et al., 2020b) have shown that one critical", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "issue for differentiable architecture search (Liu et al., 2019b) regarding the performance collapse due", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 510 + ], + "score": 1.0, + "content": "to superfluous skip connections. Accordingly, some empirical indicators for detecting the occurrence", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "of collapse have been produced. R-DARTS (Zela et al., 2020) shows that the loss landscape has", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "more curvatures (characterized by higher Hessian eigenvalues w.r.t. architectural weights) when", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 529, + 506, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 506, + 543 + ], + "score": 1.0, + "content": "the derived architecture generalizes poorly. By regularizing for a lower Hessian eigenvalue, Zela", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 540, + 505, + 554 + ], + "score": 1.0, + "content": "et al. (2020); Chen & Hsieh (2020) attempt to stabilize the search process. Meanwhile, by directly", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 104, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 104, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "constraining the number of skip connections to a fixed number (typically 2), the collapse issue", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "becomes less pronounced (Chen et al., 2019b; Liang et al., 2019). These indicator-based approaches", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 586 + ], + "score": 1.0, + "content": "have several main drawbacks. Firstly, robustness relies heavily on the quality of the indicator. An", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "imprecise indicator either inevitably accepts poor models or mistakenly reject good ones. Secondly,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "indicators impose strong priors by directly manipulating the inferred model, which is somewhat", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 606, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 505, + 620 + ], + "score": 1.0, + "content": "suspicious, akin to touching the test set. Thirdly, extra computing cost (Zela et al., 2020) or careful", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 506, + 630 + ], + "score": 1.0, + "content": "tuning of hyper-parameters (Chen et al., 2019b; Liang et al., 2019) are required. Therefore, it’s", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 628, + 262, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 262, + 641 + ], + "score": 1.0, + "content": "natural to ask the following questions:", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 32, + "bbox_fs": [ + 104, + 475, + 506, + 641 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 652, + 504, + 674 + ], + "lines": [ + { + "bbox": [ + 134, + 651, + 504, + 664 + ], + "spans": [ + { + "bbox": [ + 134, + 651, + 504, + 664 + ], + "score": 1.0, + "content": "• Can we resolve the collapse without handcrafted indicators and restrictions to interfere with", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 142, + 662, + 328, + 675 + ], + "spans": [ + { + "bbox": [ + 142, + 662, + 328, + 675 + ], + "score": 1.0, + "content": "the searching and/or discretization procedure?", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 134, + 651, + 504, + 675 + ] + }, + { + "type": "text", + "bbox": [ + 133, + 680, + 485, + 691 + ], + "lines": [ + { + "bbox": [ + 133, + 678, + 487, + 694 + ], + "spans": [ + { + "bbox": [ + 133, + 678, + 487, + 694 + ], + "score": 1.0, + "content": "• Is it possible to achieve robustness in DARTS without tuning extra hyper-parameters?", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 42, + "bbox_fs": [ + 133, + 678, + 487, + 694 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 296, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 297, + 95 + ], + "score": 1.0, + "content": "To answer the above questions, we propose", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 297, + 105 + ], + "score": 1.0, + "content": "an effective and efficient approach to stabilize", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 297, + 115 + ], + "score": 1.0, + "content": "DARTS. Our contributions can be summarized", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 153, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 153, + 127 + ], + "score": 1.0, + "content": "as follows:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 296, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 297, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 297, + 143 + ], + "score": 1.0, + "content": "New Paradigm to Stabilize DARTS. While", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 298, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 298, + 155 + ], + "score": 1.0, + "content": "empirically observing that current indicators", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 297, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 297, + 165 + ], + "score": 1.0, + "content": "(Zela et al., 2020; Chen & Hsieh, 2020) can", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 297, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 297, + 177 + ], + "score": 1.0, + "content": "avoid performance collapse at a cost of reduced", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 297, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 297, + 189 + ], + "score": 1.0, + "content": "exploration coverage in the search space, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 297, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 297, + 199 + ], + "score": 1.0, + "content": "propose a novel indicator-free approach to sta-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 297, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 297, + 209 + ], + "score": 1.0, + "content": "bilize DARTS, referred to as DARTS-1, which", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 297, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 297, + 222 + ], + "score": 1.0, + "content": "involves an auxiliary skip connection (see Fig-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 220, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 297, + 232 + ], + "score": 1.0, + "content": "ure 1) to remove the unfair advantage (Chu", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 255, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 255, + 244 + ], + "score": 1.0, + "content": "et al., 2020b) in the searching phase.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 248, + 297, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 297, + 259 + ], + "score": 1.0, + "content": "Strong Robustness and Stabilization. We", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 298, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 298, + 271 + ], + "score": 1.0, + "content": "conduct thorough experiments across seven", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 270, + 298, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 298, + 281 + ], + "score": 1.0, + "content": "search spaces and three datasets to demonstrate", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 86, + 502, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 307, + 86, + 502, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 86, + 502, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 86, + 502, + 189 + ], + "score": 0.971, + "type": "image", + "image_path": "2ca1f536bf1e323954a28c4fbbd59ffd44b2036412e387cc0ccb940f7be24143.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 307, + 86, + 502, + 100.71428571428571 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 307, + 100.71428571428571, + 502, + 115.42857142857142 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 307, + 115.42857142857142, + 502, + 130.14285714285714 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 307, + 130.14285714285714, + 502, + 144.85714285714286 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 307, + 144.85714285714286, + 502, + 159.57142857142858 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 307, + 159.57142857142858, + 502, + 174.2857142857143 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 307, + 174.2857142857143, + 502, + 189.00000000000003 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 201, + 505, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "Figure 1: Schematic illustration about (a) DARTS", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "and (b) the proposed DARTS-, featuring an auxil-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 303, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 303, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "iary skip connection (thick red line) with a decay", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 303, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 303, + 234, + 321, + 246 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 322, + 235, + 329, + 245 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "between every two nodes to remove the po-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 303, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 303, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "tential unfair advantage that leads to performance", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 303, + 255, + 342, + 269 + ], + "spans": [ + { + "bbox": [ + 303, + 255, + 342, + 269 + ], + "score": 1.0, + "content": "collapse.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + } + ], + "index": 23.25 + }, + { + "type": "text", + "bbox": [ + 108, + 281, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "the effectiveness of our method. Specifically, our approach robustly obtains state-of-the-art results", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 198, + 304 + ], + "score": 1.0, + "content": "on 4 search space with", + "type": "text" + }, + { + "bbox": [ + 199, + 292, + 213, + 302 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "fewer search cost than R-DARTS (Zela et al., 2020), which requires four", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 302, + 304, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 304, + 315 + ], + "score": 1.0, + "content": "independent runs to report the final performance.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "Seamless Plug-in Combination with DARTS Variants. We conduct experiments to demonstrate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "that our approach can work together seamlessly with other orthogonal DARTS variants by removing", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "their handcrafted indicators without extra overhead. In particular, our approach is able to improve", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 352, + 464, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 129, + 363 + ], + "score": 0.86, + "content": "0 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 353, + 244, + 365 + ], + "score": 1.0, + "content": "accuracy for P-DARTS, and", + "type": "text" + }, + { + "bbox": [ + 244, + 352, + 272, + 363 + ], + "score": 0.87, + "content": "0 . 2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 353, + 464, + 365 + ], + "score": 1.0, + "content": "accuracy for PC-DARTS on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "title", + "bbox": [ + 108, + 381, + 211, + 394 + ], + "lines": [ + { + "bbox": [ + 104, + 380, + 213, + 397 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 213, + 397 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "Neural architecture search and DARTS variants. Over the years, researchers have sought to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "automatically discover neural architectures for various deep learning tasks to relieve human from", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "the tedious effort, ranging from image classification (Zoph et al., 2018), objection detection (Ghiasi", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 441, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 504, + 453 + ], + "score": 1.0, + "content": "et al., 2019), image segmentation (Liu et al., 2019a) to machine translation (So et al., 2019) etc.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "Among many proposed approaches, Differentiable Architecture Search (Liu et al., 2019b) features", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "weight-sharing and resolves the searching problem via gradient descent, which is very efficient and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "easy to generalize. A short description of DARTS can be found in A.1. Since then, many subsequent", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "works have been dedicated to accelerating the process (Dong & Yang, 2019b), reducing memory", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 496, + 504, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 504, + 508 + ], + "score": 1.0, + "content": "cost (Xu et al., 2020), or fostering its ability such as hardware-awareness (Cai et al., 2019; Wu et al.,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "2019), finer granularity (Mei et al., 2020) and so on. However, regardless of these endeavors, a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "fundamental issue of DARTS over its searching performance collapse remains not properly solved,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 529, + 274, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 274, + 542 + ], + "score": 1.0, + "content": "which extremely prohibits its application.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43.5 + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 505, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "Robustifying DARTS. As DARTS (Liu et al., 2019b) is known to be unstable as a result of per-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "formance collapse (Chu et al., 2020b), some recent works have devoted to resolving it by either", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "designing indicators like Hessian eigenvalues for the collapse (Zela et al., 2020) or adding pertur-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 579, + 504, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 504, + 590 + ], + "score": 1.0, + "content": "bations to regularize such an indicator (Chen & Hsieh, 2020). Both methods rely heavily on the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "indicator’s accurateness, i.e., to what extent does the indicator correlate with the performance col-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 408, + 613 + ], + "score": 1.0, + "content": "lapse? Other methods like Progressive DARTS (Chen et al., 2019b) and", + "type": "text" + }, + { + "bbox": [ + 408, + 601, + 448, + 612 + ], + "score": 0.47, + "content": "{ \\mathrm { D A R T S } } +", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "(Liang et al.,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "2019) employ a strong human prior, i.e., limiting the number of skip-connections to be a fixed value.", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Fair DARTS (Chu et al., 2020b) argues that the collapse results from the unfair advantage in an ex-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "clusive competitive environment, from which skip connections overly benefit to cause an abundant", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "aggregation. To suppress such an advantage from overshooting, they convert the competition into", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "collaboration where each operation is independent of others. It is however an indirect approach.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "SGAS (Li et al., 2020), instead, circumvents the problem with a greedy strategy where the unfair", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "advantage can be prevented from taking effect. Nevertheless, potentially good operations might be", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 688, + 333, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 333, + 700 + ], + "score": 1.0, + "content": "pruned out too early because of greedy underestimation.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 56.5 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 712, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 710, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 119, + 710, + 506, + 723 + ], + "score": 1.0, + "content": "1We name it so as we undertake an inward way, as opposed to those outward ones who design new indicators,", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 721, + 294, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 294, + 732 + ], + "score": 1.0, + "content": "add extra cost and introduce new hyper-parameters.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 83, + 296, + 126 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 297, + 95 + ], + "score": 1.0, + "content": "To answer the above questions, we propose", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 297, + 105 + ], + "score": 1.0, + "content": "an effective and efficient approach to stabilize", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 297, + 115 + ], + "score": 1.0, + "content": "DARTS. Our contributions can be summarized", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 153, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 153, + 127 + ], + "score": 1.0, + "content": "as follows:", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 82, + 297, + 127 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 132, + 296, + 242 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 297, + 143 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 297, + 143 + ], + "score": 1.0, + "content": "New Paradigm to Stabilize DARTS. While", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 144, + 298, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 144, + 298, + 155 + ], + "score": 1.0, + "content": "empirically observing that current indicators", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 154, + 297, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 297, + 165 + ], + "score": 1.0, + "content": "(Zela et al., 2020; Chen & Hsieh, 2020) can", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 297, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 297, + 177 + ], + "score": 1.0, + "content": "avoid performance collapse at a cost of reduced", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 297, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 297, + 189 + ], + "score": 1.0, + "content": "exploration coverage in the search space, we", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 187, + 297, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 297, + 199 + ], + "score": 1.0, + "content": "propose a novel indicator-free approach to sta-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 297, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 297, + 209 + ], + "score": 1.0, + "content": "bilize DARTS, referred to as DARTS-1, which", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 208, + 297, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 208, + 297, + 222 + ], + "score": 1.0, + "content": "involves an auxiliary skip connection (see Fig-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 220, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 297, + 232 + ], + "score": 1.0, + "content": "ure 1) to remove the unfair advantage (Chu", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 230, + 255, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 255, + 244 + ], + "score": 1.0, + "content": "et al., 2020b) in the searching phase.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 132, + 298, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 248, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 248, + 297, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 297, + 259 + ], + "score": 1.0, + "content": "Strong Robustness and Stabilization. We", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 259, + 298, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 298, + 271 + ], + "score": 1.0, + "content": "conduct thorough experiments across seven", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 270, + 298, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 298, + 281 + ], + "score": 1.0, + "content": "search spaces and three datasets to demonstrate", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 106, + 248, + 298, + 281 + ] + }, + { + "type": "image", + "bbox": [ + 307, + 86, + 502, + 189 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 307, + 86, + 502, + 189 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 307, + 86, + 502, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 86, + 502, + 189 + ], + "score": 0.971, + "type": "image", + "image_path": "2ca1f536bf1e323954a28c4fbbd59ffd44b2036412e387cc0ccb940f7be24143.jpg" + } + ] + } + ], + "index": 20, + "virtual_lines": [ + { + "bbox": [ + 307, + 86, + 502, + 100.71428571428571 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 307, + 100.71428571428571, + 502, + 115.42857142857142 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 307, + 115.42857142857142, + 502, + 130.14285714285714 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 307, + 130.14285714285714, + 502, + 144.85714285714286 + ], + "spans": [], + "index": 20 + }, + { + "bbox": [ + 307, + 144.85714285714286, + 502, + 159.57142857142858 + ], + "spans": [], + "index": 21 + }, + { + "bbox": [ + 307, + 159.57142857142858, + 502, + 174.2857142857143 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 307, + 174.2857142857143, + 502, + 189.00000000000003 + ], + "spans": [], + "index": 23 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 304, + 201, + 505, + 267 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 505, + 213 + ], + "score": 1.0, + "content": "Figure 1: Schematic illustration about (a) DARTS", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 212, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 505, + 225 + ], + "score": 1.0, + "content": "and (b) the proposed DARTS-, featuring an auxil-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 303, + 223, + 505, + 236 + ], + "spans": [ + { + "bbox": [ + 303, + 223, + 505, + 236 + ], + "score": 1.0, + "content": "iary skip connection (thick red line) with a decay", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 303, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 303, + 234, + 321, + 246 + ], + "score": 1.0, + "content": "rate", + "type": "text" + }, + { + "bbox": [ + 322, + 235, + 329, + 245 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "between every two nodes to remove the po-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 303, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 303, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "tential unfair advantage that leads to performance", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 303, + 255, + 342, + 269 + ], + "spans": [ + { + "bbox": [ + 303, + 255, + 342, + 269 + ], + "score": 1.0, + "content": "collapse.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 26.5 + } + ], + "index": 23.25 + }, + { + "type": "text", + "bbox": [ + 108, + 281, + 504, + 314 + ], + "lines": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "spans": [ + { + "bbox": [ + 105, + 280, + 505, + 294 + ], + "score": 1.0, + "content": "the effectiveness of our method. Specifically, our approach robustly obtains state-of-the-art results", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 291, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 198, + 304 + ], + "score": 1.0, + "content": "on 4 search space with", + "type": "text" + }, + { + "bbox": [ + 199, + 292, + 213, + 302 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 291, + 505, + 304 + ], + "score": 1.0, + "content": "fewer search cost than R-DARTS (Zela et al., 2020), which requires four", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 302, + 304, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 302, + 304, + 315 + ], + "score": 1.0, + "content": "independent runs to report the final performance.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 280, + 505, + 315 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 320, + 505, + 363 + ], + "lines": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "Seamless Plug-in Combination with DARTS Variants. We conduct experiments to demonstrate", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 505, + 344 + ], + "score": 1.0, + "content": "that our approach can work together seamlessly with other orthogonal DARTS variants by removing", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 342, + 505, + 354 + ], + "score": 1.0, + "content": "their handcrafted indicators without extra overhead. In particular, our approach is able to improve", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 352, + 464, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 129, + 363 + ], + "score": 0.86, + "content": "0 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 353, + 244, + 365 + ], + "score": 1.0, + "content": "accuracy for P-DARTS, and", + "type": "text" + }, + { + "bbox": [ + 244, + 352, + 272, + 363 + ], + "score": 0.87, + "content": "0 . 2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 353, + 464, + 365 + ], + "score": 1.0, + "content": "accuracy for PC-DARTS on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 319, + 505, + 365 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 381, + 211, + 394 + ], + "lines": [ + { + "bbox": [ + 104, + 380, + 213, + 397 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 213, + 397 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 408, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 421 + ], + "score": 1.0, + "content": "Neural architecture search and DARTS variants. Over the years, researchers have sought to", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 505, + 432 + ], + "score": 1.0, + "content": "automatically discover neural architectures for various deep learning tasks to relieve human from", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 505, + 443 + ], + "score": 1.0, + "content": "the tedious effort, ranging from image classification (Zoph et al., 2018), objection detection (Ghiasi", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 441, + 504, + 453 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 504, + 453 + ], + "score": 1.0, + "content": "et al., 2019), image segmentation (Liu et al., 2019a) to machine translation (So et al., 2019) etc.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 505, + 465 + ], + "score": 1.0, + "content": "Among many proposed approaches, Differentiable Architecture Search (Liu et al., 2019b) features", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "weight-sharing and resolves the searching problem via gradient descent, which is very efficient and", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 487 + ], + "score": 1.0, + "content": "easy to generalize. A short description of DARTS can be found in A.1. Since then, many subsequent", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "works have been dedicated to accelerating the process (Dong & Yang, 2019b), reducing memory", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 496, + 504, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 504, + 508 + ], + "score": 1.0, + "content": "cost (Xu et al., 2020), or fostering its ability such as hardware-awareness (Cai et al., 2019; Wu et al.,", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "2019), finer granularity (Mei et al., 2020) and so on. However, regardless of these endeavors, a", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 104, + 517, + 506, + 531 + ], + "spans": [ + { + "bbox": [ + 104, + 517, + 506, + 531 + ], + "score": 1.0, + "content": "fundamental issue of DARTS over its searching performance collapse remains not properly solved,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 529, + 274, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 274, + 542 + ], + "score": 1.0, + "content": "which extremely prohibits its application.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 408, + 506, + 542 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 546, + 505, + 699 + ], + "lines": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 505, + 559 + ], + "score": 1.0, + "content": "Robustifying DARTS. As DARTS (Liu et al., 2019b) is known to be unstable as a result of per-", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 505, + 569 + ], + "score": 1.0, + "content": "formance collapse (Chu et al., 2020b), some recent works have devoted to resolving it by either", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "designing indicators like Hessian eigenvalues for the collapse (Zela et al., 2020) or adding pertur-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 579, + 504, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 504, + 590 + ], + "score": 1.0, + "content": "bations to regularize such an indicator (Chen & Hsieh, 2020). Both methods rely heavily on the", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 505, + 602 + ], + "score": 1.0, + "content": "indicator’s accurateness, i.e., to what extent does the indicator correlate with the performance col-", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 408, + 613 + ], + "score": 1.0, + "content": "lapse? Other methods like Progressive DARTS (Chen et al., 2019b) and", + "type": "text" + }, + { + "bbox": [ + 408, + 601, + 448, + 612 + ], + "score": 0.47, + "content": "{ \\mathrm { D A R T S } } +", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "(Liang et al.,", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 506, + 624 + ], + "score": 1.0, + "content": "2019) employ a strong human prior, i.e., limiting the number of skip-connections to be a fixed value.", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 636 + ], + "score": 1.0, + "content": "Fair DARTS (Chu et al., 2020b) argues that the collapse results from the unfair advantage in an ex-", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 646 + ], + "score": 1.0, + "content": "clusive competitive environment, from which skip connections overly benefit to cause an abundant", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 657 + ], + "score": 1.0, + "content": "aggregation. To suppress such an advantage from overshooting, they convert the competition into", + "type": "text" + } + ], + "index": 59 + }, + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "collaboration where each operation is independent of others. It is however an indirect approach.", + "type": "text" + } + ], + "index": 60 + }, + { + "bbox": [ + 106, + 666, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 679 + ], + "score": 1.0, + "content": "SGAS (Li et al., 2020), instead, circumvents the problem with a greedy strategy where the unfair", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "advantage can be prevented from taking effect. Nevertheless, potentially good operations might be", + "type": "text" + } + ], + "index": 62 + }, + { + "bbox": [ + 105, + 688, + 333, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 333, + 700 + ], + "score": 1.0, + "content": "pruned out too early because of greedy underestimation.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 56.5, + "bbox_fs": [ + 105, + 545, + 506, + 700 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 171, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 173, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 173, + 97 + ], + "score": 1.0, + "content": "3 DARTS-", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 108, + 187, + 119 + ], + "lines": [ + { + "bbox": [ + 105, + 107, + 189, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 107, + 189, + 121 + ], + "score": 1.0, + "content": "3.1 MOTIVATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "score": 1.0, + "content": "We start from the detailed analysis of the role of skip connections. Skip connections were proposed", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 139, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 104, + 139, + 506, + 154 + ], + "score": 1.0, + "content": "to construct a residual block in ResNet (He et al., 2016), which significantly improves training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "stability. It is even possible to deepen the network up to hundreds of layers without accuracy degra-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 163, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 506, + 175 + ], + "score": 1.0, + "content": "dation by simply stacking them up. In contrast, stacking the plain blocks of VGG has degenerated", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "performance when the network gets deeper. Besides, Ren et al. (2015); Wei et al. (2017); Tai et al.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "score": 1.0, + "content": "(2017); Li et al. (2018b) also empirically demonstrate that deep residual network can achieve better", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 196, + 228, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 228, + 208 + ], + "score": 1.0, + "content": "performance on various tasks.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "score": 1.0, + "content": "From the gradient flow’s perspective, skip connection is able to alleviate the gradient vanishing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 221, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 218, + 237 + ], + "score": 1.0, + "content": "problem. Given a stack of", + "type": "text" + }, + { + "bbox": [ + 218, + 226, + 225, + 234 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 221, + 365, + 237 + ], + "score": 1.0, + "content": "residual blocks, the output of the", + "type": "text" + }, + { + "bbox": [ + 366, + 223, + 402, + 235 + ], + "score": 0.93, + "content": "( i + 1 ) ^ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 221, + 464, + 237 + ], + "score": 1.0, + "content": "residual block", + "type": "text" + }, + { + "bbox": [ + 464, + 224, + 487, + 235 + ], + "score": 0.89, + "content": "X _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 221, + 506, + 237 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 233, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 104, + 233, + 172, + 247 + ], + "score": 1.0, + "content": "be computed as", + "type": "text" + }, + { + "bbox": [ + 172, + 235, + 298, + 246 + ], + "score": 0.9, + "content": "X _ { i + 1 } = f _ { i + 1 } ( X _ { i } , W _ { i + 1 } ) + X _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 233, + 329, + 247 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 329, + 235, + 348, + 246 + ], + "score": 0.91, + "content": "f _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 233, + 468, + 247 + ], + "score": 1.0, + "content": "denotes the operations of the", + "type": "text" + }, + { + "bbox": [ + 469, + 235, + 504, + 246 + ], + "score": 0.9, + "content": "( i + 1 ) ^ { \\mathrm { t h } }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 219, + 258 + ], + "score": 1.0, + "content": "residual block with weights", + "type": "text" + }, + { + "bbox": [ + 219, + 246, + 243, + 257 + ], + "score": 0.91, + "content": "W _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 244, + 415, + 258 + ], + "score": 1.0, + "content": ". Suppose the loss function of the model is", + "type": "text" + }, + { + "bbox": [ + 416, + 246, + 423, + 255 + ], + "score": 0.81, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 244, + 506, + 258 + ], + "score": 1.0, + "content": ", and the gradient of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 256, + 412, + 268 + ], + "spans": [ + { + "bbox": [ + 107, + 257, + 119, + 267 + ], + "score": 0.88, + "content": "X _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 256, + 412, + 268 + ], + "score": 1.0, + "content": "can be obtained as follows (1 denotes a tensor whose items are all ones):", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 276, + 441, + 311 + ], + "lines": [ + { + "bbox": [ + 169, + 276, + 441, + 311 + ], + "spans": [ + { + "bbox": [ + 169, + 276, + 441, + 311 + ], + "score": 0.94, + "content": "\\frac { \\partial \\mathcal { L } } { \\partial X _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + 1 } } \\cdot \\left( \\frac { \\partial f _ { i + 1 } } { \\partial X _ { i } } + 1 \\right) = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + j } } \\cdot \\prod _ { k = 1 } ^ { j } \\left( \\frac { \\partial f _ { i + k } } { \\partial X _ { i + k - 1 } } + 1 \\right)", + "type": "interline_equation", + "image_path": "74adfa405eda9baf95e46e7b3e664a2dc0b1408498bbd1833b67350e1802c276.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 169, + 276, + 441, + 287.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 169, + 287.6666666666667, + 441, + 299.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 169, + 299.33333333333337, + 441, + 311.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 506, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "We observe that the gradient of shallow layers always includes the gradient of deep layers, which", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 328, + 342, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 247, + 342 + ], + "score": 1.0, + "content": "mitigates the gradient vanishing of", + "type": "text" + }, + { + "bbox": [ + 248, + 329, + 261, + 340 + ], + "score": 0.89, + "content": "W _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 328, + 342, + 342 + ], + "score": 1.0, + "content": ". Formally we have,", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 348, + 401, + 383 + ], + "lines": [ + { + "bbox": [ + 210, + 348, + 401, + 383 + ], + "spans": [ + { + "bbox": [ + 210, + 348, + 401, + 383 + ], + "score": 0.94, + "content": "\\frac { \\partial \\mathcal { L } } { \\partial W _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + j } } \\cdot \\prod _ { k = 1 } ^ { j } \\left( \\frac { \\partial f _ { i + k } } { \\partial X _ { i + k - 1 } } + \\mathbb { 1 } \\right) \\cdot \\frac { \\partial f _ { i } } { \\partial W _ { i } }", + "type": "interline_equation", + "image_path": "738dab8bbdff37c43c38e6411228e1175c017a4e7203a00b21b8bc414a303981.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 348, + 401, + 365.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 210, + 365.5, + 401, + 383.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 391, + 504, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "To analyze how skip connect affects the performance of residual networks, we introduce a trainable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 151, + 414 + ], + "score": 1.0, + "content": "coefficient", + "type": "text" + }, + { + "bbox": [ + 151, + 402, + 158, + 413 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 402, + 403, + 414 + ], + "score": 1.0, + "content": "on all skip connections in ResNet. Therefore, the gradient of", + "type": "text" + }, + { + "bbox": [ + 404, + 402, + 416, + 413 + ], + "score": 0.89, + "content": "X _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 402, + 481, + 414 + ], + "score": 1.0, + "content": "is converted to:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 421, + 372, + 449 + ], + "lines": [ + { + "bbox": [ + 239, + 421, + 372, + 449 + ], + "spans": [ + { + "bbox": [ + 239, + 421, + 372, + 449 + ], + "score": 0.94, + "content": "\\frac { \\partial \\mathcal { L } } { \\partial X _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + 1 } } \\cdot \\left( \\frac { \\partial f _ { i + 1 } } { \\partial X _ { i } } + \\beta \\right)", + "type": "interline_equation", + "image_path": "50b18fe9a95a84ea4f34978ca622b9bd7ed2b3f6393ee93c31d8b649ce106bac.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 421, + 372, + 435.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 239, + 435.0, + 372, + 449.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 507, + 479 + ], + "lines": [ + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 130, + 469 + ], + "score": 1.0, + "content": "Once", + "type": "text" + }, + { + "bbox": [ + 131, + 457, + 158, + 468 + ], + "score": 0.9, + "content": "\\beta < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 456, + 505, + 469 + ], + "score": 1.0, + "content": ", gradients of deep layers gradually vanish during the back-propagation (BP) towards", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 467, + 498, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 191, + 480 + ], + "score": 1.0, + "content": "shallow layers. Here", + "type": "text" + }, + { + "bbox": [ + 191, + 468, + 199, + 479 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 467, + 498, + 480 + ], + "score": 1.0, + "content": "controls the memory of gradients in BP to stabilize the training procedure.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 336, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 336, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 336, + 496 + ], + "score": 1.0, + "content": "We conduct a confirmatory experiment on ResNet50", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 495, + 336, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 206, + 507 + ], + "score": 1.0, + "content": "and show the result in", + "type": "text" + }, + { + "bbox": [ + 207, + 496, + 231, + 507 + ], + "score": 0.38, + "content": "\\mathrm { F i g } \\ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 495, + 305, + 507 + ], + "score": 1.0, + "content": ". By initializing", + "type": "text" + }, + { + "bbox": [ + 306, + 496, + 313, + 507 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 495, + 336, + 507 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 505, + 338, + 520 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 157, + 518 + ], + "score": 0.92, + "content": "\\{ 0 , 0 . 5 , 1 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 505, + 302, + 520 + ], + "score": 1.0, + "content": ", we can visualize the tendency of", + "type": "text" + }, + { + "bbox": [ + 302, + 507, + 310, + 518 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 505, + 338, + 520 + ], + "score": 1.0, + "content": "along", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 517, + 336, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 269, + 529 + ], + "score": 1.0, + "content": "with training epochs. We observe that", + "type": "text" + }, + { + "bbox": [ + 270, + 518, + 277, + 529 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 517, + 336, + 529 + ], + "score": 1.0, + "content": "converges to-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 337, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 337, + 540 + ], + "score": 1.0, + "content": "wards 1 after 40 epochs no matter the initialization, which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 539, + 335, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 327, + 551 + ], + "score": 1.0, + "content": "demonstrates that the residual structure learns to push", + "type": "text" + }, + { + "bbox": [ + 328, + 540, + 335, + 551 + ], + "score": 0.79, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 548, + 321, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 321, + 564 + ], + "score": 1.0, + "content": "to a rather large value to alleviate gradient vanishing.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 336, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 336, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 336, + 579 + ], + "score": 1.0, + "content": "Similarly, DARTS (Liu et al., 2019b) utilizes a trainable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 578, + 336, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 149, + 590 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 149, + 578, + 171, + 591 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 578, + 336, + 590 + ], + "score": 1.0, + "content": "to denote the importance of skip connec-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 588, + 337, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 255, + 602 + ], + "score": 1.0, + "content": "tion. However, In the search stage,", + "type": "text" + }, + { + "bbox": [ + 255, + 589, + 278, + 601 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 588, + 337, + 602 + ], + "score": 1.0, + "content": "can generally", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 600, + 336, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 336, + 612 + ], + "score": 1.0, + "content": "increase and dominate the architecture parameters, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 611, + 337, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 337, + 623 + ], + "score": 1.0, + "content": "finally leads to performance collapse. We analyze that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 621, + 337, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 137, + 635 + ], + "score": 1.0, + "content": "a large", + "type": "text" + }, + { + "bbox": [ + 138, + 622, + 160, + 634 + ], + "score": 0.92, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 621, + 337, + 635 + ], + "score": 1.0, + "content": "in DARTS could result from two aspects:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 632, + 337, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 337, + 645 + ], + "score": 1.0, + "content": "On the one hand, as the supernet automatically learns to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 642, + 338, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 264, + 657 + ], + "score": 1.0, + "content": "alleviate gradient vanishing, it pushes", + "type": "text" + }, + { + "bbox": [ + 264, + 644, + 287, + 656 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 642, + 338, + 657 + ], + "score": 1.0, + "content": "to a proper", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 337, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 337, + 667 + ], + "score": 1.0, + "content": "large value; On the other hand, the skip connection is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 665, + 336, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 336, + 678 + ], + "score": 1.0, + "content": "indeed an important connection for the target network,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 676, + 337, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 337, + 689 + ], + "score": 1.0, + "content": "which should be selected in the discretization stage. As", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 345, + 488, + 491, + 571 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 345, + 488, + 491, + 571 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 345, + 488, + 491, + 571 + ], + "spans": [ + { + "bbox": [ + 345, + 488, + 491, + 571 + ], + "score": 0.96, + "type": "image", + "image_path": "6f78ee310b963f27f21baf3e4f1fce76a463b12ab1f9ab3ad040d3d42ec3feeb.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 345, + 488, + 491, + 529.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 345, + 529.5, + 491, + 571.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 344, + 578, + 505, + 666 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 344, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 344, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "Figure 2: Tendency of trainable coef-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 343, + 588, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 343, + 588, + 372, + 603 + ], + "score": 1.0, + "content": "ficient", + "type": "text" + }, + { + "bbox": [ + 373, + 590, + 381, + 601 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 588, + 453, + 603 + ], + "score": 1.0, + "content": "(initialized with", + "type": "text" + }, + { + "bbox": [ + 453, + 590, + 502, + 602 + ], + "score": 0.81, + "content": "\\{ 0 , \\ 0 . 5 , \\ 1 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 588, + 505, + 603 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 343, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 343, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "of the skip connection in ResNet50 and", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 343, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 343, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "test accuracy (inset figure) vs. epochs.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 343, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 343, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "The residual structure is proved to learn", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 343, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 343, + 634, + 376, + 646 + ], + "score": 1.0, + "content": "a large", + "type": "text" + }, + { + "bbox": [ + 377, + 634, + 385, + 646 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "to ease training in all three", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 343, + 645, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 343, + 645, + 505, + 656 + ], + "score": 1.0, + "content": "cases. All models are trained and tested", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 343, + 655, + 404, + 667 + ], + "spans": [ + { + "bbox": [ + 343, + 655, + 404, + 667 + ], + "score": 1.0, + "content": "on CIFAR-10.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 49.5 + } + ], + "index": 42.0 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "a consequence, the skip connection in DARTS plays two-fold roles: as an auxiliary connection to", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "stabilize the supernet training, and as a candidate operation to build the final network. Inspired by", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the above observation and analysis, we propose to stabilize the search process by distinguishing the", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 721, + 381, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 381, + 733 + ], + "score": 1.0, + "content": "two roles of skip connection and handling the issue of gradient flow.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 56.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 171, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 173, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 173, + 97 + ], + "score": 1.0, + "content": "3 DARTS-", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 107, + 108, + 187, + 119 + ], + "lines": [ + { + "bbox": [ + 105, + 107, + 189, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 107, + 189, + 121 + ], + "score": 1.0, + "content": "3.1 MOTIVATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 505, + 207 + ], + "lines": [ + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 506, + 142 + ], + "score": 1.0, + "content": "We start from the detailed analysis of the role of skip connections. Skip connections were proposed", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 139, + 506, + 154 + ], + "spans": [ + { + "bbox": [ + 104, + 139, + 506, + 154 + ], + "score": 1.0, + "content": "to construct a residual block in ResNet (He et al., 2016), which significantly improves training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 152, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 505, + 164 + ], + "score": 1.0, + "content": "stability. It is even possible to deepen the network up to hundreds of layers without accuracy degra-", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 163, + 506, + 175 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 506, + 175 + ], + "score": 1.0, + "content": "dation by simply stacking them up. In contrast, stacking the plain blocks of VGG has degenerated", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 505, + 186 + ], + "score": 1.0, + "content": "performance when the network gets deeper. Besides, Ren et al. (2015); Wei et al. (2017); Tai et al.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "spans": [ + { + "bbox": [ + 105, + 184, + 506, + 198 + ], + "score": 1.0, + "content": "(2017); Li et al. (2018b) also empirically demonstrate that deep residual network can achieve better", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 196, + 228, + 208 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 228, + 208 + ], + "score": 1.0, + "content": "performance on various tasks.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 129, + 506, + 208 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 212, + 505, + 268 + ], + "lines": [ + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 211, + 506, + 225 + ], + "score": 1.0, + "content": "From the gradient flow’s perspective, skip connection is able to alleviate the gradient vanishing", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 221, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 218, + 237 + ], + "score": 1.0, + "content": "problem. Given a stack of", + "type": "text" + }, + { + "bbox": [ + 218, + 226, + 225, + 234 + ], + "score": 0.74, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 221, + 365, + 237 + ], + "score": 1.0, + "content": "residual blocks, the output of the", + "type": "text" + }, + { + "bbox": [ + 366, + 223, + 402, + 235 + ], + "score": 0.93, + "content": "( i + 1 ) ^ { \\mathrm { t h } }", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 221, + 464, + 237 + ], + "score": 1.0, + "content": "residual block", + "type": "text" + }, + { + "bbox": [ + 464, + 224, + 487, + 235 + ], + "score": 0.89, + "content": "X _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 221, + 506, + 237 + ], + "score": 1.0, + "content": "can", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 233, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 104, + 233, + 172, + 247 + ], + "score": 1.0, + "content": "be computed as", + "type": "text" + }, + { + "bbox": [ + 172, + 235, + 298, + 246 + ], + "score": 0.9, + "content": "X _ { i + 1 } = f _ { i + 1 } ( X _ { i } , W _ { i + 1 } ) + X _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 233, + 329, + 247 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 329, + 235, + 348, + 246 + ], + "score": 0.91, + "content": "f _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 348, + 233, + 468, + 247 + ], + "score": 1.0, + "content": "denotes the operations of the", + "type": "text" + }, + { + "bbox": [ + 469, + 235, + 504, + 246 + ], + "score": 0.9, + "content": "( i + 1 ) ^ { \\mathrm { t h } }", + "type": "inline_equation" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 219, + 258 + ], + "score": 1.0, + "content": "residual block with weights", + "type": "text" + }, + { + "bbox": [ + 219, + 246, + 243, + 257 + ], + "score": 0.91, + "content": "W _ { i + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 244, + 415, + 258 + ], + "score": 1.0, + "content": ". Suppose the loss function of the model is", + "type": "text" + }, + { + "bbox": [ + 416, + 246, + 423, + 255 + ], + "score": 0.81, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 244, + 506, + 258 + ], + "score": 1.0, + "content": ", and the gradient of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 256, + 412, + 268 + ], + "spans": [ + { + "bbox": [ + 107, + 257, + 119, + 267 + ], + "score": 0.88, + "content": "X _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 119, + 256, + 412, + 268 + ], + "score": 1.0, + "content": "can be obtained as follows (1 denotes a tensor whose items are all ones):", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 104, + 211, + 506, + 268 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 276, + 441, + 311 + ], + "lines": [ + { + "bbox": [ + 169, + 276, + 441, + 311 + ], + "spans": [ + { + "bbox": [ + 169, + 276, + 441, + 311 + ], + "score": 0.94, + "content": "\\frac { \\partial \\mathcal { L } } { \\partial X _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + 1 } } \\cdot \\left( \\frac { \\partial f _ { i + 1 } } { \\partial X _ { i } } + 1 \\right) = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + j } } \\cdot \\prod _ { k = 1 } ^ { j } \\left( \\frac { \\partial f _ { i + k } } { \\partial X _ { i + k - 1 } } + 1 \\right)", + "type": "interline_equation", + "image_path": "74adfa405eda9baf95e46e7b3e664a2dc0b1408498bbd1833b67350e1802c276.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 169, + 276, + 441, + 287.6666666666667 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 169, + 287.6666666666667, + 441, + 299.33333333333337 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 169, + 299.33333333333337, + 441, + 311.00000000000006 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 506, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 106, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "We observe that the gradient of shallow layers always includes the gradient of deep layers, which", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 328, + 342, + 342 + ], + "spans": [ + { + "bbox": [ + 106, + 328, + 247, + 342 + ], + "score": 1.0, + "content": "mitigates the gradient vanishing of", + "type": "text" + }, + { + "bbox": [ + 248, + 329, + 261, + 340 + ], + "score": 0.89, + "content": "W _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 328, + 342, + 342 + ], + "score": 1.0, + "content": ". Formally we have,", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 317, + 505, + 342 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 210, + 348, + 401, + 383 + ], + "lines": [ + { + "bbox": [ + 210, + 348, + 401, + 383 + ], + "spans": [ + { + "bbox": [ + 210, + 348, + 401, + 383 + ], + "score": 0.94, + "content": "\\frac { \\partial \\mathcal { L } } { \\partial W _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + j } } \\cdot \\prod _ { k = 1 } ^ { j } \\left( \\frac { \\partial f _ { i + k } } { \\partial X _ { i + k - 1 } } + \\mathbb { 1 } \\right) \\cdot \\frac { \\partial f _ { i } } { \\partial W _ { i } }", + "type": "interline_equation", + "image_path": "738dab8bbdff37c43c38e6411228e1175c017a4e7203a00b21b8bc414a303981.jpg" + } + ] + } + ], + "index": 19.5, + "virtual_lines": [ + { + "bbox": [ + 210, + 348, + 401, + 365.5 + ], + "spans": [], + "index": 19 + }, + { + "bbox": [ + 210, + 365.5, + 401, + 383.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "text", + "bbox": [ + 105, + 391, + 504, + 414 + ], + "lines": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 106, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "To analyze how skip connect affects the performance of residual networks, we introduce a trainable", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 481, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 151, + 414 + ], + "score": 1.0, + "content": "coefficient", + "type": "text" + }, + { + "bbox": [ + 151, + 402, + 158, + 413 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 402, + 403, + 414 + ], + "score": 1.0, + "content": "on all skip connections in ResNet. Therefore, the gradient of", + "type": "text" + }, + { + "bbox": [ + 404, + 402, + 416, + 413 + ], + "score": 0.89, + "content": "X _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 402, + 481, + 414 + ], + "score": 1.0, + "content": "is converted to:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 391, + 505, + 414 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 421, + 372, + 449 + ], + "lines": [ + { + "bbox": [ + 239, + 421, + 372, + 449 + ], + "spans": [ + { + "bbox": [ + 239, + 421, + 372, + 449 + ], + "score": 0.94, + "content": "\\frac { \\partial \\mathcal { L } } { \\partial X _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + 1 } } \\cdot \\left( \\frac { \\partial f _ { i + 1 } } { \\partial X _ { i } } + \\beta \\right)", + "type": "interline_equation", + "image_path": "50b18fe9a95a84ea4f34978ca622b9bd7ed2b3f6393ee93c31d8b649ce106bac.jpg" + } + ] + } + ], + "index": 23.5, + "virtual_lines": [ + { + "bbox": [ + 239, + 421, + 372, + 435.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 239, + 435.0, + 372, + 449.0 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 456, + 507, + 479 + ], + "lines": [ + { + "bbox": [ + 106, + 456, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 130, + 469 + ], + "score": 1.0, + "content": "Once", + "type": "text" + }, + { + "bbox": [ + 131, + 457, + 158, + 468 + ], + "score": 0.9, + "content": "\\beta < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 456, + 505, + 469 + ], + "score": 1.0, + "content": ", gradients of deep layers gradually vanish during the back-propagation (BP) towards", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 467, + 498, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 467, + 191, + 480 + ], + "score": 1.0, + "content": "shallow layers. Here", + "type": "text" + }, + { + "bbox": [ + 191, + 468, + 199, + 479 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 467, + 498, + 480 + ], + "score": 1.0, + "content": "controls the memory of gradients in BP to stabilize the training procedure.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 456, + 505, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 484, + 336, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 336, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 336, + 496 + ], + "score": 1.0, + "content": "We conduct a confirmatory experiment on ResNet50", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 495, + 336, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 206, + 507 + ], + "score": 1.0, + "content": "and show the result in", + "type": "text" + }, + { + "bbox": [ + 207, + 496, + 231, + 507 + ], + "score": 0.38, + "content": "\\mathrm { F i g } \\ 2", + "type": "inline_equation" + }, + { + "bbox": [ + 231, + 495, + 305, + 507 + ], + "score": 1.0, + "content": ". By initializing", + "type": "text" + }, + { + "bbox": [ + 306, + 496, + 313, + 507 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 495, + 336, + 507 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 505, + 338, + 520 + ], + "spans": [ + { + "bbox": [ + 107, + 506, + 157, + 518 + ], + "score": 0.92, + "content": "\\{ 0 , 0 . 5 , 1 . 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 505, + 302, + 520 + ], + "score": 1.0, + "content": ", we can visualize the tendency of", + "type": "text" + }, + { + "bbox": [ + 302, + 507, + 310, + 518 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 505, + 338, + 520 + ], + "score": 1.0, + "content": "along", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 517, + 336, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 269, + 529 + ], + "score": 1.0, + "content": "with training epochs. We observe that", + "type": "text" + }, + { + "bbox": [ + 270, + 518, + 277, + 529 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 517, + 336, + 529 + ], + "score": 1.0, + "content": "converges to-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 528, + 337, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 337, + 540 + ], + "score": 1.0, + "content": "wards 1 after 40 epochs no matter the initialization, which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 539, + 335, + 551 + ], + "spans": [ + { + "bbox": [ + 106, + 539, + 327, + 551 + ], + "score": 1.0, + "content": "demonstrates that the residual structure learns to push", + "type": "text" + }, + { + "bbox": [ + 328, + 540, + 335, + 551 + ], + "score": 0.79, + "content": "\\beta", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 548, + 321, + 564 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 321, + 564 + ], + "score": 1.0, + "content": "to a rather large value to alleviate gradient vanishing.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 484, + 338, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 567, + 336, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 336, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 336, + 579 + ], + "score": 1.0, + "content": "Similarly, DARTS (Liu et al., 2019b) utilizes a trainable", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 578, + 336, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 149, + 590 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 149, + 578, + 171, + 591 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 578, + 336, + 590 + ], + "score": 1.0, + "content": "to denote the importance of skip connec-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 588, + 337, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 255, + 602 + ], + "score": 1.0, + "content": "tion. However, In the search stage,", + "type": "text" + }, + { + "bbox": [ + 255, + 589, + 278, + 601 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 588, + 337, + 602 + ], + "score": 1.0, + "content": "can generally", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 600, + 336, + 612 + ], + "spans": [ + { + "bbox": [ + 106, + 600, + 336, + 612 + ], + "score": 1.0, + "content": "increase and dominate the architecture parameters, and", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 611, + 337, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 611, + 337, + 623 + ], + "score": 1.0, + "content": "finally leads to performance collapse. We analyze that", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 621, + 337, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 137, + 635 + ], + "score": 1.0, + "content": "a large", + "type": "text" + }, + { + "bbox": [ + 138, + 622, + 160, + 634 + ], + "score": 0.92, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 621, + 337, + 635 + ], + "score": 1.0, + "content": "in DARTS could result from two aspects:", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 632, + 337, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 337, + 645 + ], + "score": 1.0, + "content": "On the one hand, as the supernet automatically learns to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 642, + 338, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 264, + 657 + ], + "score": 1.0, + "content": "alleviate gradient vanishing, it pushes", + "type": "text" + }, + { + "bbox": [ + 264, + 644, + 287, + 656 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 642, + 338, + 657 + ], + "score": 1.0, + "content": "to a proper", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 337, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 337, + 667 + ], + "score": 1.0, + "content": "large value; On the other hand, the skip connection is", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 665, + 336, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 336, + 678 + ], + "score": 1.0, + "content": "indeed an important connection for the target network,", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 676, + 337, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 337, + 689 + ], + "score": 1.0, + "content": "which should be selected in the discretization stage. As", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 700 + ], + "score": 1.0, + "content": "a consequence, the skip connection in DARTS plays two-fold roles: as an auxiliary connection to", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "stabilize the supernet training, and as a candidate operation to build the final network. Inspired by", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "the above observation and analysis, we propose to stabilize the search process by distinguishing the", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 106, + 721, + 381, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 381, + 733 + ], + "score": 1.0, + "content": "two roles of skip connection and handling the issue of gradient flow.", + "type": "text" + } + ], + "index": 58 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 567, + 338, + 689 + ] + }, + { + "type": "image", + "bbox": [ + 345, + 488, + 491, + 571 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 345, + 488, + 491, + 571 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 345, + 488, + 491, + 571 + ], + "spans": [ + { + "bbox": [ + 345, + 488, + 491, + 571 + ], + "score": 0.96, + "type": "image", + "image_path": "6f78ee310b963f27f21baf3e4f1fce76a463b12ab1f9ab3ad040d3d42ec3feeb.jpg" + } + ] + } + ], + "index": 34.5, + "virtual_lines": [ + { + "bbox": [ + 345, + 488, + 491, + 529.5 + ], + "spans": [], + "index": 34 + }, + { + "bbox": [ + 345, + 529.5, + 491, + 571.0 + ], + "spans": [], + "index": 35 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 344, + 578, + 505, + 666 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 344, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 344, + 578, + 505, + 591 + ], + "score": 1.0, + "content": "Figure 2: Tendency of trainable coef-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 343, + 588, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 343, + 588, + 372, + 603 + ], + "score": 1.0, + "content": "ficient", + "type": "text" + }, + { + "bbox": [ + 373, + 590, + 381, + 601 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 588, + 453, + 603 + ], + "score": 1.0, + "content": "(initialized with", + "type": "text" + }, + { + "bbox": [ + 453, + 590, + 502, + 602 + ], + "score": 0.81, + "content": "\\{ 0 , \\ 0 . 5 , \\ 1 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 588, + 505, + 603 + ], + "score": 0.0, + "content": "", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 343, + 600, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 343, + 600, + 505, + 613 + ], + "score": 1.0, + "content": "of the skip connection in ResNet50 and", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 343, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 343, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "test accuracy (inset figure) vs. epochs.", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 343, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 343, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "The residual structure is proved to learn", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 343, + 634, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 343, + 634, + 376, + 646 + ], + "score": 1.0, + "content": "a large", + "type": "text" + }, + { + "bbox": [ + 377, + 634, + 385, + 646 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 634, + 505, + 646 + ], + "score": 1.0, + "content": "to ease training in all three", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 343, + 645, + 505, + 656 + ], + "spans": [ + { + "bbox": [ + 343, + 645, + 505, + 656 + ], + "score": 1.0, + "content": "cases. All models are trained and tested", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 343, + 655, + 404, + 667 + ], + "spans": [ + { + "bbox": [ + 343, + 655, + 404, + 667 + ], + "score": 1.0, + "content": "on CIFAR-10.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 49.5 + } + ], + "index": 42.0 + }, + { + "type": "text", + "bbox": [ + 107, + 688, + 504, + 732 + ], + "lines": [], + "index": 56.5, + "bbox_fs": [ + 105, + 687, + 506, + 733 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 344, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 345, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 345, + 95 + ], + "score": 1.0, + "content": "3.2 STEPPING OUT OF THE PERFORMANCE COLLAPSE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 103, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 104, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 115 + ], + "score": 1.0, + "content": "To distinguish the two roles, we introduce an auxiliary skip connection between every two nodes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "score": 1.0, + "content": "in a cell, see Fig. 1 (b). On the one hand, the fixed auxiliary skip connection carries the function", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 302, + 138 + ], + "score": 1.0, + "content": "of stabilizing the supernet training, even when", + "type": "text" + }, + { + "bbox": [ + 302, + 126, + 325, + 137 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 124, + 506, + 138 + ], + "score": 1.0, + "content": "is rather small. On the other hand, it also", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "breaks the unfair advantage (Chu et al., 2020b) as the advantageous contribution from the residual", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 145, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 393, + 160 + ], + "score": 1.0, + "content": "block is factored out. Consequently, the learned architectural parameter", + "type": "text" + }, + { + "bbox": [ + 394, + 147, + 416, + 159 + ], + "score": 0.92, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 145, + 506, + 160 + ], + "score": 1.0, + "content": "can be freed from the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "role of controlling the memory of gradients, and is more precisely aimed to represent the relative", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 168, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 104, + 168, + 506, + 183 + ], + "score": 1.0, + "content": "importance of skip connection as a candidate operation. In contrast to Eq. 7, the output feature map", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 138, + 199 + ], + "score": 1.0, + "content": "of edge", + "type": "text" + }, + { + "bbox": [ + 139, + 182, + 160, + 194 + ], + "score": 0.89, + "content": "e ^ { ( i , j ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 182, + 312, + 199 + ], + "score": 1.0, + "content": "can now be obtained by Eq. 4, where", + "type": "text" + }, + { + "bbox": [ + 312, + 180, + 409, + 202 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\beta _ { o } ^ { i , j } = \\frac { \\exp ( \\alpha _ { o } ^ { ( i , j ) } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\alpha _ { o ^ { \\prime } } ^ { ( i , j ) } ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "denotes the normalized", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 200, + 429, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 173, + 214 + ], + "score": 1.0, + "content": "importance, and", + "type": "text" + }, + { + "bbox": [ + 173, + 201, + 181, + 212 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 200, + 429, + 214 + ], + "score": 1.0, + "content": "is a coefficient independent from the architecture parameters.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "score": 1.0, + "content": "Moreover, to eliminate the impact of auxiliary connection on the discretization procedure, we pro-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 174, + 240 + ], + "score": 1.0, + "content": "pose to decrease", + "type": "text" + }, + { + "bbox": [ + 174, + 229, + 182, + 240 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 229, + 505, + 240 + ], + "score": 1.0, + "content": "to 0 in the search phase, and our method can be degenerated to standard DARTS", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "at the end of the search. Note that our method is insensitive to the type of decay strategy, so we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 289, + 263 + ], + "score": 1.0, + "content": "choose linear decay by default for simplicity.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 267, + 448, + 297 + ], + "lines": [ + { + "bbox": [ + 164, + 267, + 448, + 297 + ], + "spans": [ + { + "bbox": [ + 164, + 267, + 448, + 297 + ], + "score": 0.92, + "content": "\\bar { o } ^ { ( i , j ) } ( x ) = \\beta x + \\sum _ { o \\in \\mathcal { O } } \\beta _ { o } ^ { ( i , j ) } o ( x ) = \\left( \\beta + \\beta _ { s k i p } ^ { ( i , j ) } \\right) x + \\sum _ { o \\neq s k i p } \\beta _ { o } ^ { ( i , j ) } o ( x )", + "type": "interline_equation", + "image_path": "ef30e40957ecc1317803f71e12afe92f345077ca07c8a7c9704646edd25fa470.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 164, + 267, + 448, + 277.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 164, + 277.0, + 448, + 287.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 164, + 287.0, + 448, + 297.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 322 + ], + "score": 1.0, + "content": "We then analyze how the auxiliary skip connection handles the issue of gradient flow. Referring", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 479, + 332 + ], + "score": 1.0, + "content": "to the theorem of a recent work by Zhou et al. (2020), the convergence of network weight", + "type": "text" + }, + { + "bbox": [ + 479, + 319, + 493, + 330 + ], + "score": 0.28, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 330, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 253, + 344 + ], + "score": 1.0, + "content": "the supernet can heavily depend on", + "type": "text" + }, + { + "bbox": [ + 253, + 330, + 275, + 342 + ], + "score": 0.93, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 330, + 506, + 344 + ], + "score": 1.0, + "content": ". Specifically, suppose only three operations (none, skip", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "score": 1.0, + "content": "connection, and convolution) are included in the search space and MSE loss is utilized as training", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 349, + 507, + 368 + ], + "spans": [ + { + "bbox": [ + 104, + 349, + 248, + 368 + ], + "score": 1.0, + "content": "loss, when architecture parameters", + "type": "text" + }, + { + "bbox": [ + 249, + 352, + 264, + 366 + ], + "score": 0.91, + "content": "\\beta _ { i , j } ^ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 349, + 352, + 368 + ], + "score": 1.0, + "content": "are fixed to optimize", + "type": "text" + }, + { + "bbox": [ + 352, + 352, + 366, + 362 + ], + "score": 0.33, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 349, + 507, + 368 + ], + "score": 1.0, + "content": "via gradient descent, training loss", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 194, + 378 + ], + "score": 1.0, + "content": "can decrease by ratio", + "type": "text" + }, + { + "bbox": [ + 194, + 364, + 239, + 376 + ], + "score": 0.91, + "content": "( 1 - \\eta \\lambda / 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 363, + 383, + 378 + ], + "score": 1.0, + "content": "at one step with probability at least", + "type": "text" + }, + { + "bbox": [ + 384, + 364, + 407, + 375 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 363, + 437, + 378 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 438, + 366, + 445, + 376 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 363, + 505, + 378 + ], + "score": 1.0, + "content": "is the learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 325, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 232, + 388 + ], + "score": 1.0, + "content": "rate that should be bounded by", + "type": "text" + }, + { + "bbox": [ + 232, + 376, + 237, + 385 + ], + "score": 0.79, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 375, + 258, + 388 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 258, + 376, + 265, + 385 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 375, + 325, + 388 + ], + "score": 1.0, + "content": "follows Eq. 5.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20 + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 393, + 381, + 427 + ], + "lines": [ + { + "bbox": [ + 228, + 393, + 381, + 427 + ], + "spans": [ + { + "bbox": [ + 228, + 393, + 381, + 427 + ], + "score": 0.95, + "content": "\\lambda \\propto \\sum _ { i = 0 } ^ { h - 2 } \\left[ \\left( \\beta _ { c o n v } ^ { ( i , h - 1 ) } \\right) ^ { 2 } \\prod _ { t = 0 } ^ { i - 1 } \\left( \\beta _ { s k i p } ^ { ( t , i ) } \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "072057d2ed0eea14b0b25ca3b4288785cb274135e8d106c4157eef5ccdd7dbfe.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 393, + 381, + 410.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 228, + 410.0, + 381, + 427.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 133, + 447 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 433, + 140, + 442 + ], + "score": 0.82, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 429, + 412, + 447 + ], + "score": 1.0, + "content": "is the number of layers of the supernet. From Eq. 5, we observe that", + "type": "text" + }, + { + "bbox": [ + 413, + 433, + 420, + 442 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 429, + 481, + 447 + ], + "score": 1.0, + "content": "relies much on", + "type": "text" + }, + { + "bbox": [ + 482, + 433, + 504, + 444 + ], + "score": 0.84, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 441, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 127, + 458 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 444, + 151, + 455 + ], + "score": 0.92, + "content": "\\beta _ { c o n v }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 441, + 328, + 458 + ], + "score": 1.0, + "content": ", which indicates that the network weights", + "type": "text" + }, + { + "bbox": [ + 328, + 444, + 342, + 453 + ], + "score": 0.31, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 441, + 479, + 458 + ], + "score": 1.0, + "content": "can converge faster with a large", + "type": "text" + }, + { + "bbox": [ + 479, + 444, + 501, + 456 + ], + "score": 0.87, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 441, + 506, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 454, + 504, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 364, + 468 + ], + "score": 1.0, + "content": "However, by involving an auxiliary skip connection weighted as", + "type": "text" + }, + { + "bbox": [ + 365, + 455, + 372, + 466 + ], + "score": 0.79, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 454, + 504, + 468 + ], + "score": 1.0, + "content": ", Eq. 5 can be refined as follows:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 471, + 391, + 506 + ], + "lines": [ + { + "bbox": [ + 219, + 471, + 391, + 506 + ], + "spans": [ + { + "bbox": [ + 219, + 471, + 391, + 506 + ], + "score": 0.94, + "content": "\\lambda \\propto \\sum _ { i = 0 } ^ { h - 2 } \\left[ \\left( \\beta _ { c o n v } ^ { ( i , h - 1 ) } \\right) ^ { 2 } \\prod _ { t = 0 } ^ { i - 1 } \\left( \\beta _ { s k i p } ^ { ( t , i ) } + \\beta \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "7b1fa59568be3fb2d300fb92d3069ed49a329b9a359bcf38f8021ab1cf72c9c9.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 471, + 391, + 488.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 219, + 488.5, + 391, + 506.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 288, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 135, + 531 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 518, + 185, + 529 + ], + "score": 0.91, + "content": "\\beta \\ \\gg \\ \\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 516, + 222, + 531 + ], + "score": 1.0, + "content": "making", + "type": "text" + }, + { + "bbox": [ + 222, + 518, + 230, + 527 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 516, + 290, + 531 + ], + "score": 1.0, + "content": "insensitive to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 527, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 129, + 540 + ], + "score": 0.9, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 527, + 289, + 542 + ], + "score": 1.0, + "content": ", so that the convergence of network", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 538, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 141, + 552 + ], + "score": 1.0, + "content": "weights", + "type": "text" + }, + { + "bbox": [ + 141, + 540, + 155, + 550 + ], + "score": 0.4, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 538, + 230, + 552 + ], + "score": 1.0, + "content": "depends more on", + "type": "text" + }, + { + "bbox": [ + 231, + 540, + 254, + 551 + ], + "score": 0.91, + "content": "\\beta _ { c o n v }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 538, + 290, + 552 + ], + "score": 1.0, + "content": ". In the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 550, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 289, + 563 + ], + "score": 1.0, + "content": "beginning of the search, the common value", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 560, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 120, + 574 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 120, + 561, + 142, + 573 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 560, + 195, + 574 + ], + "score": 1.0, + "content": "is 0.15 while", + "type": "text" + }, + { + "bbox": [ + 195, + 561, + 203, + 573 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 560, + 290, + 574 + ], + "score": 1.0, + "content": "is 1.0. From the view", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 572, + 289, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 289, + 585 + ], + "score": 1.0, + "content": "of convergence theorem (Zhou et al., 2020),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 583, + 289, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 289, + 595 + ], + "score": 1.0, + "content": "the auxiliary skip connection alleviates the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 594, + 289, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 158, + 607 + ], + "score": 1.0, + "content": "privilege of", + "type": "text" + }, + { + "bbox": [ + 159, + 594, + 181, + 606 + ], + "score": 0.93, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 594, + 289, + 607 + ], + "score": 1.0, + "content": "and equalize the compe-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 605, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 289, + 616 + ], + "score": 1.0, + "content": "tition among architecture parameters. Even", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 616, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 131, + 628 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 132, + 616, + 140, + 627 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 616, + 289, + 628 + ], + "score": 1.0, + "content": "gradually decays, the fair competi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 289, + 639 + ], + "score": 1.0, + "content": "tion still holds since network weights W have", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 639, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 289, + 650 + ], + "score": 1.0, + "content": "been converged to an optimal point. Con-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 289, + 660 + ], + "score": 1.0, + "content": "sequently, DARTS- is able to stabilize the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 660, + 205, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 205, + 672 + ], + "score": 1.0, + "content": "search stage of DARTS.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 37.5 + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 289, + 710 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 289, + 689 + ], + "score": 1.0, + "content": "Extensive experiments are performed to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 289, + 700 + ], + "score": 1.0, + "content": "demonstrate the efficiency of the proposed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 289, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 289, + 711 + ], + "score": 1.0, + "content": "auxiliary skip connection, and we emphasize", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + }, + { + "type": "title", + "bbox": [ + 297, + 523, + 386, + 534 + ], + "lines": [ + { + "bbox": [ + 295, + 522, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 295, + 522, + 388, + 537 + ], + "score": 1.0, + "content": "Algorithm 1 DARTS-", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "title", + "bbox": [ + 297, + 537, + 335, + 547 + ], + "lines": [ + { + "bbox": [ + 295, + 536, + 338, + 550 + ], + "spans": [ + { + "bbox": [ + 295, + 536, + 338, + 550 + ], + "score": 1.0, + "content": "Require:", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + }, + { + "type": "text", + "bbox": [ + 312, + 548, + 504, + 581 + ], + "lines": [ + { + "bbox": [ + 313, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 383, + 560 + ], + "score": 1.0, + "content": "Network weights", + "type": "text" + }, + { + "bbox": [ + 383, + 549, + 392, + 557 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 546, + 493, + 560 + ], + "score": 1.0, + "content": "; Architecture parameters", + "type": "text" + }, + { + "bbox": [ + 493, + 549, + 501, + 557 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 546, + 505, + 560 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 558, + 431, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 417, + 569 + ], + "score": 1.0, + "content": "Number of search epochs", + "type": "text" + }, + { + "bbox": [ + 417, + 558, + 426, + 568 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 558, + 431, + 569 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 312, + 569, + 475, + 582 + ], + "spans": [ + { + "bbox": [ + 312, + 569, + 389, + 582 + ], + "score": 1.0, + "content": "Decay strategy for", + "type": "text" + }, + { + "bbox": [ + 389, + 569, + 471, + 581 + ], + "score": 0.91, + "content": "\\beta _ { e } , e \\in \\{ 1 , 2 , . . . , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 569, + 475, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 52 + } + ], + "index": 51 + }, + { + "type": "title", + "bbox": [ + 297, + 581, + 331, + 590 + ], + "lines": [ + { + "bbox": [ + 296, + 579, + 333, + 592 + ], + "spans": [ + { + "bbox": [ + 296, + 579, + 333, + 592 + ], + "score": 1.0, + "content": "Ensure:", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "text", + "bbox": [ + 301, + 591, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 313, + 591, + 460, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 448, + 603 + ], + "score": 1.0, + "content": "Searched architecture parameters", + "type": "text" + }, + { + "bbox": [ + 448, + 594, + 456, + 601 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 591, + 460, + 603 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 54 + }, + { + "bbox": [ + 301, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 301, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "1: Construct a super-network by stacking cells in", + "type": "text" + } + ], + "index": 55 + }, + { + "bbox": [ + 312, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 312, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "which there is an auxiliary skip connection be-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 312, + 624, + 447, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 447, + 635 + ], + "score": 1.0, + "content": "tween every two nodes of choice", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 302, + 635, + 406, + 647 + ], + "spans": [ + { + "bbox": [ + 302, + 636, + 311, + 645 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 312, + 635, + 349, + 647 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 349, + 635, + 390, + 646 + ], + "score": 0.92, + "content": "e \\in [ 1 , E ]", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 635, + 406, + 647 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 58 + }, + { + "bbox": [ + 301, + 645, + 494, + 658 + ], + "spans": [ + { + "bbox": [ + 301, + 646, + 311, + 657 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 322, + 645, + 388, + 658 + ], + "score": 1.0, + "content": "Update weights", + "type": "text" + }, + { + "bbox": [ + 388, + 648, + 397, + 655 + ], + "score": 0.76, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 645, + 411, + 658 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 411, + 646, + 494, + 657 + ], + "score": 0.86, + "content": "\\nabla _ { w } \\mathcal { L } _ { t r a i n } ( w , \\alpha , \\beta _ { e } )", + "type": "inline_equation" + } + ], + "index": 59 + }, + { + "bbox": [ + 301, + 655, + 496, + 670 + ], + "spans": [ + { + "bbox": [ + 301, + 657, + 311, + 668 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 322, + 655, + 401, + 670 + ], + "score": 1.0, + "content": "Update parameters", + "type": "text" + }, + { + "bbox": [ + 401, + 658, + 409, + 666 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 655, + 422, + 670 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 423, + 657, + 496, + 668 + ], + "score": 0.9, + "content": "\\nabla _ { \\alpha } \\mathcal { L } _ { v a l } ( w , \\alpha , \\beta _ { e } )", + "type": "inline_equation" + } + ], + "index": 60 + }, + { + "bbox": [ + 301, + 668, + 347, + 678 + ], + "spans": [ + { + "bbox": [ + 301, + 668, + 312, + 678 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 313, + 668, + 347, + 678 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 61 + }, + { + "bbox": [ + 301, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 301, + 677, + 496, + 689 + ], + "score": 1.0, + "content": "6: Derive the final architecture based on learned", + "type": "text" + }, + { + "bbox": [ + 496, + 680, + 504, + 688 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 62 + }, + { + "bbox": [ + 313, + 689, + 450, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 450, + 702 + ], + "score": 1.0, + "content": "from the best validation supernet.", + "type": "text" + } + ], + "index": 63 + } + ], + "index": 58.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 710, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 721 + ], + "score": 1.0, + "content": "that our method is flexible to combine with other methods to further improve the stabilization and", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 719, + 366, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 366, + 734 + ], + "score": 1.0, + "content": "searching performance. The overall algorithm is given in Alg. 1.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 344, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 345, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 345, + 95 + ], + "score": 1.0, + "content": "3.2 STEPPING OUT OF THE PERFORMANCE COLLAPSE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 103, + 505, + 213 + ], + "lines": [ + { + "bbox": [ + 106, + 104, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 505, + 115 + ], + "score": 1.0, + "content": "To distinguish the two roles, we introduce an auxiliary skip connection between every two nodes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 506, + 126 + ], + "score": 1.0, + "content": "in a cell, see Fig. 1 (b). On the one hand, the fixed auxiliary skip connection carries the function", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 124, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 302, + 138 + ], + "score": 1.0, + "content": "of stabilizing the supernet training, even when", + "type": "text" + }, + { + "bbox": [ + 302, + 126, + 325, + 137 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 124, + 506, + 138 + ], + "score": 1.0, + "content": "is rather small. On the other hand, it also", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 505, + 149 + ], + "score": 1.0, + "content": "breaks the unfair advantage (Chu et al., 2020b) as the advantageous contribution from the residual", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 145, + 506, + 160 + ], + "spans": [ + { + "bbox": [ + 104, + 145, + 393, + 160 + ], + "score": 1.0, + "content": "block is factored out. Consequently, the learned architectural parameter", + "type": "text" + }, + { + "bbox": [ + 394, + 147, + 416, + 159 + ], + "score": 0.92, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 145, + 506, + 160 + ], + "score": 1.0, + "content": "can be freed from the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 171 + ], + "score": 1.0, + "content": "role of controlling the memory of gradients, and is more precisely aimed to represent the relative", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 168, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 104, + 168, + 506, + 183 + ], + "score": 1.0, + "content": "importance of skip connection as a candidate operation. In contrast to Eq. 7, the output feature map", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 180, + 505, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 138, + 199 + ], + "score": 1.0, + "content": "of edge", + "type": "text" + }, + { + "bbox": [ + 139, + 182, + 160, + 194 + ], + "score": 0.89, + "content": "e ^ { ( i , j ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 182, + 312, + 199 + ], + "score": 1.0, + "content": "can now be obtained by Eq. 4, where", + "type": "text" + }, + { + "bbox": [ + 312, + 180, + 409, + 202 + ], + "score": 0.94, + "content": "\\begin{array} { r } { \\beta _ { o } ^ { i , j } = \\frac { \\exp ( \\alpha _ { o } ^ { ( i , j ) } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\alpha _ { o ^ { \\prime } } ^ { ( i , j ) } ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 184, + 505, + 196 + ], + "score": 1.0, + "content": "denotes the normalized", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 200, + 429, + 214 + ], + "spans": [ + { + "bbox": [ + 105, + 200, + 173, + 214 + ], + "score": 1.0, + "content": "importance, and", + "type": "text" + }, + { + "bbox": [ + 173, + 201, + 181, + 212 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 200, + 429, + 214 + ], + "score": 1.0, + "content": "is a coefficient independent from the architecture parameters.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 104, + 104, + 506, + 214 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 217, + 505, + 262 + ], + "lines": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 216, + 505, + 231 + ], + "score": 1.0, + "content": "Moreover, to eliminate the impact of auxiliary connection on the discretization procedure, we pro-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 229, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 229, + 174, + 240 + ], + "score": 1.0, + "content": "pose to decrease", + "type": "text" + }, + { + "bbox": [ + 174, + 229, + 182, + 240 + ], + "score": 0.85, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 229, + 505, + 240 + ], + "score": 1.0, + "content": "to 0 in the search phase, and our method can be degenerated to standard DARTS", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 505, + 252 + ], + "score": 1.0, + "content": "at the end of the search. Note that our method is insensitive to the type of decay strategy, so we", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 250, + 289, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 250, + 289, + 263 + ], + "score": 1.0, + "content": "choose linear decay by default for simplicity.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 216, + 505, + 263 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 164, + 267, + 448, + 297 + ], + "lines": [ + { + "bbox": [ + 164, + 267, + 448, + 297 + ], + "spans": [ + { + "bbox": [ + 164, + 267, + 448, + 297 + ], + "score": 0.92, + "content": "\\bar { o } ^ { ( i , j ) } ( x ) = \\beta x + \\sum _ { o \\in \\mathcal { O } } \\beta _ { o } ^ { ( i , j ) } o ( x ) = \\left( \\beta + \\beta _ { s k i p } ^ { ( i , j ) } \\right) x + \\sum _ { o \\neq s k i p } \\beta _ { o } ^ { ( i , j ) } o ( x )", + "type": "interline_equation", + "image_path": "ef30e40957ecc1317803f71e12afe92f345077ca07c8a7c9704646edd25fa470.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 164, + 267, + 448, + 277.0 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 164, + 277.0, + 448, + 287.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 164, + 287.0, + 448, + 297.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 505, + 387 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 506, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 506, + 322 + ], + "score": 1.0, + "content": "We then analyze how the auxiliary skip connection handles the issue of gradient flow. Referring", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 319, + 479, + 332 + ], + "score": 1.0, + "content": "to the theorem of a recent work by Zhou et al. (2020), the convergence of network weight", + "type": "text" + }, + { + "bbox": [ + 479, + 319, + 493, + 330 + ], + "score": 0.28, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 330, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 105, + 330, + 253, + 344 + ], + "score": 1.0, + "content": "the supernet can heavily depend on", + "type": "text" + }, + { + "bbox": [ + 253, + 330, + 275, + 342 + ], + "score": 0.93, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 330, + 506, + 344 + ], + "score": 1.0, + "content": ". Specifically, suppose only three operations (none, skip", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 506, + 355 + ], + "score": 1.0, + "content": "connection, and convolution) are included in the search space and MSE loss is utilized as training", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 349, + 507, + 368 + ], + "spans": [ + { + "bbox": [ + 104, + 349, + 248, + 368 + ], + "score": 1.0, + "content": "loss, when architecture parameters", + "type": "text" + }, + { + "bbox": [ + 249, + 352, + 264, + 366 + ], + "score": 0.91, + "content": "\\beta _ { i , j } ^ { o }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 349, + 352, + 368 + ], + "score": 1.0, + "content": "are fixed to optimize", + "type": "text" + }, + { + "bbox": [ + 352, + 352, + 366, + 362 + ], + "score": 0.33, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 366, + 349, + 507, + 368 + ], + "score": 1.0, + "content": "via gradient descent, training loss", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 363, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 194, + 378 + ], + "score": 1.0, + "content": "can decrease by ratio", + "type": "text" + }, + { + "bbox": [ + 194, + 364, + 239, + 376 + ], + "score": 0.91, + "content": "( 1 - \\eta \\lambda / 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 363, + 383, + 378 + ], + "score": 1.0, + "content": "at one step with probability at least", + "type": "text" + }, + { + "bbox": [ + 384, + 364, + 407, + 375 + ], + "score": 0.87, + "content": "1 - \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 363, + 437, + 378 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 438, + 366, + 445, + 376 + ], + "score": 0.8, + "content": "\\eta", + "type": "inline_equation" + }, + { + "bbox": [ + 445, + 363, + 505, + 378 + ], + "score": 1.0, + "content": "is the learning", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 375, + 325, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 232, + 388 + ], + "score": 1.0, + "content": "rate that should be bounded by", + "type": "text" + }, + { + "bbox": [ + 232, + 376, + 237, + 385 + ], + "score": 0.79, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 375, + 258, + 388 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 258, + 376, + 265, + 385 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 375, + 325, + 388 + ], + "score": 1.0, + "content": "follows Eq. 5.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20, + "bbox_fs": [ + 104, + 306, + 507, + 388 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 228, + 393, + 381, + 427 + ], + "lines": [ + { + "bbox": [ + 228, + 393, + 381, + 427 + ], + "spans": [ + { + "bbox": [ + 228, + 393, + 381, + 427 + ], + "score": 0.95, + "content": "\\lambda \\propto \\sum _ { i = 0 } ^ { h - 2 } \\left[ \\left( \\beta _ { c o n v } ^ { ( i , h - 1 ) } \\right) ^ { 2 } \\prod _ { t = 0 } ^ { i - 1 } \\left( \\beta _ { s k i p } ^ { ( t , i ) } \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "072057d2ed0eea14b0b25ca3b4288785cb274135e8d106c4157eef5ccdd7dbfe.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 228, + 393, + 381, + 410.0 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 228, + 410.0, + 381, + 427.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 432, + 504, + 466 + ], + "lines": [ + { + "bbox": [ + 105, + 429, + 504, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 133, + 447 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 433, + 140, + 442 + ], + "score": 0.82, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 429, + 412, + 447 + ], + "score": 1.0, + "content": "is the number of layers of the supernet. From Eq. 5, we observe that", + "type": "text" + }, + { + "bbox": [ + 413, + 433, + 420, + 442 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 429, + 481, + 447 + ], + "score": 1.0, + "content": "relies much on", + "type": "text" + }, + { + "bbox": [ + 482, + 433, + 504, + 444 + ], + "score": 0.84, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 441, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 441, + 127, + 458 + ], + "score": 1.0, + "content": "than", + "type": "text" + }, + { + "bbox": [ + 127, + 444, + 151, + 455 + ], + "score": 0.92, + "content": "\\beta _ { c o n v }", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 441, + 328, + 458 + ], + "score": 1.0, + "content": ", which indicates that the network weights", + "type": "text" + }, + { + "bbox": [ + 328, + 444, + 342, + 453 + ], + "score": 0.31, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 441, + 479, + 458 + ], + "score": 1.0, + "content": "can converge faster with a large", + "type": "text" + }, + { + "bbox": [ + 479, + 444, + 501, + 456 + ], + "score": 0.87, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 441, + 506, + 458 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 454, + 504, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 364, + 468 + ], + "score": 1.0, + "content": "However, by involving an auxiliary skip connection weighted as", + "type": "text" + }, + { + "bbox": [ + 365, + 455, + 372, + 466 + ], + "score": 0.79, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 372, + 454, + 504, + 468 + ], + "score": 1.0, + "content": ", Eq. 5 can be refined as follows:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 429, + 506, + 468 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 219, + 471, + 391, + 506 + ], + "lines": [ + { + "bbox": [ + 219, + 471, + 391, + 506 + ], + "spans": [ + { + "bbox": [ + 219, + 471, + 391, + 506 + ], + "score": 0.94, + "content": "\\lambda \\propto \\sum _ { i = 0 } ^ { h - 2 } \\left[ \\left( \\beta _ { c o n v } ^ { ( i , h - 1 ) } \\right) ^ { 2 } \\prod _ { t = 0 } ^ { i - 1 } \\left( \\beta _ { s k i p } ^ { ( t , i ) } + \\beta \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "7b1fa59568be3fb2d300fb92d3069ed49a329b9a359bcf38f8021ab1cf72c9c9.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 219, + 471, + 391, + 488.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 219, + 488.5, + 391, + 506.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 288, + 671 + ], + "lines": [ + { + "bbox": [ + 106, + 516, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 106, + 516, + 135, + 531 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 135, + 518, + 185, + 529 + ], + "score": 0.91, + "content": "\\beta \\ \\gg \\ \\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 516, + 222, + 531 + ], + "score": 1.0, + "content": "making", + "type": "text" + }, + { + "bbox": [ + 222, + 518, + 230, + 527 + ], + "score": 0.8, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 516, + 290, + 531 + ], + "score": 1.0, + "content": "insensitive to", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 527, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 529, + 129, + 540 + ], + "score": 0.9, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 129, + 527, + 289, + 542 + ], + "score": 1.0, + "content": ", so that the convergence of network", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 538, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 141, + 552 + ], + "score": 1.0, + "content": "weights", + "type": "text" + }, + { + "bbox": [ + 141, + 540, + 155, + 550 + ], + "score": 0.4, + "content": "\\mathbf { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 538, + 230, + 552 + ], + "score": 1.0, + "content": "depends more on", + "type": "text" + }, + { + "bbox": [ + 231, + 540, + 254, + 551 + ], + "score": 0.91, + "content": "\\beta _ { c o n v }", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 538, + 290, + 552 + ], + "score": 1.0, + "content": ". In the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 550, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 106, + 550, + 289, + 563 + ], + "score": 1.0, + "content": "beginning of the search, the common value", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 560, + 290, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 120, + 574 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 120, + 561, + 142, + 573 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 560, + 195, + 574 + ], + "score": 1.0, + "content": "is 0.15 while", + "type": "text" + }, + { + "bbox": [ + 195, + 561, + 203, + 573 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 560, + 290, + 574 + ], + "score": 1.0, + "content": "is 1.0. From the view", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 572, + 289, + 585 + ], + "spans": [ + { + "bbox": [ + 106, + 572, + 289, + 585 + ], + "score": 1.0, + "content": "of convergence theorem (Zhou et al., 2020),", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 583, + 289, + 595 + ], + "spans": [ + { + "bbox": [ + 106, + 583, + 289, + 595 + ], + "score": 1.0, + "content": "the auxiliary skip connection alleviates the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 594, + 289, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 158, + 607 + ], + "score": 1.0, + "content": "privilege of", + "type": "text" + }, + { + "bbox": [ + 159, + 594, + 181, + 606 + ], + "score": 0.93, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 594, + 289, + 607 + ], + "score": 1.0, + "content": "and equalize the compe-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 605, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 289, + 616 + ], + "score": 1.0, + "content": "tition among architecture parameters. Even", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 616, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 131, + 628 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 132, + 616, + 140, + 627 + ], + "score": 0.81, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 140, + 616, + 289, + 628 + ], + "score": 1.0, + "content": "gradually decays, the fair competi-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 627, + 289, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 289, + 639 + ], + "score": 1.0, + "content": "tion still holds since network weights W have", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 639, + 289, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 289, + 650 + ], + "score": 1.0, + "content": "been converged to an optimal point. Con-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 289, + 660 + ], + "score": 1.0, + "content": "sequently, DARTS- is able to stabilize the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 660, + 205, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 205, + 672 + ], + "score": 1.0, + "content": "search stage of DARTS.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 516, + 290, + 672 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 289, + 710 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 289, + 689 + ], + "score": 1.0, + "content": "Extensive experiments are performed to", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 687, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 687, + 289, + 700 + ], + "score": 1.0, + "content": "demonstrate the efficiency of the proposed", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 699, + 289, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 289, + 711 + ], + "score": 1.0, + "content": "auxiliary skip connection, and we emphasize", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 106, + 677, + 289, + 711 + ] + }, + { + "type": "title", + "bbox": [ + 297, + 523, + 386, + 534 + ], + "lines": [ + { + "bbox": [ + 295, + 522, + 388, + 537 + ], + "spans": [ + { + "bbox": [ + 295, + 522, + 388, + 537 + ], + "score": 1.0, + "content": "Algorithm 1 DARTS-", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 48 + }, + { + "type": "title", + "bbox": [ + 297, + 537, + 335, + 547 + ], + "lines": [ + { + "bbox": [ + 295, + 536, + 338, + 550 + ], + "spans": [ + { + "bbox": [ + 295, + 536, + 338, + 550 + ], + "score": 1.0, + "content": "Require:", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 49 + }, + { + "type": "list", + "bbox": [ + 312, + 548, + 504, + 581 + ], + "lines": [ + { + "bbox": [ + 313, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 383, + 560 + ], + "score": 1.0, + "content": "Network weights", + "type": "text" + }, + { + "bbox": [ + 383, + 549, + 392, + 557 + ], + "score": 0.77, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 392, + 546, + 493, + 560 + ], + "score": 1.0, + "content": "; Architecture parameters", + "type": "text" + }, + { + "bbox": [ + 493, + 549, + 501, + 557 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 546, + 505, + 560 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 50, + "is_list_end_line": true + }, + { + "bbox": [ + 313, + 558, + 431, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 417, + 569 + ], + "score": 1.0, + "content": "Number of search epochs", + "type": "text" + }, + { + "bbox": [ + 417, + 558, + 426, + 568 + ], + "score": 0.82, + "content": "E", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 558, + 431, + 569 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 51, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 312, + 569, + 475, + 582 + ], + "spans": [ + { + "bbox": [ + 312, + 569, + 389, + 582 + ], + "score": 1.0, + "content": "Decay strategy for", + "type": "text" + }, + { + "bbox": [ + 389, + 569, + 471, + 581 + ], + "score": 0.91, + "content": "\\beta _ { e } , e \\in \\{ 1 , 2 , . . . , E \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 569, + 475, + 582 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 52, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 51, + "bbox_fs": [ + 312, + 546, + 505, + 582 + ] + }, + { + "type": "title", + "bbox": [ + 297, + 581, + 331, + 590 + ], + "lines": [ + { + "bbox": [ + 296, + 579, + 333, + 592 + ], + "spans": [ + { + "bbox": [ + 296, + 579, + 333, + 592 + ], + "score": 1.0, + "content": "Ensure:", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 53 + }, + { + "type": "list", + "bbox": [ + 301, + 591, + 504, + 701 + ], + "lines": [ + { + "bbox": [ + 313, + 591, + 460, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 448, + 603 + ], + "score": 1.0, + "content": "Searched architecture parameters", + "type": "text" + }, + { + "bbox": [ + 448, + 594, + 456, + 601 + ], + "score": 0.75, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 591, + 460, + 603 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 54, + "is_list_end_line": true + }, + { + "bbox": [ + 301, + 602, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 301, + 602, + 505, + 614 + ], + "score": 1.0, + "content": "1: Construct a super-network by stacking cells in", + "type": "text" + } + ], + "index": 55, + "is_list_start_line": true + }, + { + "bbox": [ + 312, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 312, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "which there is an auxiliary skip connection be-", + "type": "text" + } + ], + "index": 56 + }, + { + "bbox": [ + 312, + 624, + 447, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 447, + 635 + ], + "score": 1.0, + "content": "tween every two nodes of choice", + "type": "text" + } + ], + "index": 57, + "is_list_end_line": true + }, + { + "bbox": [ + 302, + 635, + 406, + 647 + ], + "spans": [ + { + "bbox": [ + 302, + 636, + 311, + 645 + ], + "score": 1.0, + "content": "2:", + "type": "text" + }, + { + "bbox": [ + 312, + 635, + 349, + 647 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 349, + 635, + 390, + 646 + ], + "score": 0.92, + "content": "e \\in [ 1 , E ]", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 635, + 406, + 647 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 58, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 301, + 645, + 494, + 658 + ], + "spans": [ + { + "bbox": [ + 301, + 646, + 311, + 657 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 322, + 645, + 388, + 658 + ], + "score": 1.0, + "content": "Update weights", + "type": "text" + }, + { + "bbox": [ + 388, + 648, + 397, + 655 + ], + "score": 0.76, + "content": "w", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 645, + 411, + 658 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 411, + 646, + 494, + 657 + ], + "score": 0.86, + "content": "\\nabla _ { w } \\mathcal { L } _ { t r a i n } ( w , \\alpha , \\beta _ { e } )", + "type": "inline_equation" + } + ], + "index": 59, + "is_list_start_line": true + }, + { + "bbox": [ + 301, + 655, + 496, + 670 + ], + "spans": [ + { + "bbox": [ + 301, + 657, + 311, + 668 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 322, + 655, + 401, + 670 + ], + "score": 1.0, + "content": "Update parameters", + "type": "text" + }, + { + "bbox": [ + 401, + 658, + 409, + 666 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 655, + 422, + 670 + ], + "score": 1.0, + "content": "by", + "type": "text" + }, + { + "bbox": [ + 423, + 657, + 496, + 668 + ], + "score": 0.9, + "content": "\\nabla _ { \\alpha } \\mathcal { L } _ { v a l } ( w , \\alpha , \\beta _ { e } )", + "type": "inline_equation" + } + ], + "index": 60, + "is_list_start_line": true + }, + { + "bbox": [ + 301, + 668, + 347, + 678 + ], + "spans": [ + { + "bbox": [ + 301, + 668, + 312, + 678 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 313, + 668, + 347, + 678 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 61, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 301, + 677, + 504, + 689 + ], + "spans": [ + { + "bbox": [ + 301, + 677, + 496, + 689 + ], + "score": 1.0, + "content": "6: Derive the final architecture based on learned", + "type": "text" + }, + { + "bbox": [ + 496, + 680, + 504, + 688 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 62, + "is_list_start_line": true + }, + { + "bbox": [ + 313, + 689, + 450, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 450, + 702 + ], + "score": 1.0, + "content": "from the best validation supernet.", + "type": "text" + } + ], + "index": 63, + "is_list_end_line": true + } + ], + "index": 58.5, + "bbox_fs": [ + 301, + 591, + 505, + 702 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 269, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 271, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 271, + 95 + ], + "score": 1.0, + "content": "3.3 RELATIONSHIP TO PRIOR WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "Our method is aimed to address the performance collapse in differentiable neural architecture search.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "Most previous works (Zela et al., 2020; Chen & Hsieh, 2020; Liang et al., 2019) concentrate on de-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "veloping various criteria or indicators characterizing the occurrence of collapse. Whereas, we don’t", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "study or rely on these indicators because they can mistakenly reject good models. Inspired by Chu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "et al. (2020b), our method focuses on calibrating the biased searching process. The underlying phi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "losophy is simple: if the biased process is rectified, the searching result will be better. In summary,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "score": 1.0, + "content": "our method differs from others in two aspects: being process-oriented and indicator-free. Distinct", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "from Chu et al. (2020b) that tweaks the competitive environment, our method can be viewed as", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "one that breaks the unfair advantage. Moreover, we don’t introduce any handcrafted indicators to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 482, + 216 + ], + "score": 1.0, + "content": "represent performance collapse, thus greatly reducing the burden of shifting to different tasks.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 108, + 235, + 200, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 201, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 201, + 249 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 262, + 315, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 316, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 316, + 275 + ], + "score": 1.0, + "content": "4.1 SEARCH SPACES AND TRAINING SETTINGS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "For searching and evaluation in the standard DARTS space (we name it as S0 for simplicity), we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "keep the same settings as in DARTS (Liu et al., 2019b). We follow R-DARTS (Zela et al., 2020)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "for their proposed reduced spaces S1- S4 (harder than S0). However, the inferred models are trained", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "with two different settings from R-DARTS (Zela et al., 2020) and SDARTS (Chen & Hsieh, 2020).", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "The difference lies in the number of layers and initial channels for evaluation on CIFAR-100. R-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "DARTS sets 8 layers and 16 initial channels. Instead, SDARTS uses 20 and 36 respectively. For the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 350, + 507, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 507, + 363 + ], + "score": 1.0, + "content": "proxyless searching on ImageNet, we instead search in MobileNetV2-like search space (we name it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 361, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 466, + 374 + ], + "score": 1.0, + "content": "S5) proposed in FBNet (Wu et al., 2019). We use the SGD optimizer for weight and Adam", + "type": "text" + }, + { + "bbox": [ + 467, + 361, + 504, + 372 + ], + "score": 0.88, + "content": "\\beta _ { 1 } = 0 . 5", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 371, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 124, + 386 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 372, + 174, + 384 + ], + "score": 0.87, + "content": "\\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 371, + 506, + 386 + ], + "score": 1.0, + "content": ", 0.001 learning rate) for architecture parameters with the batch-size of 768. The", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "initial learning rate is 0.045 and decayed to 0 within 30 epochs following the cosine schedule. We", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "also use L2 regularization with 1e-4. It takes about 4.5 GPU days on Tesla V100. More details are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "provided in the appendix. We also use NAS-Bench-201 (S6) since DARTS performs severely bad.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "In total, we use 7 different search spaces to conduct the experiments, which involves three datasets.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 108, + 445, + 223, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 225, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 225, + 458 + ], + "score": 1.0, + "content": "4.2 SEARCHING RESULTS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 310, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 465, + 311, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 311, + 480 + ], + "score": 1.0, + "content": "CIFAR-10 and CIFAR-100. Following the set-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 478, + 311, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 311, + 490 + ], + "score": 1.0, + "content": "tings of R-DARTS (Zela et al., 2020), we obtain an", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 311, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 216, + 501 + ], + "score": 1.0, + "content": "average top-1 accuracy of", + "type": "text" + }, + { + "bbox": [ + 216, + 489, + 249, + 500 + ], + "score": 0.9, + "content": "9 7 . 4 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 489, + 311, + 501 + ], + "score": 1.0, + "content": "on CIFAR-10,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 500, + 311, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 311, + 513 + ], + "score": 1.0, + "content": "as shown in Table 2. Moreover, our method is very", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 511, + 311, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 311, + 522 + ], + "score": 1.0, + "content": "robust since out of 5 independent runs the search-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 523, + 311, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 311, + 534 + ], + "score": 1.0, + "content": "ing results are quite stable. The best cells found on", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31.5 + }, + { + "type": "table", + "bbox": [ + 321, + 495, + 500, + 528 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 317, + 470, + 504, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 318, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 318, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "Table 1: Comparison of searched CNN in the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 318, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "DARTS search space on two different datasets.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.0 + }, + { + "type": "table_body", + "bbox": [ + 321, + 495, + 500, + 528 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 321, + 495, + 500, + 528 + ], + "spans": [ + { + "bbox": [ + 321, + 495, + 500, + 528 + ], + "score": 0.963, + "html": "
DatasetDARTSR-DARTS (L2)Ours
C10 (SO)2.91±0.252.95±0.212.63±0.07
C100 (S0)20.58±0.4418.01±0.2617.51±0.25
", + "type": "table", + "image_path": "e9c58bdd00daf5c365ed4c835a640ea64168343db0815053209fc543703af0b3.jpg" + } + ] + } + ], + "index": 34.0, + "virtual_lines": [ + { + "bbox": [ + 321, + 495, + 500, + 511.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 321, + 511.5, + 500, + 528.0 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 151, + 545 + ], + "score": 1.0, + "content": "CIFAR-10", + "type": "text" + }, + { + "bbox": [ + 151, + 533, + 183, + 545 + ], + "score": 0.81, + "content": "( 9 7 . 5 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "are shown in Figure 9 (B). Results on CIFAR-100 are presented in Table 10 (see", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 357, + 556 + ], + "score": 1.0, + "content": "A.2.1). Moreover, our method has a much lower searching cost", + "type": "text" + }, + { + "bbox": [ + 357, + 545, + 394, + 555 + ], + "score": 0.55, + "content": "( 3 \\times \\mathrm { l e s s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "than R-DARTS (Zela et al.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "2020), where four independent searches with different regularization settings are needed to generate", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 565, + 494, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 428, + 578 + ], + "score": 1.0, + "content": "the best architecture. In other words, its robustness comes from the cost of more", + "type": "text" + }, + { + "bbox": [ + 428, + 566, + 450, + 577 + ], + "score": 0.91, + "content": "C O _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 565, + 494, + 578 + ], + "score": 1.0, + "content": "emissions.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 106, + 583, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "ImageNet. To further verify the efficiency of DARTS-, we directly search on ImageNet in S5", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "score": 1.0, + "content": "and compare our results with the state-of-the-art models under the mobile setting in Table 2. The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 387, + 618 + ], + "score": 1.0, + "content": "visualization of the architecture is given in Fig 10. DARTS-A obtains", + "type": "text" + }, + { + "bbox": [ + 388, + 605, + 415, + 615 + ], + "score": 0.88, + "content": "7 6 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "top-1 accuracy on the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "ImageNet validation dataset. By contrast, direct applying DARTS on this search space only obtains", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 625, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 134, + 638 + ], + "score": 0.87, + "content": "6 6 . { \\bar { 4 \\% } }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 625, + 306, + 640 + ], + "score": 1.0, + "content": "(Chu et al., 2020b). Moreover, it obtains", + "type": "text" + }, + { + "bbox": [ + 306, + 627, + 334, + 638 + ], + "score": 0.88, + "content": "7 7 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 625, + 505, + 640 + ], + "score": 1.0, + "content": "top-1 accuracy after being equipped with", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "auto-augmentation (Cubuk et al., 2019) and squeeze-and-excitation (Hu et al., 2018), which are also", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 649, + 191, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 191, + 660 + ], + "score": 1.0, + "content": "used in EfficientNet.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "NAS-Bench-201. Apart from standard search spaces, benchmarking with known optimal in a lim-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "ited setting is also recommended. NAS-Bench-201 (Dong & Yang, 2020) consists of 15,625 archi-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "tectures in a reduced DARTS-like search space, where it has 4 internal nodes and 5 operations per", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "node. We compare our method with prior work in Table 3. We search on CIFAR-10 and look up the", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "ground-truth performance with found genotypes on various test sets. Remarkably, we achieve a new", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 721, + 351, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 351, + 732 + ], + "score": 1.0, + "content": "state of the art, the best of which almost touches the optimal.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 269, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 271, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 271, + 95 + ], + "score": 1.0, + "content": "3.3 RELATIONSHIP TO PRIOR WORK", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "Our method is aimed to address the performance collapse in differentiable neural architecture search.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 505, + 127 + ], + "score": 1.0, + "content": "Most previous works (Zela et al., 2020; Chen & Hsieh, 2020; Liang et al., 2019) concentrate on de-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 140 + ], + "score": 1.0, + "content": "veloping various criteria or indicators characterizing the occurrence of collapse. Whereas, we don’t", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "study or rely on these indicators because they can mistakenly reject good models. Inspired by Chu", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 161 + ], + "score": 1.0, + "content": "et al. (2020b), our method focuses on calibrating the biased searching process. The underlying phi-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "losophy is simple: if the biased process is rectified, the searching result will be better. In summary,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 170, + 506, + 183 + ], + "score": 1.0, + "content": "our method differs from others in two aspects: being process-oriented and indicator-free. Distinct", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 182, + 505, + 194 + ], + "score": 1.0, + "content": "from Chu et al. (2020b) that tweaks the competitive environment, our method can be viewed as", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "one that breaks the unfair advantage. Moreover, we don’t introduce any handcrafted indicators to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 204, + 482, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 482, + 216 + ], + "score": 1.0, + "content": "represent performance collapse, thus greatly reducing the burden of shifting to different tasks.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 104, + 506, + 216 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 235, + 200, + 247 + ], + "lines": [ + { + "bbox": [ + 105, + 234, + 201, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 201, + 249 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 262, + 315, + 273 + ], + "lines": [ + { + "bbox": [ + 105, + 262, + 316, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 316, + 275 + ], + "score": 1.0, + "content": "4.1 SEARCH SPACES AND TRAINING SETTINGS", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 298 + ], + "score": 1.0, + "content": "For searching and evaluation in the standard DARTS space (we name it as S0 for simplicity), we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "keep the same settings as in DARTS (Liu et al., 2019b). We follow R-DARTS (Zela et al., 2020)", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "for their proposed reduced spaces S1- S4 (harder than S0). However, the inferred models are trained", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 505, + 330 + ], + "score": 1.0, + "content": "with two different settings from R-DARTS (Zela et al., 2020) and SDARTS (Chen & Hsieh, 2020).", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "spans": [ + { + "bbox": [ + 105, + 328, + 505, + 341 + ], + "score": 1.0, + "content": "The difference lies in the number of layers and initial channels for evaluation on CIFAR-100. R-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 506, + 352 + ], + "score": 1.0, + "content": "DARTS sets 8 layers and 16 initial channels. Instead, SDARTS uses 20 and 36 respectively. For the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 350, + 507, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 350, + 507, + 363 + ], + "score": 1.0, + "content": "proxyless searching on ImageNet, we instead search in MobileNetV2-like search space (we name it", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 361, + 504, + 374 + ], + "spans": [ + { + "bbox": [ + 105, + 361, + 466, + 374 + ], + "score": 1.0, + "content": "S5) proposed in FBNet (Wu et al., 2019). We use the SGD optimizer for weight and Adam", + "type": "text" + }, + { + "bbox": [ + 467, + 361, + 504, + 372 + ], + "score": 0.88, + "content": "\\beta _ { 1 } = 0 . 5", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 371, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 371, + 124, + 386 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 372, + 174, + 384 + ], + "score": 0.87, + "content": "\\beta _ { 2 } = 0 . 9 9 9", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 371, + 506, + 386 + ], + "score": 1.0, + "content": ", 0.001 learning rate) for architecture parameters with the batch-size of 768. The", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 383, + 506, + 396 + ], + "score": 1.0, + "content": "initial learning rate is 0.045 and decayed to 0 within 30 epochs following the cosine schedule. We", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 407 + ], + "score": 1.0, + "content": "also use L2 regularization with 1e-4. It takes about 4.5 GPU days on Tesla V100. More details are", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 405, + 505, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 505, + 417 + ], + "score": 1.0, + "content": "provided in the appendix. We also use NAS-Bench-201 (S6) since DARTS performs severely bad.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "In total, we use 7 different search spaces to conduct the experiments, which involves three datasets.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 284, + 507, + 429 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 445, + 223, + 456 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 225, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 225, + 458 + ], + "score": 1.0, + "content": "4.2 SEARCHING RESULTS", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 467, + 310, + 533 + ], + "lines": [ + { + "bbox": [ + 106, + 465, + 311, + 480 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 311, + 480 + ], + "score": 1.0, + "content": "CIFAR-10 and CIFAR-100. Following the set-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 478, + 311, + 490 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 311, + 490 + ], + "score": 1.0, + "content": "tings of R-DARTS (Zela et al., 2020), we obtain an", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 311, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 216, + 501 + ], + "score": 1.0, + "content": "average top-1 accuracy of", + "type": "text" + }, + { + "bbox": [ + 216, + 489, + 249, + 500 + ], + "score": 0.9, + "content": "9 7 . 4 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 489, + 311, + 501 + ], + "score": 1.0, + "content": "on CIFAR-10,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 500, + 311, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 311, + 513 + ], + "score": 1.0, + "content": "as shown in Table 2. Moreover, our method is very", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 511, + 311, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 311, + 522 + ], + "score": 1.0, + "content": "robust since out of 5 independent runs the search-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 523, + 311, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 311, + 534 + ], + "score": 1.0, + "content": "ing results are quite stable. The best cells found on", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 465, + 311, + 534 + ] + }, + { + "type": "table", + "bbox": [ + 321, + 495, + 500, + 528 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 317, + 470, + 504, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 318, + 470, + 505, + 482 + ], + "spans": [ + { + "bbox": [ + 318, + 470, + 505, + 482 + ], + "score": 1.0, + "content": "Table 1: Comparison of searched CNN in the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 318, + 481, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 318, + 481, + 505, + 493 + ], + "score": 1.0, + "content": "DARTS search space on two different datasets.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.0 + }, + { + "type": "table_body", + "bbox": [ + 321, + 495, + 500, + 528 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 321, + 495, + 500, + 528 + ], + "spans": [ + { + "bbox": [ + 321, + 495, + 500, + 528 + ], + "score": 0.963, + "html": "
DatasetDARTSR-DARTS (L2)Ours
C10 (SO)2.91±0.252.95±0.212.63±0.07
C100 (S0)20.58±0.4418.01±0.2617.51±0.25
", + "type": "table", + "image_path": "e9c58bdd00daf5c365ed4c835a640ea64168343db0815053209fc543703af0b3.jpg" + } + ] + } + ], + "index": 34.0, + "virtual_lines": [ + { + "bbox": [ + 321, + 495, + 500, + 511.5 + ], + "spans": [], + "index": 33 + }, + { + "bbox": [ + 321, + 511.5, + 500, + 528.0 + ], + "spans": [], + "index": 35 + } + ] + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 151, + 545 + ], + "score": 1.0, + "content": "CIFAR-10", + "type": "text" + }, + { + "bbox": [ + 151, + 533, + 183, + 545 + ], + "score": 0.81, + "content": "( 9 7 . 5 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "are shown in Figure 9 (B). Results on CIFAR-100 are presented in Table 10 (see", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 544, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 357, + 556 + ], + "score": 1.0, + "content": "A.2.1). Moreover, our method has a much lower searching cost", + "type": "text" + }, + { + "bbox": [ + 357, + 545, + 394, + 555 + ], + "score": 0.55, + "content": "( 3 \\times \\mathrm { l e s s } )", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 544, + 505, + 556 + ], + "score": 1.0, + "content": "than R-DARTS (Zela et al.,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 505, + 568 + ], + "score": 1.0, + "content": "2020), where four independent searches with different regularization settings are needed to generate", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 565, + 494, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 428, + 578 + ], + "score": 1.0, + "content": "the best architecture. In other words, its robustness comes from the cost of more", + "type": "text" + }, + { + "bbox": [ + 428, + 566, + 450, + 577 + ], + "score": 0.91, + "content": "C O _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 565, + 494, + 578 + ], + "score": 1.0, + "content": "emissions.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 533, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 583, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 505, + 596 + ], + "score": 1.0, + "content": "ImageNet. To further verify the efficiency of DARTS-, we directly search on ImageNet in S5", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 504, + 605 + ], + "score": 1.0, + "content": "and compare our results with the state-of-the-art models under the mobile setting in Table 2. The", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 387, + 618 + ], + "score": 1.0, + "content": "visualization of the architecture is given in Fig 10. DARTS-A obtains", + "type": "text" + }, + { + "bbox": [ + 388, + 605, + 415, + 615 + ], + "score": 0.88, + "content": "7 6 . 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "top-1 accuracy on the", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "ImageNet validation dataset. By contrast, direct applying DARTS on this search space only obtains", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 625, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 134, + 638 + ], + "score": 0.87, + "content": "6 6 . { \\bar { 4 \\% } }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 625, + 306, + 640 + ], + "score": 1.0, + "content": "(Chu et al., 2020b). Moreover, it obtains", + "type": "text" + }, + { + "bbox": [ + 306, + 627, + 334, + 638 + ], + "score": 0.88, + "content": "7 7 . 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 334, + 625, + 505, + 640 + ], + "score": 1.0, + "content": "top-1 accuracy after being equipped with", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 505, + 650 + ], + "score": 1.0, + "content": "auto-augmentation (Cubuk et al., 2019) and squeeze-and-excitation (Hu et al., 2018), which are also", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 649, + 191, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 191, + 660 + ], + "score": 1.0, + "content": "used in EfficientNet.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 582, + 505, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "NAS-Bench-201. Apart from standard search spaces, benchmarking with known optimal in a lim-", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "ited setting is also recommended. NAS-Bench-201 (Dong & Yang, 2020) consists of 15,625 archi-", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 104, + 687, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 701 + ], + "score": 1.0, + "content": "tectures in a reduced DARTS-like search space, where it has 4 internal nodes and 5 operations per", + "type": "text" + } + ], + "index": 50 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "node. We compare our method with prior work in Table 3. We search on CIFAR-10 and look up the", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "ground-truth performance with found genotypes on various test sets. Remarkably, we achieve a new", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 106, + 721, + 351, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 351, + 732 + ], + "score": 1.0, + "content": "state of the art, the best of which almost touches the optimal.", + "type": "text" + } + ], + "index": 53 + } + ], + "index": 50.5, + "bbox_fs": [ + 104, + 665, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 145, + 493, + 343 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 144 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "score": 1.0, + "content": "Table 2: Comparison of the state-of-the-art models on CIFAR-10 (left) and ImageNet (right). On", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "CIFAR-10 dataset, our average result is obtained on 5 independently searched models to assure the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "robustness. For ImageNet, networks in the top block are directly searched on ImageNet; the middle", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 135 + ], + "score": 1.0, + "content": "indicates that architectures are searched on CIFAR-10 and then transferred to ImageNet; the bottom", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 471, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 471, + 146 + ], + "score": 1.0, + "content": "indicates models have SE and Swish. We search in S0 for CIFAR-10 and S5 for ImageNet.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 109, + 145, + 493, + 343 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 145, + 493, + 341 + ], + "spans": [ + { + "bbox": [ + 109, + 145, + 493, + 341 + ], + "score": 0.713, + "html": "
ModelsFLOPs Params Top-1 Top-5Cost (GPU days)
ModelsParams (M)FLOPs Acc (M)(%)Cost(M)(M)(%)(%)
NASNet-A (2018)608tGPU Days AmoebaNet-A (2019) MnasNet-92 (2019)555 3885.1 3.974.592.0 74.79 92.13150 3791
3.3626†97.352000FBNet-C (2019)3755.574.992.39
ENAS (2018) DARTS (2019b)4.697.110.5FairNAS-A (2019b)3884.675.392.412
3.3528t97.00±0.14*0.4SCARLET-C(2019a)3656.776.993.410
SNAS (2019)2.8422t97.15±0.02*1.5FairDARTS-D (2020b) 4404.375.692.63
GDAS (2019b)3.4519t97.070.2PC-DARTS (2020)5975.375.892.73.8
P-DARTS (2019b)3.4532t97.50.3DARTS- (ours)4674.976.293.04.5
PC-DARTS (2020) DARTS- (best)3.6558t 56897.43 97.50.1 0.4NASNet-A (2018)5645.374.091.62000
3.5DARTS (2019b)5744.773.391.30.4
P-DARTS (2019b)3.3±0.21 540±34 97.19±0.140.3SNAS (2019)5224.372.790.81.5
R-DARTS (2020)197.05±0.211.6PC-DARTS (2020)5865.374.992.20.1
SDARTS-ADV (2020) 3.397.39±0.02 3.5±0.13 583±22 97.41±0.081.3 0.4FairDARTS-B (2020b) 5414.875.192.50.4
DARTS- (avg.)MobileNetV3 (2019)5.475.292.2~3000
† Based on the provided genotypes. ↑MoGA-A (2020a)2195.175.992.812
5 independent searches using their released code.MixNet-M (2019)3045.077.093.3~3000
EfficientNet B0 (2019) 3903605.376.393.2~3000
* Training the best searched model for several times NoisyDARTS-A4495.577.994.012
(whose average doesn't indicate the stability of the DARTS- (ours) method)4705.577.893.94.5
", + "type": "table", + "image_path": "3c9ac279d1fa6cb2aa9ef41e42b9b1893179a4e5f786b78337f81090d44e6e2d.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 145, + 493, + 211.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 211.0, + 493, + 277.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 277.0, + 493, + 343.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 307, + 337, + 403, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 401, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 401, + 347 + ], + "score": 1.0, + "content": "‡ Estimated by Wu et al. (2019).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 342, + 405, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 405, + 356 + ], + "score": 1.0, + "content": "\u0005 SE modules and Swish enabled.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 109, + 385, + 501, + 489 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 104, + 362, + 503, + 384 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 362, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 504, + 375 + ], + "score": 1.0, + "content": "Table 3: Searching performance on NAS-Bench-201 (Dong & Yang, 2020). Our method robustly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 370, + 458, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 458, + 387 + ], + "score": 1.0, + "content": "obtains new SOTA. Averaged on 4 runs of searching. 1st: first-order, 2nd: second-order", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 385, + 501, + 489 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 385, + 501, + 489 + ], + "spans": [ + { + "bbox": [ + 109, + 385, + 501, + 489 + ], + "score": 0.981, + "html": "
MethodCost (hours)CIFAR-10CIFAR-100ImageNet16-120
validtestvalidtestvalidtest
DARTS1st (2019b)3.239.77±0.00 54.30±0.0015.03±0.00 15.61±0.0016.43±0.00 16.32±0.00
DARTS2nd (2019b)10.239.77±0.0054.30±0.0015.03±0.0015.61±0.0016.43±0.00 16.32±0.00
GDAS (2019b)8.789.89±0.08 93.61±0.0971.34±0.04：70.70±0.30 41.59±1.33 41.71±0.98
SETN (2019a)9.584.04±0.2887.64±0.0058.86±0.0659.05±0.2433.06±0.0232.52±0.21
DARTS-3.291.03±0.44 93.80±0.4071.36±1.51 71.53±1.5144.87±1.46 45.12±0.82
DARTS- (best)3.291.5594.3673.4973.5146.3746.34
optimaln/a91.6194.3773.4973.5146.7747.31
", + "type": "table", + "image_path": "0b513fc0af68a942755131d0bfbb3ef7d718b4a2039cfe14d21aab52bad82ee3.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 109, + 385, + 501, + 419.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 109, + 419.6666666666667, + 501, + 454.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 109, + 454.33333333333337, + 501, + 489.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.75 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Transfer results on objection detection We further evaluate the transfer-ability of our models on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "score": 1.0, + "content": "down-stream object detection task by replacing the backbone of RetinaNet (Lin et al., 2017) on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "MMDetection toolbox platform (Chen et al., 2019a). Specifically, with the same training setting as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 263, + 550 + ], + "score": 1.0, + "content": "Chu et al. (2020b), our model achieves", + "type": "text" + }, + { + "bbox": [ + 263, + 537, + 290, + 547 + ], + "score": 0.85, + "content": "3 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "mAP on the COCO dataset, surpassing other similar-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "sized models such as MobileNetV3, MixNet, and FairDARTS. The detailed results are shown in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 558, + 194, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 194, + 571 + ], + "score": 1.0, + "content": "Appendix (Table 11).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 107, + 584, + 362, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 363, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 363, + 596 + ], + "score": 1.0, + "content": "4.3 ORTHOGONAL COMBINATION WITH OTHER VARIANTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 605, + 503, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "score": 1.0, + "content": "Our method can be flexibly adapted to combine with prior work for further improvements. Here we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 616, + 412, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 412, + 628 + ], + "score": 1.0, + "content": "investigate the joint outcome with two methods: P-DARTS and PC-DARTS.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 338, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 340, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 340, + 645 + ], + "score": 1.0, + "content": "Progressive DARTS (P-DARTS). P-DARTS (Chen et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 642, + 340, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 340, + 657 + ], + "score": 1.0, + "content": "2019b) proposes a progressive approach to search gradu-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 655, + 339, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 339, + 667 + ], + "score": 1.0, + "content": "ally with deeper depths while pruning out the uncompeti-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "score": 1.0, + "content": "tive paths. Additionally, it makes use of some handcrafted", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 677, + 340, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 340, + 689 + ], + "score": 1.0, + "content": "criteria to address the collapse (the progressive idea itself", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 340, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 340, + 701 + ], + "score": 1.0, + "content": "cannot deal with it), for instance, they impose two strong", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 699, + 338, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 326, + 711 + ], + "score": 1.0, + "content": "priors by regularizing the number of skip connections", + "type": "text" + }, + { + "bbox": [ + 326, + 699, + 338, + 709 + ], + "score": 0.75, + "content": "M", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 340, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 340, + 721 + ], + "score": 1.0, + "content": "as 2 as well as dropout. To be fair, we remove such a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 720, + 340, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 340, + 732 + ], + "score": 1.0, + "content": "carefully handcrafted trick and run P-DARTS for several", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 28 + }, + { + "type": "table", + "bbox": [ + 361, + 691, + 488, + 734 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 347, + 635, + 504, + 690 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 345, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 345, + 634, + 505, + 648 + ], + "score": 1.0, + "content": "Table 4: We remove the strong con-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 346, + 647, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 346, + 647, + 505, + 658 + ], + "score": 1.0, + "content": "straints on the number of skip connec-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 345, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 345, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "tions as 2 and dropout (priors) for P-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 345, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "DARTS and compare its performance", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 346, + 679, + 433, + 691 + ], + "spans": [ + { + "bbox": [ + 346, + 679, + 433, + 691 + ], + "score": 1.0, + "content": "w/ and w/o DARTS-.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "table_body", + "bbox": [ + 361, + 691, + 488, + 734 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 361, + 691, + 488, + 734 + ], + "spans": [ + { + "bbox": [ + 361, + 691, + 488, + 734 + ], + "score": 0.964, + "html": "
MethodSetting商银 Acc(%)
P-DARTS w/o priors 96.48±0.55
P-DARTS- w/o priors 97.28±0.04
", + "type": "table", + "image_path": "7f5ce7128509e78c6ebfe7c5d79fd24dd79a10a6ee3f1747e68e4777ae3fdff8.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 361, + 691, + 488, + 712.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 361, + 712.5, + 488, + 734.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 36.75 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 39 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 145, + 493, + 343 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 144 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 90, + 505, + 101 + ], + "score": 1.0, + "content": "Table 2: Comparison of the state-of-the-art models on CIFAR-10 (left) and ImageNet (right). On", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "CIFAR-10 dataset, our average result is obtained on 5 independently searched models to assure the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 112, + 505, + 124 + ], + "score": 1.0, + "content": "robustness. For ImageNet, networks in the top block are directly searched on ImageNet; the middle", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 506, + 135 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 506, + 135 + ], + "score": 1.0, + "content": "indicates that architectures are searched on CIFAR-10 and then transferred to ImageNet; the bottom", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 132, + 471, + 146 + ], + "spans": [ + { + "bbox": [ + 105, + 132, + 471, + 146 + ], + "score": 1.0, + "content": "indicates models have SE and Swish. We search in S0 for CIFAR-10 and S5 for ImageNet.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "table_body", + "bbox": [ + 109, + 145, + 493, + 343 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 109, + 145, + 493, + 341 + ], + "spans": [ + { + "bbox": [ + 109, + 145, + 493, + 341 + ], + "score": 0.713, + "html": "
ModelsFLOPs Params Top-1 Top-5Cost (GPU days)
ModelsParams (M)FLOPs Acc (M)(%)Cost(M)(M)(%)(%)
NASNet-A (2018)608tGPU Days AmoebaNet-A (2019) MnasNet-92 (2019)555 3885.1 3.974.592.0 74.79 92.13150 3791
3.3626†97.352000FBNet-C (2019)3755.574.992.39
ENAS (2018) DARTS (2019b)4.697.110.5FairNAS-A (2019b)3884.675.392.412
3.3528t97.00±0.14*0.4SCARLET-C(2019a)3656.776.993.410
SNAS (2019)2.8422t97.15±0.02*1.5FairDARTS-D (2020b) 4404.375.692.63
GDAS (2019b)3.4519t97.070.2PC-DARTS (2020)5975.375.892.73.8
P-DARTS (2019b)3.4532t97.50.3DARTS- (ours)4674.976.293.04.5
PC-DARTS (2020) DARTS- (best)3.6558t 56897.43 97.50.1 0.4NASNet-A (2018)5645.374.091.62000
3.5DARTS (2019b)5744.773.391.30.4
P-DARTS (2019b)3.3±0.21 540±34 97.19±0.140.3SNAS (2019)5224.372.790.81.5
R-DARTS (2020)197.05±0.211.6PC-DARTS (2020)5865.374.992.20.1
SDARTS-ADV (2020) 3.397.39±0.02 3.5±0.13 583±22 97.41±0.081.3 0.4FairDARTS-B (2020b) 5414.875.192.50.4
DARTS- (avg.)MobileNetV3 (2019)5.475.292.2~3000
† Based on the provided genotypes. ↑MoGA-A (2020a)2195.175.992.812
5 independent searches using their released code.MixNet-M (2019)3045.077.093.3~3000
EfficientNet B0 (2019) 3903605.376.393.2~3000
* Training the best searched model for several times NoisyDARTS-A4495.577.994.012
(whose average doesn't indicate the stability of the DARTS- (ours) method)4705.577.893.94.5
", + "type": "table", + "image_path": "3c9ac279d1fa6cb2aa9ef41e42b9b1893179a4e5f786b78337f81090d44e6e2d.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 109, + 145, + 493, + 211.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 109, + 211.0, + 493, + 277.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 109, + 277.0, + 493, + 343.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_footnote", + "bbox": [ + 307, + 337, + 403, + 353 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 306, + 335, + 401, + 347 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 401, + 347 + ], + "score": 1.0, + "content": "‡ Estimated by Wu et al. (2019).", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 342, + 405, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 405, + 356 + ], + "score": 1.0, + "content": "\u0005 SE modules and Swish enabled.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 109, + 385, + 501, + 489 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 104, + 362, + 503, + 384 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 362, + 504, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 504, + 375 + ], + "score": 1.0, + "content": "Table 3: Searching performance on NAS-Bench-201 (Dong & Yang, 2020). Our method robustly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 370, + 458, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 458, + 387 + ], + "score": 1.0, + "content": "obtains new SOTA. Averaged on 4 runs of searching. 1st: first-order, 2nd: second-order", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "table_body", + "bbox": [ + 109, + 385, + 501, + 489 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 109, + 385, + 501, + 489 + ], + "spans": [ + { + "bbox": [ + 109, + 385, + 501, + 489 + ], + "score": 0.981, + "html": "
MethodCost (hours)CIFAR-10CIFAR-100ImageNet16-120
validtestvalidtestvalidtest
DARTS1st (2019b)3.239.77±0.00 54.30±0.0015.03±0.00 15.61±0.0016.43±0.00 16.32±0.00
DARTS2nd (2019b)10.239.77±0.0054.30±0.0015.03±0.0015.61±0.0016.43±0.00 16.32±0.00
GDAS (2019b)8.789.89±0.08 93.61±0.0971.34±0.04：70.70±0.30 41.59±1.33 41.71±0.98
SETN (2019a)9.584.04±0.2887.64±0.0058.86±0.0659.05±0.2433.06±0.0232.52±0.21
DARTS-3.291.03±0.44 93.80±0.4071.36±1.51 71.53±1.5144.87±1.46 45.12±0.82
DARTS- (best)3.291.5594.3673.4973.5146.3746.34
optimaln/a91.6194.3773.4973.5146.7747.31
", + "type": "table", + "image_path": "0b513fc0af68a942755131d0bfbb3ef7d718b4a2039cfe14d21aab52bad82ee3.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 109, + 385, + 501, + 419.6666666666667 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 109, + 419.6666666666667, + 501, + 454.33333333333337 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 109, + 454.33333333333337, + 501, + 489.00000000000006 + ], + "spans": [], + "index": 14 + } + ] + } + ], + "index": 11.75 + }, + { + "type": "text", + "bbox": [ + 107, + 503, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 503, + 505, + 516 + ], + "score": 1.0, + "content": "Transfer results on objection detection We further evaluate the transfer-ability of our models on", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 528 + ], + "score": 1.0, + "content": "down-stream object detection task by replacing the backbone of RetinaNet (Lin et al., 2017) on", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 505, + 539 + ], + "score": 1.0, + "content": "MMDetection toolbox platform (Chen et al., 2019a). Specifically, with the same training setting as", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 536, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 263, + 550 + ], + "score": 1.0, + "content": "Chu et al. (2020b), our model achieves", + "type": "text" + }, + { + "bbox": [ + 263, + 537, + 290, + 547 + ], + "score": 0.85, + "content": "3 2 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 536, + 505, + 550 + ], + "score": 1.0, + "content": "mAP on the COCO dataset, surpassing other similar-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "sized models such as MobileNetV3, MixNet, and FairDARTS. The detailed results are shown in", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 558, + 194, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 194, + 571 + ], + "score": 1.0, + "content": "Appendix (Table 11).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 503, + 505, + 571 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 584, + 362, + 596 + ], + "lines": [ + { + "bbox": [ + 106, + 584, + 363, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 363, + 596 + ], + "score": 1.0, + "content": "4.3 ORTHOGONAL COMBINATION WITH OTHER VARIANTS", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 108, + 605, + 503, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 619 + ], + "score": 1.0, + "content": "Our method can be flexibly adapted to combine with prior work for further improvements. Here we", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 616, + 412, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 616, + 412, + 628 + ], + "score": 1.0, + "content": "investigate the joint outcome with two methods: P-DARTS and PC-DARTS.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 603, + 505, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 632, + 338, + 731 + ], + "lines": [ + { + "bbox": [ + 105, + 632, + 340, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 340, + 645 + ], + "score": 1.0, + "content": "Progressive DARTS (P-DARTS). P-DARTS (Chen et al.,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 642, + 340, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 340, + 657 + ], + "score": 1.0, + "content": "2019b) proposes a progressive approach to search gradu-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 655, + 339, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 339, + 667 + ], + "score": 1.0, + "content": "ally with deeper depths while pruning out the uncompeti-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 340, + 677 + ], + "score": 1.0, + "content": "tive paths. Additionally, it makes use of some handcrafted", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 677, + 340, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 340, + 689 + ], + "score": 1.0, + "content": "criteria to address the collapse (the progressive idea itself", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 687, + 340, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 340, + 701 + ], + "score": 1.0, + "content": "cannot deal with it), for instance, they impose two strong", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 699, + 338, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 326, + 711 + ], + "score": 1.0, + "content": "priors by regularizing the number of skip connections", + "type": "text" + }, + { + "bbox": [ + 326, + 699, + 338, + 709 + ], + "score": 0.75, + "content": "M", + "type": "inline_equation" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 709, + 340, + 721 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 340, + 721 + ], + "score": 1.0, + "content": "as 2 as well as dropout. To be fair, we remove such a", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 720, + 340, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 340, + 732 + ], + "score": 1.0, + "content": "carefully handcrafted trick and run P-DARTS for several", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "times. As a natural control group, we also combine DARTS- with P-DARTS. We run both exper-", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "iments for 3 times on CIFAR-10 dataset in Table 4. Without the strong priors, P-DARTS severely", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "suffers from the collapse, where the inferred models contain an excessive number of skip connec-", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 293, + 312 + ], + "score": 1.0, + "content": "tions. Specifically, it has a very high test error", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 294, + 299, + 324, + 310 + ], + "score": 0.86, + "content": "( 3 . 4 2 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 324, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "on average), even worse than DARTS. How-", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 309, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 506, + 323 + ], + "score": 1.0, + "content": "ever, P-DARTS can benefit greatly from the combination with DARTS-. The improved version (we", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 318, + 334 + ], + "score": 1.0, + "content": "call P-DARTS-) obtains much higher top-1 accuracy", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 319, + 321, + 354, + 332 + ], + "score": 0.88, + "content": "( + \\mathbf { 0 . 8 \\% } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 354, + 321, + 481, + 334 + ], + "score": 1.0, + "content": "on CIFAR-10 than its baseline.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 632, + 340, + 732 + ] + }, + { + "type": "table", + "bbox": [ + 361, + 691, + 488, + 734 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 347, + 635, + 504, + 690 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 345, + 634, + 505, + 648 + ], + "spans": [ + { + "bbox": [ + 345, + 634, + 505, + 648 + ], + "score": 1.0, + "content": "Table 4: We remove the strong con-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 346, + 647, + 505, + 658 + ], + "spans": [ + { + "bbox": [ + 346, + 647, + 505, + 658 + ], + "score": 1.0, + "content": "straints on the number of skip connec-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 345, + 657, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 345, + 657, + 505, + 669 + ], + "score": 1.0, + "content": "tions as 2 and dropout (priors) for P-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 345, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 345, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "DARTS and compare its performance", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 346, + 679, + 433, + 691 + ], + "spans": [ + { + "bbox": [ + 346, + 679, + 433, + 691 + ], + "score": 1.0, + "content": "w/ and w/o DARTS-.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "table_body", + "bbox": [ + 361, + 691, + 488, + 734 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 361, + 691, + 488, + 734 + ], + "spans": [ + { + "bbox": [ + 361, + 691, + 488, + 734 + ], + "score": 0.964, + "html": "
MethodSetting商银 Acc(%)
P-DARTS w/o priors 96.48±0.55
P-DARTS- w/o priors 97.28±0.04
", + "type": "table", + "image_path": "7f5ce7128509e78c6ebfe7c5d79fd24dd79a10a6ee3f1747e68e4777ae3fdff8.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 361, + 691, + 488, + 712.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 361, + 712.5, + 488, + 734.0 + ], + "spans": [], + "index": 39 + } + ] + } + ], + "index": 36.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 136, + 487, + 259 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 133 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "score": 1.0, + "content": "Table 7: Comparison in various search spaces. We report the lowest error rate of 3 found archi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "tectures. †: under Chen & Hsieh (2020)’s training settings where all models have 20 layers and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "36 initial channels (the best is shown in boldface). ‡: under Zela et al. (2020)’s settings where", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 498, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 498, + 134 + ], + "score": 1.0, + "content": "CIFAR-100 models have 8 layers and 16 initial channels (The best is in boldface and underlined).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 122, + 136, + 487, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 136, + 487, + 259 + ], + "spans": [ + { + "bbox": [ + 122, + 136, + 487, + 259 + ], + "score": 0.984, + "html": "
Benchmark DARTS‡R-DARTS‡DARTSOursPC-DARTS+SDARTS+Ourst
DPL2ESADARSADV
C10S13.843.112.783.013.102.683.112.782.732.68
S24.853.483.313.263.352.633.022.752.652.63
S33.342.932.512.742.592.422.512.532.492.42
S47.203.583.563.714.842.863.022.932.872.86
C100S129.4625.9324.2528.3724.0322.4118.8717.0216.8816.92
S226.0522.3022.2423.2523.5221.6118.2317.5617.2416.14
S328.9022.3623.9923.7323.3721.1318.0517.7317.1215.86
S422.8522.1821.9421.2623.2021.5517.1617.1715.4617.48
", + "type": "table", + "image_path": "aa305b66636e358ec535bafa2706cb424f53534e8f09638a403c57dd4964e9ef.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 122, + 136, + 487, + 177.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 122, + 177.0, + 487, + 218.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 122, + 218.0, + 487, + 259.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 505, + 333 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "times. As a natural control group, we also combine DARTS- with P-DARTS. We run both exper-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "score": 1.0, + "content": "iments for 3 times on CIFAR-10 dataset in Table 4. Without the strong priors, P-DARTS severely", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "suffers from the collapse, where the inferred models contain an excessive number of skip connec-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 299, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 293, + 312 + ], + "score": 1.0, + "content": "tions. Specifically, it has a very high test error", + "type": "text" + }, + { + "bbox": [ + 294, + 299, + 324, + 310 + ], + "score": 0.86, + "content": "( 3 . 4 2 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 299, + 505, + 312 + ], + "score": 1.0, + "content": "on average), even worse than DARTS. How-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 309, + 506, + 323 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 506, + 323 + ], + "score": 1.0, + "content": "ever, P-DARTS can benefit greatly from the combination with DARTS-. The improved version (we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 321, + 481, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 318, + 334 + ], + "score": 1.0, + "content": "call P-DARTS-) obtains much higher top-1 accuracy", + "type": "text" + }, + { + "bbox": [ + 319, + 321, + 354, + 332 + ], + "score": 0.88, + "content": "( + \\mathbf { 0 . 8 \\% } )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 321, + 481, + 334 + ], + "score": 1.0, + "content": "on CIFAR-10 than its baseline.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 316, + 437 + ], + "lines": [ + { + "bbox": [ + 106, + 338, + 316, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 316, + 349 + ], + "score": 1.0, + "content": "Memory Friendly DARTS (PC-DARTS). To alle-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 349, + 316, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 316, + 361 + ], + "score": 1.0, + "content": "viate the large memory overhead from the whole su-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 359, + 316, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 189, + 372 + ], + "score": 1.0, + "content": "pernet, PC-DARTS", + "type": "text" + }, + { + "bbox": [ + 190, + 360, + 203, + 370 + ], + "score": 0.43, + "content": "\\mathrm { { X u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 359, + 316, + 372 + ], + "score": 1.0, + "content": "et al., 2020) selects the par-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 370, + 316, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 316, + 383 + ], + "score": 1.0, + "content": "tial channel for searching. The proportion hyper-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 381, + 317, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 381, + 149, + 394 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 150, + 382, + 160, + 392 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 381, + 317, + 394 + ], + "score": 1.0, + "content": "needs careful calibration to achieve a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 317, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 317, + 405 + ], + "score": 1.0, + "content": "good result for specific tasks. As a byproduct, the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 403, + 316, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 285, + 416 + ], + "score": 1.0, + "content": "search time is also reduced to 0.1 GPU days", + "type": "text" + }, + { + "bbox": [ + 286, + 404, + 312, + 415 + ], + "score": 0.81, + "content": "\\left( K { = } 4 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 403, + 316, + 416 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 414, + 316, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 316, + 427 + ], + "score": 1.0, + "content": "We use their released code and run repeated experi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 317, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 317, + 439 + ], + "score": 1.0, + "content": "ments across different seeds under the same settings.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 324, + 398, + 509, + 440 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 323, + 340, + 505, + 395 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 324, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 324, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "Table 5: Comparison of PC-DARTS remov-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 324, + 352, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 324, + 352, + 505, + 363 + ], + "score": 1.0, + "content": "ing the strong prior (i.e. channel shuffle) and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 323, + 362, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 323, + 362, + 505, + 373 + ], + "score": 1.0, + "content": "combining DARTS-. The results are from 3", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 322, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 322, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "independent runs on CIFAR-10. The GPU", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 323, + 385, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 323, + 385, + 479, + 396 + ], + "score": 1.0, + "content": "memory cost is on a batch size of 256.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "table_body", + "bbox": [ + 324, + 398, + 509, + 440 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 324, + 398, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 324, + 398, + 506, + 440 + ], + "score": 0.963, + "html": "
MethodSettingAcc (%）Memory Cost
PC-DARTS K = 2 97.09±0.1419.9G 股银 3.75h
PC-DARTS- K = 2 97.35±0.0220.8G 3.41h
", + "type": "table", + "image_path": "1d67fc707d5b0e793f7535aa262e7620f1ad64da77a1e0366a7d9dd73cc88f01.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 324, + 398, + 509, + 412.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 324, + 412.0, + 509, + 426.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 324, + 426.0, + 509, + 440.0 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 26.0 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 339, + 455 + ], + "score": 1.0, + "content": "To accurately evaluate the role of our method, we choose", + "type": "text" + }, + { + "bbox": [ + 340, + 443, + 361, + 453 + ], + "score": 0.85, + "content": "K { = } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "(a bad configuration in the original", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "score": 1.0, + "content": "paper). We make comparisons between the original PC-DARTS and its combination with ours", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 463, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 479 + ], + "score": 1.0, + "content": "(named PC-DARTS-) in Table 5. PC-DARTS- can marginally boost the CIFAR-10 top-1 accuracy", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 107, + 475, + 143, + 487 + ], + "score": 0.9, + "content": "( + 0 . 2 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "on average). The result also confirms that our method can make PC-DARTS less sensitive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 486, + 460, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 196, + 500 + ], + "score": 1.0, + "content": "to its hyper-parameter", + "type": "text" + }, + { + "bbox": [ + 197, + 487, + 207, + 496 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 486, + 460, + 500 + ], + "score": 1.0, + "content": "while keeping its advantage of less memory cost and run time.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 108, + 512, + 208, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 210, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 210, + 524 + ], + "score": 1.0, + "content": "4.4 ABLATION STUDY", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 475, + 546 + ], + "score": 1.0, + "content": "Robustness to Decay Strategy. Our method is insensitive to the type of decay policy on", + "type": "text" + }, + { + "bbox": [ + 475, + 534, + 483, + 545 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 533, + 505, + 546 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "design extra two strategies as comparisons: cosine and step decay. They both have the similar", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 554, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 237, + 569 + ], + "score": 1.0, + "content": "performance. Specifically, when", + "type": "text" + }, + { + "bbox": [ + 238, + 556, + 246, + 567 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 554, + 505, + 569 + ], + "score": 1.0, + "content": "is scheduled to zero by the cosine strategy, the average accuracy", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 563, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 286, + 580 + ], + "score": 1.0, + "content": "of four searched CIFAR-10 models in S3 is", + "type": "text" + }, + { + "bbox": [ + 287, + 566, + 344, + 577 + ], + "score": 0.89, + "content": "9 7 . 3 3 \\% \\pm 0 . 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 563, + 392, + 580 + ], + "score": 1.0, + "content": ", the best is", + "type": "text" + }, + { + "bbox": [ + 393, + 566, + 425, + 577 + ], + "score": 0.89, + "content": "9 7 . 4 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 563, + 506, + 580 + ], + "score": 1.0, + "content": ". The step decay at", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 576, + 420, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 176, + 591 + ], + "score": 1.0, + "content": "epoch 45 obtains", + "type": "text" + }, + { + "bbox": [ + 177, + 577, + 209, + 588 + ], + "score": 0.87, + "content": "9 7 . 3 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 576, + 420, + 591 + ], + "score": 1.0, + "content": "top-1 accuracy on average in the same search space.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 353, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 353, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 353, + 605 + ], + "score": 1.0, + "content": "Robustness Comparison on C10 and C100 in S0-S4. To", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 605, + 353, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 353, + 617 + ], + "score": 1.0, + "content": "verify the robustness, it is required to search several times to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 616, + 353, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 353, + 627 + ], + "score": 1.0, + "content": "report the average performance of derived models (Yu et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 626, + 354, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 354, + 638 + ], + "score": 1.0, + "content": "2020; Antoine et al., 2020). As shown in Table 1, Table 9", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 637, + 353, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 353, + 650 + ], + "score": 1.0, + "content": "and Table 7, our method outperforms the recent SOTAs across", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 648, + 354, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 354, + 661 + ], + "score": 1.0, + "content": "several spaces and datasets. Note that SDARTS-ADV utilizes", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 660, + 353, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 242, + 672 + ], + "score": 1.0, + "content": "adversarial training and requires", + "type": "text" + }, + { + "bbox": [ + 242, + 660, + 257, + 671 + ], + "score": 0.9, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 660, + 353, + 672 + ], + "score": 1.0, + "content": "more search time than", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 671, + 353, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 277, + 682 + ], + "score": 1.0, + "content": "ours. Especially, we find a good model in", + "type": "text" + }, + { + "bbox": [ + 278, + 671, + 290, + 682 + ], + "score": 0.88, + "content": "S _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 671, + 353, + 682 + ], + "score": 1.0, + "content": "on CIFAR-100", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 681, + 354, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 233, + 693 + ], + "score": 1.0, + "content": "with the lowest top-1 test error", + "type": "text" + }, + { + "bbox": [ + 233, + 681, + 265, + 692 + ], + "score": 0.89, + "content": "1 5 . 8 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 681, + 354, + 693 + ], + "score": 1.0, + "content": ". The architectures of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 692, + 281, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 281, + 705 + ], + "score": 1.0, + "content": "these models can be found in the appendix.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 45.5 + }, + { + "type": "table", + "bbox": [ + 390, + 640, + 473, + 719 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 360, + 594, + 505, + 637 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 360, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 360, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "Table 6: Searching performance on", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 360, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 360, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "CIFAR-10 in S3 w.r.t the initial lin-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 360, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 360, + 615, + 422, + 627 + ], + "score": 1.0, + "content": "ear decay rate", + "type": "text" + }, + { + "bbox": [ + 422, + 616, + 433, + 627 + ], + "score": 0.87, + "content": "\\beta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 615, + 505, + 627 + ], + "score": 1.0, + "content": ". Each setting is", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 360, + 626, + 440, + 638 + ], + "spans": [ + { + "bbox": [ + 360, + 626, + 440, + 638 + ], + "score": 1.0, + "content": "run for three times.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52.5 + }, + { + "type": "table_body", + "bbox": [ + 390, + 640, + 473, + 719 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 390, + 640, + 473, + 719 + ], + "spans": [ + { + "bbox": [ + 390, + 640, + 473, + 719 + ], + "score": 0.964, + "html": "
βError (%)
12.65±0.04
0.7 0.42.76±0.16
0.13.04±0.19 3.11±0.16
04.58±1.3
", + "type": "table", + "image_path": "6315ddaee34b410657358ca8e734579c395e25b4c1ee358fa2166235db5b644c.jpg" + } + ] + } + ], + "index": 55.5, + "virtual_lines": [ + { + "bbox": [ + 390, + 640, + 473, + 679.5 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 390, + 679.5, + 473, + 719.0 + ], + "spans": [], + "index": 56 + } + ] + } + ], + "index": 54.0 + }, + { + "type": "text", + "bbox": [ + 105, + 721, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 362, + 734 + ], + "score": 1.0, + "content": "connection branch can be discounted by setting a lower initial", + "type": "text" + }, + { + "bbox": [ + 362, + 721, + 369, + 732 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 719, + 505, + 734 + ], + "score": 1.0, + "content": ". We now evaluate how sensitive", + "type": "text" + } + ], + "index": 57 + } + ], + "index": 57 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 122, + 136, + 487, + 259 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 505, + 133 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 105, + 88, + 505, + 101 + ], + "score": 1.0, + "content": "Table 7: Comparison in various search spaces. We report the lowest error rate of 3 found archi-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "tectures. †: under Chen & Hsieh (2020)’s training settings where all models have 20 layers and", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 110, + 505, + 123 + ], + "score": 1.0, + "content": "36 initial channels (the best is shown in boldface). ‡: under Zela et al. (2020)’s settings where", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 122, + 498, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 122, + 498, + 134 + ], + "score": 1.0, + "content": "CIFAR-100 models have 8 layers and 16 initial channels (The best is in boldface and underlined).", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 122, + 136, + 487, + 259 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 136, + 487, + 259 + ], + "spans": [ + { + "bbox": [ + 122, + 136, + 487, + 259 + ], + "score": 0.984, + "html": "
Benchmark DARTS‡R-DARTS‡DARTSOursPC-DARTS+SDARTS+Ourst
DPL2ESADARSADV
C10S13.843.112.783.013.102.683.112.782.732.68
S24.853.483.313.263.352.633.022.752.652.63
S33.342.932.512.742.592.422.512.532.492.42
S47.203.583.563.714.842.863.022.932.872.86
C100S129.4625.9324.2528.3724.0322.4118.8717.0216.8816.92
S226.0522.3022.2423.2523.5221.6118.2317.5617.2416.14
S328.9022.3623.9923.7323.3721.1318.0517.7317.1215.86
S422.8522.1821.9421.2623.2021.5517.1617.1715.4617.48
", + "type": "table", + "image_path": "aa305b66636e358ec535bafa2706cb424f53534e8f09638a403c57dd4964e9ef.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 122, + 136, + 487, + 177.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 122, + 177.0, + 487, + 218.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 122, + 218.0, + 487, + 259.0 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 106, + 266, + 505, + 333 + ], + "lines": [], + "index": 9.5, + "bbox_fs": [ + 104, + 266, + 506, + 334 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 316, + 437 + ], + "lines": [ + { + "bbox": [ + 106, + 338, + 316, + 349 + ], + "spans": [ + { + "bbox": [ + 106, + 338, + 316, + 349 + ], + "score": 1.0, + "content": "Memory Friendly DARTS (PC-DARTS). To alle-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 349, + 316, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 349, + 316, + 361 + ], + "score": 1.0, + "content": "viate the large memory overhead from the whole su-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 359, + 316, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 359, + 189, + 372 + ], + "score": 1.0, + "content": "pernet, PC-DARTS", + "type": "text" + }, + { + "bbox": [ + 190, + 360, + 203, + 370 + ], + "score": 0.43, + "content": "\\mathrm { { X u } }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 359, + 316, + 372 + ], + "score": 1.0, + "content": "et al., 2020) selects the par-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 370, + 316, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 316, + 383 + ], + "score": 1.0, + "content": "tial channel for searching. The proportion hyper-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 381, + 317, + 394 + ], + "spans": [ + { + "bbox": [ + 104, + 381, + 149, + 394 + ], + "score": 1.0, + "content": "parameter", + "type": "text" + }, + { + "bbox": [ + 150, + 382, + 160, + 392 + ], + "score": 0.78, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 161, + 381, + 317, + 394 + ], + "score": 1.0, + "content": "needs careful calibration to achieve a", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 317, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 317, + 405 + ], + "score": 1.0, + "content": "good result for specific tasks. As a byproduct, the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 403, + 316, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 403, + 285, + 416 + ], + "score": 1.0, + "content": "search time is also reduced to 0.1 GPU days", + "type": "text" + }, + { + "bbox": [ + 286, + 404, + 312, + 415 + ], + "score": 0.81, + "content": "\\left( K { = } 4 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 403, + 316, + 416 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 414, + 316, + 427 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 316, + 427 + ], + "score": 1.0, + "content": "We use their released code and run repeated experi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 317, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 317, + 439 + ], + "score": 1.0, + "content": "ments across different seeds under the same settings.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 338, + 317, + 439 + ] + }, + { + "type": "table", + "bbox": [ + 324, + 398, + 509, + 440 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 323, + 340, + 505, + 395 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 324, + 340, + 505, + 352 + ], + "spans": [ + { + "bbox": [ + 324, + 340, + 505, + 352 + ], + "score": 1.0, + "content": "Table 5: Comparison of PC-DARTS remov-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 324, + 352, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 324, + 352, + 505, + 363 + ], + "score": 1.0, + "content": "ing the strong prior (i.e. channel shuffle) and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 323, + 362, + 505, + 373 + ], + "spans": [ + { + "bbox": [ + 323, + 362, + 505, + 373 + ], + "score": 1.0, + "content": "combining DARTS-. The results are from 3", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 322, + 373, + 506, + 385 + ], + "spans": [ + { + "bbox": [ + 322, + 373, + 506, + 385 + ], + "score": 1.0, + "content": "independent runs on CIFAR-10. The GPU", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 323, + 385, + 479, + 396 + ], + "spans": [ + { + "bbox": [ + 323, + 385, + 479, + 396 + ], + "score": 1.0, + "content": "memory cost is on a batch size of 256.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "table_body", + "bbox": [ + 324, + 398, + 509, + 440 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 324, + 398, + 506, + 440 + ], + "spans": [ + { + "bbox": [ + 324, + 398, + 506, + 440 + ], + "score": 0.963, + "html": "
MethodSettingAcc (%）Memory Cost
PC-DARTS K = 2 97.09±0.1419.9G 股银 3.75h
PC-DARTS- K = 2 97.35±0.0220.8G 3.41h
", + "type": "table", + "image_path": "1d67fc707d5b0e793f7535aa262e7620f1ad64da77a1e0366a7d9dd73cc88f01.jpg" + } + ] + } + ], + "index": 28, + "virtual_lines": [ + { + "bbox": [ + 324, + 398, + 509, + 412.0 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 324, + 412.0, + 509, + 426.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 324, + 426.0, + 509, + 440.0 + ], + "spans": [], + "index": 29 + } + ] + } + ], + "index": 26.0 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 339, + 455 + ], + "score": 1.0, + "content": "To accurately evaluate the role of our method, we choose", + "type": "text" + }, + { + "bbox": [ + 340, + 443, + 361, + 453 + ], + "score": 0.85, + "content": "K { = } 2", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 443, + 505, + 455 + ], + "score": 1.0, + "content": "(a bad configuration in the original", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 506, + 467 + ], + "score": 1.0, + "content": "paper). We make comparisons between the original PC-DARTS and its combination with ours", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 463, + 505, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 479 + ], + "score": 1.0, + "content": "(named PC-DARTS-) in Table 5. PC-DARTS- can marginally boost the CIFAR-10 top-1 accuracy", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 107, + 475, + 143, + 487 + ], + "score": 0.9, + "content": "( + 0 . 2 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "on average). The result also confirms that our method can make PC-DARTS less sensitive", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 486, + 460, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 196, + 500 + ], + "score": 1.0, + "content": "to its hyper-parameter", + "type": "text" + }, + { + "bbox": [ + 197, + 487, + 207, + 496 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 486, + 460, + 500 + ], + "score": 1.0, + "content": "while keeping its advantage of less memory cost and run time.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 443, + 506, + 500 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 512, + 208, + 523 + ], + "lines": [ + { + "bbox": [ + 106, + 511, + 210, + 524 + ], + "spans": [ + { + "bbox": [ + 106, + 511, + 210, + 524 + ], + "score": 1.0, + "content": "4.4 ABLATION STUDY", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 533, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 533, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 475, + 546 + ], + "score": 1.0, + "content": "Robustness to Decay Strategy. Our method is insensitive to the type of decay policy on", + "type": "text" + }, + { + "bbox": [ + 475, + 534, + 483, + 545 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 533, + 505, + 546 + ], + "score": 1.0, + "content": ". We", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 544, + 505, + 557 + ], + "score": 1.0, + "content": "design extra two strategies as comparisons: cosine and step decay. They both have the similar", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 554, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 237, + 569 + ], + "score": 1.0, + "content": "performance. Specifically, when", + "type": "text" + }, + { + "bbox": [ + 238, + 556, + 246, + 567 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 554, + 505, + 569 + ], + "score": 1.0, + "content": "is scheduled to zero by the cosine strategy, the average accuracy", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 563, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 286, + 580 + ], + "score": 1.0, + "content": "of four searched CIFAR-10 models in S3 is", + "type": "text" + }, + { + "bbox": [ + 287, + 566, + 344, + 577 + ], + "score": 0.89, + "content": "9 7 . 3 3 \\% \\pm 0 . 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 344, + 563, + 392, + 580 + ], + "score": 1.0, + "content": ", the best is", + "type": "text" + }, + { + "bbox": [ + 393, + 566, + 425, + 577 + ], + "score": 0.89, + "content": "9 7 . 4 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 563, + 506, + 580 + ], + "score": 1.0, + "content": ". The step decay at", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 576, + 420, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 576, + 176, + 591 + ], + "score": 1.0, + "content": "epoch 45 obtains", + "type": "text" + }, + { + "bbox": [ + 177, + 577, + 209, + 588 + ], + "score": 0.87, + "content": "9 7 . 3 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 209, + 576, + 420, + 591 + ], + "score": 1.0, + "content": "top-1 accuracy on average in the same search space.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 533, + 506, + 591 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 594, + 353, + 704 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 353, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 353, + 605 + ], + "score": 1.0, + "content": "Robustness Comparison on C10 and C100 in S0-S4. To", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 605, + 353, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 353, + 617 + ], + "score": 1.0, + "content": "verify the robustness, it is required to search several times to", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 616, + 353, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 353, + 627 + ], + "score": 1.0, + "content": "report the average performance of derived models (Yu et al.,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 626, + 354, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 626, + 354, + 638 + ], + "score": 1.0, + "content": "2020; Antoine et al., 2020). As shown in Table 1, Table 9", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 637, + 353, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 637, + 353, + 650 + ], + "score": 1.0, + "content": "and Table 7, our method outperforms the recent SOTAs across", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 648, + 354, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 354, + 661 + ], + "score": 1.0, + "content": "several spaces and datasets. Note that SDARTS-ADV utilizes", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 660, + 353, + 672 + ], + "spans": [ + { + "bbox": [ + 106, + 660, + 242, + 672 + ], + "score": 1.0, + "content": "adversarial training and requires", + "type": "text" + }, + { + "bbox": [ + 242, + 660, + 257, + 671 + ], + "score": 0.9, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 660, + 353, + 672 + ], + "score": 1.0, + "content": "more search time than", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 671, + 353, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 277, + 682 + ], + "score": 1.0, + "content": "ours. Especially, we find a good model in", + "type": "text" + }, + { + "bbox": [ + 278, + 671, + 290, + 682 + ], + "score": 0.88, + "content": "S _ { 3 }", + "type": "inline_equation" + }, + { + "bbox": [ + 290, + 671, + 353, + 682 + ], + "score": 1.0, + "content": "on CIFAR-100", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 106, + 681, + 354, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 233, + 693 + ], + "score": 1.0, + "content": "with the lowest top-1 test error", + "type": "text" + }, + { + "bbox": [ + 233, + 681, + 265, + 692 + ], + "score": 0.89, + "content": "1 5 . 8 6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 681, + 354, + 693 + ], + "score": 1.0, + "content": ". The architectures of", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 692, + 281, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 692, + 281, + 705 + ], + "score": 1.0, + "content": "these models can be found in the appendix.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 594, + 354, + 705 + ] + }, + { + "type": "table", + "bbox": [ + 390, + 640, + 473, + 719 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 360, + 594, + 505, + 637 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 360, + 592, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 360, + 592, + 505, + 606 + ], + "score": 1.0, + "content": "Table 6: Searching performance on", + "type": "text" + } + ], + "index": 51 + }, + { + "bbox": [ + 360, + 604, + 505, + 616 + ], + "spans": [ + { + "bbox": [ + 360, + 604, + 505, + 616 + ], + "score": 1.0, + "content": "CIFAR-10 in S3 w.r.t the initial lin-", + "type": "text" + } + ], + "index": 52 + }, + { + "bbox": [ + 360, + 615, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 360, + 615, + 422, + 627 + ], + "score": 1.0, + "content": "ear decay rate", + "type": "text" + }, + { + "bbox": [ + 422, + 616, + 433, + 627 + ], + "score": 0.87, + "content": "\\beta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 615, + 505, + 627 + ], + "score": 1.0, + "content": ". Each setting is", + "type": "text" + } + ], + "index": 53 + }, + { + "bbox": [ + 360, + 626, + 440, + 638 + ], + "spans": [ + { + "bbox": [ + 360, + 626, + 440, + 638 + ], + "score": 1.0, + "content": "run for three times.", + "type": "text" + } + ], + "index": 54 + } + ], + "index": 52.5 + }, + { + "type": "table_body", + "bbox": [ + 390, + 640, + 473, + 719 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 390, + 640, + 473, + 719 + ], + "spans": [ + { + "bbox": [ + 390, + 640, + 473, + 719 + ], + "score": 0.964, + "html": "
βError (%)
12.65±0.04
0.7 0.42.76±0.16
0.13.04±0.19 3.11±0.16
04.58±1.3
", + "type": "table", + "image_path": "6315ddaee34b410657358ca8e734579c395e25b4c1ee358fa2166235db5b644c.jpg" + } + ] + } + ], + "index": 55.5, + "virtual_lines": [ + { + "bbox": [ + 390, + 640, + 473, + 679.5 + ], + "spans": [], + "index": 55 + }, + { + "bbox": [ + 390, + 679.5, + 473, + 719.0 + ], + "spans": [], + "index": 56 + } + ] + } + ], + "index": 54.0 + }, + { + "type": "text", + "bbox": [ + 105, + 721, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 505, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 362, + 734 + ], + "score": 1.0, + "content": "connection branch can be discounted by setting a lower initial", + "type": "text" + }, + { + "bbox": [ + 362, + 721, + 369, + 732 + ], + "score": 0.82, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 719, + 505, + 734 + ], + "score": 1.0, + "content": ". We now evaluate how sensitive", + "type": "text" + } + ], + "index": 57 + }, + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 236, + 95 + ], + "score": 1.0, + "content": "our approach is to the value of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 236, + 83, + 244, + 94 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 244, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ". It’s trivial to see that our approach degrades to DARTS when", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 94, + 134, + 105 + ], + "score": 0.9, + "content": "\\beta = 0", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 134, + 93, + 308, + 106 + ], + "score": 1.0, + "content": ". We compare results when searching with", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 309, + 93, + 406, + 106 + ], + "score": 0.91, + "content": "\\beta \\in \\{ 1 , 0 . 7 , 0 . 4 , 0 . 1 , 0 \\}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 407, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "in Table 6, which show", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 342, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 158, + 117 + ], + "score": 1.0, + "content": "that a bigger", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 159, + 105, + 170, + 116 + ], + "score": 0.88, + "content": "\\beta _ { 0 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 171, + 105, + 342, + 117 + ], + "score": 1.0, + "content": "is advantageous to obtain better networks.", + "type": "text", + "cross_page": true + } + ], + "index": 2 + } + ], + "index": 57, + "bbox_fs": [ + 105, + 719, + 505, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 236, + 95 + ], + "score": 1.0, + "content": "our approach is to the value of", + "type": "text" + }, + { + "bbox": [ + 236, + 83, + 244, + 94 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ". It’s trivial to see that our approach degrades to DARTS when", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 107, + 94, + 134, + 105 + ], + "score": 0.9, + "content": "\\beta = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 93, + 308, + 106 + ], + "score": 1.0, + "content": ". We compare results when searching with", + "type": "text" + }, + { + "bbox": [ + 309, + 93, + 406, + 106 + ], + "score": 0.91, + "content": "\\beta \\in \\{ 1 , 0 . 7 , 0 . 4 , 0 . 1 , 0 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 407, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "in Table 6, which show", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 105, + 342, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 158, + 117 + ], + "score": 1.0, + "content": "that a bigger", + "type": "text" + }, + { + "bbox": [ + 159, + 105, + 170, + 116 + ], + "score": 0.88, + "content": "\\beta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 105, + 342, + 117 + ], + "score": 1.0, + "content": "is advantageous to obtain better networks.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "The choice of auxiliary branch Apart from the default skip connection serving as the auxiliary", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 389, + 145 + ], + "score": 1.0, + "content": "branch, we show that it is also effective to replace it with a learnable", + "type": "text" + }, + { + "bbox": [ + 389, + 132, + 408, + 143 + ], + "score": 0.87, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "convolution projection,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "which is initialized with an identity tensor. The average accuracy of three searched CIFAR-10", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 171, + 167 + ], + "score": 1.0, + "content": "models in S3 is", + "type": "text" + }, + { + "bbox": [ + 172, + 154, + 228, + 165 + ], + "score": 0.89, + "content": "9 7 . 2 5 \\% \\pm 0 . 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 153, + 505, + 167 + ], + "score": 1.0, + "content": ". Akin to the ablation in He et al. (2016), the projection convolution", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "here works in a similar way as the proposed skip connection. This proves the necessity of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 176, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 176, + 188 + ], + "score": 1.0, + "content": "auxiliary branch.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + }, + { + "type": "text", + "bbox": [ + 107, + 192, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "Performance with longer epochs. It is claimed by Bi et al. (2019) that much longer epochs lead", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "to better convergence of the supernet, supposedly beneficial for inferring the final models. However,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 216, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 227 + ], + "score": 1.0, + "content": "many of DARTS variants fail since their final models are full of skip connections. We thus evaluate", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "how our method behaves in such a situation. Specifically, we extend the standard 50 epochs to 150,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "score": 1.0, + "content": "200 and we search 3 independent times each for S0, S2 and S3. Due to the longer epochs, we slightly", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 248, + 260 + ], + "score": 1.0, + "content": "change our decay strategy, we keep", + "type": "text" + }, + { + "bbox": [ + 248, + 248, + 273, + 259 + ], + "score": 0.91, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 248, + 451, + 260 + ], + "score": 1.0, + "content": "all the way until for last 50 epochs we decay", + "type": "text" + }, + { + "bbox": [ + 451, + 248, + 459, + 259 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "to 0. Other", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "hyper-parameters are kept unchanged. The results are shown in Table 8 and the found genotypes are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 268, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 283 + ], + "score": 1.0, + "content": "listed in Figure 17, 18, 19, 20 and 21. It indicates that DARTS- doesn’t suffer from longer epochs", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 246, + 293 + ], + "score": 1.0, + "content": "since it has reasonable values for", + "type": "text" + }, + { + "bbox": [ + 247, + 281, + 264, + 292 + ], + "score": 0.83, + "content": "\\# P", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 281, + 357, + 293 + ], + "score": 1.0, + "content": "compared with those", + "type": "text" + }, + { + "bbox": [ + 357, + 281, + 400, + 292 + ], + "score": 0.87, + "content": "( \\# P = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 281, + 504, + 293 + ], + "score": 1.0, + "content": ") investigated by Bi et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "(2019). Notice S2 and S3 are harder cases where DARTS suffers more severely from the collapse", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "than S0. As a result, DARTS- can successfully survive longer epochs even in challenging search", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "spaces. Noticeably, it is still unclear whether longer epochs can truly boost searching performance.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 325, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 436, + 336 + ], + "score": 1.0, + "content": "Although we achieve a new state-of-the-art result in S2 where the best model has", + "type": "text" + }, + { + "bbox": [ + 437, + 325, + 464, + 335 + ], + "score": 0.87, + "content": "2 . 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 325, + 505, + 336 + ], + "score": 1.0, + "content": "error rate", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 155, + 348 + ], + "score": 1.0, + "content": "(previously", + "type": "text" + }, + { + "bbox": [ + 155, + 336, + 183, + 347 + ], + "score": 0.87, + "content": "2 . 6 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 335, + 352, + 348 + ], + "score": 1.0, + "content": "), it still has worse average performance", + "type": "text" + }, + { + "bbox": [ + 353, + 336, + 408, + 346 + ], + "score": 0.87, + "content": "( 2 . 7 1 { \\pm } 0 . 1 1 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "in S0 than the models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 214, + 358 + ], + "score": 1.0, + "content": "searched with 50 epochs", + "type": "text" + }, + { + "bbox": [ + 214, + 347, + 270, + 358 + ], + "score": 0.89, + "content": "( 2 . 5 9 { \\pm } 0 . 0 8 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 346, + 387, + 358 + ], + "score": 1.0, + "content": ", and the best model in S3", + "type": "text" + }, + { + "bbox": [ + 388, + 347, + 419, + 357 + ], + "score": 0.8, + "content": "( 2 . 5 3 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "is also weaker than", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 357, + 172, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 135, + 370 + ], + "score": 1.0, + "content": "before", + "type": "text" + }, + { + "bbox": [ + 135, + 358, + 167, + 369 + ], + "score": 0.84, + "content": "( 2 . 4 2 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 357, + 172, + 370 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16.5 + }, + { + "type": "table", + "bbox": [ + 106, + 435, + 508, + 492 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 398, + 503, + 432 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "Table 8: Searching performance on CIFAR-10 in S0, S2 and S3 using longer epochs. Following Bi", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 159, + 422 + ], + "score": 1.0, + "content": "et al. (2019),", + "type": "text" + }, + { + "bbox": [ + 160, + 410, + 177, + 421 + ], + "score": 0.79, + "content": "\\# P", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 408, + 505, + 422 + ], + "score": 1.0, + "content": "means the number of parametric operators in the normal cell. Averaged on 3 runs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 420, + 148, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 148, + 432 + ], + "score": 1.0, + "content": "of search.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "table_body", + "bbox": [ + 106, + 435, + 508, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 435, + 508, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 508, + 492 + ], + "score": 0.975, + "html": "
EpochlS0S2S3
#PParams (M)Error (%)#PParams (M)Error (%)#PParams (M)Error (%)
1506.6±1.13.3±0.32.74±0.066.0±0.03.9±0.32.58±0.11|6.0±1.03.6±0.32.55±0.03
2007.3±0.63.2±0.32.71±0.118.0±0.04.3±0.12.65±0.217.6±0.54.3±0.22.66±0.09
", + "type": "table", + "image_path": "873306efd32bdfeaca53a6375bb3173bf3c3b5ce7c9cfe899c8703aa2c321075.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 106, + 435, + 508, + 454.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 106, + 454.0, + 508, + 473.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 106, + 473.0, + 508, + 492.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "Besides, compared with first-order DARTS with a cost of 0.4 GPU days in S0, Amended-DARTS", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "(Bi et al., 2019), particularly designed to survive longer epochs, reports 1.7 GPU days even with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 536, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 504, + 549 + ], + "score": 1.0, + "content": "pruned edges in S0. Our approach has the same cost as first-order DARTS, which is more efficient.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "title", + "bbox": [ + 108, + 576, + 276, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 279, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 279, + 591 + ], + "score": 1.0, + "content": "5 ANALYSIS AND DISCUSSIONS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 107, + 607, + 282, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 284, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 284, + 620 + ], + "score": 1.0, + "content": "5.1 FAILURE OF HESSIAN EIGENVALUE", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 403, + 645 + ], + "score": 1.0, + "content": "The maximal Hessian eigenvalue calculated from the validation loss w.r.t", + "type": "text" + }, + { + "bbox": [ + 404, + 635, + 412, + 643 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "is regarded as an indi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "score": 1.0, + "content": "cator of performance collapse (Zela et al., 2020; Chen & Hsieh, 2020). Surprisingly, our method", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "develops a growing eigenvalue in the majority of configurations, which conflicts with the previous", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "observations. We visualize these statistics across different search space and datasets in Figure 4", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "(A.2.2). Although eigenvalues increase almost monotonously and reach a relatively large value in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "score": 1.0, + "content": "the end, the final models still have good performance that matches with state-of-the-art ones (see Ta-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 504, + 712 + ], + "score": 1.0, + "content": "ble 9). These models can be mistakenly deemed as bad ones or never visited according to the eigen-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "value criteria. Our observations disclose one fatal drawback of these indicator-based approaches:", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 428, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 428, + 733 + ], + "score": 1.0, + "content": "they are prone to rejecting good models. Further analysis can be found in A.2.2.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 115 + ], + "lines": [], + "index": 1, + "bbox_fs": [ + 105, + 82, + 505, + 117 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 121, + 505, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 106, + 120, + 505, + 134 + ], + "score": 1.0, + "content": "The choice of auxiliary branch Apart from the default skip connection serving as the auxiliary", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 131, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 105, + 131, + 389, + 145 + ], + "score": 1.0, + "content": "branch, we show that it is also effective to replace it with a learnable", + "type": "text" + }, + { + "bbox": [ + 389, + 132, + 408, + 143 + ], + "score": 0.87, + "content": "1 \\times 1", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 131, + 505, + 145 + ], + "score": 1.0, + "content": "convolution projection,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "which is initialized with an identity tensor. The average accuracy of three searched CIFAR-10", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 153, + 505, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 153, + 171, + 167 + ], + "score": 1.0, + "content": "models in S3 is", + "type": "text" + }, + { + "bbox": [ + 172, + 154, + 228, + 165 + ], + "score": 0.89, + "content": "9 7 . 2 5 \\% \\pm 0 . 0 9", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 153, + 505, + 167 + ], + "score": 1.0, + "content": ". Akin to the ablation in He et al. (2016), the projection convolution", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 505, + 178 + ], + "score": 1.0, + "content": "here works in a similar way as the proposed skip connection. This proves the necessity of the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 176, + 176, + 188 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 176, + 188 + ], + "score": 1.0, + "content": "auxiliary branch.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5, + "bbox_fs": [ + 105, + 120, + 506, + 188 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 192, + 505, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 206 + ], + "score": 1.0, + "content": "Performance with longer epochs. It is claimed by Bi et al. (2019) that much longer epochs lead", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "to better convergence of the supernet, supposedly beneficial for inferring the final models. However,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 216, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 227 + ], + "score": 1.0, + "content": "many of DARTS variants fail since their final models are full of skip connections. We thus evaluate", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "spans": [ + { + "bbox": [ + 106, + 226, + 505, + 239 + ], + "score": 1.0, + "content": "how our method behaves in such a situation. Specifically, we extend the standard 50 epochs to 150,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 235, + 505, + 250 + ], + "score": 1.0, + "content": "200 and we search 3 independent times each for S0, S2 and S3. Due to the longer epochs, we slightly", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 248, + 506, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 248, + 248, + 260 + ], + "score": 1.0, + "content": "change our decay strategy, we keep", + "type": "text" + }, + { + "bbox": [ + 248, + 248, + 273, + 259 + ], + "score": 0.91, + "content": "\\beta = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 274, + 248, + 451, + 260 + ], + "score": 1.0, + "content": "all the way until for last 50 epochs we decay", + "type": "text" + }, + { + "bbox": [ + 451, + 248, + 459, + 259 + ], + "score": 0.84, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 459, + 248, + 506, + 260 + ], + "score": 1.0, + "content": "to 0. Other", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "hyper-parameters are kept unchanged. The results are shown in Table 8 and the found genotypes are", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 268, + 505, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 283 + ], + "score": 1.0, + "content": "listed in Figure 17, 18, 19, 20 and 21. It indicates that DARTS- doesn’t suffer from longer epochs", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 281, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 246, + 293 + ], + "score": 1.0, + "content": "since it has reasonable values for", + "type": "text" + }, + { + "bbox": [ + 247, + 281, + 264, + 292 + ], + "score": 0.83, + "content": "\\# P", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 281, + 357, + 293 + ], + "score": 1.0, + "content": "compared with those", + "type": "text" + }, + { + "bbox": [ + 357, + 281, + 400, + 292 + ], + "score": 0.87, + "content": "( \\# P = 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 281, + 504, + 293 + ], + "score": 1.0, + "content": ") investigated by Bi et al.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 505, + 304 + ], + "score": 1.0, + "content": "(2019). Notice S2 and S3 are harder cases where DARTS suffers more severely from the collapse", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 316 + ], + "score": 1.0, + "content": "than S0. As a result, DARTS- can successfully survive longer epochs even in challenging search", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 327 + ], + "score": 1.0, + "content": "spaces. Noticeably, it is still unclear whether longer epochs can truly boost searching performance.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 325, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 106, + 325, + 436, + 336 + ], + "score": 1.0, + "content": "Although we achieve a new state-of-the-art result in S2 where the best model has", + "type": "text" + }, + { + "bbox": [ + 437, + 325, + 464, + 335 + ], + "score": 0.87, + "content": "2 . 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 325, + 505, + 336 + ], + "score": 1.0, + "content": "error rate", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 155, + 348 + ], + "score": 1.0, + "content": "(previously", + "type": "text" + }, + { + "bbox": [ + 155, + 336, + 183, + 347 + ], + "score": 0.87, + "content": "2 . 6 3 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 184, + 335, + 352, + 348 + ], + "score": 1.0, + "content": "), it still has worse average performance", + "type": "text" + }, + { + "bbox": [ + 353, + 336, + 408, + 346 + ], + "score": 0.87, + "content": "( 2 . 7 1 { \\pm } 0 . 1 1 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "in S0 than the models", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 346, + 505, + 358 + ], + "spans": [ + { + "bbox": [ + 105, + 346, + 214, + 358 + ], + "score": 1.0, + "content": "searched with 50 epochs", + "type": "text" + }, + { + "bbox": [ + 214, + 347, + 270, + 358 + ], + "score": 0.89, + "content": "( 2 . 5 9 { \\pm } 0 . 0 8 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 270, + 346, + 387, + 358 + ], + "score": 1.0, + "content": ", and the best model in S3", + "type": "text" + }, + { + "bbox": [ + 388, + 347, + 419, + 357 + ], + "score": 0.8, + "content": "( 2 . 5 3 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 346, + 505, + 358 + ], + "score": 1.0, + "content": "is also weaker than", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 357, + 172, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 135, + 370 + ], + "score": 1.0, + "content": "before", + "type": "text" + }, + { + "bbox": [ + 135, + 358, + 167, + 369 + ], + "score": 0.84, + "content": "( 2 . 4 2 \\% )", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 357, + 172, + 370 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 192, + 506, + 370 + ] + }, + { + "type": "table", + "bbox": [ + 106, + 435, + 508, + 492 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 108, + 398, + 503, + 432 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "Table 8: Searching performance on CIFAR-10 in S0, S2 and S3 using longer epochs. Following Bi", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 408, + 505, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 159, + 422 + ], + "score": 1.0, + "content": "et al. (2019),", + "type": "text" + }, + { + "bbox": [ + 160, + 410, + 177, + 421 + ], + "score": 0.79, + "content": "\\# P", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 408, + 505, + 422 + ], + "score": 1.0, + "content": "means the number of parametric operators in the normal cell. Averaged on 3 runs", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 420, + 148, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 148, + 432 + ], + "score": 1.0, + "content": "of search.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26 + }, + { + "type": "table_body", + "bbox": [ + 106, + 435, + 508, + 492 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 435, + 508, + 492 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 508, + 492 + ], + "score": 0.975, + "html": "
EpochlS0S2S3
#PParams (M)Error (%)#PParams (M)Error (%)#PParams (M)Error (%)
1506.6±1.13.3±0.32.74±0.066.0±0.03.9±0.32.58±0.11|6.0±1.03.6±0.32.55±0.03
2007.3±0.63.2±0.32.71±0.118.0±0.04.3±0.12.65±0.217.6±0.54.3±0.22.66±0.09
", + "type": "table", + "image_path": "873306efd32bdfeaca53a6375bb3173bf3c3b5ce7c9cfe899c8703aa2c321075.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 106, + 435, + 508, + 454.0 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 106, + 454.0, + 508, + 473.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 106, + 473.0, + 508, + 492.0 + ], + "spans": [], + "index": 30 + } + ] + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 107, + 515, + 505, + 548 + ], + "lines": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 527 + ], + "score": 1.0, + "content": "Besides, compared with first-order DARTS with a cost of 0.4 GPU days in S0, Amended-DARTS", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 505, + 538 + ], + "score": 1.0, + "content": "(Bi et al., 2019), particularly designed to survive longer epochs, reports 1.7 GPU days even with", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 536, + 504, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 536, + 504, + 549 + ], + "score": 1.0, + "content": "pruned edges in S0. Our approach has the same cost as first-order DARTS, which is more efficient.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 514, + 505, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 576, + 276, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 279, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 279, + 591 + ], + "score": 1.0, + "content": "5 ANALYSIS AND DISCUSSIONS", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 107, + 607, + 282, + 619 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 284, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 284, + 620 + ], + "score": 1.0, + "content": "5.1 FAILURE OF HESSIAN EIGENVALUE", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 106, + 632, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 403, + 645 + ], + "score": 1.0, + "content": "The maximal Hessian eigenvalue calculated from the validation loss w.r.t", + "type": "text" + }, + { + "bbox": [ + 404, + 635, + 412, + 643 + ], + "score": 0.76, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "is regarded as an indi-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "score": 1.0, + "content": "cator of performance collapse (Zela et al., 2020; Chen & Hsieh, 2020). Surprisingly, our method", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "develops a growing eigenvalue in the majority of configurations, which conflicts with the previous", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "observations. We visualize these statistics across different search space and datasets in Figure 4", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "(A.2.2). Although eigenvalues increase almost monotonously and reach a relatively large value in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 686, + 505, + 701 + ], + "score": 1.0, + "content": "the end, the final models still have good performance that matches with state-of-the-art ones (see Ta-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 698, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 504, + 712 + ], + "score": 1.0, + "content": "ble 9). These models can be mistakenly deemed as bad ones or never visited according to the eigen-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "value criteria. Our observations disclose one fatal drawback of these indicator-based approaches:", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 720, + 428, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 428, + 733 + ], + "score": 1.0, + "content": "they are prone to rejecting good models. Further analysis can be found in A.2.2.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 291, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 293, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 293, + 95 + ], + "score": 1.0, + "content": "5.2 VALIDATION ACCURACY LANDSCAPE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "Recent works, R-DARTS (Zela et al., 2020) and SDARTS (Chen & Hsieh, 2020) point that the ar-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "chitectural weights are expected to converge to an optimal point where accuracy is insensitive to per-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "turbations to obtain a stable architecture after the discretization process, i.e., the convergence point", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "should have a smooth landscape. SDARTS proposes a perturbation-based regularization, which", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "further stabilizes the searching process of DARTS. However, the perturbation regularization dis-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 173 + ], + "score": 1.0, + "content": "turbs the training procedure and thus misleads the update of architectural weights. Different from", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "score": 1.0, + "content": "SDARTS that explicitly smooths landscape by perturbation, DARTS- can implicitly do the same", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 304, + 196 + ], + "score": 1.0, + "content": "without directly perturbing architectural weights.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5 + }, + { + "type": "image", + "bbox": [ + 114, + 221, + 497, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 221, + 497, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 221, + 497, + 302 + ], + "spans": [ + { + "bbox": [ + 114, + 221, + 497, + 302 + ], + "score": 0.967, + "type": "image", + "image_path": "9209d9b128c595cbb45eaac137aa78868321675acd149e11fd78d5b77f3ab1a8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 114, + 221, + 497, + 248.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 114, + 248.0, + 497, + 275.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 114, + 275.0, + 497, + 302.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 314, + 505, + 359 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 314, + 504, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 496, + 327 + ], + "score": 1.0, + "content": "Figure 3: Comparison of the validation accuracy landscape of (a) DARTS and (b) DARTS- w.r.t.", + "type": "text" + }, + { + "bbox": [ + 496, + 317, + 504, + 325 + ], + "score": 0.6, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "on CIFAR-10 in S3. Their contour maps are shown respectively in (c) and (d), where we set the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 377, + 349 + ], + "score": 1.0, + "content": "step of contour map as 0.1. The accuracy of the derived models are", + "type": "text" + }, + { + "bbox": [ + 377, + 336, + 409, + 347 + ], + "score": 0.88, + "content": "9 4 . 8 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 336, + 447, + 349 + ], + "score": 1.0, + "content": "(a,c) and", + "type": "text" + }, + { + "bbox": [ + 448, + 336, + 480, + 347 + ], + "score": 0.86, + "content": "9 7 . 5 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "(b,d),", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 347, + 407, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 407, + 360 + ], + "score": 1.0, + "content": "while the maximum Hessian eigenvalues are similarly high (0.52 and 0.42)", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + } + ], + "index": 11.75 + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "To analyze the efficacy of DARTS-, we plot the validation accuracy landscape w.r.t architectural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 140, + 401 + ], + "score": 1.0, + "content": "weights", + "type": "text" + }, + { + "bbox": [ + 140, + 390, + 148, + 398 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 388, + 505, + 401 + ], + "score": 1.0, + "content": ", and find that auxiliary connection smooths the landscape and thus stabilizes the search-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 397, + 504, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 397, + 496, + 413 + ], + "score": 1.0, + "content": "ing stage. Specifically, we choose two random directions and apply normalized perturbation on", + "type": "text" + }, + { + "bbox": [ + 496, + 401, + 504, + 409 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "(following Li et al. 2018a). As shown in Figure 3, DARTS- is less sensitive to the perturbation than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 420, + 366, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 366, + 434 + ], + "score": 1.0, + "content": "DARTS, and the contour map of DARTS- descends more gently.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18 + }, + { + "type": "title", + "bbox": [ + 107, + 456, + 195, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 454, + 197, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 197, + 471 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "We propose a simple and effective approach named DARTS- to address the performance collapse in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "differentiable architecture search. Its core idea is to make use of an auxiliary skip connection branch", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "to take over the gradient advantage role from the candidate skip connection operation. This can", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "create a fair competition where the bi-level optimization process can easily differentiate good oper-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "ations from the bad. As a result, the search process is more stable and the collapse seldom happens", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 541, + 504, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 504, + 553 + ], + "score": 1.0, + "content": "across various search spaces and different datasets. Under strictly controlled settings, it steadily", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 412, + 563 + ], + "score": 1.0, + "content": "outperforms recent state-of-the-art RobustDARTS (Zela et al., 2020) with", + "type": "text" + }, + { + "bbox": [ + 412, + 552, + 427, + 562 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "fewer search cost.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "Moreover, our method disapproves of various handcrafted regularization tricks. Last but not least, it", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 572, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 482, + 587 + ], + "score": 1.0, + "content": "can be used stand-alone or in cooperation with various orthogonal improvements if necessary.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "This paper conveys two important messages for future research. On the one hand, the Hessian", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "eigenvalue indicator for performance collapse (Zela et al., 2020; Chen & Hsieh, 2020) is not ideal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "because it has a risk of rejecting good models. On the other hand, handcraft regularization tricks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "score": 1.0, + "content": "(Chen et al., 2019b) seem more critical to search a good model instead of the proposed methods.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "Then what’s the solution? In principle, it’s difficult to find a perfect indicator of the collapse. Our", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "approach shows the potential to control the search process and doesn’t impose limitations or priors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 656, + 381, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 381, + 668 + ], + "score": 1.0, + "content": "on the final model. We hope more attention be paid in this direction.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34 + }, + { + "type": "title", + "bbox": [ + 108, + 691, + 237, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 690, + 239, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 239, + 705 + ], + "score": 1.0, + "content": "7 ACKNOWLEDGEMENT", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 720, + 271, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 718, + 273, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 273, + 735 + ], + "score": 1.0, + "content": "This research was supported by Meituan.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 291, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 293, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 293, + 95 + ], + "score": 1.0, + "content": "5.2 VALIDATION ACCURACY LANDSCAPE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "Recent works, R-DARTS (Zela et al., 2020) and SDARTS (Chen & Hsieh, 2020) point that the ar-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 129 + ], + "score": 1.0, + "content": "chitectural weights are expected to converge to an optimal point where accuracy is insensitive to per-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "turbations to obtain a stable architecture after the discretization process, i.e., the convergence point", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 139, + 505, + 151 + ], + "score": 1.0, + "content": "should have a smooth landscape. SDARTS proposes a perturbation-based regularization, which", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 506, + 163 + ], + "score": 1.0, + "content": "further stabilizes the searching process of DARTS. However, the perturbation regularization dis-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 161, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 161, + 506, + 173 + ], + "score": 1.0, + "content": "turbs the training procedure and thus misleads the update of architectural weights. Different from", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 506, + 185 + ], + "score": 1.0, + "content": "SDARTS that explicitly smooths landscape by perturbation, DARTS- can implicitly do the same", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 304, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 304, + 196 + ], + "score": 1.0, + "content": "without directly perturbing architectural weights.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 105, + 506, + 196 + ] + }, + { + "type": "image", + "bbox": [ + 114, + 221, + 497, + 302 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 221, + 497, + 302 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 221, + 497, + 302 + ], + "spans": [ + { + "bbox": [ + 114, + 221, + 497, + 302 + ], + "score": 0.967, + "type": "image", + "image_path": "9209d9b128c595cbb45eaac137aa78868321675acd149e11fd78d5b77f3ab1a8.jpg" + } + ] + } + ], + "index": 10, + "virtual_lines": [ + { + "bbox": [ + 114, + 221, + 497, + 248.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 114, + 248.0, + 497, + 275.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 114, + 275.0, + 497, + 302.0 + ], + "spans": [], + "index": 11 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 314, + 505, + 359 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 314, + 504, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 496, + 327 + ], + "score": 1.0, + "content": "Figure 3: Comparison of the validation accuracy landscape of (a) DARTS and (b) DARTS- w.r.t.", + "type": "text" + }, + { + "bbox": [ + 496, + 317, + 504, + 325 + ], + "score": 0.6, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "on CIFAR-10 in S3. Their contour maps are shown respectively in (c) and (d), where we set the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 336, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 377, + 349 + ], + "score": 1.0, + "content": "step of contour map as 0.1. The accuracy of the derived models are", + "type": "text" + }, + { + "bbox": [ + 377, + 336, + 409, + 347 + ], + "score": 0.88, + "content": "9 4 . 8 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 336, + 447, + 349 + ], + "score": 1.0, + "content": "(a,c) and", + "type": "text" + }, + { + "bbox": [ + 448, + 336, + 480, + 347 + ], + "score": 0.86, + "content": "9 7 . 5 8 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 480, + 336, + 505, + 349 + ], + "score": 1.0, + "content": "(b,d),", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 347, + 407, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 407, + 360 + ], + "score": 1.0, + "content": "while the maximum Hessian eigenvalues are similarly high (0.52 and 0.42)", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 13.5 + } + ], + "index": 11.75 + }, + { + "type": "text", + "bbox": [ + 107, + 376, + 505, + 432 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "To analyze the efficacy of DARTS-, we plot the validation accuracy landscape w.r.t architectural", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 106, + 388, + 140, + 401 + ], + "score": 1.0, + "content": "weights", + "type": "text" + }, + { + "bbox": [ + 140, + 390, + 148, + 398 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 388, + 505, + 401 + ], + "score": 1.0, + "content": ", and find that auxiliary connection smooths the landscape and thus stabilizes the search-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 397, + 504, + 413 + ], + "spans": [ + { + "bbox": [ + 104, + 397, + 496, + 413 + ], + "score": 1.0, + "content": "ing stage. Specifically, we choose two random directions and apply normalized perturbation on", + "type": "text" + }, + { + "bbox": [ + 496, + 401, + 504, + 409 + ], + "score": 0.73, + "content": "\\alpha", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 505, + 423 + ], + "score": 1.0, + "content": "(following Li et al. 2018a). As shown in Figure 3, DARTS- is less sensitive to the perturbation than", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 420, + 366, + 434 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 366, + 434 + ], + "score": 1.0, + "content": "DARTS, and the contour map of DARTS- descends more gently.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18, + "bbox_fs": [ + 104, + 376, + 505, + 434 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 456, + 195, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 454, + 197, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 197, + 471 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 485, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "We propose a simple and effective approach named DARTS- to address the performance collapse in", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 505, + 509 + ], + "score": 1.0, + "content": "differentiable architecture search. Its core idea is to make use of an auxiliary skip connection branch", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 505, + 520 + ], + "score": 1.0, + "content": "to take over the gradient advantage role from the candidate skip connection operation. This can", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "create a fair competition where the bi-level optimization process can easily differentiate good oper-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "ations from the bad. As a result, the search process is more stable and the collapse seldom happens", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 541, + 504, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 504, + 553 + ], + "score": 1.0, + "content": "across various search spaces and different datasets. Under strictly controlled settings, it steadily", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 551, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 412, + 563 + ], + "score": 1.0, + "content": "outperforms recent state-of-the-art RobustDARTS (Zela et al., 2020) with", + "type": "text" + }, + { + "bbox": [ + 412, + 552, + 427, + 562 + ], + "score": 0.86, + "content": "3 \\times", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 551, + 505, + 563 + ], + "score": 1.0, + "content": "fewer search cost.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 575 + ], + "score": 1.0, + "content": "Moreover, our method disapproves of various handcrafted regularization tricks. Last but not least, it", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 572, + 482, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 482, + 587 + ], + "score": 1.0, + "content": "can be used stand-alone or in cooperation with various orthogonal improvements if necessary.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 486, + 506, + 587 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 590, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "This paper conveys two important messages for future research. On the one hand, the Hessian", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "eigenvalue indicator for performance collapse (Zela et al., 2020; Chen & Hsieh, 2020) is not ideal", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 505, + 624 + ], + "score": 1.0, + "content": "because it has a risk of rejecting good models. On the other hand, handcraft regularization tricks", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 505, + 635 + ], + "score": 1.0, + "content": "(Chen et al., 2019b) seem more critical to search a good model instead of the proposed methods.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 505, + 646 + ], + "score": 1.0, + "content": "Then what’s the solution? In principle, it’s difficult to find a perfect indicator of the collapse. Our", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "approach shows the potential to control the search process and doesn’t impose limitations or priors", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 656, + 381, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 381, + 668 + ], + "score": 1.0, + "content": "on the final model. We hope more attention be paid in this direction.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 34, + "bbox_fs": [ + 105, + 590, + 506, + 668 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 691, + 237, + 703 + ], + "lines": [ + { + "bbox": [ + 105, + 690, + 239, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 239, + 705 + ], + "score": 1.0, + "content": "7 ACKNOWLEDGEMENT", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 108, + 720, + 271, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 718, + 273, + 735 + ], + "spans": [ + { + "bbox": [ + 106, + 718, + 273, + 735 + ], + "score": 1.0, + "content": "This research was supported by Meituan.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 39, + "bbox_fs": [ + 106, + 718, + 273, + 735 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 99, + 504, + 122 + ], + "lines": [ + { + "bbox": [ + 106, + 99, + 504, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 504, + 112 + ], + "score": 1.0, + "content": "Yang Antoine, Esperanc¸a Pedro M., and Carlucci Fabio M. NAS Evaluation is Frustratingly Hard.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 179, + 122 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 179, + 122 + ], + "score": 1.0, + "content": "In ICLR, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 107, + 128, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "Kaifeng Bi, Changping Hu, Lingxi Xie, Xin Chen, Longhui Wei, and Qi Tian. Stabilizing darts", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 116, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "with amended gradient estimation on architectural parameters. arXiv preprint arXiv:1910.11831,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 150, + 142, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 142, + 163 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 504, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 182 + ], + "score": 1.0, + "content": "Han Cai, Ligeng Zhu, and Song Han. ProxylessNAS: Direct Neural Architecture Search on Target", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 180, + 262, + 191 + ], + "spans": [ + { + "bbox": [ + 116, + 180, + 262, + 191 + ], + "score": 1.0, + "content": "Task and Hardware. In ICLR, 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 506, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 210, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 116, + 210, + 504, + 221 + ], + "score": 1.0, + "content": "Feng, Ziwei Liu, Jiarui Xu, et al. MMDetection: Open mmlab detection toolbox and benchmark.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 220, + 283, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 220, + 283, + 232 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1906.07155, 2019a.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 503, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Xiangning Chen and Cho-Jui Hsieh. Stabilizing differentiable architecture search via perturbation-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 267, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 267, + 262 + ], + "score": 1.0, + "content": "based regularization. In ICML, 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 281 + ], + "score": 1.0, + "content": "Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive Differentiable Architecture Search: Bridg-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 279, + 392, + 291 + ], + "spans": [ + { + "bbox": [ + 115, + 279, + 392, + 291 + ], + "score": 1.0, + "content": "ing the Depth Gap between Search and Evaluation. In ICCV, 2019b.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 311 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Bo Zhang, Jixiang Li, Qingyuan Li, and Ruijun Xu. Scarletnas: Bridging the gap", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "between scalability and fairness in neural architecture search. arXiv preprint arXiv:1908.06022,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 317, + 147, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 147, + 331 + ], + "score": 1.0, + "content": "2019a.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. FairNAS: Rethinking Evaluation Fairness", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 348, + 476, + 360 + ], + "spans": [ + { + "bbox": [ + 116, + 348, + 476, + 360 + ], + "score": 1.0, + "content": "of Weight Sharing Neural Architecture Search. arXiv preprint. arXiv:1907.01845, 2019b.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Bo Zhang, and Ruijun Xu. MoGA: Searching Beyond MobileNetV3. In ICASSP,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 377, + 211, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 377, + 211, + 389 + ], + "score": 1.0, + "content": "pp. 4042–4046, 2020a.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 409 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Tianbao Zhou, Bo Zhang, and Jixiang Li. Fair darts: Eliminating unfair advantages", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 406, + 335, + 417 + ], + "spans": [ + { + "bbox": [ + 117, + 406, + 335, + 417 + ], + "score": 1.0, + "content": "in differentiable architecture search. In ECCV, 2020b.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "score": 1.0, + "content": "Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. AutoAugment:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 436, + 361, + 448 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 361, + 448 + ], + "score": 1.0, + "content": "Learning Augmentation Policies from Data. In CVPR, 2019.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 453, + 504, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Xuanyi Dong and Yi Yang. One-shot neural architecture search via self-evaluated template network.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE International Conference on Computer Vision, pp. 3681–3690, 2019a.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "Xuanyi Dong and Yi Yang. Searching for a Robust Neural Architecture in Four GPU Hours. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 494, + 241, + 506 + ], + "spans": [ + { + "bbox": [ + 116, + 494, + 241, + 506 + ], + "score": 1.0, + "content": "CVPR, pp. 1761–1770, 2019b.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Xuanyi Dong and Yi Yang. Nas-bench-201: Extending the scope of reproducible neural architecture", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "search. In International Conference on Learning Representations (ICLR), 2020. URL https:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 117, + 534, + 336, + 547 + ], + "spans": [ + { + "bbox": [ + 117, + 534, + 266, + 547 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 536, + 273, + 544 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 534, + 336, + 547 + ], + "score": 1.0, + "content": "HJxyZkBKDr.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. NAS-FPN: Learning scalable feature pyramid archi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 564, + 360, + 576 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 360, + 576 + ], + "score": 1.0, + "content": "tecture for object detection. In CVPR, pp. 7036–7045, 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 504, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 596 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 592, + 292, + 605 + ], + "spans": [ + { + "bbox": [ + 116, + 592, + 292, + 605 + ], + "score": 1.0, + "content": "Recognition. In CVPR, pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for MobileNetV3. In ICCV,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 632, + 144, + 645 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 144, + 645 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 104, + 650, + 503, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "Jie Hu, Li Shen, and Gang Sun. Squeeze-and-Excitation Networks. In CVPR, pp. 7132–7141, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 502, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 504, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 504, + 682 + ], + "score": 1.0, + "content": "Guohao Li, Guocheng Qian, Itzel C Delgadillo, Matthias Muller, Ali Thabet, and Bernard Ghanem. ¨", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 680, + 362, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 362, + 693 + ], + "score": 1.0, + "content": "Sgas: Sequential greedy architecture search. In CVPR, 2020.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss land-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "scape of neural nets. In Advances in Neural Information Processing Systems, pp. 6389–6399,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 718, + 147, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 718, + 147, + 733 + ], + "score": 1.0, + "content": "2018a.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 99, + 504, + 122 + ], + "lines": [ + { + "bbox": [ + 106, + 99, + 504, + 112 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 504, + 112 + ], + "score": 1.0, + "content": "Yang Antoine, Esperanc¸a Pedro M., and Carlucci Fabio M. NAS Evaluation is Frustratingly Hard.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 110, + 179, + 122 + ], + "spans": [ + { + "bbox": [ + 115, + 110, + 179, + 122 + ], + "score": 1.0, + "content": "In ICLR, 2020.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 106, + 99, + 504, + 122 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 128, + 505, + 162 + ], + "lines": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 129, + 505, + 141 + ], + "score": 1.0, + "content": "Kaifeng Bi, Changping Hu, Lingxi Xie, Xin Chen, Longhui Wei, and Qi Tian. Stabilizing darts", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 140, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 116, + 140, + 505, + 152 + ], + "score": 1.0, + "content": "with amended gradient estimation on architectural parameters. arXiv preprint arXiv:1910.11831,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 150, + 142, + 163 + ], + "spans": [ + { + "bbox": [ + 115, + 150, + 142, + 163 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 106, + 129, + 505, + 163 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 168, + 504, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 167, + 505, + 182 + ], + "spans": [ + { + "bbox": [ + 105, + 167, + 505, + 182 + ], + "score": 1.0, + "content": "Han Cai, Ligeng Zhu, and Song Han. ProxylessNAS: Direct Neural Architecture Search on Target", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 180, + 262, + 191 + ], + "spans": [ + { + "bbox": [ + 116, + 180, + 262, + 191 + ], + "score": 1.0, + "content": "Task and Hardware. In ICLR, 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 167, + 505, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 198, + 506, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 198, + 505, + 211 + ], + "score": 1.0, + "content": "Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 210, + 504, + 221 + ], + "spans": [ + { + "bbox": [ + 116, + 210, + 504, + 221 + ], + "score": 1.0, + "content": "Feng, Ziwei Liu, Jiarui Xu, et al. MMDetection: Open mmlab detection toolbox and benchmark.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 220, + 283, + 232 + ], + "spans": [ + { + "bbox": [ + 116, + 220, + 283, + 232 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1906.07155, 2019a.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 198, + 505, + 232 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 503, + 261 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 252 + ], + "score": 1.0, + "content": "Xiangning Chen and Cho-Jui Hsieh. Stabilizing differentiable architecture search via perturbation-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 267, + 262 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 267, + 262 + ], + "score": 1.0, + "content": "based regularization. In ICML, 2020.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 237, + 505, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 267, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 105, + 266, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 505, + 281 + ], + "score": 1.0, + "content": "Xin Chen, Lingxi Xie, Jun Wu, and Qi Tian. Progressive Differentiable Architecture Search: Bridg-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 279, + 392, + 291 + ], + "spans": [ + { + "bbox": [ + 115, + 279, + 392, + 291 + ], + "score": 1.0, + "content": "ing the Depth Gap between Search and Evaluation. In ICCV, 2019b.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 266, + 505, + 291 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 296, + 505, + 330 + ], + "lines": [ + { + "bbox": [ + 105, + 295, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 295, + 505, + 311 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Bo Zhang, Jixiang Li, Qingyuan Li, and Ruijun Xu. Scarletnas: Bridging the gap", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 308, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 115, + 308, + 505, + 320 + ], + "score": 1.0, + "content": "between scalability and fairness in neural architecture search. arXiv preprint arXiv:1908.06022,", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 317, + 147, + 331 + ], + "spans": [ + { + "bbox": [ + 115, + 317, + 147, + 331 + ], + "score": 1.0, + "content": "2019a.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 295, + 505, + 331 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 336, + 504, + 360 + ], + "lines": [ + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "spans": [ + { + "bbox": [ + 106, + 337, + 506, + 350 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Bo Zhang, Ruijun Xu, and Jixiang Li. FairNAS: Rethinking Evaluation Fairness", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 116, + 348, + 476, + 360 + ], + "spans": [ + { + "bbox": [ + 116, + 348, + 476, + 360 + ], + "score": 1.0, + "content": "of Weight Sharing Neural Architecture Search. arXiv preprint. arXiv:1907.01845, 2019b.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 337, + 506, + 360 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 506, + 379 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Bo Zhang, and Ruijun Xu. MoGA: Searching Beyond MobileNetV3. In ICASSP,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 377, + 211, + 389 + ], + "spans": [ + { + "bbox": [ + 115, + 377, + 211, + 389 + ], + "score": 1.0, + "content": "pp. 4042–4046, 2020a.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 365, + 506, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 504, + 418 + ], + "lines": [ + { + "bbox": [ + 105, + 394, + 506, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 506, + 409 + ], + "score": 1.0, + "content": "Xiangxiang Chu, Tianbao Zhou, Bo Zhang, and Jixiang Li. Fair darts: Eliminating unfair advantages", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 117, + 406, + 335, + 417 + ], + "spans": [ + { + "bbox": [ + 117, + 406, + 335, + 417 + ], + "score": 1.0, + "content": "in differentiable architecture search. In ECCV, 2020b.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 394, + 506, + 417 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 504, + 447 + ], + "lines": [ + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 504, + 438 + ], + "score": 1.0, + "content": "Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. AutoAugment:", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 436, + 361, + 448 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 361, + 448 + ], + "score": 1.0, + "content": "Learning Augmentation Policies from Data. In CVPR, 2019.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 425, + 504, + 448 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 453, + 504, + 477 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "Xuanyi Dong and Yi Yang. One-shot neural architecture search via self-evaluated template network.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 465, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 116, + 465, + 505, + 478 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE International Conference on Computer Vision, pp. 3681–3690, 2019a.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 106, + 453, + 505, + 478 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 482, + 504, + 506 + ], + "lines": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "Xuanyi Dong and Yi Yang. Searching for a Robust Neural Architecture in Four GPU Hours. In", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 494, + 241, + 506 + ], + "spans": [ + { + "bbox": [ + 116, + 494, + 241, + 506 + ], + "score": 1.0, + "content": "CVPR, pp. 1761–1770, 2019b.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 482, + 506, + 506 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 546 + ], + "lines": [ + { + "bbox": [ + 106, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 106, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Xuanyi Dong and Yi Yang. Nas-bench-201: Extending the scope of reproducible neural architecture", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 523, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 115, + 523, + 505, + 537 + ], + "score": 1.0, + "content": "search. In International Conference on Learning Representations (ICLR), 2020. URL https:", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 117, + 534, + 336, + 547 + ], + "spans": [ + { + "bbox": [ + 117, + 534, + 266, + 547 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 536, + 273, + 544 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 534, + 336, + 547 + ], + "score": 1.0, + "content": "HJxyZkBKDr.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 106, + 512, + 505, + 547 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 551, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V Le. NAS-FPN: Learning scalable feature pyramid archi-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 564, + 360, + 576 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 360, + 576 + ], + "score": 1.0, + "content": "tecture for object detection. In CVPR, pp. 7036–7045, 2019.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 106, + 552, + 505, + 576 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 581, + 504, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 580, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 505, + 596 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 592, + 292, + 605 + ], + "spans": [ + { + "bbox": [ + 116, + 592, + 292, + 605 + ], + "score": 1.0, + "content": "Recognition. In CVPR, pp. 770–778, 2016.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 580, + 505, + 605 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 610, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 505, + 624 + ], + "score": 1.0, + "content": "Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 621, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 506, + 635 + ], + "score": 1.0, + "content": "Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for MobileNetV3. In ICCV,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 115, + 632, + 144, + 645 + ], + "spans": [ + { + "bbox": [ + 115, + 632, + 144, + 645 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 609, + 506, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 650, + 503, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 505, + 664 + ], + "score": 1.0, + "content": "Jie Hu, Li Shen, and Gang Sun. Squeeze-and-Excitation Networks. In CVPR, pp. 7132–7141, 2018.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 650, + 505, + 664 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 668, + 502, + 692 + ], + "lines": [ + { + "bbox": [ + 105, + 668, + 504, + 682 + ], + "spans": [ + { + "bbox": [ + 105, + 668, + 504, + 682 + ], + "score": 1.0, + "content": "Guohao Li, Guocheng Qian, Itzel C Delgadillo, Matthias Muller, Ali Thabet, and Bernard Ghanem. ¨", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 116, + 680, + 362, + 693 + ], + "spans": [ + { + "bbox": [ + 116, + 680, + 362, + 693 + ], + "score": 1.0, + "content": "Sgas: Sequential greedy architecture search. In CVPR, 2020.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41.5, + "bbox_fs": [ + 105, + 668, + 504, + 693 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 504, + 711 + ], + "score": 1.0, + "content": "Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss land-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 115, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "scape of neural nets. In Advances in Neural Information Processing Systems, pp. 6389–6399,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 718, + 147, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 718, + 147, + 733 + ], + "score": 1.0, + "content": "2018a.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 698, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "score": 1.0, + "content": "Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 108 + ], + "score": 1.0, + "content": "super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 186, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 186, + 116 + ], + "score": 1.0, + "content": "517–532, 2018b.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 123, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Hanwen Liang, Shifeng Zhang, Jiacheng Sun, Xingqiu He, Weiran Huang, Kechen Zhuang, and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 202, + 147 + ], + "score": 1.0, + "content": "Zhenguo Li. DARTS", + "type": "text" + }, + { + "bbox": [ + 202, + 136, + 208, + 144 + ], + "score": 0.32, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 134, + 505, + 147 + ], + "score": 1.0, + "content": ": Improved Differentiable Architecture Search with Early Stopping. arXiv", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 146, + 254, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 146, + 254, + 157 + ], + "score": 1.0, + "content": "preprint arXiv:1909.06035, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 503, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 319, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 319, + 187 + ], + "score": 1.0, + "content": "Object Detection. In ICCV, pp. 2980–2988, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "Yuille, Jonathan Huang, and Kevin Murphy. Progressive Neural Architecture Search. In ECCV,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 216, + 187, + 228 + ], + "spans": [ + { + "bbox": [ + 114, + 216, + 187, + 228 + ], + "score": 1.0, + "content": "pp. 19–34, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 106, + 235, + 506, + 269 + ], + "lines": [ + { + "bbox": [ + 107, + 235, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 107, + 235, + 504, + 246 + ], + "score": 1.0, + "content": "Chenxi Liu, Liang-Chieh Chen, Florian Schroff, Hartwig Adam, Wei Hua, Alan L Yuille, and Li Fei-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "Fei. Auto-deeplab: Hierarchical neural architecture search for semantic image segmentation. In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 257, + 221, + 270 + ], + "spans": [ + { + "bbox": [ + 117, + 257, + 221, + 270 + ], + "score": 1.0, + "content": "CVPR, pp. 82–92, 2019a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 504, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable Architecture Search. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 286, + 173, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 286, + 173, + 299 + ], + "score": 1.0, + "content": "ICLR, 2019b.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 504, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Jieru Mei, Yingwei Li, Xiaochen Lian, Xiaojie Jin, Linjie Yang, Alan Yuille, and Yang Jianchao.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 317, + 445, + 329 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 445, + 329 + ], + "score": 1.0, + "content": "AtomNAS: Fine-Grained End-to-End Neural Architecture Search. In ICLR, 2020.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 359 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "Hieu Pham, Melody Y Guan, Barret Zoph, Quoc V Le, and Jeff Dean. Efficient Neural Architecture", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 347, + 304, + 358 + ], + "spans": [ + { + "bbox": [ + 117, + 347, + 304, + 358 + ], + "score": 1.0, + "content": "Search via Parameter Sharing. In ICML, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 380 + ], + "score": 1.0, + "content": "Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 376, + 410, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 376, + 410, + 389 + ], + "score": 1.0, + "content": "classifier architecture search. In AAAI, volume 33, pp. 4780–4789, 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 406, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 506, + 421 + ], + "score": 1.0, + "content": "detection with region proposal networks. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 417, + 187, + 430 + ], + "spans": [ + { + "bbox": [ + 114, + 417, + 187, + 430 + ], + "score": 1.0, + "content": "pp. 91–99, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 104, + 436, + 503, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mo-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 447, + 469, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 469, + 460 + ], + "score": 1.0, + "content": "bileNetV2: Inverted Residuals and Linear Bottlenecks. In CVPR, pp. 4510–4520, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 491, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 491, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 491, + 480 + ], + "score": 1.0, + "content": "David So, Quoc Le, and Chen Liang. The evolved transformer. In ICML, pp. 5877–5886, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 502, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 504, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 504, + 498 + ], + "score": 1.0, + "content": "Dimitrios Stamoulis, Ruizhou Ding, Di Wang, Dimitrios Lymberopoulos, Bodhi Priyantha, Jie Liu,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 116, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "and Diana Marculescu. Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 508, + 233, + 519 + ], + "spans": [ + { + "bbox": [ + 116, + 508, + 233, + 519 + ], + "score": 1.0, + "content": "Hours. ECML PKDD, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "Ying Tai, Jian Yang, and Xiaoming Liu. Image super-resolution via deep recursive residual network.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3147–", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 548, + 167, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 167, + 560 + ], + "score": 1.0, + "content": "3155, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 104, + 567, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "Mingxing Tan and Quoc V Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 579, + 226, + 590 + ], + "spans": [ + { + "bbox": [ + 116, + 579, + 226, + 590 + ], + "score": 1.0, + "content": "Networks. In ICML, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 104, + 597, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "Mingxing Tan and Quoc V. Le. MixConv: Mixed Depthwise Convolutional Kernels. In BMVC,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 607, + 144, + 621 + ], + "spans": [ + { + "bbox": [ + 114, + 607, + 144, + 621 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 104, + 627, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, and Quoc V Le. Mnasnet: Platform-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 639, + 370, + 650 + ], + "spans": [ + { + "bbox": [ + 116, + 639, + 370, + 650 + ], + "score": 1.0, + "content": "Aware Neural Architecture Search for Mobile. In CVPR, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "Yancong Wei, Qiangqiang Yuan, Huanfeng Shen, and Liangpei Zhang. Boosting the accuracy of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "multispectral image pansharpening by learning a deep residual network. IEEE Geoscience and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 680, + 319, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 319, + 691 + ], + "score": 1.0, + "content": "Remote Sensing Letters, 14(10):1795–1799, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Peter Vajda, Yangqing Jia, and Kurt Keutzer. FBNet: Hardware-Aware Efficient ConvNet Design", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 117, + 721, + 370, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 370, + 732 + ], + "score": 1.0, + "content": "via Differentiable Neural Architecture Search. In CVPR, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 80, + 505, + 96 + ], + "score": 1.0, + "content": "Juncheng Li, Faming Fang, Kangfu Mei, and Guixu Zhang. Multi-scale residual network for image", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 108 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 108 + ], + "score": 1.0, + "content": "super-resolution. In Proceedings of the European Conference on Computer Vision (ECCV), pp.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 104, + 186, + 116 + ], + "spans": [ + { + "bbox": [ + 116, + 104, + 186, + 116 + ], + "score": 1.0, + "content": "517–532, 2018b.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 80, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 123, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "spans": [ + { + "bbox": [ + 105, + 123, + 505, + 136 + ], + "score": 1.0, + "content": "Hanwen Liang, Shifeng Zhang, Jiacheng Sun, Xingqiu He, Weiran Huang, Kechen Zhuang, and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 505, + 147 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 202, + 147 + ], + "score": 1.0, + "content": "Zhenguo Li. DARTS", + "type": "text" + }, + { + "bbox": [ + 202, + 136, + 208, + 144 + ], + "score": 0.32, + "content": "^ +", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 134, + 505, + 147 + ], + "score": 1.0, + "content": ": Improved Differentiable Architecture Search with Early Stopping. arXiv", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 146, + 254, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 146, + 254, + 157 + ], + "score": 1.0, + "content": "preprint arXiv:1909.06035, 2019.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 123, + 505, + 157 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 163, + 503, + 187 + ], + "lines": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 177 + ], + "score": 1.0, + "content": "Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal Loss for Dense ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 175, + 319, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 175, + 319, + 187 + ], + "score": 1.0, + "content": "Object Detection. In ICCV, pp. 2980–2988, 2017.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 106, + 164, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 194, + 505, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 505, + 206 + ], + "score": 1.0, + "content": "Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 505, + 217 + ], + "score": 1.0, + "content": "Yuille, Jonathan Huang, and Kevin Murphy. Progressive Neural Architecture Search. In ECCV,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 216, + 187, + 228 + ], + "spans": [ + { + "bbox": [ + 114, + 216, + 187, + 228 + ], + "score": 1.0, + "content": "pp. 19–34, 2018.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 194, + 505, + 228 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 235, + 506, + 269 + ], + "lines": [ + { + "bbox": [ + 107, + 235, + 504, + 246 + ], + "spans": [ + { + "bbox": [ + 107, + 235, + 504, + 246 + ], + "score": 1.0, + "content": "Chenxi Liu, Liang-Chieh Chen, Florian Schroff, Hartwig Adam, Wei Hua, Alan L Yuille, and Li Fei-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 115, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "Fei. Auto-deeplab: Hierarchical neural architecture search for semantic image segmentation. In", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 257, + 221, + 270 + ], + "spans": [ + { + "bbox": [ + 117, + 257, + 221, + 270 + ], + "score": 1.0, + "content": "CVPR, pp. 82–92, 2019a.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 107, + 235, + 505, + 270 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 275, + 504, + 299 + ], + "lines": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "Hanxiao Liu, Karen Simonyan, and Yiming Yang. DARTS: Differentiable Architecture Search. In", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 286, + 173, + 299 + ], + "spans": [ + { + "bbox": [ + 116, + 286, + 173, + 299 + ], + "score": 1.0, + "content": "ICLR, 2019b.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 275, + 505, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 305, + 504, + 329 + ], + "lines": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 306, + 505, + 318 + ], + "score": 1.0, + "content": "Jieru Mei, Yingwei Li, Xiaochen Lian, Xiaojie Jin, Linjie Yang, Alan Yuille, and Yang Jianchao.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 116, + 317, + 445, + 329 + ], + "spans": [ + { + "bbox": [ + 116, + 317, + 445, + 329 + ], + "score": 1.0, + "content": "AtomNAS: Fine-Grained End-to-End Neural Architecture Search. In ICLR, 2020.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 306, + 505, + 329 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 505, + 359 + ], + "lines": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "Hieu Pham, Melody Y Guan, Barret Zoph, Quoc V Le, and Jeff Dean. Efficient Neural Architecture", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 117, + 347, + 304, + 358 + ], + "spans": [ + { + "bbox": [ + 117, + 347, + 304, + 358 + ], + "score": 1.0, + "content": "Search via Parameter Sharing. In ICML, 2018.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 335, + 505, + 358 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 366, + 504, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 506, + 380 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 506, + 380 + ], + "score": 1.0, + "content": "Esteban Real, Alok Aggarwal, Yanping Huang, and Quoc V Le. Regularized evolution for image", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 116, + 376, + 410, + 389 + ], + "spans": [ + { + "bbox": [ + 116, + 376, + 410, + 389 + ], + "score": 1.0, + "content": "classifier architecture search. In AAAI, volume 33, pp. 4780–4789, 2019.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 364, + 506, + 389 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 395, + 504, + 429 + ], + "lines": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 395, + 505, + 409 + ], + "score": 1.0, + "content": "Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 115, + 406, + 506, + 421 + ], + "spans": [ + { + "bbox": [ + 115, + 406, + 506, + 421 + ], + "score": 1.0, + "content": "detection with region proposal networks. In Advances in neural information processing systems,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 417, + 187, + 430 + ], + "spans": [ + { + "bbox": [ + 114, + 417, + 187, + 430 + ], + "score": 1.0, + "content": "pp. 91–99, 2015.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 395, + 506, + 430 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 436, + 503, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 450 + ], + "score": 1.0, + "content": "Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. Mo-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 116, + 447, + 469, + 460 + ], + "spans": [ + { + "bbox": [ + 116, + 447, + 469, + 460 + ], + "score": 1.0, + "content": "bileNetV2: Inverted Residuals and Linear Bottlenecks. In CVPR, pp. 4510–4520, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 437, + 505, + 460 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 466, + 491, + 479 + ], + "lines": [ + { + "bbox": [ + 105, + 465, + 491, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 465, + 491, + 480 + ], + "score": 1.0, + "content": "David So, Quoc Le, and Chen Liang. The evolved transformer. In ICML, pp. 5877–5886, 2019.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 465, + 491, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 502, + 519 + ], + "lines": [ + { + "bbox": [ + 106, + 486, + 504, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 504, + 498 + ], + "score": 1.0, + "content": "Dimitrios Stamoulis, Ruizhou Ding, Di Wang, Dimitrios Lymberopoulos, Bodhi Priyantha, Jie Liu,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 116, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 116, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "and Diana Marculescu. Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 116, + 508, + 233, + 519 + ], + "spans": [ + { + "bbox": [ + 116, + 508, + 233, + 519 + ], + "score": 1.0, + "content": "Hours. ECML PKDD, 2019.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 106, + 486, + 505, + 519 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 505, + 540 + ], + "score": 1.0, + "content": "Ying Tai, Jian Yang, and Xiaoming Liu. Image super-resolution via deep recursive residual network.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 537, + 505, + 550 + ], + "spans": [ + { + "bbox": [ + 115, + 537, + 505, + 550 + ], + "score": 1.0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 3147–", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 548, + 167, + 560 + ], + "spans": [ + { + "bbox": [ + 115, + 548, + 167, + 560 + ], + "score": 1.0, + "content": "3155, 2017.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 526, + 505, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 567, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "Mingxing Tan and Quoc V Le. EfficientNet: Rethinking Model Scaling for Convolutional Neural", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 579, + 226, + 590 + ], + "spans": [ + { + "bbox": [ + 116, + 579, + 226, + 590 + ], + "score": 1.0, + "content": "Networks. In ICML, 2019.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 106, + 567, + 505, + 590 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 597, + 505, + 621 + ], + "lines": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 596, + 506, + 610 + ], + "score": 1.0, + "content": "Mingxing Tan and Quoc V. Le. MixConv: Mixed Depthwise Convolutional Kernels. In BMVC,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 607, + 144, + 621 + ], + "spans": [ + { + "bbox": [ + 114, + 607, + 144, + 621 + ], + "score": 1.0, + "content": "2019.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 105, + 596, + 506, + 621 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 627, + 504, + 650 + ], + "lines": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, and Quoc V Le. Mnasnet: Platform-", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 639, + 370, + 650 + ], + "spans": [ + { + "bbox": [ + 116, + 639, + 370, + 650 + ], + "score": 1.0, + "content": "Aware Neural Architecture Search for Mobile. In CVPR, 2019.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 627, + 505, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 657, + 504, + 691 + ], + "lines": [ + { + "bbox": [ + 106, + 657, + 506, + 671 + ], + "spans": [ + { + "bbox": [ + 106, + 657, + 506, + 671 + ], + "score": 1.0, + "content": "Yancong Wei, Qiangqiang Yuan, Huanfeng Shen, and Liangpei Zhang. Boosting the accuracy of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 668, + 506, + 682 + ], + "spans": [ + { + "bbox": [ + 115, + 668, + 506, + 682 + ], + "score": 1.0, + "content": "multispectral image pansharpening by learning a deep residual network. IEEE Geoscience and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 680, + 319, + 691 + ], + "spans": [ + { + "bbox": [ + 115, + 680, + 319, + 691 + ], + "score": 1.0, + "content": "Remote Sensing Letters, 14(10):1795–1799, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 106, + 657, + 506, + 691 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 699, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Bichen Wu, Xiaoliang Dai, Peizhao Zhang, Yanghan Wang, Fei Sun, Yiming Wu, Yuandong Tian,", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Peter Vajda, Yangqing Jia, and Kurt Keutzer. FBNet: Hardware-Aware Efficient ConvNet Design", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 117, + 721, + 370, + 732 + ], + "spans": [ + { + "bbox": [ + 117, + 721, + 370, + 732 + ], + "score": 1.0, + "content": "via Differentiable Neural Architecture Search. In CVPR, 2019.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 106, + 698, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 81, + 506, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. SNAS: Stochastic Neural Architecture", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 94, + 212, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 212, + 106 + ], + "score": 1.0, + "content": "Search. In ICLR, 2019.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 127 + ], + "score": 1.0, + "content": "Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian, and Hongkai Xiong. PC-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 505, + 137 + ], + "score": 1.0, + "content": "DARTS: Partial Channel Connections for Memory-Efficient Architecture Search. In ICLR, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 325, + 148 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 136, + 331, + 144 + ], + "score": 0.54, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 135, + 395, + 148 + ], + "score": 1.0, + "content": "BJlS634tPr.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 151, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 104, + 151, + 506, + 169 + ], + "score": 1.0, + "content": "Kaicheng Yu, Christian Sciuto, Martin Jaggi, Claudiu Musat, and Mathieu Salzmann. Evaluating", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 165, + 504, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 165, + 504, + 178 + ], + "score": 1.0, + "content": "the search phase of neural architecture search. In ICLR, 2020. URL https://openreview.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 174, + 259, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 188, + 189 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 189, + 177, + 195, + 184 + ], + "score": 0.4, + "content": "{ . } = { }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 174, + 259, + 189 + ], + "score": 1.0, + "content": "H1loF2NFwr.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "Arber Zela, Thomas Elsken, Tonmoy Saikia, Yassine Marrakchi, Thomas Brox, and Frank Hutter.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 204, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 219 + ], + "score": 1.0, + "content": "Understanding and robustifying differentiable architecture search. In ICLR, 2020. URL https:", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 337, + 230 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 266, + 230 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 218, + 273, + 226 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 216, + 337, + 230 + ], + "score": 1.0, + "content": "H1gDNyrKDS.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 234, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 234, + 505, + 249 + ], + "score": 1.0, + "content": "Pan Zhou, Caiming Xiong, Richard Socher, and Steven CH Hoi. Theory-inspired path-regularized", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 245, + 441, + 260 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 441, + 260 + ], + "score": 1.0, + "content": "differential network architecture search. arXiv preprint arXiv:2006.16537, 2020.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning Transferable Architec-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 379, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 379, + 289 + ], + "score": 1.0, + "content": "tures for Scalable Image Recognition. In CVPR, volume 2, 2018.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 7 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 81, + 506, + 289 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 505, + 96 + ], + "score": 1.0, + "content": "Sirui Xie, Hehui Zheng, Chunxiao Liu, and Liang Lin. SNAS: Stochastic Neural Architecture", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 94, + 212, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 94, + 212, + 106 + ], + "score": 1.0, + "content": "Search. In ICLR, 2019.", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 111, + 505, + 127 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 505, + 127 + ], + "score": 1.0, + "content": "Yuhui Xu, Lingxi Xie, Xiaopeng Zhang, Xin Chen, Guo-Jun Qi, Qi Tian, and Hongkai Xiong. PC-", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 124, + 505, + 137 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 505, + 137 + ], + "score": 1.0, + "content": "DARTS: Partial Channel Connections for Memory-Efficient Architecture Search. In ICLR, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 135, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 116, + 135, + 325, + 148 + ], + "score": 1.0, + "content": "URL https://openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 325, + 136, + 331, + 144 + ], + "score": 0.54, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 332, + 135, + 395, + 148 + ], + "score": 1.0, + "content": "BJlS634tPr.", + "type": "text" + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 151, + 506, + 169 + ], + "spans": [ + { + "bbox": [ + 104, + 151, + 506, + 169 + ], + "score": 1.0, + "content": "Kaicheng Yu, Christian Sciuto, Martin Jaggi, Claudiu Musat, and Mathieu Salzmann. Evaluating", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 165, + 504, + 178 + ], + "spans": [ + { + "bbox": [ + 116, + 165, + 504, + 178 + ], + "score": 1.0, + "content": "the search phase of neural architecture search. In ICLR, 2020. URL https://openreview.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 174, + 259, + 189 + ], + "spans": [ + { + "bbox": [ + 116, + 174, + 188, + 189 + ], + "score": 1.0, + "content": "net/forum?id", + "type": "text" + }, + { + "bbox": [ + 189, + 177, + 195, + 184 + ], + "score": 0.4, + "content": "{ . } = { }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 174, + 259, + 189 + ], + "score": 1.0, + "content": "H1loF2NFwr.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 195, + 505, + 207 + ], + "spans": [ + { + "bbox": [ + 106, + 195, + 505, + 207 + ], + "score": 1.0, + "content": "Arber Zela, Thomas Elsken, Tonmoy Saikia, Yassine Marrakchi, Thomas Brox, and Frank Hutter.", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 204, + 505, + 219 + ], + "spans": [ + { + "bbox": [ + 115, + 204, + 505, + 219 + ], + "score": 1.0, + "content": "Understanding and robustifying differentiable architecture search. In ICLR, 2020. URL https:", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 116, + 216, + 337, + 230 + ], + "spans": [ + { + "bbox": [ + 116, + 216, + 266, + 230 + ], + "score": 1.0, + "content": "//openreview.net/forum?id", + "type": "text" + }, + { + "bbox": [ + 267, + 218, + 273, + 226 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 216, + 337, + 230 + ], + "score": 1.0, + "content": "H1gDNyrKDS.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 234, + 505, + 249 + ], + "spans": [ + { + "bbox": [ + 104, + 234, + 505, + 249 + ], + "score": 1.0, + "content": "Pan Zhou, Caiming Xiong, Richard Socher, and Steven CH Hoi. Theory-inspired path-regularized", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 245, + 441, + 260 + ], + "spans": [ + { + "bbox": [ + 115, + 245, + 441, + 260 + ], + "score": 1.0, + "content": "differential network architecture search. arXiv preprint arXiv:2006.16537, 2020.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 505, + 278 + ], + "score": 1.0, + "content": "Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning Transferable Architec-", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 276, + 379, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 379, + 289 + ], + "score": 1.0, + "content": "tures for Scalable Image Recognition. In CVPR, volume 2, 2018.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + } + ], + "index": 7, + "bbox_fs": [ + 104, + 82, + 506, + 289 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 183, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 263, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 264, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 264, + 119 + ], + "score": 1.0, + "content": "A.1 PRELIMINARY ABOUT DARTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "In differentiable architecture search (Liu et al., 2019b), a cell-based search space in the form of Di-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "rected Acyclic Graph (DAG) is constructed. The DAG has two input nodes from the previous layers,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "four intermediate nodes, and one output node. There are several paralleling operators (denoted as", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 160, + 462, + 174 + ], + "spans": [ + { + "bbox": [ + 107, + 162, + 116, + 171 + ], + "score": 0.61, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 160, + 239, + 174 + ], + "score": 1.0, + "content": ") between each two nodes (say", + "type": "text" + }, + { + "bbox": [ + 239, + 161, + 254, + 173 + ], + "score": 0.53, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 160, + 316, + 174 + ], + "score": 1.0, + "content": "), whose output", + "type": "text" + }, + { + "bbox": [ + 317, + 160, + 338, + 171 + ], + "score": 0.9, + "content": "\\bar { o } ^ { ( i , j ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 160, + 398, + 174 + ], + "score": 1.0, + "content": "given an input", + "type": "text" + }, + { + "bbox": [ + 399, + 163, + 405, + 171 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 160, + 462, + 174 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 178, + 388, + 212 + ], + "lines": [ + { + "bbox": [ + 223, + 178, + 388, + 212 + ], + "spans": [ + { + "bbox": [ + 223, + 178, + 388, + 212 + ], + "score": 0.95, + "content": "\\bar { o } ^ { ( i , j ) } ( x ) = \\sum _ { o \\in \\mathcal { O } } \\frac { \\exp ( \\alpha _ { o } ^ { ( i , j ) } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\alpha _ { o ^ { \\prime } } ^ { ( i , j ) } ) } o ( x )", + "type": "interline_equation", + "image_path": "f4033e8ae8c25ecf744f15ba953032691b4929d2c808b638a8554c8fdbc0beaf.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 178, + 388, + 195.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 223, + 195.0, + 388, + 212.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 504, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "It is essentially applying softmax over all operators where each operator is assigned with an archi-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 168, + 240 + ], + "score": 1.0, + "content": "tectural weight", + "type": "text" + }, + { + "bbox": [ + 168, + 230, + 175, + 238 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 228, + 505, + 240 + ], + "score": 1.0, + "content": ". A supernet is built on two kinds of such cells, so-called normal cells and reduction", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 239, + 504, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 504, + 253 + ], + "score": 1.0, + "content": "cells (for down-sampling). The architectural search is then characterized as a bi-level optimization:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 256, + 381, + 293 + ], + "lines": [ + { + "bbox": [ + 229, + 256, + 381, + 293 + ], + "spans": [ + { + "bbox": [ + 229, + 256, + 381, + 293 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\underset { \\alpha } { \\operatorname* { m i n } } } & { \\mathcal { L } _ { v a l } ( w ^ { * } ( \\alpha ) , \\alpha ) } \\\\ { s . t . } & { w ^ { * } ( \\alpha ) = \\arg \\underset { w } { \\operatorname* { m i n } } \\mathcal { L } _ { t r a i n } ( w , \\alpha ) } \\end{array}", + "type": "interline_equation", + "image_path": "36bc2f352c6ecfa2c064f4d6bbc12da58888ea4e044d64c0e9147466090b282f.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 256, + 381, + 274.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 229, + 274.5, + 381, + 293.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "This indicates that the training of such a cell-based supernet should be interleaved, where at each step", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "the network weights and architectural weights are updated iteratively. The final model is determined", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 321, + 383, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 383, + 333 + ], + "score": 1.0, + "content": "by simply choosing operations with the largest architectural weights.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 191, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 192, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 192, + 358 + ], + "score": 1.0, + "content": "A.2 EXPERIMENT", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 366, + 226, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 227, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 227, + 379 + ], + "score": 1.0, + "content": "A.2.1 TRAINING DETAILS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 399 + ], + "score": 1.0, + "content": "CIFAR-10 and CIFAR-100. Table 9 gives the averaged performance in reduced search spaces", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 397, + 466, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 466, + 410 + ], + "score": 1.0, + "content": "S1-S4, as well with maximum eigenvalues. Table 10 reports the CIFAR-100 results in S0.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "table", + "bbox": [ + 134, + 475, + 474, + 582 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 428, + 505, + 473 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "Table 9: Comparison of searched CNN architectures in four reduced search spaces S1-S4 (Zela", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "et al., 2020) on CIFAR-10 and CIFAR-100. We report the mean±std of test error over 3 found", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 463 + ], + "score": 1.0, + "content": "architectures retrained from scratch, alongside with eigenvalue (EV) that corresponds to the best", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 461, + 388, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 388, + 474 + ], + "score": 1.0, + "content": "validation accuracy. We follow the same settings as Zela et al. (2020).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "table_body", + "bbox": [ + 134, + 475, + 474, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 475, + 474, + 582 + ], + "spans": [ + { + "bbox": [ + 134, + 475, + 474, + 582 + ], + "score": 0.983, + "html": "
BenchmarkDARTSDARTS-ESDARTS-ADAoursours (EV)
C10S14.66±0.713.05±0.073.03±0.082.76±0.070.37±0.19
S24.42±0.403.41±0.143.59±0.312.79±0.040.41±0.08
S34.12±0.853.71±1.142.99±0.342.65±0.040.31±0.09
S46.95±0.184.17±0.213.89±0.672.91±0.040.30±0.11
C100S129.93±0.4128.90±0.8124.94±0.8123.26±0.590.60±0.15
S228.75±0.9224.68±1.4326.88±1.1122.31±0.650.54±0.15
S329.01±0.2426.99±1.7924.55±0.6321.47±0.400.40±0.03
S424.77±1.5123.90±2.0123.66±0.9021.75±0.260.93±0.15
", + "type": "table", + "image_path": "0bbceae89cb54113b1bcf7af347aaaf9ab1ace25850057aa3e9b8605dca47e53.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 134, + 475, + 474, + 510.6666666666667 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 134, + 510.6666666666667, + 474, + 546.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 134, + 546.3333333333334, + 474, + 582.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 23.25 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "score": 1.0, + "content": "ImageNet classification. For training on ImageNet, we use the same setting as MnasNet (Tan et al.,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "2019). To be comparable with EfficientNet (Tan & Le, 2019), we also use squeeze-and-excitation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "(Hu et al., 2018). Furthermore, we don’t include methods trained using large model distillation,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "because it can boost final validation accuracy marginally. To be fair, we don’t use the efficient head", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 637, + 430, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 430, + 652 + ], + "score": 1.0, + "content": "in MobileNetV3 (Howard et al., 2019) although it can reduce FLOPs marginally.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 504, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "COCO object detection. All models are trained and evaluated on MS COCO dataset for 12 epochs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 666, + 482, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 482, + 679 + ], + "score": 1.0, + "content": "with a batch size of 16. The initial learning rate is 0.01 and reduced by 0.1 at epoch 8 and 11.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "title", + "bbox": [ + 107, + 690, + 372, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 373, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 373, + 703 + ], + "score": 1.0, + "content": "A.2.2 FURTHER DISCUSSIONS ON FAILURE OF EIGENVALUE", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "To further explore the relationship between the searching performance and Hessian Eigenvalue, we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 504, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 504, + 734 + ], + "score": 1.0, + "content": "plot the performance trajectory of the searched models in Figure 4 (b). Specifically, we sample", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 26, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 183, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 79, + 185, + 97 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 263, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 264, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 264, + 119 + ], + "score": 1.0, + "content": "A.1 PRELIMINARY ABOUT DARTS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 126, + 505, + 173 + ], + "lines": [ + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 505, + 139 + ], + "score": 1.0, + "content": "In differentiable architecture search (Liu et al., 2019b), a cell-based search space in the form of Di-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 151 + ], + "score": 1.0, + "content": "rected Acyclic Graph (DAG) is constructed. The DAG has two input nodes from the previous layers,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 162 + ], + "score": 1.0, + "content": "four intermediate nodes, and one output node. There are several paralleling operators (denoted as", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 160, + 462, + 174 + ], + "spans": [ + { + "bbox": [ + 107, + 162, + 116, + 171 + ], + "score": 0.61, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 160, + 239, + 174 + ], + "score": 1.0, + "content": ") between each two nodes (say", + "type": "text" + }, + { + "bbox": [ + 239, + 161, + 254, + 173 + ], + "score": 0.53, + "content": "i , j", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 160, + 316, + 174 + ], + "score": 1.0, + "content": "), whose output", + "type": "text" + }, + { + "bbox": [ + 317, + 160, + 338, + 171 + ], + "score": 0.9, + "content": "\\bar { o } ^ { ( i , j ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 160, + 398, + 174 + ], + "score": 1.0, + "content": "given an input", + "type": "text" + }, + { + "bbox": [ + 399, + 163, + 405, + 171 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 160, + 462, + 174 + ], + "score": 1.0, + "content": "is defined as,", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 127, + 506, + 174 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 223, + 178, + 388, + 212 + ], + "lines": [ + { + "bbox": [ + 223, + 178, + 388, + 212 + ], + "spans": [ + { + "bbox": [ + 223, + 178, + 388, + 212 + ], + "score": 0.95, + "content": "\\bar { o } ^ { ( i , j ) } ( x ) = \\sum _ { o \\in \\mathcal { O } } \\frac { \\exp ( \\alpha _ { o } ^ { ( i , j ) } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\alpha _ { o ^ { \\prime } } ^ { ( i , j ) } ) } o ( x )", + "type": "interline_equation", + "image_path": "f4033e8ae8c25ecf744f15ba953032691b4929d2c808b638a8554c8fdbc0beaf.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 223, + 178, + 388, + 195.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 223, + 195.0, + 388, + 212.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 217, + 504, + 251 + ], + "lines": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 505, + 230 + ], + "score": 1.0, + "content": "It is essentially applying softmax over all operators where each operator is assigned with an archi-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 228, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 168, + 240 + ], + "score": 1.0, + "content": "tectural weight", + "type": "text" + }, + { + "bbox": [ + 168, + 230, + 175, + 238 + ], + "score": 0.74, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 228, + 505, + 240 + ], + "score": 1.0, + "content": ". A supernet is built on two kinds of such cells, so-called normal cells and reduction", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 239, + 504, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 504, + 253 + ], + "score": 1.0, + "content": "cells (for down-sampling). The architectural search is then characterized as a bi-level optimization:", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 218, + 505, + 253 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 229, + 256, + 381, + 293 + ], + "lines": [ + { + "bbox": [ + 229, + 256, + 381, + 293 + ], + "spans": [ + { + "bbox": [ + 229, + 256, + 381, + 293 + ], + "score": 0.91, + "content": "\\begin{array} { r l } { \\underset { \\alpha } { \\operatorname* { m i n } } } & { \\mathcal { L } _ { v a l } ( w ^ { * } ( \\alpha ) , \\alpha ) } \\\\ { s . t . } & { w ^ { * } ( \\alpha ) = \\arg \\underset { w } { \\operatorname* { m i n } } \\mathcal { L } _ { t r a i n } ( w , \\alpha ) } \\end{array}", + "type": "interline_equation", + "image_path": "36bc2f352c6ecfa2c064f4d6bbc12da58888ea4e044d64c0e9147466090b282f.jpg" + } + ] + } + ], + "index": 11.5, + "virtual_lines": [ + { + "bbox": [ + 229, + 256, + 381, + 274.5 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 229, + 274.5, + 381, + 293.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 298, + 505, + 332 + ], + "lines": [ + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "This indicates that the training of such a cell-based supernet should be interleaved, where at each step", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "the network weights and architectural weights are updated iteratively. The final model is determined", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 321, + 383, + 333 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 383, + 333 + ], + "score": 1.0, + "content": "by simply choosing operations with the largest architectural weights.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 297, + 505, + 333 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 191, + 357 + ], + "lines": [ + { + "bbox": [ + 106, + 346, + 192, + 358 + ], + "spans": [ + { + "bbox": [ + 106, + 346, + 192, + 358 + ], + "score": 1.0, + "content": "A.2 EXPERIMENT", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 366, + 226, + 378 + ], + "lines": [ + { + "bbox": [ + 106, + 365, + 227, + 379 + ], + "spans": [ + { + "bbox": [ + 106, + 365, + 227, + 379 + ], + "score": 1.0, + "content": "A.2.1 TRAINING DETAILS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 385, + 504, + 408 + ], + "lines": [ + { + "bbox": [ + 106, + 385, + 505, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 505, + 399 + ], + "score": 1.0, + "content": "CIFAR-10 and CIFAR-100. Table 9 gives the averaged performance in reduced search spaces", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 397, + 466, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 466, + 410 + ], + "score": 1.0, + "content": "S1-S4, as well with maximum eigenvalues. Table 10 reports the CIFAR-100 results in S0.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 385, + 505, + 410 + ] + }, + { + "type": "table", + "bbox": [ + 134, + 475, + 474, + 582 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 428, + 505, + 473 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 505, + 441 + ], + "score": 1.0, + "content": "Table 9: Comparison of searched CNN architectures in four reduced search spaces S1-S4 (Zela", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "et al., 2020) on CIFAR-10 and CIFAR-100. We report the mean±std of test error over 3 found", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 451, + 506, + 463 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 463 + ], + "score": 1.0, + "content": "architectures retrained from scratch, alongside with eigenvalue (EV) that corresponds to the best", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 461, + 388, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 388, + 474 + ], + "score": 1.0, + "content": "validation accuracy. We follow the same settings as Zela et al. (2020).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "table_body", + "bbox": [ + 134, + 475, + 474, + 582 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 475, + 474, + 582 + ], + "spans": [ + { + "bbox": [ + 134, + 475, + 474, + 582 + ], + "score": 0.983, + "html": "
BenchmarkDARTSDARTS-ESDARTS-ADAoursours (EV)
C10S14.66±0.713.05±0.073.03±0.082.76±0.070.37±0.19
S24.42±0.403.41±0.143.59±0.312.79±0.040.41±0.08
S34.12±0.853.71±1.142.99±0.342.65±0.040.31±0.09
S46.95±0.184.17±0.213.89±0.672.91±0.040.30±0.11
C100S129.93±0.4128.90±0.8124.94±0.8123.26±0.590.60±0.15
S228.75±0.9224.68±1.4326.88±1.1122.31±0.650.54±0.15
S329.01±0.2426.99±1.7924.55±0.6321.47±0.400.40±0.03
S424.77±1.5123.90±2.0123.66±0.9021.75±0.260.93±0.15
", + "type": "table", + "image_path": "0bbceae89cb54113b1bcf7af347aaaf9ab1ace25850057aa3e9b8605dca47e53.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 134, + 475, + 474, + 510.6666666666667 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 134, + 510.6666666666667, + 474, + 546.3333333333334 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 134, + 546.3333333333334, + 474, + 582.0 + ], + "spans": [], + "index": 26 + } + ] + } + ], + "index": 23.25 + }, + { + "type": "text", + "bbox": [ + 106, + 594, + 505, + 650 + ], + "lines": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "score": 1.0, + "content": "ImageNet classification. For training on ImageNet, we use the same setting as MnasNet (Tan et al.,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "2019). To be comparable with EfficientNet (Tan & Le, 2019), we also use squeeze-and-excitation", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 630 + ], + "score": 1.0, + "content": "(Hu et al., 2018). Furthermore, we don’t include methods trained using large model distillation,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 106, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "because it can boost final validation accuracy marginally. To be fair, we don’t use the efficient head", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 637, + 430, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 430, + 652 + ], + "score": 1.0, + "content": "in MobileNetV3 (Howard et al., 2019) although it can reduce FLOPs marginally.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 593, + 505, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 655, + 504, + 678 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 668 + ], + "score": 1.0, + "content": "COCO object detection. All models are trained and evaluated on MS COCO dataset for 12 epochs", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 666, + 482, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 482, + 679 + ], + "score": 1.0, + "content": "with a batch size of 16. The initial learning rate is 0.01 and reduced by 0.1 at epoch 8 and 11.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 654, + 505, + 679 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 690, + 372, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 689, + 373, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 373, + 703 + ], + "score": 1.0, + "content": "A.2.2 FURTHER DISCUSSIONS ON FAILURE OF EIGENVALUE", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 504, + 722 + ], + "score": 1.0, + "content": "To further explore the relationship between the searching performance and Hessian Eigenvalue, we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 721, + 504, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 504, + 734 + ], + "score": 1.0, + "content": "plot the performance trajectory of the searched models in Figure 4 (b). Specifically, we sample", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 382, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 397 + ], + "score": 1.0, + "content": "models every 10 epochs and train these models from scratch using the same setting as above (Fig-", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "ure 5). The performance of the inferred models continues growing, where the accuracy is boosted", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 129, + 418 + ], + "score": 1.0, + "content": "from", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 129, + 405, + 157, + 416 + ], + "score": 0.88, + "content": "9 6 . 5 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 157, + 405, + 169, + 418 + ], + "score": 1.0, + "content": "to", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 170, + 405, + 198, + 416 + ], + "score": 0.9, + "content": "9 7 . 4 \\%", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 198, + 405, + 505, + 418 + ], + "score": 1.0, + "content": ". This affirms the validity of searching using our method. In contrast, the", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "early-stopping strategies based on eigenvalues (Zela et al., 2020) would fail in this setting. We ar-", + "type": "text", + "cross_page": true + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 426, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 104, + 426, + 506, + 441 + ], + "score": 1.0, + "content": "gue that the proposed auxiliary skip branch can regularize the overfitting of the supernet, leaving", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "the architectural weights to represent the ability of candidate operations. This experiment poses as", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "a counterexample to R-DARTS, where good models can appear albeit Hessian eigenvalues change", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 460, + 310, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 310, + 473 + ], + "score": 1.0, + "content": "fast. It again denies the need for a costly indicator.", + "type": "text", + "cross_page": true + } + ], + "index": 23 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 709, + 504, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 112, + 399, + 244 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 502, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 87, + 505, + 103 + ], + "spans": [ + { + "bbox": [ + 105, + 87, + 505, + 103 + ], + "score": 1.0, + "content": "Table 10: Comparison of searched models on CIFAR-100. \u0005: Reported by Dong & Yang (2019b),", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 320, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 241, + 113 + ], + "score": 1.0, + "content": "?: Reported by Zela et al. (2020),", + "type": "text" + }, + { + "bbox": [ + 242, + 100, + 247, + 110 + ], + "score": 0.46, + "content": "^ \\ddag", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 99, + 320, + 113 + ], + "score": 1.0, + "content": ":Rerun their code.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 212, + 112, + 399, + 244 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 112, + 399, + 244 + ], + "spans": [ + { + "bbox": [ + 212, + 112, + 399, + 244 + ], + "score": 0.924, + "html": "
ModelsParams Error (M) (%)Cost GPU Days
ResNet (2016)1.722.10-
AmoebaNet (2019) 3.118.933150
PNAS (2018)3.219.53150
ENAS (2018)4.619.430.45
DARTS (2019b)120.58±0.44*(*0.4
GDAS (2019b)3.418.380.2
P-DARTS (2019b） 3.617.49‡0.3
R-DARTS (2020)118.01±0.261.6
DARTS- (avg.)3.317.51±0.250.4
DARTS- (best)3.417.160.4
", + "type": "table", + "image_path": "b9557b2ced7056ec8f767a11c6ccf4169f7f71609a3c43ab3aa10328e291ddf8.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 112, + 399, + 125.2 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 212, + 125.2, + 399, + 138.4 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 212, + 138.4, + 399, + 151.6 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 212, + 151.6, + 399, + 164.79999999999998 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 212, + 164.79999999999998, + 399, + 177.99999999999997 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 212, + 177.99999999999997, + 399, + 191.19999999999996 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 212, + 191.19999999999996, + 399, + 204.39999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 212, + 204.39999999999995, + 399, + 217.59999999999994 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 212, + 217.59999999999994, + 399, + 230.79999999999993 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 212, + 230.79999999999993, + 399, + 243.99999999999991 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 123, + 276, + 487, + 368 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 154, + 264, + 455, + 276 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 153, + 262, + 457, + 279 + ], + "spans": [ + { + "bbox": [ + 153, + 262, + 457, + 279 + ], + "score": 1.0, + "content": "Table 11: Transfer results on COCO datasets of various drop-in backbones.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 123, + 276, + 487, + 368 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 276, + 487, + 368 + ], + "spans": [ + { + "bbox": [ + 123, + 276, + 487, + 368 + ], + "score": 0.983, + "html": "
BackbonesParams (M)AccAPAP50AP75APsAPMAPL
MobileNetV2 (2018)3.472.028.346.729.314.830.738.1
SingPath NAS (2019)4.375.030.749.832.215.433.941.6
MnasNet-A2 (2019)4.875.630.550.232.016.634.141.1
MobileNetV3 (2019)5.475.229.949.330.814.933.341.1
MixNet-M (2019)5.077.031.351.732.417.035.041.9
FairDARTS-C (2020b)5.377.231.951.933.017.435.343.0
DARTS-A (Ours)5.577.832.552.834.118.036.143.4
", + "type": "table", + "image_path": "d8f5dc75f26cebaeab72bec4d962517dc7f4f92bdbac62af19ce4f0a726e13ff.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 123, + 276, + 487, + 306.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 123, + 306.6666666666667, + 487, + 337.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 123, + 337.33333333333337, + 487, + 368.00000000000006 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.0 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 505, + 472 + ], + "lines": [ + { + "bbox": [ + 105, + 382, + 505, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 382, + 505, + 397 + ], + "score": 1.0, + "content": "models every 10 epochs and train these models from scratch using the same setting as above (Fig-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "spans": [ + { + "bbox": [ + 106, + 394, + 506, + 407 + ], + "score": 1.0, + "content": "ure 5). The performance of the inferred models continues growing, where the accuracy is boosted", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 405, + 505, + 418 + ], + "spans": [ + { + "bbox": [ + 105, + 405, + 129, + 418 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 129, + 405, + 157, + 416 + ], + "score": 0.88, + "content": "9 6 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 405, + 169, + 418 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 170, + 405, + 198, + 416 + ], + "score": 0.9, + "content": "9 7 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 405, + 505, + 418 + ], + "score": 1.0, + "content": ". This affirms the validity of searching using our method. In contrast, the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "early-stopping strategies based on eigenvalues (Zela et al., 2020) would fail in this setting. We ar-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 426, + 506, + 441 + ], + "spans": [ + { + "bbox": [ + 104, + 426, + 506, + 441 + ], + "score": 1.0, + "content": "gue that the proposed auxiliary skip branch can regularize the overfitting of the supernet, leaving", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 438, + 506, + 451 + ], + "score": 1.0, + "content": "the architectural weights to represent the ability of candidate operations. This experiment poses as", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 505, + 462 + ], + "score": 1.0, + "content": "a counterexample to R-DARTS, where good models can appear albeit Hessian eigenvalues change", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 460, + 310, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 310, + 473 + ], + "score": 1.0, + "content": "fast. It again denies the need for a costly indicator.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "image", + "bbox": [ + 114, + 488, + 500, + 655 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 488, + 500, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 488, + 500, + 655 + ], + "spans": [ + { + "bbox": [ + 114, + 488, + 500, + 655 + ], + "score": 0.932, + "type": "image", + "image_path": "2ee7c0de76fb6d88dacfb8ce4a9475a0bcf887a3438690ae0338da626d09b87d.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 114, + 488, + 500, + 543.6666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 114, + 543.6666666666666, + 500, + 599.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 114, + 599.3333333333333, + 500, + 654.9999999999999 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 667, + 506, + 723 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "Figure 4: The evolution of maximal eigenvalues of DARTS- when searching in different search", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "spaces S0-S4 on CIFAR-10 (a) and CIFAR-100 (c). We run each experiment 3 times on different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "score": 1.0, + "content": "seeds. (b) DARTS-’s growing Hessian eigenvalues don’t induce poor performance. Among the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "sampled five models, the one corresponding to the highest eigenvalue has the best performance.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 712, + 275, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 275, + 724 + ], + "score": 1.0, + "content": "This example is done in S0 on CIFAR-10.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "index": 27.0 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 108, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 212, + 112, + 399, + 244 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 502, + 111 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 87, + 505, + 103 + ], + "spans": [ + { + "bbox": [ + 105, + 87, + 505, + 103 + ], + "score": 1.0, + "content": "Table 10: Comparison of searched models on CIFAR-100. \u0005: Reported by Dong & Yang (2019b),", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 99, + 320, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 241, + 113 + ], + "score": 1.0, + "content": "?: Reported by Zela et al. (2020),", + "type": "text" + }, + { + "bbox": [ + 242, + 100, + 247, + 110 + ], + "score": 0.46, + "content": "^ \\ddag", + "type": "inline_equation" + }, + { + "bbox": [ + 248, + 99, + 320, + 113 + ], + "score": 1.0, + "content": ":Rerun their code.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 212, + 112, + 399, + 244 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 112, + 399, + 244 + ], + "spans": [ + { + "bbox": [ + 212, + 112, + 399, + 244 + ], + "score": 0.924, + "html": "
ModelsParams Error (M) (%)Cost GPU Days
ResNet (2016)1.722.10-
AmoebaNet (2019) 3.118.933150
PNAS (2018)3.219.53150
ENAS (2018)4.619.430.45
DARTS (2019b)120.58±0.44*(*0.4
GDAS (2019b)3.418.380.2
P-DARTS (2019b） 3.617.49‡0.3
R-DARTS (2020)118.01±0.261.6
DARTS- (avg.)3.317.51±0.250.4
DARTS- (best)3.417.160.4
", + "type": "table", + "image_path": "b9557b2ced7056ec8f767a11c6ccf4169f7f71609a3c43ab3aa10328e291ddf8.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 212, + 112, + 399, + 125.2 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 212, + 125.2, + 399, + 138.4 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 212, + 138.4, + 399, + 151.6 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 212, + 151.6, + 399, + 164.79999999999998 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 212, + 164.79999999999998, + 399, + 177.99999999999997 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 212, + 177.99999999999997, + 399, + 191.19999999999996 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 212, + 191.19999999999996, + 399, + 204.39999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 212, + 204.39999999999995, + 399, + 217.59999999999994 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 212, + 217.59999999999994, + 399, + 230.79999999999993 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 212, + 230.79999999999993, + 399, + 243.99999999999991 + ], + "spans": [], + "index": 11 + } + ] + } + ], + "index": 3.5 + }, + { + "type": "table", + "bbox": [ + 123, + 276, + 487, + 368 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 154, + 264, + 455, + 276 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 153, + 262, + 457, + 279 + ], + "spans": [ + { + "bbox": [ + 153, + 262, + 457, + 279 + ], + "score": 1.0, + "content": "Table 11: Transfer results on COCO datasets of various drop-in backbones.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 123, + 276, + 487, + 368 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 276, + 487, + 368 + ], + "spans": [ + { + "bbox": [ + 123, + 276, + 487, + 368 + ], + "score": 0.983, + "html": "
BackbonesParams (M)AccAPAP50AP75APsAPMAPL
MobileNetV2 (2018)3.472.028.346.729.314.830.738.1
SingPath NAS (2019)4.375.030.749.832.215.433.941.6
MnasNet-A2 (2019)4.875.630.550.232.016.634.141.1
MobileNetV3 (2019)5.475.229.949.330.814.933.341.1
MixNet-M (2019)5.077.031.351.732.417.035.041.9
FairDARTS-C (2020b)5.377.231.951.933.017.435.343.0
DARTS-A (Ours)5.577.832.552.834.118.036.143.4
", + "type": "table", + "image_path": "d8f5dc75f26cebaeab72bec4d962517dc7f4f92bdbac62af19ce4f0a726e13ff.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 123, + 276, + 487, + 306.6666666666667 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 123, + 306.6666666666667, + 487, + 337.33333333333337 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 123, + 337.33333333333337, + 487, + 368.00000000000006 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.0 + }, + { + "type": "text", + "bbox": [ + 106, + 383, + 505, + 472 + ], + "lines": [], + "index": 19.5, + "bbox_fs": [ + 104, + 382, + 506, + 473 + ], + "lines_deleted": true + }, + { + "type": "image", + "bbox": [ + 114, + 488, + 500, + 655 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 114, + 488, + 500, + 655 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 488, + 500, + 655 + ], + "spans": [ + { + "bbox": [ + 114, + 488, + 500, + 655 + ], + "score": 0.932, + "type": "image", + "image_path": "2ee7c0de76fb6d88dacfb8ce4a9475a0bcf887a3438690ae0338da626d09b87d.jpg" + } + ] + } + ], + "index": 25, + "virtual_lines": [ + { + "bbox": [ + 114, + 488, + 500, + 543.6666666666666 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 114, + 543.6666666666666, + 500, + 599.3333333333333 + ], + "spans": [], + "index": 25 + }, + { + "bbox": [ + 114, + 599.3333333333333, + 500, + 654.9999999999999 + ], + "spans": [], + "index": 26 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 667, + 506, + 723 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 668, + 505, + 680 + ], + "score": 1.0, + "content": "Figure 4: The evolution of maximal eigenvalues of DARTS- when searching in different search", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "spaces S0-S4 on CIFAR-10 (a) and CIFAR-100 (c). We run each experiment 3 times on different", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 505, + 703 + ], + "score": 1.0, + "content": "seeds. (b) DARTS-’s growing Hessian eigenvalues don’t induce poor performance. Among the", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 700, + 505, + 714 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 714 + ], + "score": 1.0, + "content": "sampled five models, the one corresponding to the highest eigenvalue has the best performance.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 712, + 275, + 724 + ], + "spans": [ + { + "bbox": [ + 106, + 712, + 275, + 724 + ], + "score": 1.0, + "content": "This example is done in S0 on CIFAR-10.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + } + ], + "index": 27.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 215, + 85, + 387, + 216 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 215, + 85, + 387, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 85, + 387, + 216 + ], + "spans": [ + { + "bbox": [ + 215, + 85, + 387, + 216 + ], + "score": 0.968, + "type": "image", + "image_path": "1f60b3396478452c014f262ef127284e35887acc7e47cddc13310a1ce6cf0b66.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 85, + 387, + 98.1 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 215, + 98.1, + 387, + 111.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 215, + 111.19999999999999, + 387, + 124.29999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 215, + 124.29999999999998, + 387, + 137.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 215, + 137.39999999999998, + 387, + 150.49999999999997 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 215, + 150.49999999999997, + 387, + 163.59999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 215, + 163.59999999999997, + 387, + 176.69999999999996 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 215, + 176.69999999999996, + 387, + 189.79999999999995 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 215, + 189.79999999999995, + 387, + 202.89999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 215, + 202.89999999999995, + 387, + 215.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 227, + 505, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "Figure 5: Training five models sampled every 10 epochs during DARTS- searching process. See", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 238, + 283, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 283, + 251 + ], + "score": 1.0, + "content": "Fig 4 (b) for the corresponding eigenvalues.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 108, + 270, + 257, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 258, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 258, + 282 + ], + "score": 1.0, + "content": "A.2.3 MORE ABLATION STUDIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "score": 1.0, + "content": "To supplement the sensitivity analysis in Section 4.4, Figure 6 shows the training loss curve when", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 300, + 248, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 152, + 313 + ], + "score": 1.0, + "content": "initializing", + "type": "text" + }, + { + "bbox": [ + 152, + 301, + 160, + 312 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 300, + 248, + 313 + ], + "score": 1.0, + "content": "with different values.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "image", + "bbox": [ + 243, + 327, + 366, + 455 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 243, + 327, + 366, + 455 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 243, + 327, + 366, + 455 + ], + "spans": [ + { + "bbox": [ + 243, + 327, + 366, + 455 + ], + "score": 0.96, + "type": "image", + "image_path": "4c385a3466eb060cf91292bd589e967c287ffa63d137801d0e272d3eb19c20bf.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 327, + 366, + 391.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 243, + 391.0, + 366, + 455.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 475, + 504, + 498 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "Figure 6: The training loss curve of the over-parameterized network on CIFAR-10 in S3 with differ-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 486, + 163, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 147, + 499 + ], + "score": 1.0, + "content": "ent initial", + "type": "text" + }, + { + "bbox": [ + 147, + 487, + 158, + 498 + ], + "score": 0.88, + "content": "\\beta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 486, + 163, + 499 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 523, + 212, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 213, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 213, + 536 + ], + "score": 1.0, + "content": "A.3 LOSS LANDSCAPE", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "To show the consistent smoothing capacity of DARTS-, we draw more loss landscapes in S0 and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "S3 (on several seeds) w.r.t architectural parameters and their contours in Figure 7 and 8. Generally,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "DARTS-’s slopes are more inflated if we consider them as camping tents, which suggest better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 577, + 300, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 300, + 588 + ], + "score": 1.0, + "content": "convergence of the over-parameterized network.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "title", + "bbox": [ + 108, + 601, + 232, + 613 + ], + "lines": [ + { + "bbox": [ + 107, + 602, + 232, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 602, + 232, + 614 + ], + "score": 1.0, + "content": "A.4 LIST OF EXPERIMENTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 622, + 481, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 482, + 635 + ], + "score": 1.0, + "content": "We summarize all the conducted experiments with their related figures and tables in Table 12.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 108, + 649, + 255, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 257, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 257, + 665 + ], + "score": 1.0, + "content": "B FIGURES OF GENOTYPES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 215, + 85, + 387, + 216 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 215, + 85, + 387, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 215, + 85, + 387, + 216 + ], + "spans": [ + { + "bbox": [ + 215, + 85, + 387, + 216 + ], + "score": 0.968, + "type": "image", + "image_path": "1f60b3396478452c014f262ef127284e35887acc7e47cddc13310a1ce6cf0b66.jpg" + } + ] + } + ], + "index": 4.5, + "virtual_lines": [ + { + "bbox": [ + 215, + 85, + 387, + 98.1 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 215, + 98.1, + 387, + 111.19999999999999 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 215, + 111.19999999999999, + 387, + 124.29999999999998 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 215, + 124.29999999999998, + 387, + 137.39999999999998 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 215, + 137.39999999999998, + 387, + 150.49999999999997 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 215, + 150.49999999999997, + 387, + 163.59999999999997 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 215, + 163.59999999999997, + 387, + 176.69999999999996 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 215, + 176.69999999999996, + 387, + 189.79999999999995 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 215, + 189.79999999999995, + 387, + 202.89999999999995 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 215, + 202.89999999999995, + 387, + 215.99999999999994 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 227, + 505, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "Figure 5: Training five models sampled every 10 epochs during DARTS- searching process. See", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 238, + 283, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 283, + 251 + ], + "score": 1.0, + "content": "Fig 4 (b) for the corresponding eigenvalues.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + } + ], + "index": 7.5 + }, + { + "type": "title", + "bbox": [ + 108, + 270, + 257, + 281 + ], + "lines": [ + { + "bbox": [ + 106, + 269, + 258, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 258, + 282 + ], + "score": 1.0, + "content": "A.2.3 MORE ABLATION STUDIES", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 289, + 505, + 312 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "score": 1.0, + "content": "To supplement the sensitivity analysis in Section 4.4, Figure 6 shows the training loss curve when", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 300, + 248, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 300, + 152, + 313 + ], + "score": 1.0, + "content": "initializing", + "type": "text" + }, + { + "bbox": [ + 152, + 301, + 160, + 312 + ], + "score": 0.83, + "content": "\\beta", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 300, + 248, + 313 + ], + "score": 1.0, + "content": "with different values.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 287, + 505, + 313 + ] + }, + { + "type": "image", + "bbox": [ + 243, + 327, + 366, + 455 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 243, + 327, + 366, + 455 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 243, + 327, + 366, + 455 + ], + "spans": [ + { + "bbox": [ + 243, + 327, + 366, + 455 + ], + "score": 0.96, + "type": "image", + "image_path": "4c385a3466eb060cf91292bd589e967c287ffa63d137801d0e272d3eb19c20bf.jpg" + } + ] + } + ], + "index": 15.5, + "virtual_lines": [ + { + "bbox": [ + 243, + 327, + 366, + 391.0 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 243, + 391.0, + 366, + 455.0 + ], + "spans": [], + "index": 16 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 475, + 504, + 498 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 505, + 488 + ], + "score": 1.0, + "content": "Figure 6: The training loss curve of the over-parameterized network on CIFAR-10 in S3 with differ-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 486, + 163, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 147, + 499 + ], + "score": 1.0, + "content": "ent initial", + "type": "text" + }, + { + "bbox": [ + 147, + 487, + 158, + 498 + ], + "score": 0.88, + "content": "\\beta _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 486, + 163, + 499 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + } + ], + "index": 16.5 + }, + { + "type": "title", + "bbox": [ + 108, + 523, + 212, + 534 + ], + "lines": [ + { + "bbox": [ + 106, + 523, + 213, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 213, + 536 + ], + "score": 1.0, + "content": "A.3 LOSS LANDSCAPE", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 543, + 505, + 588 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 505, + 556 + ], + "score": 1.0, + "content": "To show the consistent smoothing capacity of DARTS-, we draw more loss landscapes in S0 and", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 554, + 505, + 567 + ], + "score": 1.0, + "content": "S3 (on several seeds) w.r.t architectural parameters and their contours in Figure 7 and 8. Generally,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 505, + 578 + ], + "score": 1.0, + "content": "DARTS-’s slopes are more inflated if we consider them as camping tents, which suggest better", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 577, + 300, + 588 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 300, + 588 + ], + "score": 1.0, + "content": "convergence of the over-parameterized network.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 543, + 505, + 588 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 601, + 232, + 613 + ], + "lines": [ + { + "bbox": [ + 107, + 602, + 232, + 614 + ], + "spans": [ + { + "bbox": [ + 107, + 602, + 232, + 614 + ], + "score": 1.0, + "content": "A.4 LIST OF EXPERIMENTS", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 622, + 481, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 482, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 482, + 635 + ], + "score": 1.0, + "content": "We summarize all the conducted experiments with their related figures and tables in Table 12.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 622, + 482, + 635 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 649, + 255, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 648, + 257, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 257, + 665 + ], + "score": 1.0, + "content": "B FIGURES OF GENOTYPES", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 648, + 257, + 665 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 77, + 497, + 252 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 77, + 497, + 252 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 77, + 497, + 252 + ], + "spans": [ + { + "bbox": [ + 119, + 77, + 497, + 252 + ], + "score": 0.974, + "type": "image", + "image_path": "2a9495f4bd817a16ca6ed904008ed76e11422c9c3eb2094c924bde1a84f3ae8c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 77, + 497, + 135.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 135.33333333333334, + 497, + 193.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 193.66666666666669, + 497, + 252.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 264, + 505, + 309 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "Figure 7: More visualization of validation accuracy landscapes of DARTS (a,b) and DARTS- (e,f)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 232, + 288 + ], + "score": 1.0, + "content": "w.r.t. the architectural weights", + "type": "text" + }, + { + "bbox": [ + 233, + 278, + 241, + 286 + ], + "score": 0.69, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "on CIFAR-10 in S0. Their contour maps are shown respectively", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 288, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 157, + 299 + ], + "score": 1.0, + "content": "in (c,d) and", + "type": "text" + }, + { + "bbox": [ + 158, + 288, + 177, + 298 + ], + "score": 0.26, + "content": "^ { ( \\mathrm { g , h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 288, + 505, + 299 + ], + "score": 1.0, + "content": ". The step of contour map is 0.1. The inferred models by DARTS- have higher", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 298, + 358, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 154, + 310 + ], + "score": 1.0, + "content": "accuracies (", + "type": "text" + }, + { + "bbox": [ + 154, + 298, + 185, + 309 + ], + "score": 0.77, + "content": "9 7 . 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 298, + 189, + 310 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 189, + 298, + 222, + 308 + ], + "score": 0.79, + "content": "9 7 . 4 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 298, + 281, + 310 + ], + "score": 1.0, + "content": ") than DARTS", + "type": "text" + }, + { + "bbox": [ + 282, + 298, + 315, + 309 + ], + "score": 0.71, + "content": "( 9 7 . 1 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 298, + 318, + 310 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 318, + 298, + 351, + 309 + ], + "score": 0.77, + "content": "9 7 . 2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 298, + 358, + 310 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 115, + 315, + 498, + 489 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 315, + 498, + 489 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 315, + 498, + 489 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 498, + 489 + ], + "score": 0.977, + "type": "image", + "image_path": "29923a592a8d5157ef46cf1e03033d7003870bed18aa6bb44063ecd4855316ca.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 115, + 315, + 498, + 373.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 373.0, + 498, + 431.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 431.0, + 498, + 489.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 501, + 505, + 535 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "Figure 8: More visualization of validation accuracy landscapes of DARTS (a,b) and DARTS- (e,f)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 229, + 525 + ], + "score": 1.0, + "content": "w.r.t. the architectural weights", + "type": "text" + }, + { + "bbox": [ + 230, + 514, + 237, + 522 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "on CIFAR-10 in S3. Their contour maps are shown respectively in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 523, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 144, + 536 + ], + "score": 1.0, + "content": "(c,d) and", + "type": "text" + }, + { + "bbox": [ + 145, + 524, + 164, + 535 + ], + "score": 0.53, + "content": "^ { ( \\mathrm { g , h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 523, + 296, + 536 + ], + "score": 1.0, + "content": ". The step of contour map is 0.1.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 107, + 566, + 501, + 729 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 198, + 552, + 414, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 197, + 551, + 414, + 566 + ], + "spans": [ + { + "bbox": [ + 197, + 551, + 414, + 566 + ], + "score": 1.0, + "content": "Table 12: List of experiments conducted in this paper", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "table_body", + "bbox": [ + 107, + 566, + 501, + 729 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 566, + 501, + 729 + ], + "spans": [ + { + "bbox": [ + 107, + 566, + 501, + 729 + ], + "score": 0.982, + "html": "
MethodSearch SpaceDatasetFiguresTables
Main textSupp.Main textSupp.
DARTSS0CIFAR-107
DARTSS3CIFAR-103
DARTS-S0CIFAR-104,5,9,71,2
DARTS-S1-S4CIFAR-1034,6,9,87,96
DARTS-S5ImageNet5,102
DARTS-S5MS COCO11
DARTS-S6CIFAR-103
DARTS-S1-S4CIFAR-1004,1179
DARTS-S0CIFAR-1004,111,10
P-DARTS w/o M=2S0CIFAR-10124
P-DARTS w/ auxiliary skipS0CIFAR-10134
PC-DARTS w/o channel shufflingS0CIFAR-10145
PC-DARTS w/ auxiliary skipS0CIFAR-10155
", + "type": "table", + "image_path": "fc65dabc67dcc1ebf186d66f1ee6d607d7585eeeaf7ce0f267418961afcb1a93.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 107, + 566, + 501, + 620.3333333333334 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 107, + 620.3333333333334, + 501, + 674.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 107, + 674.6666666666667, + 501, + 729.0000000000001 + ], + "spans": [], + "index": 16 + } + ] + } + ], + "index": 14.0 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 119, + 77, + 497, + 252 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 119, + 77, + 497, + 252 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 119, + 77, + 497, + 252 + ], + "spans": [ + { + "bbox": [ + 119, + 77, + 497, + 252 + ], + "score": 0.974, + "type": "image", + "image_path": "2a9495f4bd817a16ca6ed904008ed76e11422c9c3eb2094c924bde1a84f3ae8c.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 119, + 77, + 497, + 135.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 119, + 135.33333333333334, + 497, + 193.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 119, + 193.66666666666669, + 497, + 252.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 264, + 505, + 309 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "Figure 7: More visualization of validation accuracy landscapes of DARTS (a,b) and DARTS- (e,f)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 276, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 232, + 288 + ], + "score": 1.0, + "content": "w.r.t. the architectural weights", + "type": "text" + }, + { + "bbox": [ + 233, + 278, + 241, + 286 + ], + "score": 0.69, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 276, + 505, + 288 + ], + "score": 1.0, + "content": "on CIFAR-10 in S0. Their contour maps are shown respectively", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 288, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 157, + 299 + ], + "score": 1.0, + "content": "in (c,d) and", + "type": "text" + }, + { + "bbox": [ + 158, + 288, + 177, + 298 + ], + "score": 0.26, + "content": "^ { ( \\mathrm { g , h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 177, + 288, + 505, + 299 + ], + "score": 1.0, + "content": ". The step of contour map is 0.1. The inferred models by DARTS- have higher", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 298, + 358, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 154, + 310 + ], + "score": 1.0, + "content": "accuracies (", + "type": "text" + }, + { + "bbox": [ + 154, + 298, + 185, + 309 + ], + "score": 0.77, + "content": "9 7 . 5 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 298, + 189, + 310 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 189, + 298, + 222, + 308 + ], + "score": 0.79, + "content": "9 7 . 4 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 298, + 281, + 310 + ], + "score": 1.0, + "content": ") than DARTS", + "type": "text" + }, + { + "bbox": [ + 282, + 298, + 315, + 309 + ], + "score": 0.71, + "content": "( 9 7 . 1 9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 298, + 318, + 310 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 318, + 298, + 351, + 309 + ], + "score": 0.77, + "content": "9 7 . 2 0 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 298, + 358, + 310 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 4.5 + } + ], + "index": 2.75 + }, + { + "type": "image", + "bbox": [ + 115, + 315, + 498, + 489 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 315, + 498, + 489 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 315, + 498, + 489 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 498, + 489 + ], + "score": 0.977, + "type": "image", + "image_path": "29923a592a8d5157ef46cf1e03033d7003870bed18aa6bb44063ecd4855316ca.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 115, + 315, + 498, + 373.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 115, + 373.0, + 498, + 431.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 115, + 431.0, + 498, + 489.0 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 501, + 505, + 535 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 106, + 500, + 505, + 514 + ], + "score": 1.0, + "content": "Figure 8: More visualization of validation accuracy landscapes of DARTS (a,b) and DARTS- (e,f)", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 229, + 525 + ], + "score": 1.0, + "content": "w.r.t. the architectural weights", + "type": "text" + }, + { + "bbox": [ + 230, + 514, + 237, + 522 + ], + "score": 0.67, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "on CIFAR-10 in S3. Their contour maps are shown respectively in", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 523, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 144, + 536 + ], + "score": 1.0, + "content": "(c,d) and", + "type": "text" + }, + { + "bbox": [ + 145, + 524, + 164, + 535 + ], + "score": 0.53, + "content": "^ { ( \\mathrm { g , h } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 523, + 296, + 536 + ], + "score": 1.0, + "content": ". The step of contour map is 0.1.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "table", + "bbox": [ + 107, + 566, + 501, + 729 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 198, + 552, + 414, + 564 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 197, + 551, + 414, + 566 + ], + "spans": [ + { + "bbox": [ + 197, + 551, + 414, + 566 + ], + "score": 1.0, + "content": "Table 12: List of experiments conducted in this paper", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "table_body", + "bbox": [ + 107, + 566, + 501, + 729 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 566, + 501, + 729 + ], + "spans": [ + { + "bbox": [ + 107, + 566, + 501, + 729 + ], + "score": 0.982, + "html": "
MethodSearch SpaceDatasetFiguresTables
Main textSupp.Main textSupp.
DARTSS0CIFAR-107
DARTSS3CIFAR-103
DARTS-S0CIFAR-104,5,9,71,2
DARTS-S1-S4CIFAR-1034,6,9,87,96
DARTS-S5ImageNet5,102
DARTS-S5MS COCO11
DARTS-S6CIFAR-103
DARTS-S1-S4CIFAR-1004,1179
DARTS-S0CIFAR-1004,111,10
P-DARTS w/o M=2S0CIFAR-10124
P-DARTS w/ auxiliary skipS0CIFAR-10134
PC-DARTS w/o channel shufflingS0CIFAR-10145
PC-DARTS w/ auxiliary skipS0CIFAR-10155
", + "type": "table", + "image_path": "fc65dabc67dcc1ebf186d66f1ee6d607d7585eeeaf7ce0f267418961afcb1a93.jpg" + } + ] + } + ], + "index": 15, + "virtual_lines": [ + { + "bbox": [ + 107, + 566, + 501, + 620.3333333333334 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 107, + 620.3333333333334, + 501, + 674.6666666666667 + ], + "spans": [], + "index": 15 + }, + { + "bbox": [ + 107, + 674.6666666666667, + 501, + 729.0000000000001 + ], + "spans": [], + "index": 16 + } + ] + } + ], + "index": 14.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 128, + 497, + 433 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 128, + 497, + 433 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 128, + 497, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 128, + 497, + 433 + ], + "score": 0.972, + "type": "image", + "image_path": "8ce376a98a358cfa0571a1d2c1837de2ad9b9ab517055b2a8eaa14526d983be0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 128, + 497, + 229.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 229.66666666666669, + 497, + 331.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 331.33333333333337, + 497, + 433.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 446, + 503, + 458 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "score": 1.0, + "content": "Figure 9: The best found normal cell and reduction cell in search spaces S0-S4 on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 478, + 506, + 641 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 478, + 506, + 641 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 478, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 107, + 478, + 506, + 641 + ], + "score": 0.494, + "type": "image", + "image_path": "b93e5bebc4be38b06a9a3b3f81f8cfbacd0c45fdf1315c783022ec1efcd1df70.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 478, + 506, + 532.3333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 532.3333333333334, + 506, + 586.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 586.6666666666667, + 506, + 641.0000000000001 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 167, + 667, + 444, + 680 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 665, + 444, + 682 + ], + "spans": [ + { + "bbox": [ + 165, + 665, + 444, + 682 + ], + "score": 1.0, + "content": "Figure 10: Architecture of DARTS-A searched on ImageNet dataset.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "17", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 128, + 497, + 433 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 128, + 497, + 433 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 128, + 497, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 128, + 497, + 433 + ], + "score": 0.972, + "type": "image", + "image_path": "8ce376a98a358cfa0571a1d2c1837de2ad9b9ab517055b2a8eaa14526d983be0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 128, + 497, + 229.66666666666669 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 229.66666666666669, + 497, + 331.33333333333337 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 331.33333333333337, + 497, + 433.00000000000006 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 446, + 503, + 458 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 460 + ], + "score": 1.0, + "content": "Figure 9: The best found normal cell and reduction cell in search spaces S0-S4 on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 478, + 506, + 641 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 478, + 506, + 641 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 478, + 506, + 641 + ], + "spans": [ + { + "bbox": [ + 107, + 478, + 506, + 641 + ], + "score": 0.494, + "type": "image", + "image_path": "b93e5bebc4be38b06a9a3b3f81f8cfbacd0c45fdf1315c783022ec1efcd1df70.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 478, + 506, + 532.3333333333334 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 532.3333333333334, + 506, + 586.6666666666667 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 586.6666666666667, + 506, + 641.0000000000001 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 167, + 667, + 444, + 680 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 165, + 665, + 444, + 682 + ], + "spans": [ + { + "bbox": [ + 165, + 665, + 444, + 682 + ], + "score": 1.0, + "content": "Figure 10: Architecture of DARTS-A searched on ImageNet dataset.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 89, + 495, + 395 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 89, + 495, + 395 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 89, + 495, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 89, + 495, + 395 + ], + "score": 0.972, + "type": "image", + "image_path": "bb4cac4e0acfb86c3427107e6b51ea45d471616b65c10e9dd877548cfb58daff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 89, + 495, + 191.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 191.0, + 495, + 293.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 293.0, + 495, + 395.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 406, + 506, + 430 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Figure 11: The best found normal cell and reduction cell in search spaces S0-S4 on CIFAR-100", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 416, + 140, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 140, + 431 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 123, + 476, + 476, + 681 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 476, + 476, + 681 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 476, + 476, + 681 + ], + "spans": [ + { + "bbox": [ + 123, + 476, + 476, + 681 + ], + "score": 0.962, + "type": "image", + "image_path": "59168baee3e57e3abbee41b713ad516736821a38c792ebd8b62b612c11983e4a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 123, + 476, + 476, + 544.3333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 123, + 544.3333333333334, + 476, + 612.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 123, + 612.6666666666667, + 476, + 681.0000000000001 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 692, + 504, + 716 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 693, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 505, + 705 + ], + "score": 1.0, + "content": "Figure 12: Found normal cells and reduction cells by P-DARTS (Chen et al., 2019b) without prior", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 703, + 375, + 717 + ], + "spans": [ + { + "bbox": [ + 108, + 704, + 133, + 715 + ], + "score": 0.65, + "content": "\\left( \\mathrm { M } { = } 2 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 703, + 375, + 717 + ], + "score": 1.0, + "content": "in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 89, + 495, + 395 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 89, + 495, + 395 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 89, + 495, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 89, + 495, + 395 + ], + "score": 0.972, + "type": "image", + "image_path": "bb4cac4e0acfb86c3427107e6b51ea45d471616b65c10e9dd877548cfb58daff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 89, + 495, + 191.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 191.0, + 495, + 293.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 293.0, + 495, + 395.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 406, + 506, + 430 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 106, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "Figure 11: The best found normal cell and reduction cell in search spaces S0-S4 on CIFAR-100", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 416, + 140, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 140, + 431 + ], + "score": 1.0, + "content": "dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 123, + 476, + 476, + 681 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 476, + 476, + 681 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 123, + 476, + 476, + 681 + ], + "spans": [ + { + "bbox": [ + 123, + 476, + 476, + 681 + ], + "score": 0.962, + "type": "image", + "image_path": "59168baee3e57e3abbee41b713ad516736821a38c792ebd8b62b612c11983e4a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 123, + 476, + 476, + 544.3333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 123, + 544.3333333333334, + 476, + 612.6666666666667 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 123, + 612.6666666666667, + 476, + 681.0000000000001 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 692, + 504, + 716 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 693, + 505, + 705 + ], + "spans": [ + { + "bbox": [ + 105, + 693, + 505, + 705 + ], + "score": 1.0, + "content": "Figure 12: Found normal cells and reduction cells by P-DARTS (Chen et al., 2019b) without prior", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 703, + 375, + 717 + ], + "spans": [ + { + "bbox": [ + 108, + 704, + 133, + 715 + ], + "score": 0.65, + "content": "\\left( \\mathrm { M } { = } 2 \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 703, + 375, + 717 + ], + "score": 1.0, + "content": "in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 131, + 479, + 295 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 131, + 479, + 295 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 131, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 124, + 131, + 479, + 295 + ], + "score": 0.916, + "type": "image", + "image_path": "4aa6caab7efecd691b10ad7b69ddf96431f90381de5f91f8d8e1de988b899a55.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 131, + 479, + 185.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 185.66666666666666, + 479, + 240.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 240.33333333333331, + 479, + 295.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 307, + 505, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "Figure 13: Found normal cells and reduction cells by P-DARTS (Chen et al., 2019b) with the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 318, + 493, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 493, + 332 + ], + "score": 1.0, + "content": "proposed auxiliary skip connections in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 125, + 422, + 479, + 657 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 422, + 479, + 657 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 125, + 422, + 479, + 657 + ], + "spans": [ + { + "bbox": [ + 125, + 422, + 479, + 657 + ], + "score": 0.947, + "type": "image", + "image_path": "9eb5847b3748618f79e68edd3e7290ba0b0664ce86d6cf8255072dc4ce5d8afe.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 125, + 422, + 479, + 500.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 125, + 500.3333333333333, + 479, + 578.6666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 125, + 578.6666666666666, + 479, + 657.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 670, + 505, + 693 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 683 + ], + "score": 1.0, + "content": "Figure 14: Found normal cells and reduction cells by PC-DARTS (Xu et al., 2020) without channel", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 681, + 385, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 385, + 694 + ], + "score": 1.0, + "content": "shuffling in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 131, + 479, + 295 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 131, + 479, + 295 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 131, + 479, + 295 + ], + "spans": [ + { + "bbox": [ + 124, + 131, + 479, + 295 + ], + "score": 0.916, + "type": "image", + "image_path": "4aa6caab7efecd691b10ad7b69ddf96431f90381de5f91f8d8e1de988b899a55.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 131, + 479, + 185.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 185.66666666666666, + 479, + 240.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 240.33333333333331, + 479, + 295.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 307, + 505, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 307, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 320 + ], + "score": 1.0, + "content": "Figure 13: Found normal cells and reduction cells by P-DARTS (Chen et al., 2019b) with the", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 318, + 493, + 332 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 493, + 332 + ], + "score": 1.0, + "content": "proposed auxiliary skip connections in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 125, + 422, + 479, + 657 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 125, + 422, + 479, + 657 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 125, + 422, + 479, + 657 + ], + "spans": [ + { + "bbox": [ + 125, + 422, + 479, + 657 + ], + "score": 0.947, + "type": "image", + "image_path": "9eb5847b3748618f79e68edd3e7290ba0b0664ce86d6cf8255072dc4ce5d8afe.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 125, + 422, + 479, + 500.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 125, + 500.3333333333333, + 479, + 578.6666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 125, + 578.6666666666666, + 479, + 657.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 670, + 505, + 693 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 670, + 505, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 505, + 683 + ], + "score": 1.0, + "content": "Figure 14: Found normal cells and reduction cells by PC-DARTS (Xu et al., 2020) without channel", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 681, + 385, + 694 + ], + "spans": [ + { + "bbox": [ + 106, + 681, + 385, + 694 + ], + "score": 1.0, + "content": "shuffling in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 101, + 488, + 303 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 101, + 488, + 303 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 101, + 488, + 303 + ], + "spans": [ + { + "bbox": [ + 118, + 101, + 488, + 303 + ], + "score": 0.954, + "type": "image", + "image_path": "87fda05d569e1baffc9c8528bca5132ed0ff26ec88d2ecc4d7d3f62b6ada74bf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 101, + 488, + 168.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 168.33333333333331, + 488, + 235.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 235.66666666666663, + 488, + 302.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 315, + 504, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "Figure 15: Found normal cells and reduction cells by PC-DARTS (Xu et al., 2020) with the pro-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 326, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 480, + 340 + ], + "score": 1.0, + "content": "posed auxiliary skip connections in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 124, + 376, + 483, + 520 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 376, + 483, + 520 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 376, + 483, + 520 + ], + "spans": [ + { + "bbox": [ + 124, + 376, + 483, + 520 + ], + "score": 0.97, + "type": "image", + "image_path": "28cb38bdb83fa8425399f9a319693b7f8899d497151c37ee46600553d55deca8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 124, + 376, + 483, + 424.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 124, + 424.0, + 483, + 472.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 472.0, + 483, + 520.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 533, + 504, + 556 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 173, + 547 + ], + "score": 1.0, + "content": "Figure 16: Keep", + "type": "text" + }, + { + "bbox": [ + 174, + 533, + 215, + 546 + ], + "score": 0.92, + "content": "\\beta _ { s k i p } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 531, + 506, + 547 + ], + "score": 1.0, + "content": "throughout the DARTS- searching in the DARTS’ standard search space", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 544, + 196, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 196, + 556 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 123, + 590, + 479, + 682 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 590, + 479, + 682 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 590, + 479, + 682 + ], + "spans": [ + { + "bbox": [ + 123, + 590, + 479, + 682 + ], + "score": 0.937, + "type": "image", + "image_path": "68f00aff2b2a8afb574bdce394963b31ea9113387ccc939bc6a9eb4dc8829d7f.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 123, + 590, + 479, + 620.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 123, + 620.6666666666666, + 479, + 651.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 123, + 651.3333333333333, + 479, + 681.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 694, + 505, + 718 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 694, + 506, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 278, + 708 + ], + "score": 1.0, + "content": "Figure 17: Best cells found when decaying", + "type": "text" + }, + { + "bbox": [ + 279, + 695, + 300, + 707 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 694, + 506, + 708 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 705, + 424, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 424, + 718 + ], + "score": 1.0, + "content": "for 150 and 200 epochs respectively in the DARTS search space on CIFAR-10.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 118, + 101, + 488, + 303 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 118, + 101, + 488, + 303 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 118, + 101, + 488, + 303 + ], + "spans": [ + { + "bbox": [ + 118, + 101, + 488, + 303 + ], + "score": 0.954, + "type": "image", + "image_path": "87fda05d569e1baffc9c8528bca5132ed0ff26ec88d2ecc4d7d3f62b6ada74bf.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 118, + 101, + 488, + 168.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 118, + 168.33333333333331, + 488, + 235.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 118, + 235.66666666666663, + 488, + 302.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 315, + 504, + 339 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 505, + 329 + ], + "score": 1.0, + "content": "Figure 15: Found normal cells and reduction cells by PC-DARTS (Xu et al., 2020) with the pro-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 326, + 480, + 340 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 480, + 340 + ], + "score": 1.0, + "content": "posed auxiliary skip connections in the DARTS’ standard search space on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 124, + 376, + 483, + 520 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 376, + 483, + 520 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 124, + 376, + 483, + 520 + ], + "spans": [ + { + "bbox": [ + 124, + 376, + 483, + 520 + ], + "score": 0.97, + "type": "image", + "image_path": "28cb38bdb83fa8425399f9a319693b7f8899d497151c37ee46600553d55deca8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 124, + 376, + 483, + 424.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 124, + 424.0, + 483, + 472.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 124, + 472.0, + 483, + 520.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 533, + 504, + 556 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 531, + 506, + 547 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 173, + 547 + ], + "score": 1.0, + "content": "Figure 16: Keep", + "type": "text" + }, + { + "bbox": [ + 174, + 533, + 215, + 546 + ], + "score": 0.92, + "content": "\\beta _ { s k i p } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 531, + 506, + 547 + ], + "score": 1.0, + "content": "throughout the DARTS- searching in the DARTS’ standard search space", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 544, + 196, + 556 + ], + "spans": [ + { + "bbox": [ + 106, + 544, + 196, + 556 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 123, + 590, + 479, + 682 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 590, + 479, + 682 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 123, + 590, + 479, + 682 + ], + "spans": [ + { + "bbox": [ + 123, + 590, + 479, + 682 + ], + "score": 0.937, + "type": "image", + "image_path": "68f00aff2b2a8afb574bdce394963b31ea9113387ccc939bc6a9eb4dc8829d7f.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 123, + 590, + 479, + 620.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 123, + 620.6666666666666, + 479, + 651.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 123, + 651.3333333333333, + 479, + 681.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 694, + 505, + 718 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 694, + 506, + 708 + ], + "spans": [ + { + "bbox": [ + 105, + 694, + 278, + 708 + ], + "score": 1.0, + "content": "Figure 17: Best cells found when decaying", + "type": "text" + }, + { + "bbox": [ + 279, + 695, + 300, + 707 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 694, + 506, + 708 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 705, + 424, + 718 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 424, + 718 + ], + "score": 1.0, + "content": "for 150 and 200 epochs respectively in the DARTS search space on CIFAR-10.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 102, + 481, + 215 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 102, + 481, + 215 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 102, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 122, + 102, + 481, + 215 + ], + "score": 0.942, + "type": "image", + "image_path": "fea26212834d7f4e1cfe56c801661a7736ef15af5724913154e8d655e7fb4005.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 102, + 481, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 139.66666666666666, + 481, + 177.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 177.33333333333331, + 481, + 214.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 227, + 504, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 191, + 240 + ], + "score": 1.0, + "content": "Figure 18: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 228, + 214, + 240 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 150 epochs in S2", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 238, + 196, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 196, + 251 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 122, + 293, + 482, + 407 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 293, + 482, + 407 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 293, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 122, + 293, + 482, + 407 + ], + "score": 0.941, + "type": "image", + "image_path": "38adf322fed77dcecab1707bc625a43ecd1647d879a59ad0e45618f7a4352828.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 122, + 293, + 482, + 331.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 122, + 331.0, + 482, + 369.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 122, + 369.0, + 482, + 407.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 419, + 505, + 443 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 191, + 433 + ], + "score": 1.0, + "content": "Figure 19: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 420, + 214, + 432 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 200 epochs in S2", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 430, + 197, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 197, + 443 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 126, + 488, + 479, + 679 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 488, + 479, + 679 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 126, + 488, + 479, + 679 + ], + "spans": [ + { + "bbox": [ + 126, + 488, + 479, + 679 + ], + "score": 0.957, + "type": "image", + "image_path": "3105c03082ffb033f6da5cfd8c399b4caf7f314f6bb471176ef9eef614cdd86f.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 126, + 488, + 479, + 551.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 126, + 551.6666666666666, + 479, + 615.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 126, + 615.3333333333333, + 479, + 678.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 691, + 505, + 714 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 191, + 704 + ], + "score": 1.0, + "content": "Figure 20: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 692, + 214, + 704 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 150 epochs in S3", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 702, + 196, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 196, + 714 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 105, + 25, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 14 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 122, + 102, + 481, + 215 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 102, + 481, + 215 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 122, + 102, + 481, + 215 + ], + "spans": [ + { + "bbox": [ + 122, + 102, + 481, + 215 + ], + "score": 0.942, + "type": "image", + "image_path": "fea26212834d7f4e1cfe56c801661a7736ef15af5724913154e8d655e7fb4005.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 122, + 102, + 481, + 139.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 122, + 139.66666666666666, + 481, + 177.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 122, + 177.33333333333331, + 481, + 214.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 227, + 504, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 104, + 226, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 226, + 191, + 240 + ], + "score": 1.0, + "content": "Figure 18: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 228, + 214, + 240 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 226, + 505, + 240 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 150 epochs in S2", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 238, + 196, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 196, + 251 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "image", + "bbox": [ + 122, + 293, + 482, + 407 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 122, + 293, + 482, + 407 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 122, + 293, + 482, + 407 + ], + "spans": [ + { + "bbox": [ + 122, + 293, + 482, + 407 + ], + "score": 0.941, + "type": "image", + "image_path": "38adf322fed77dcecab1707bc625a43ecd1647d879a59ad0e45618f7a4352828.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 122, + 293, + 482, + 331.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 122, + 331.0, + 482, + 369.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 122, + 369.0, + 482, + 407.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 419, + 505, + 443 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 419, + 505, + 433 + ], + "spans": [ + { + "bbox": [ + 105, + 419, + 191, + 433 + ], + "score": 1.0, + "content": "Figure 19: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 420, + 214, + 432 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 419, + 505, + 433 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 200 epochs in S2", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 430, + 197, + 443 + ], + "spans": [ + { + "bbox": [ + 105, + 430, + 197, + 443 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + } + ], + "index": 7.25 + }, + { + "type": "image", + "bbox": [ + 126, + 488, + 479, + 679 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 126, + 488, + 479, + 679 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 126, + 488, + 479, + 679 + ], + "spans": [ + { + "bbox": [ + 126, + 488, + 479, + 679 + ], + "score": 0.957, + "type": "image", + "image_path": "3105c03082ffb033f6da5cfd8c399b4caf7f314f6bb471176ef9eef614cdd86f.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 126, + 488, + 479, + 551.6666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 126, + 551.6666666666666, + 479, + 615.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 126, + 615.3333333333333, + 479, + 678.9999999999999 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 691, + 505, + 714 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 105, + 690, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 690, + 191, + 704 + ], + "score": 1.0, + "content": "Figure 20: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 692, + 214, + 704 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 690, + 505, + 704 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 150 epochs in S3", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 702, + 196, + 714 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 196, + 714 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + } + ], + "index": 12.25 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 51, + 482, + 498 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 51, + 482, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 51, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 124, + 51, + 482, + 498 + ], + "score": 0.134, + "type": "image", + "image_path": "481adfe5288a09b65c0c085885b5bd095d167481e63b455d2b4e4f3ca5e83a57.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 51, + 482, + 200.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 200.0, + 482, + 349.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 349.0, + 482, + 498.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 499, + 505, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 191, + 512 + ], + "score": 1.0, + "content": "Figure 21: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 500, + 214, + 511 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 200 epochs in S3", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 510, + 196, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 196, + 522 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 21, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 292, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 293, + 38 + ], + "score": 1.0, + "content": "Published as a conference paper at ICLR 2021", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 124, + 51, + 482, + 498 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 124, + 51, + 482, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 124, + 51, + 482, + 498 + ], + "spans": [ + { + "bbox": [ + 124, + 51, + 482, + 498 + ], + "score": 0.134, + "type": "image", + "image_path": "481adfe5288a09b65c0c085885b5bd095d167481e63b455d2b4e4f3ca5e83a57.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 124, + 51, + 482, + 200.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 124, + 200.0, + 482, + 349.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 124, + 349.0, + 482, + 498.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 104, + 499, + 505, + 522 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 498, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 191, + 512 + ], + "score": 1.0, + "content": "Figure 21: Decaying", + "type": "text" + }, + { + "bbox": [ + 191, + 500, + 214, + 511 + ], + "score": 0.91, + "content": "\\beta _ { s k i p }", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 498, + 505, + 512 + ], + "score": 1.0, + "content": "in the last 50 epochs during the DARTS- searching for 200 epochs in S3", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 510, + 196, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 196, + 522 + ], + "score": 1.0, + "content": "on CIFAR-10 dataset.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/KLH36ELmwIB/KLH36ELmwIB_model.json b/parse/train/KLH36ELmwIB/KLH36ELmwIB_model.json new file mode 100644 index 0000000000000000000000000000000000000000..614632b7aef0d3b733b16480a5dc4f2fbc6a2bbb --- /dev/null +++ b/parse/train/KLH36ELmwIB/KLH36ELmwIB_model.json @@ -0,0 +1,41603 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 686, + 1302, + 686, + 1302, + 1173, + 398, + 1173 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1319, + 1404, + 1319, + 1404, + 1778, + 298, + 1778 + ], + "score": 0.983 + }, + { + "category_id": 0, + "poly": [ + 300, + 219, + 1398, + 219, + 1398, + 324, + 300, + 324 + ], + "score": 0.961 + }, + { + "category_id": 2, + "poly": [ + 334, + 1975, + 731, + 1975, + 731, + 2033, + 334, + 2033 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 302, + 1246, + 573, + 1246, + 573, + 1282, + 302, + 1282 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.878 + }, + { + "category_id": 0, + "poly": [ + 773, + 612, + 926, + 612, + 926, + 645, + 773, + 645 + ], + "score": 0.873 + }, + { + "category_id": 1, + "poly": [ + 370, + 1812, + 1402, + 1812, + 1402, + 1873, + 370, + 1873 + ], + "score": 0.816 + }, + { + "category_id": 1, + "poly": [ + 372, + 1890, + 1349, + 1890, + 1349, + 1922, + 372, + 1922 + ], + "score": 0.793 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.701 + }, + { + "category_id": 1, + "poly": [ + 313, + 372, + 1385, + 372, + 1385, + 530, + 313, + 530 + ], + "score": 0.486 + }, + { + "category_id": 13, + "poly": [ + 456, + 375, + 521, + 375, + 521, + 407, + 456, + 407 + ], + "score": 0.73, + "latex": "\\mathbf { C h u ^ { 1 } }" + }, + { + "category_id": 13, + "poly": [ + 968, + 375, + 1046, + 375, + 1046, + 407, + 968, + 407 + ], + "score": 0.71, + "latex": "\\mathbf { L u ^ { 1 , 3 * } }" + }, + { + "category_id": 13, + "poly": [ + 1310, + 373, + 1384, + 373, + 1384, + 407, + 1310, + 407 + ], + "score": 0.36, + "latex": "\\mathbf { Y a n } ^ { 2 \\dagger }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 218.0, + 1404.0, + 218.0, + 1404.0, + 273.0, + 294.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 273.0, + 1167.0, + 273.0, + 1167.0, + 330.0, + 291.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1970.0, + 736.0, + 1970.0, + 736.0, + 2006.0, + 330.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 2000.0, + 578.0, + 2000.0, + 578.0, + 2039.0, + 330.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1243.0, + 579.0, + 1243.0, + 579.0, + 1290.0, + 294.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 611.0, + 932.0, + 611.0, + 932.0, + 648.0, + 769.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 685.0, + 1306.0, + 685.0, + 1306.0, + 722.0, + 393.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 713.0, + 1305.0, + 713.0, + 1305.0, + 755.0, + 392.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 748.0, + 1305.0, + 748.0, + 1305.0, + 783.0, + 393.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 778.0, + 1306.0, + 778.0, + 1306.0, + 812.0, + 392.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 808.0, + 1305.0, + 808.0, + 1305.0, + 842.0, + 394.0, + 842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 840.0, + 1306.0, + 840.0, + 1306.0, + 871.0, + 395.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 869.0, + 1306.0, + 869.0, + 1306.0, + 903.0, + 393.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 901.0, + 1306.0, + 901.0, + 1306.0, + 932.0, + 393.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 932.0, + 1306.0, + 932.0, + 1306.0, + 963.0, + 394.0, + 963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 961.0, + 1305.0, + 961.0, + 1305.0, + 995.0, + 393.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 990.0, + 1306.0, + 990.0, + 1306.0, + 1026.0, + 392.0, + 1026.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1021.0, + 1305.0, + 1021.0, + 1305.0, + 1057.0, + 394.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1051.0, + 1306.0, + 1051.0, + 1306.0, + 1089.0, + 393.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1081.0, + 1307.0, + 1081.0, + 1307.0, + 1118.0, + 393.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1114.0, + 1305.0, + 1114.0, + 1305.0, + 1146.0, + 395.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1145.0, + 1105.0, + 1145.0, + 1105.0, + 1176.0, + 394.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1320.0, + 1404.0, + 1320.0, + 1404.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1349.0, + 1404.0, + 1349.0, + 1404.0, + 1386.0, + 293.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1379.0, + 1405.0, + 1379.0, + 1405.0, + 1419.0, + 292.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1410.0, + 1404.0, + 1410.0, + 1404.0, + 1447.0, + 293.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1442.0, + 1405.0, + 1442.0, + 1405.0, + 1477.0, + 292.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1471.0, + 1408.0, + 1471.0, + 1408.0, + 1510.0, + 293.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1501.0, + 1404.0, + 1501.0, + 1404.0, + 1539.0, + 292.0, + 1539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1531.0, + 1406.0, + 1531.0, + 1406.0, + 1571.0, + 291.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1563.0, + 1404.0, + 1563.0, + 1404.0, + 1598.0, + 294.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1595.0, + 1405.0, + 1595.0, + 1405.0, + 1630.0, + 293.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1625.0, + 1404.0, + 1625.0, + 1404.0, + 1660.0, + 296.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1693.0, + 293.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1723.0, + 293.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1714.0, + 1406.0, + 1714.0, + 1406.0, + 1752.0, + 293.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1746.0, + 728.0, + 1746.0, + 728.0, + 1782.0, + 293.0, + 1782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1810.0, + 1401.0, + 1810.0, + 1401.0, + 1845.0, + 373.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1840.0, + 912.0, + 1840.0, + 912.0, + 1875.0, + 395.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1886.0, + 1354.0, + 1886.0, + 1354.0, + 1928.0, + 370.0, + 1928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 308.0, + 374.0, + 455.0, + 374.0, + 455.0, + 411.0, + 308.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 374.0, + 967.0, + 374.0, + 967.0, + 411.0, + 522.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 374.0, + 1309.0, + 374.0, + 1309.0, + 411.0, + 1047.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 401.0, + 1308.0, + 401.0, + 1308.0, + 446.0, + 310.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 437.0, + 1149.0, + 437.0, + 1149.0, + 475.0, + 312.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 467.0, + 891.0, + 467.0, + 891.0, + 505.0, + 312.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 313.0, + 500.0, + 736.0, + 500.0, + 736.0, + 533.0, + 313.0, + 533.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1135, + 1404, + 1135, + 1404, + 1502, + 298, + 1502 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 368, + 824, + 368, + 824, + 674, + 298, + 674 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1517, + 1404, + 1517, + 1404, + 1943, + 297, + 1943 + ], + "score": 0.978 + }, + { + "category_id": 4, + "poly": [ + 846, + 559, + 1403, + 559, + 1403, + 744, + 846, + 744 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 854, + 240, + 1395, + 240, + 1395, + 526, + 854, + 526 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 231, + 823, + 231, + 823, + 351, + 298, + 351 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 298, + 889, + 1404, + 889, + 1404, + 1011, + 298, + 1011 + ], + "score": 0.97 + }, + { + "category_id": 1, + "poly": [ + 300, + 782, + 1400, + 782, + 1400, + 874, + 300, + 874 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 297, + 690, + 824, + 690, + 824, + 781, + 297, + 781 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 295, + 1978, + 1402, + 1978, + 1402, + 2034, + 295, + 2034 + ], + "score": 0.936 + }, + { + "category_id": 0, + "poly": [ + 301, + 1061, + 587, + 1061, + 587, + 1097, + 301, + 1097 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.747 + }, + { + "category_id": 13, + "poly": [ + 680, + 980, + 756, + 980, + 756, + 1010, + 680, + 1010 + ], + "score": 0.87, + "latex": "0 . 2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 553, + 813, + 593, + 813, + 593, + 841, + 553, + 841 + ], + "score": 0.86, + "latex": "3 \\times" + }, + { + "category_id": 13, + "poly": [ + 297, + 980, + 360, + 980, + 360, + 1011, + 297, + 1011 + ], + "score": 0.86, + "latex": "0 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 895, + 653, + 916, + 653, + 916, + 683, + 895, + 683 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1670, + 1246, + 1670, + 1246, + 1700, + 1136, + 1700 + ], + "score": 0.47, + "latex": "{ \\mathrm { D A R T S } } +" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 559.0, + 1403.0, + 559.0, + 1403.0, + 592.0, + 845.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 591.0, + 1403.0, + 591.0, + 1403.0, + 625.0, + 846.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 621.0, + 1404.0, + 621.0, + 1404.0, + 656.0, + 844.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 650.0, + 894.0, + 650.0, + 894.0, + 686.0, + 844.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 650.0, + 1403.0, + 650.0, + 1403.0, + 686.0, + 917.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 680.0, + 1404.0, + 680.0, + 1404.0, + 717.0, + 843.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 711.0, + 952.0, + 711.0, + 952.0, + 749.0, + 844.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 337.0, + 898.0, + 337.0, + 898.0, + 360.0, + 856.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 338.0, + 949.0, + 338.0, + 949.0, + 359.0, + 903.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 337.0, + 999.0, + 337.0, + 999.0, + 362.0, + 953.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 336.0, + 1084.0, + 336.0, + 1084.0, + 361.0, + 1006.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 335.0, + 1142.0, + 335.0, + 1142.0, + 364.0, + 1096.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 338.0, + 1190.0, + 338.0, + 1190.0, + 359.0, + 1144.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 337.0, + 1238.0, + 337.0, + 1238.0, + 362.0, + 1193.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 336.0, + 1323.0, + 336.0, + 1323.0, + 362.0, + 1246.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 325.0, + 1400.0, + 325.0, + 1400.0, + 356.0, + 1335.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 374.0, + 901.0, + 374.0, + 901.0, + 394.0, + 878.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 378.0, + 932.0, + 378.0, + 932.0, + 387.0, + 922.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 376.0, + 964.0, + 376.0, + 964.0, + 389.0, + 951.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 374.0, + 1012.0, + 374.0, + 1012.0, + 394.0, + 987.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 373.0, + 1059.0, + 373.0, + 1059.0, + 395.0, + 1030.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1121.0, + 374.0, + 1146.0, + 374.0, + 1146.0, + 394.0, + 1121.0, + 394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 378.0, + 1203.0, + 378.0, + 1203.0, + 387.0, + 1193.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1226.0, + 373.0, + 1255.0, + 373.0, + 1255.0, + 395.0, + 1226.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 373.0, + 1290.0, + 373.0, + 1290.0, + 395.0, + 1262.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 430.0, + 1241.0, + 430.0, + 1241.0, + 447.0, + 1225.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 494.0, + 1018.0, + 494.0, + 1018.0, + 525.0, + 897.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 493.0, + 1276.0, + 493.0, + 1276.0, + 527.0, + 1144.0, + 527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 342.0, + 1384.0, + 342.0, + 1384.0, + 371.5, + 1346.0, + 371.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.25, + 403.0, + 967.25, + 403.0, + 967.25, + 419.0, + 942.25, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 417.5, + 1279.0, + 417.5, + 1279.0, + 444.5, + 1237.0, + 444.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1973.0, + 1406.0, + 1973.0, + 1406.0, + 2011.0, + 331.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 817.0, + 2004.0, + 817.0, + 2035.0, + 295.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1058.0, + 594.0, + 1058.0, + 594.0, + 1105.0, + 291.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1136.0, + 1405.0, + 1136.0, + 1405.0, + 1170.0, + 294.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1166.0, + 1405.0, + 1166.0, + 1405.0, + 1201.0, + 293.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1197.0, + 1404.0, + 1197.0, + 1404.0, + 1231.0, + 294.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1227.0, + 1402.0, + 1227.0, + 1402.0, + 1261.0, + 293.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1255.0, + 1405.0, + 1255.0, + 1405.0, + 1292.0, + 292.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1288.0, + 1405.0, + 1288.0, + 1405.0, + 1322.0, + 294.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1317.0, + 1406.0, + 1317.0, + 1406.0, + 1354.0, + 292.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 1404.0, + 1345.0, + 1404.0, + 1387.0, + 292.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1379.0, + 1402.0, + 1379.0, + 1402.0, + 1413.0, + 294.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1406.0, + 1408.0, + 1406.0, + 1408.0, + 1446.0, + 292.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1438.0, + 1406.0, + 1438.0, + 1406.0, + 1477.0, + 291.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1472.0, + 763.0, + 1472.0, + 763.0, + 1506.0, + 296.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 367.0, + 827.0, + 367.0, + 827.0, + 399.0, + 295.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 400.0, + 828.0, + 400.0, + 828.0, + 431.0, + 294.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 428.0, + 827.0, + 428.0, + 827.0, + 461.0, + 296.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 460.0, + 827.0, + 460.0, + 827.0, + 493.0, + 293.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 492.0, + 827.0, + 492.0, + 827.0, + 525.0, + 295.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 522.0, + 826.0, + 522.0, + 826.0, + 554.0, + 293.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 551.0, + 826.0, + 551.0, + 826.0, + 582.0, + 295.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 580.0, + 826.0, + 580.0, + 826.0, + 619.0, + 293.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 612.0, + 825.0, + 612.0, + 825.0, + 645.0, + 295.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 639.0, + 710.0, + 639.0, + 710.0, + 678.0, + 293.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1514.0, + 1404.0, + 1514.0, + 1404.0, + 1553.0, + 292.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1545.0, + 1405.0, + 1545.0, + 1405.0, + 1583.0, + 292.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1577.0, + 1404.0, + 1577.0, + 1404.0, + 1612.0, + 294.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1610.0, + 1402.0, + 1610.0, + 1402.0, + 1641.0, + 295.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1636.0, + 1404.0, + 1636.0, + 1404.0, + 1673.0, + 292.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1669.0, + 1135.0, + 1669.0, + 1135.0, + 1705.0, + 294.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 1669.0, + 1404.0, + 1669.0, + 1404.0, + 1705.0, + 1247.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1696.0, + 1406.0, + 1696.0, + 1406.0, + 1736.0, + 292.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1767.0, + 292.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1759.0, + 1406.0, + 1759.0, + 1406.0, + 1796.0, + 292.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 1404.0, + 1820.0, + 1404.0, + 1856.0, + 294.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1887.0, + 295.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1880.0, + 1405.0, + 1880.0, + 1405.0, + 1918.0, + 294.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1913.0, + 926.0, + 1913.0, + 926.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 827.0, + 228.0, + 827.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 827.0, + 261.0, + 827.0, + 294.0, + 294.0, + 294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 290.0, + 827.0, + 290.0, + 827.0, + 322.0, + 294.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 319.0, + 426.0, + 319.0, + 426.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 888.0, + 1405.0, + 888.0, + 1405.0, + 923.0, + 294.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 918.0, + 1405.0, + 918.0, + 1405.0, + 956.0, + 292.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 950.0, + 1404.0, + 950.0, + 1404.0, + 986.0, + 293.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 981.0, + 679.0, + 981.0, + 679.0, + 1014.0, + 361.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 981.0, + 1289.0, + 981.0, + 1289.0, + 1014.0, + 757.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 780.0, + 1404.0, + 780.0, + 1404.0, + 817.0, + 294.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 810.0, + 552.0, + 810.0, + 552.0, + 846.0, + 294.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 810.0, + 1404.0, + 810.0, + 1404.0, + 846.0, + 594.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 841.0, + 845.0, + 841.0, + 845.0, + 877.0, + 295.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 689.0, + 827.0, + 689.0, + 827.0, + 722.0, + 296.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 721.0, + 828.0, + 721.0, + 828.0, + 754.0, + 296.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 750.0, + 828.0, + 750.0, + 828.0, + 783.0, + 295.0, + 783.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 361, + 1405, + 361, + 1405, + 576, + 297, + 576 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 590, + 1405, + 590, + 1405, + 746, + 298, + 746 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 1347, + 935, + 1347, + 935, + 1561, + 299, + 1561 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1575, + 934, + 1575, + 934, + 1911, + 298, + 1911 + ], + "score": 0.977 + }, + { + "category_id": 4, + "poly": [ + 957, + 1608, + 1403, + 1608, + 1403, + 1852, + 957, + 1852 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 299, + 1913, + 1402, + 1913, + 1402, + 2034, + 299, + 2034 + ], + "score": 0.96 + }, + { + "category_id": 3, + "poly": [ + 961, + 1358, + 1366, + 1358, + 1366, + 1587, + 961, + 1587 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 584, + 974, + 1114, + 974, + 1114, + 1065, + 584, + 1065 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 667, + 1170, + 1031, + 1170, + 1031, + 1251, + 667, + 1251 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 294, + 1087, + 1402, + 1087, + 1402, + 1150, + 294, + 1150 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 472, + 772, + 1225, + 772, + 1225, + 862, + 472, + 862 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 296, + 884, + 1406, + 884, + 1406, + 948, + 296, + 948 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 299, + 1268, + 1409, + 1268, + 1409, + 1332, + 299, + 1332 + ], + "score": 0.949 + }, + { + "category_id": 0, + "poly": [ + 300, + 225, + 476, + 225, + 476, + 261, + 300, + 261 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 299, + 300, + 521, + 300, + 521, + 333, + 299, + 333 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.898 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1195, + 1400, + 1195, + 1400, + 1225, + 1366, + 1225 + ], + "score": 0.89 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1003, + 1400, + 1003, + 1400, + 1033, + 1366, + 1033 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 800, + 1400, + 800, + 1400, + 830, + 1366, + 830 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.614 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.458 + }, + { + "category_id": 14, + "poly": [ + 472, + 767, + 1227, + 767, + 1227, + 864, + 472, + 864 + ], + "score": 0.94, + "latex": "\\frac { \\partial \\mathcal { L } } { \\partial X _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + 1 } } \\cdot \\left( \\frac { \\partial f _ { i + 1 } } { \\partial X _ { i } } + 1 \\right) = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + j } } \\cdot \\prod _ { k = 1 } ^ { j } \\left( \\frac { \\partial f _ { i + k } } { \\partial X _ { i + k - 1 } } + 1 \\right)" + }, + { + "category_id": 14, + "poly": [ + 666, + 1171, + 1034, + 1171, + 1034, + 1249, + 666, + 1249 + ], + "score": 0.94, + "latex": "\\frac { \\partial \\mathcal { L } } { \\partial X _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + 1 } } \\cdot \\left( \\frac { \\partial f _ { i + 1 } } { \\partial X _ { i } } + \\beta \\right)" + }, + { + "category_id": 14, + "poly": [ + 584, + 969, + 1116, + 969, + 1116, + 1066, + 584, + 1066 + ], + "score": 0.94, + "latex": "\\frac { \\partial \\mathcal { L } } { \\partial W _ { i } } = \\frac { \\partial \\mathcal { L } } { \\partial X _ { i + j } } \\cdot \\prod _ { k = 1 } ^ { j } \\left( \\frac { \\partial f _ { i + k } } { \\partial X _ { i + k - 1 } } + \\mathbb { 1 } \\right) \\cdot \\frac { \\partial f _ { i } } { \\partial W _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 1017, + 621, + 1117, + 621, + 1117, + 655, + 1017, + 655 + ], + "score": 0.93, + "latex": "( i + 1 ) ^ { \\mathrm { t h } }" + }, + { + "category_id": 13, + "poly": [ + 299, + 1407, + 437, + 1407, + 437, + 1441, + 299, + 1441 + ], + "score": 0.92, + "latex": "\\{ 0 , 0 . 5 , 1 . 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 384, + 1729, + 447, + 1729, + 447, + 1763, + 384, + 1763 + ], + "score": 0.92, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 915, + 654, + 968, + 654, + 968, + 685, + 915, + 685 + ], + "score": 0.91, + "latex": "f _ { i + 1 }" + }, + { + "category_id": 13, + "poly": [ + 735, + 1790, + 798, + 1790, + 798, + 1824, + 735, + 1824 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 415, + 1608, + 476, + 1608, + 476, + 1642, + 415, + 1642 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 710, + 1638, + 773, + 1638, + 773, + 1671, + 710, + 1671 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 611, + 686, + 676, + 686, + 676, + 716, + 611, + 716 + ], + "score": 0.91, + "latex": "W _ { i + 1 }" + }, + { + "category_id": 13, + "poly": [ + 479, + 653, + 828, + 653, + 828, + 685, + 479, + 685 + ], + "score": 0.9, + "latex": "X _ { i + 1 } = f _ { i + 1 } ( X _ { i } , W _ { i + 1 } ) + X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1303, + 653, + 1400, + 653, + 1400, + 686, + 1303, + 686 + ], + "score": 0.9, + "latex": "( i + 1 ) ^ { \\mathrm { t h } }" + }, + { + "category_id": 13, + "poly": [ + 364, + 1270, + 441, + 1270, + 441, + 1302, + 364, + 1302 + ], + "score": 0.9, + "latex": "\\beta < 1" + }, + { + "category_id": 13, + "poly": [ + 1290, + 623, + 1353, + 623, + 1353, + 654, + 1290, + 654 + ], + "score": 0.89, + "latex": "X _ { i + 1 }" + }, + { + "category_id": 13, + "poly": [ + 1123, + 1119, + 1157, + 1119, + 1157, + 1149, + 1123, + 1149 + ], + "score": 0.89, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 689, + 916, + 726, + 916, + 726, + 946, + 689, + 946 + ], + "score": 0.89, + "latex": "W _ { i }" + }, + { + "category_id": 13, + "poly": [ + 298, + 715, + 332, + 715, + 332, + 744, + 298, + 744 + ], + "score": 0.88, + "latex": "X _ { i }" + }, + { + "category_id": 13, + "poly": [ + 421, + 1119, + 441, + 1119, + 441, + 1149, + 421, + 1149 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 532, + 1301, + 553, + 1301, + 553, + 1331, + 532, + 1331 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 840, + 1410, + 862, + 1410, + 862, + 1439, + 840, + 1439 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 750, + 1439, + 772, + 1439, + 772, + 1470, + 750, + 1470 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1048, + 1763, + 1070, + 1763, + 1070, + 1795, + 1048, + 1795 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1641, + 1060, + 1641, + 1060, + 1672, + 1037, + 1672 + ], + "score": 0.82, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 850, + 1378, + 872, + 1378, + 872, + 1409, + 850, + 1409 + ], + "score": 0.81, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1156, + 685, + 1177, + 685, + 1177, + 711, + 1156, + 711 + ], + "score": 0.81, + "latex": "\\mathcal { L }" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1640, + 1397, + 1640, + 1397, + 1673, + 1260, + 1673 + ], + "score": 0.81, + "latex": "\\{ 0 , \\ 0 . 5 , \\ 1 \\} )" + }, + { + "category_id": 13, + "poly": [ + 912, + 1500, + 933, + 1500, + 933, + 1532, + 912, + 1532 + ], + "score": 0.79, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 607, + 628, + 627, + 628, + 627, + 650, + 607, + 650 + ], + "score": 0.74, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 575, + 1378, + 643, + 1378, + 643, + 1410, + 575, + 1410 + ], + "score": 0.38, + "latex": "\\mathrm { F i g } \\ 2" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1608.0, + 1403.0, + 1608.0, + 1403.0, + 1642.0, + 956.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1635.0, + 1036.0, + 1635.0, + 1036.0, + 1675.0, + 953.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1061.0, + 1635.0, + 1259.0, + 1635.0, + 1259.0, + 1675.0, + 1061.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1675.0, + 1398.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1669.0, + 1405.0, + 1669.0, + 1405.0, + 1703.0, + 955.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1702.0, + 1403.0, + 1702.0, + 1403.0, + 1735.0, + 954.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1765.0, + 953.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1762.0, + 1047.0, + 1762.0, + 1047.0, + 1795.0, + 954.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1762.0, + 1405.0, + 1762.0, + 1405.0, + 1795.0, + 1071.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1793.0, + 1405.0, + 1793.0, + 1405.0, + 1823.0, + 954.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1821.0, + 1123.0, + 1821.0, + 1123.0, + 1853.0, + 953.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1364.0, + 1004.0, + 1364.0, + 1004.0, + 1383.0, + 976.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1376.0, + 1059.0, + 1376.0, + 1059.0, + 1396.0, + 1014.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1374.0, + 1181.0, + 1374.0, + 1181.0, + 1384.0, + 1153.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 1375.0, + 1222.0, + 1375.0, + 1222.0, + 1382.0, + 1205.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1397.0, + 1003.0, + 1397.0, + 1003.0, + 1420.0, + 974.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1419.0, + 1221.0, + 1419.0, + 1221.0, + 1439.0, + 1199.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1427.0, + 1202.0, + 1427.0, + 1202.0, + 1504.0, + 1182.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1434.0, + 1004.0, + 1434.0, + 1004.0, + 1454.0, + 975.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1444.0, + 979.0, + 1444.0, + 979.0, + 1479.0, + 958.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1455.0, + 1220.0, + 1455.0, + 1220.0, + 1478.0, + 1199.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1469.0, + 1004.0, + 1469.0, + 1004.0, + 1489.0, + 975.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1501.0, + 1009.0, + 1501.0, + 1009.0, + 1527.0, + 972.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1485.0, + 1095.0, + 1485.0, + 1095.0, + 1526.0, + 1042.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1497.0, + 1336.0, + 1497.0, + 1336.0, + 1528.0, + 1200.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1520.0, + 1094.0, + 1520.0, + 1094.0, + 1544.0, + 1042.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1522.0, + 1302.0, + 1522.0, + 1302.0, + 1548.0, + 1257.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1538.0, + 1013.0, + 1538.0, + 1013.0, + 1571.0, + 974.0, + 1571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1545.0, + 1062.0, + 1545.0, + 1062.0, + 1575.0, + 1034.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 1548.0, + 1105.0, + 1548.0, + 1105.0, + 1573.0, + 1080.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1552.0, + 1149.0, + 1552.0, + 1149.0, + 1572.0, + 1124.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1552.0, + 1192.0, + 1552.0, + 1192.0, + 1572.0, + 1168.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1209.0, + 1551.0, + 1240.0, + 1551.0, + 1240.0, + 1573.0, + 1209.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1551.0, + 1284.0, + 1551.0, + 1284.0, + 1573.0, + 1253.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1298.0, + 1551.0, + 1327.0, + 1551.0, + 1327.0, + 1573.0, + 1298.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1553.0, + 1371.0, + 1553.0, + 1371.0, + 1572.0, + 1342.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1568.0, + 1201.0, + 1568.0, + 1201.0, + 1589.0, + 1157.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.25, + 1370.0, + 1155.25, + 1370.0, + 1155.25, + 1390.0, + 1066.25, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 222.0, + 483.0, + 222.0, + 483.0, + 270.0, + 289.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 298.0, + 526.0, + 298.0, + 526.0, + 337.0, + 293.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 360.0, + 1406.0, + 360.0, + 1406.0, + 397.0, + 294.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 388.0, + 1406.0, + 388.0, + 1406.0, + 430.0, + 291.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 423.0, + 1405.0, + 423.0, + 1405.0, + 457.0, + 294.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 454.0, + 1406.0, + 454.0, + 1406.0, + 488.0, + 295.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 483.0, + 1405.0, + 483.0, + 1405.0, + 518.0, + 294.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 512.0, + 1408.0, + 512.0, + 1408.0, + 551.0, + 292.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 546.0, + 636.0, + 546.0, + 636.0, + 578.0, + 292.0, + 578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 588.0, + 1406.0, + 588.0, + 1406.0, + 627.0, + 293.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 616.0, + 606.0, + 616.0, + 606.0, + 660.0, + 292.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 616.0, + 1016.0, + 616.0, + 1016.0, + 660.0, + 628.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 616.0, + 1289.0, + 616.0, + 1289.0, + 660.0, + 1118.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 616.0, + 1406.0, + 616.0, + 1406.0, + 660.0, + 1354.0, + 660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 648.0, + 478.0, + 648.0, + 478.0, + 688.0, + 291.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 648.0, + 914.0, + 648.0, + 914.0, + 688.0, + 829.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 648.0, + 1302.0, + 648.0, + 1302.0, + 688.0, + 969.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 648.0, + 1406.0, + 648.0, + 1406.0, + 688.0, + 1401.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 679.0, + 610.0, + 679.0, + 610.0, + 719.0, + 292.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 679.0, + 1155.0, + 679.0, + 1155.0, + 719.0, + 677.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 679.0, + 1408.0, + 679.0, + 1408.0, + 719.0, + 1178.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 712.0, + 1146.0, + 712.0, + 1146.0, + 745.0, + 333.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1346.0, + 936.0, + 1346.0, + 936.0, + 1380.0, + 296.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1375.0, + 574.0, + 1375.0, + 574.0, + 1409.0, + 295.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1375.0, + 849.0, + 1375.0, + 849.0, + 1409.0, + 644.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 1375.0, + 935.0, + 1375.0, + 935.0, + 1409.0, + 873.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1404.0, + 298.0, + 1404.0, + 298.0, + 1445.0, + 294.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1404.0, + 839.0, + 1404.0, + 839.0, + 1445.0, + 438.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 1404.0, + 940.0, + 1404.0, + 940.0, + 1445.0, + 863.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1437.0, + 749.0, + 1437.0, + 749.0, + 1472.0, + 296.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1437.0, + 936.0, + 1437.0, + 936.0, + 1472.0, + 773.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1467.0, + 937.0, + 1467.0, + 937.0, + 1500.0, + 294.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1498.0, + 911.0, + 1498.0, + 911.0, + 1533.0, + 296.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1498.0, + 938.0, + 1498.0, + 938.0, + 1533.0, + 934.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1524.0, + 893.0, + 1524.0, + 893.0, + 1568.0, + 292.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1576.0, + 936.0, + 1576.0, + 936.0, + 1609.0, + 297.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1608.0, + 414.0, + 1608.0, + 414.0, + 1641.0, + 294.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1608.0, + 936.0, + 1608.0, + 936.0, + 1641.0, + 477.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 709.0, + 1635.0, + 709.0, + 1674.0, + 293.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1635.0, + 937.0, + 1635.0, + 937.0, + 1674.0, + 774.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1668.0, + 936.0, + 1668.0, + 936.0, + 1700.0, + 295.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1698.0, + 937.0, + 1698.0, + 937.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1727.0, + 383.0, + 1727.0, + 383.0, + 1764.0, + 294.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1727.0, + 937.0, + 1727.0, + 937.0, + 1764.0, + 448.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1757.0, + 937.0, + 1757.0, + 937.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1786.0, + 734.0, + 1786.0, + 734.0, + 1826.0, + 293.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 1786.0, + 939.0, + 1786.0, + 939.0, + 1826.0, + 799.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1820.0, + 938.0, + 1820.0, + 938.0, + 1853.0, + 294.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 936.0, + 1849.0, + 936.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1879.0, + 937.0, + 1879.0, + 937.0, + 1915.0, + 295.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1911.0, + 1407.0, + 1911.0, + 1407.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2005.0, + 1060.0, + 2005.0, + 1060.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1122.0, + 296.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1117.0, + 420.0, + 1117.0, + 420.0, + 1152.0, + 294.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1117.0, + 1122.0, + 1117.0, + 1122.0, + 1152.0, + 442.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1117.0, + 1337.0, + 1117.0, + 1337.0, + 1152.0, + 1158.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 883.0, + 1403.0, + 883.0, + 1403.0, + 919.0, + 296.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 913.0, + 688.0, + 913.0, + 688.0, + 952.0, + 295.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 913.0, + 951.0, + 913.0, + 951.0, + 952.0, + 727.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1268.0, + 363.0, + 1268.0, + 363.0, + 1304.0, + 296.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 442.0, + 1268.0, + 1404.0, + 1268.0, + 1404.0, + 1304.0, + 442.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1299.0, + 531.0, + 1299.0, + 531.0, + 1335.0, + 296.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 1299.0, + 1386.0, + 1299.0, + 1386.0, + 1335.0, + 554.0, + 1335.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 287, + 1405, + 287, + 1405, + 592, + 296, + 592 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 855, + 1405, + 855, + 1405, + 1077, + 297, + 1077 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1438, + 802, + 1438, + 802, + 1866, + 299, + 1866 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1202, + 1402, + 1202, + 1402, + 1296, + 299, + 1296 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 604, + 1404, + 604, + 1404, + 729, + 297, + 729 + ], + "score": 0.976 + }, + { + "category_id": 8, + "poly": [ + 611, + 1312, + 1084, + 1312, + 1084, + 1407, + 611, + 1407 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 636, + 1092, + 1058, + 1092, + 1058, + 1188, + 636, + 1188 + ], + "score": 0.959 + }, + { + "category_id": 8, + "poly": [ + 454, + 744, + 1246, + 744, + 1246, + 823, + 454, + 823 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 298, + 1883, + 804, + 1883, + 804, + 1974, + 298, + 1974 + ], + "score": 0.931 + }, + { + "category_id": 1, + "poly": [ + 838, + 1644, + 1402, + 1644, + 1402, + 1948, + 838, + 1948 + ], + "score": 0.916 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1125, + 1400, + 1125, + 1400, + 1156, + 1366, + 1156 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 827, + 1454, + 1073, + 1454, + 1073, + 1484, + 827, + 1484 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1345, + 1400, + 1345, + 1400, + 1374, + 1366, + 1374 + ], + "score": 0.885 + }, + { + "category_id": 1, + "poly": [ + 868, + 1523, + 1400, + 1523, + 1400, + 1615, + 868, + 1615 + ], + "score": 0.874 + }, + { + "category_id": 9, + "poly": [ + 1366, + 760, + 1400, + 760, + 1400, + 789, + 1366, + 789 + ], + "score": 0.867 + }, + { + "category_id": 0, + "poly": [ + 826, + 1493, + 932, + 1493, + 932, + 1520, + 826, + 1520 + ], + "score": 0.866 + }, + { + "category_id": 0, + "poly": [ + 826, + 1614, + 922, + 1614, + 922, + 1640, + 826, + 1640 + ], + "score": 0.839 + }, + { + "category_id": 2, + "poly": [ + 299, + 1974, + 1403, + 1974, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.794 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.766 + }, + { + "category_id": 0, + "poly": [ + 297, + 229, + 957, + 229, + 957, + 262, + 297, + 262 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 957, + 229, + 957, + 262, + 297, + 262 + ], + "score": 0.159 + }, + { + "category_id": 1, + "poly": [ + 299, + 1974, + 1403, + 1974, + 1403, + 2035, + 299, + 2035 + ], + "score": 0.105 + }, + { + "category_id": 14, + "poly": [ + 636, + 1092, + 1060, + 1092, + 1060, + 1188, + 636, + 1188 + ], + "score": 0.95, + "latex": "\\lambda \\propto \\sum _ { i = 0 } ^ { h - 2 } \\left[ \\left( \\beta _ { c o n v } ^ { ( i , h - 1 ) } \\right) ^ { 2 } \\prod _ { t = 0 } ^ { i - 1 } \\left( \\beta _ { s k i p } ^ { ( t , i ) } \\right) ^ { 2 } \\right]" + }, + { + "category_id": 13, + "poly": [ + 868, + 502, + 1137, + 502, + 1137, + 563, + 868, + 563 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\beta _ { o } ^ { i , j } = \\frac { \\exp ( \\alpha _ { o } ^ { ( i , j ) } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\alpha _ { o ^ { \\prime } } ^ { ( i , j ) } ) } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 610, + 1309, + 1088, + 1309, + 1088, + 1408, + 610, + 1408 + ], + "score": 0.94, + "latex": "\\lambda \\propto \\sum _ { i = 0 } ^ { h - 2 } \\left[ \\left( \\beta _ { c o n v } ^ { ( i , h - 1 ) } \\right) ^ { 2 } \\prod _ { t = 0 } ^ { i - 1 } \\left( \\beta _ { s k i p } ^ { ( t , i ) } + \\beta \\right) ^ { 2 } \\right]" + }, + { + "category_id": 13, + "poly": [ + 442, + 1652, + 504, + 1652, + 504, + 1685, + 442, + 1685 + ], + "score": 0.93, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 704, + 918, + 765, + 918, + 765, + 952, + 704, + 952 + ], + "score": 0.93, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 354, + 1234, + 421, + 1234, + 421, + 1265, + 354, + 1265 + ], + "score": 0.92, + "latex": "\\beta _ { c o n v }" + }, + { + "category_id": 13, + "poly": [ + 971, + 1764, + 1084, + 1764, + 1084, + 1797, + 971, + 1797 + ], + "score": 0.92, + "latex": "e \\in [ 1 , E ]" + }, + { + "category_id": 13, + "poly": [ + 1095, + 411, + 1157, + 411, + 1157, + 444, + 1095, + 444 + ], + "score": 0.92, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 14, + "poly": [ + 456, + 744, + 1245, + 744, + 1245, + 826, + 456, + 826 + ], + "score": 0.92, + "latex": "\\bar { o } ^ { ( i , j ) } ( x ) = \\beta x + \\sum _ { o \\in \\mathcal { O } } \\beta _ { o } ^ { ( i , j ) } o ( x ) = \\left( \\beta + \\beta _ { s k i p } ^ { ( i , j ) } \\right) x + \\sum _ { o \\neq s k i p } \\beta _ { o } ^ { ( i , j ) } o ( x )" + }, + { + "category_id": 13, + "poly": [ + 841, + 350, + 903, + 350, + 903, + 383, + 841, + 383 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 692, + 980, + 736, + 980, + 736, + 1017, + 692, + 1017 + ], + "score": 0.91, + "latex": "\\beta _ { i , j } ^ { o }" + }, + { + "category_id": 13, + "poly": [ + 336, + 1561, + 397, + 1561, + 397, + 1594, + 336, + 1594 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 377, + 1439, + 516, + 1439, + 516, + 1472, + 377, + 1472 + ], + "score": 0.91, + "latex": "\\beta \\ \\gg \\ \\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 642, + 1500, + 708, + 1500, + 708, + 1532, + 642, + 1532 + ], + "score": 0.91, + "latex": "\\beta _ { c o n v }" + }, + { + "category_id": 13, + "poly": [ + 1083, + 1582, + 1310, + 1582, + 1310, + 1615, + 1083, + 1615 + ], + "score": 0.91, + "latex": "\\beta _ { e } , e \\in \\{ 1 , 2 , . . . , E \\}" + }, + { + "category_id": 13, + "poly": [ + 540, + 1013, + 666, + 1013, + 666, + 1046, + 540, + 1046 + ], + "score": 0.91, + "latex": "( 1 - \\eta \\lambda / 4 )" + }, + { + "category_id": 13, + "poly": [ + 1175, + 1827, + 1380, + 1827, + 1380, + 1858, + 1175, + 1858 + ], + "score": 0.9, + "latex": "\\nabla _ { \\alpha } \\mathcal { L } _ { v a l } ( w , \\alpha , \\beta _ { e } )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1470, + 359, + 1470, + 359, + 1502, + 297, + 1502 + ], + "score": 0.9, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 387, + 508, + 445, + 508, + 445, + 541, + 387, + 541 + ], + "score": 0.89, + "latex": "e ^ { ( i , j ) }" + }, + { + "category_id": 13, + "poly": [ + 1332, + 1235, + 1393, + 1235, + 1393, + 1267, + 1332, + 1267 + ], + "score": 0.87, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 1067, + 1013, + 1132, + 1013, + 1132, + 1043, + 1067, + 1043 + ], + "score": 0.87, + "latex": "1 - \\delta" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1795, + 1374, + 1795, + 1374, + 1827, + 1143, + 1827 + ], + "score": 0.86, + "latex": "\\nabla _ { w } \\mathcal { L } _ { t r a i n } ( w , \\alpha , \\beta _ { e } )" + }, + { + "category_id": 13, + "poly": [ + 482, + 559, + 503, + 559, + 503, + 590, + 482, + 590 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 485, + 638, + 506, + 638, + 506, + 668, + 485, + 668 + ], + "score": 0.85, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1339, + 1204, + 1400, + 1204, + 1400, + 1235, + 1339, + 1235 + ], + "score": 0.84, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 544, + 1561, + 564, + 1561, + 564, + 1592, + 544, + 1592 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 372, + 1204, + 391, + 1204, + 391, + 1230, + 372, + 1230 + ], + "score": 0.82, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 1161, + 1552, + 1186, + 1552, + 1186, + 1579, + 1161, + 1579 + ], + "score": 0.82, + "latex": "E" + }, + { + "category_id": 13, + "poly": [ + 367, + 1713, + 390, + 1713, + 390, + 1744, + 367, + 1744 + ], + "score": 0.81, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 719, + 1045, + 738, + 1045, + 738, + 1071, + 719, + 1071 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1148, + 1204, + 1167, + 1204, + 1167, + 1230, + 1148, + 1230 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 619, + 1440, + 640, + 1440, + 640, + 1466, + 619, + 1466 + ], + "score": 0.8, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1018, + 1237, + 1018, + 1237, + 1046, + 1217, + 1046 + ], + "score": 0.8, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 1014, + 1265, + 1034, + 1265, + 1034, + 1296, + 1014, + 1296 + ], + "score": 0.79, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 646, + 1047, + 661, + 1047, + 661, + 1071, + 646, + 1071 + ], + "score": 0.79, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 1066, + 1526, + 1090, + 1526, + 1090, + 1549, + 1066, + 1549 + ], + "score": 0.77, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 1080, + 1800, + 1104, + 1800, + 1104, + 1822, + 1080, + 1822 + ], + "score": 0.76, + "latex": "w" + }, + { + "category_id": 13, + "poly": [ + 1380, + 1891, + 1402, + 1891, + 1402, + 1913, + 1380, + 1913 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1372, + 1527, + 1393, + 1527, + 1393, + 1548, + 1372, + 1548 + ], + "score": 0.75, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1247, + 1650, + 1268, + 1650, + 1268, + 1670, + 1247, + 1670 + ], + "score": 0.75, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1830, + 1137, + 1830, + 1137, + 1852, + 1115, + 1852 + ], + "score": 0.74, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 393, + 1500, + 431, + 1500, + 431, + 1528, + 393, + 1528 + ], + "score": 0.4, + "latex": "\\mathbf { W }" + }, + { + "category_id": 13, + "poly": [ + 980, + 979, + 1017, + 979, + 1017, + 1007, + 980, + 1007 + ], + "score": 0.33, + "latex": "\\mathbf { W }" + }, + { + "category_id": 13, + "poly": [ + 913, + 1234, + 951, + 1234, + 951, + 1261, + 913, + 1261 + ], + "score": 0.31, + "latex": "\\mathbf { W }" + }, + { + "category_id": 13, + "poly": [ + 1333, + 888, + 1371, + 888, + 1371, + 917, + 1333, + 917 + ], + "score": 0.28, + "latex": "\\mathbf { W }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1451.0, + 1078.0, + 1451.0, + 1078.0, + 1493.0, + 822.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1489.0, + 939.0, + 1489.0, + 939.0, + 1528.0, + 821.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1610.0, + 926.0, + 1610.0, + 926.0, + 1646.0, + 823.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2004.0, + 297.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 1019.0, + 1999.0, + 1019.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 960.0, + 228.0, + 960.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 289.0, + 1403.0, + 289.0, + 1403.0, + 321.0, + 296.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 318.0, + 1406.0, + 318.0, + 1406.0, + 352.0, + 293.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 347.0, + 840.0, + 347.0, + 840.0, + 386.0, + 293.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 347.0, + 1407.0, + 347.0, + 1407.0, + 386.0, + 904.0, + 386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 379.0, + 1405.0, + 379.0, + 1405.0, + 414.0, + 294.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 405.0, + 1094.0, + 405.0, + 1094.0, + 447.0, + 291.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 405.0, + 1406.0, + 405.0, + 1406.0, + 447.0, + 1158.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 441.0, + 1405.0, + 441.0, + 1405.0, + 476.0, + 294.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 467.0, + 1406.0, + 467.0, + 1406.0, + 509.0, + 291.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 507.0, + 386.0, + 507.0, + 386.0, + 555.0, + 293.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 507.0, + 867.0, + 507.0, + 867.0, + 555.0, + 446.0, + 555.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 513.0, + 1404.0, + 513.0, + 1404.0, + 547.0, + 1138.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 556.0, + 481.0, + 556.0, + 481.0, + 595.0, + 294.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 504.0, + 556.0, + 1193.0, + 556.0, + 1193.0, + 595.0, + 504.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 851.0, + 1406.0, + 851.0, + 1406.0, + 897.0, + 294.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 888.0, + 1332.0, + 888.0, + 1332.0, + 923.0, + 294.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 888.0, + 1405.0, + 888.0, + 1405.0, + 923.0, + 1372.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 917.0, + 703.0, + 917.0, + 703.0, + 956.0, + 294.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 917.0, + 1406.0, + 917.0, + 1406.0, + 956.0, + 766.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 945.0, + 1406.0, + 945.0, + 1406.0, + 987.0, + 292.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 972.0, + 691.0, + 972.0, + 691.0, + 1024.0, + 290.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 972.0, + 979.0, + 972.0, + 979.0, + 1024.0, + 737.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 972.0, + 1410.0, + 972.0, + 1410.0, + 1024.0, + 1018.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1011.0, + 539.0, + 1011.0, + 539.0, + 1052.0, + 292.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1011.0, + 1066.0, + 1011.0, + 1066.0, + 1052.0, + 667.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1011.0, + 1216.0, + 1011.0, + 1216.0, + 1052.0, + 1133.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1011.0, + 1405.0, + 1011.0, + 1405.0, + 1052.0, + 1238.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 645.0, + 1044.0, + 645.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1044.0, + 718.0, + 1044.0, + 718.0, + 1079.0, + 662.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1044.0, + 904.0, + 1044.0, + 904.0, + 1079.0, + 739.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1435.0, + 376.0, + 1435.0, + 376.0, + 1475.0, + 296.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1435.0, + 618.0, + 1435.0, + 618.0, + 1475.0, + 517.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1435.0, + 806.0, + 1435.0, + 806.0, + 1475.0, + 641.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1466.0, + 805.0, + 1466.0, + 805.0, + 1506.0, + 360.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1496.0, + 392.0, + 1496.0, + 392.0, + 1534.0, + 293.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 1496.0, + 641.0, + 1496.0, + 641.0, + 1534.0, + 432.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1496.0, + 806.0, + 1496.0, + 806.0, + 1534.0, + 709.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1528.0, + 804.0, + 1528.0, + 804.0, + 1564.0, + 295.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1558.0, + 335.0, + 1558.0, + 335.0, + 1597.0, + 293.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1558.0, + 543.0, + 1558.0, + 543.0, + 1597.0, + 398.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 1558.0, + 806.0, + 1558.0, + 806.0, + 1597.0, + 565.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1589.0, + 805.0, + 1589.0, + 805.0, + 1625.0, + 295.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1621.0, + 804.0, + 1621.0, + 804.0, + 1653.0, + 295.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 441.0, + 1650.0, + 441.0, + 1687.0, + 293.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1650.0, + 805.0, + 1650.0, + 805.0, + 1687.0, + 505.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1681.0, + 803.0, + 1681.0, + 803.0, + 1713.0, + 295.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 366.0, + 1712.0, + 366.0, + 1747.0, + 295.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1712.0, + 804.0, + 1712.0, + 804.0, + 1747.0, + 391.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 804.0, + 1742.0, + 804.0, + 1776.0, + 294.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1775.0, + 803.0, + 1775.0, + 803.0, + 1807.0, + 296.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1802.0, + 805.0, + 1802.0, + 805.0, + 1836.0, + 293.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1834.0, + 571.0, + 1834.0, + 571.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1193.0, + 371.0, + 1193.0, + 371.0, + 1244.0, + 292.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1193.0, + 1147.0, + 1193.0, + 1147.0, + 1244.0, + 392.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1168.0, + 1193.0, + 1338.0, + 1193.0, + 1338.0, + 1244.0, + 1168.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1193.0, + 1406.0, + 1193.0, + 1406.0, + 1244.0, + 1401.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1227.0, + 353.0, + 1227.0, + 353.0, + 1273.0, + 292.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1227.0, + 912.0, + 1227.0, + 912.0, + 1273.0, + 422.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1227.0, + 1331.0, + 1227.0, + 1331.0, + 1273.0, + 952.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1227.0, + 1407.0, + 1227.0, + 1407.0, + 1273.0, + 1394.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1262.0, + 1013.0, + 1262.0, + 1013.0, + 1300.0, + 294.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1262.0, + 1401.0, + 1262.0, + 1401.0, + 1300.0, + 1035.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 602.0, + 1404.0, + 602.0, + 1404.0, + 642.0, + 292.0, + 642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 637.0, + 484.0, + 637.0, + 484.0, + 668.0, + 292.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 637.0, + 1404.0, + 637.0, + 1404.0, + 668.0, + 507.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 665.0, + 1404.0, + 665.0, + 1404.0, + 702.0, + 292.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 695.0, + 803.0, + 695.0, + 803.0, + 733.0, + 292.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1881.0, + 805.0, + 1881.0, + 805.0, + 1915.0, + 296.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 805.0, + 1910.0, + 805.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 804.0, + 1942.0, + 804.0, + 1976.0, + 295.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1643.0, + 1246.0, + 1643.0, + 1246.0, + 1675.0, + 871.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1643.0, + 1278.0, + 1643.0, + 1278.0, + 1675.0, + 1269.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1673.0, + 1404.0, + 1673.0, + 1404.0, + 1707.0, + 838.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1702.0, + 1403.0, + 1702.0, + 1403.0, + 1738.0, + 869.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1734.0, + 1242.0, + 1734.0, + 1242.0, + 1765.0, + 869.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1768.0, + 865.0, + 1768.0, + 865.0, + 1793.0, + 840.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1764.0, + 970.0, + 1764.0, + 970.0, + 1798.0, + 867.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1764.0, + 1128.0, + 1764.0, + 1128.0, + 1798.0, + 1085.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1796.0, + 866.0, + 1796.0, + 866.0, + 1826.0, + 837.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1792.0, + 1079.0, + 1792.0, + 1079.0, + 1830.0, + 895.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1105.0, + 1792.0, + 1142.0, + 1792.0, + 1142.0, + 1830.0, + 1105.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1826.0, + 866.0, + 1826.0, + 866.0, + 1856.0, + 837.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 1822.0, + 1114.0, + 1822.0, + 1114.0, + 1862.0, + 896.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 1822.0, + 1174.0, + 1822.0, + 1174.0, + 1862.0, + 1138.0, + 1862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1857.0, + 867.0, + 1857.0, + 867.0, + 1885.0, + 837.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1856.0, + 965.0, + 1856.0, + 965.0, + 1885.0, + 871.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1883.0, + 1379.0, + 1883.0, + 1379.0, + 1916.0, + 838.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1914.0, + 1250.0, + 1914.0, + 1250.0, + 1951.0, + 871.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1517.0, + 1065.0, + 1517.0, + 1065.0, + 1557.0, + 870.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1517.0, + 1371.0, + 1517.0, + 1371.0, + 1557.0, + 1091.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1517.0, + 1405.0, + 1517.0, + 1405.0, + 1557.0, + 1394.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 1551.0, + 1160.0, + 1551.0, + 1160.0, + 1583.0, + 871.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 1551.0, + 1198.0, + 1551.0, + 1198.0, + 1583.0, + 1187.0, + 1583.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1581.0, + 1082.0, + 1581.0, + 1082.0, + 1617.0, + 869.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 1581.0, + 1320.0, + 1581.0, + 1320.0, + 1617.0, + 1311.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 960.0, + 228.0, + 960.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1973.0, + 1403.0, + 1973.0, + 1403.0, + 2004.0, + 297.0, + 2004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1999.0, + 1019.0, + 1999.0, + 1019.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 791, + 1404, + 791, + 1404, + 1189, + 298, + 1189 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 292, + 1403, + 292, + 1403, + 598, + 298, + 598 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2033, + 298, + 2033 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1620, + 1404, + 1620, + 1404, + 1834, + 297, + 1834 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1298, + 862, + 1298, + 862, + 1482, + 298, + 1482 + ], + "score": 0.97 + }, + { + "category_id": 5, + "poly": [ + 892, + 1375, + 1389, + 1375, + 1389, + 1467, + 892, + 1467 + ], + "score": 0.963, + "html": "
DatasetDARTSR-DARTS (L2)Ours
C10 (SO)2.91±0.252.95±0.212.63±0.07
C100 (S0)20.58±0.4418.01±0.2617.51±0.25
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1483, + 1404, + 1483, + 1404, + 1604, + 299, + 1604 + ], + "score": 0.963 + }, + { + "category_id": 6, + "poly": [ + 883, + 1306, + 1402, + 1306, + 1402, + 1368, + 883, + 1368 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 298, + 729, + 876, + 729, + 876, + 761, + 298, + 761 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 301, + 1238, + 622, + 1238, + 622, + 1269, + 301, + 1269 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 301, + 654, + 557, + 654, + 557, + 688, + 301, + 688 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 302, + 229, + 748, + 229, + 748, + 261, + 302, + 261 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.687 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2111, + 841, + 2111 + ], + "score": 0.109 + }, + { + "category_id": 13, + "poly": [ + 1191, + 1574, + 1250, + 1574, + 1250, + 1605, + 1191, + 1605 + ], + "score": 0.91, + "latex": "C O _ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 601, + 1360, + 692, + 1360, + 692, + 1390, + 601, + 1390 + ], + "score": 0.9, + "latex": "9 7 . 4 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 852, + 1743, + 928, + 1743, + 928, + 1773, + 852, + 1773 + ], + "score": 0.88, + "latex": "7 7 . 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1681, + 1154, + 1681, + 1154, + 1711, + 1078, + 1711 + ], + "score": 0.88, + "latex": "7 6 . 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 1298, + 1005, + 1402, + 1005, + 1402, + 1036, + 1298, + 1036 + ], + "score": 0.88, + "latex": "\\beta _ { 1 } = 0 . 5" + }, + { + "category_id": 13, + "poly": [ + 297, + 1742, + 374, + 1742, + 374, + 1773, + 297, + 1773 + ], + "score": 0.87, + "latex": "6 6 . { \\bar { 4 \\% } }" + }, + { + "category_id": 13, + "poly": [ + 346, + 1036, + 484, + 1036, + 484, + 1067, + 346, + 1067 + ], + "score": 0.87, + "latex": "\\beta _ { 2 } = 0 . 9 9 9" + }, + { + "category_id": 13, + "poly": [ + 422, + 1483, + 511, + 1483, + 511, + 1514, + 422, + 1514 + ], + "score": 0.81, + "latex": "( 9 7 . 5 \\% )" + }, + { + "category_id": 13, + "poly": [ + 994, + 1514, + 1095, + 1514, + 1095, + 1544, + 994, + 1544 + ], + "score": 0.55, + "latex": "( 3 \\times \\mathrm { l e s s } )" + }, + { + "category_id": 13, + "poly": [ + 997, + 1515, + 1041, + 1515, + 1041, + 1544, + 997, + 1544 + ], + "score": 0.32, + "latex": "( 3 \\times" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1306.0, + 1404.0, + 1306.0, + 1404.0, + 1340.0, + 885.0, + 1340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1337.0, + 1404.0, + 1337.0, + 1404.0, + 1371.0, + 884.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 729.0, + 879.0, + 729.0, + 879.0, + 765.0, + 294.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1235.0, + 627.0, + 1235.0, + 627.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 651.0, + 561.0, + 651.0, + 561.0, + 693.0, + 293.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 754.0, + 228.0, + 754.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 790.0, + 1405.0, + 790.0, + 1405.0, + 828.0, + 295.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 821.0, + 1405.0, + 821.0, + 1405.0, + 856.0, + 293.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 854.0, + 1405.0, + 854.0, + 1405.0, + 886.0, + 296.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 919.0, + 294.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 912.0, + 1405.0, + 912.0, + 1405.0, + 948.0, + 294.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 942.0, + 1406.0, + 942.0, + 1406.0, + 980.0, + 293.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 973.0, + 1409.0, + 973.0, + 1409.0, + 1011.0, + 292.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1004.0, + 1297.0, + 1004.0, + 1297.0, + 1041.0, + 294.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1033.0, + 345.0, + 1033.0, + 345.0, + 1073.0, + 293.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1033.0, + 1408.0, + 1033.0, + 1408.0, + 1073.0, + 485.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1065.0, + 1406.0, + 1065.0, + 1406.0, + 1102.0, + 294.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1095.0, + 1405.0, + 1095.0, + 1405.0, + 1131.0, + 294.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1126.0, + 1404.0, + 1126.0, + 1404.0, + 1160.0, + 293.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1155.0, + 1403.0, + 1155.0, + 1403.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1403.0, + 291.0, + 1403.0, + 327.0, + 294.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 323.0, + 1403.0, + 323.0, + 1403.0, + 355.0, + 296.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 354.0, + 1407.0, + 354.0, + 1407.0, + 389.0, + 294.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 382.0, + 1407.0, + 382.0, + 1407.0, + 420.0, + 292.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 413.0, + 1403.0, + 413.0, + 1403.0, + 449.0, + 293.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 444.0, + 1407.0, + 444.0, + 1407.0, + 481.0, + 293.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 474.0, + 1407.0, + 474.0, + 1407.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 507.0, + 1405.0, + 507.0, + 1405.0, + 539.0, + 296.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 535.0, + 1407.0, + 535.0, + 1407.0, + 571.0, + 294.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 567.0, + 1341.0, + 567.0, + 1341.0, + 601.0, + 293.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 1404.0, + 1848.0, + 1404.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1403.0, + 1881.0, + 1403.0, + 1916.0, + 293.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1949.0, + 291.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2003.0, + 975.0, + 2003.0, + 975.0, + 2035.0, + 296.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1618.0, + 1405.0, + 1618.0, + 1405.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1652.0, + 1402.0, + 1652.0, + 1402.0, + 1683.0, + 296.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 1077.0, + 1678.0, + 1077.0, + 1718.0, + 294.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1678.0, + 1404.0, + 1678.0, + 1404.0, + 1718.0, + 1155.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1404.0, + 1711.0, + 1404.0, + 1748.0, + 294.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1738.0, + 296.0, + 1738.0, + 296.0, + 1779.0, + 292.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 1738.0, + 851.0, + 1738.0, + 851.0, + 1779.0, + 375.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 1738.0, + 1405.0, + 1738.0, + 1405.0, + 1779.0, + 929.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1774.0, + 1405.0, + 1774.0, + 1405.0, + 1808.0, + 294.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1804.0, + 532.0, + 1804.0, + 532.0, + 1834.0, + 293.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1294.0, + 865.0, + 1294.0, + 865.0, + 1334.0, + 296.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1330.0, + 866.0, + 1330.0, + 866.0, + 1363.0, + 295.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1360.0, + 600.0, + 1360.0, + 600.0, + 1393.0, + 293.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 693.0, + 1360.0, + 864.0, + 1360.0, + 864.0, + 1393.0, + 693.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1390.0, + 865.0, + 1390.0, + 865.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1422.0, + 864.0, + 1422.0, + 864.0, + 1452.0, + 295.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1453.0, + 864.0, + 1453.0, + 864.0, + 1485.0, + 295.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 421.0, + 1481.0, + 421.0, + 1516.0, + 294.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 1481.0, + 1404.0, + 1481.0, + 1404.0, + 1516.0, + 512.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1512.0, + 993.0, + 1512.0, + 993.0, + 1547.0, + 294.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 1512.0, + 1404.0, + 1512.0, + 1404.0, + 1547.0, + 1096.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1542.0, + 1404.0, + 1542.0, + 1404.0, + 1578.0, + 294.0, + 1578.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1572.0, + 1190.0, + 1572.0, + 1190.0, + 1608.0, + 293.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1572.0, + 1374.0, + 1572.0, + 1374.0, + 1608.0, + 1251.0, + 1608.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 305, + 1070, + 1394, + 1070, + 1394, + 1359, + 305, + 1359 + ], + "score": 0.981, + "html": "
MethodCost (hours)CIFAR-10CIFAR-100ImageNet16-120
validtestvalidtestvalidtest
DARTS1st (2019b)3.239.77±0.00 54.30±0.0015.03±0.00 15.61±0.0016.43±0.00 16.32±0.00
DARTS2nd (2019b)10.239.77±0.0054.30±0.0015.03±0.0015.61±0.0016.43±0.00 16.32±0.00
GDAS (2019b)8.789.89±0.08 93.61±0.0971.34±0.04：70.70±0.30 41.59±1.33 41.71±0.98
SETN (2019a)9.584.04±0.2887.64±0.0058.86±0.0659.05±0.2433.06±0.0232.52±0.21
DARTS-3.291.03±0.44 93.80±0.4071.36±1.51 71.53±1.5144.87±1.46 45.12±0.82
DARTS- (best)3.291.5594.3673.4973.5146.3746.34
optimaln/a91.6194.3773.4973.5146.7747.31
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1758, + 940, + 1758, + 940, + 2033, + 298, + 2033 + ], + "score": 0.981 + }, + { + "category_id": 6, + "poly": [ + 296, + 248, + 1403, + 248, + 1403, + 402, + 296, + 402 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 1003, + 1920, + 1356, + 1920, + 1356, + 2040, + 1003, + 2040 + ], + "score": 0.964, + "html": "
MethodSetting商银 Acc(%)
P-DARTS w/o priors 96.48±0.55
P-DARTS- w/o priors 97.28±0.04
" + }, + { + "category_id": 1, + "poly": [ + 300, + 1681, + 1399, + 1681, + 1399, + 1743, + 300, + 1743 + ], + "score": 0.947 + }, + { + "category_id": 6, + "poly": [ + 291, + 1008, + 1398, + 1008, + 1398, + 1069, + 291, + 1069 + ], + "score": 0.931 + }, + { + "category_id": 7, + "poly": [ + 853, + 937, + 1121, + 937, + 1121, + 983, + 853, + 983 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 298, + 1399, + 1404, + 1399, + 1404, + 1584, + 298, + 1584 + ], + "score": 0.884 + }, + { + "category_id": 0, + "poly": [ + 299, + 1623, + 1008, + 1623, + 1008, + 1656, + 299, + 1656 + ], + "score": 0.881 + }, + { + "category_id": 6, + "poly": [ + 964, + 1766, + 1401, + 1766, + 1401, + 1918, + 964, + 1918 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.794 + }, + { + "category_id": 5, + "poly": [ + 305, + 404, + 1372, + 404, + 1372, + 948, + 305, + 948 + ], + "score": 0.713, + "html": "
ModelsFLOPs Params Top-1 Top-5Cost (GPU days)
ModelsParams (M)FLOPs Acc (M)(%)Cost(M)(M)(%)(%)
NASNet-A (2018)608tGPU Days AmoebaNet-A (2019) MnasNet-92 (2019)555 3885.1 3.974.592.0 74.79 92.13150 3791
3.3626†97.352000FBNet-C (2019)3755.574.992.39
ENAS (2018) DARTS (2019b)4.697.110.5FairNAS-A (2019b)3884.675.392.412
3.3528t97.00±0.14*0.4SCARLET-C(2019a)3656.776.993.410
SNAS (2019)2.8422t97.15±0.02*1.5FairDARTS-D (2020b) 4404.375.692.63
GDAS (2019b)3.4519t97.070.2PC-DARTS (2020)5975.375.892.73.8
P-DARTS (2019b)3.4532t97.50.3DARTS- (ours)4674.976.293.04.5
PC-DARTS (2020) DARTS- (best)3.6558t 56897.43 97.50.1 0.4NASNet-A (2018)5645.374.091.62000
3.5DARTS (2019b)5744.773.391.30.4
P-DARTS (2019b)3.3±0.21 540±34 97.19±0.140.3SNAS (2019)5224.372.790.81.5
R-DARTS (2020)197.05±0.211.6PC-DARTS (2020)5865.374.992.20.1
SDARTS-ADV (2020) 3.397.39±0.02 3.5±0.13 583±22 97.41±0.081.3 0.4FairDARTS-B (2020b) 5414.875.192.50.4
DARTS- (avg.)MobileNetV3 (2019)5.475.292.2~3000
† Based on the provided genotypes. ↑MoGA-A (2020a)2195.175.992.812
5 independent searches using their released code.MixNet-M (2019)3045.077.093.3~3000
EfficientNet B0 (2019) 3903605.376.393.2~3000
* Training the best searched model for several times NoisyDARTS-A4495.577.994.012
(whose average doesn't indicate the stability of the DARTS- (ours) method)4705.577.893.94.5
" + }, + { + "category_id": 7, + "poly": [ + 319, + 805, + 829, + 805, + 829, + 953, + 319, + 953 + ], + "score": 0.417 + }, + { + "category_id": 13, + "poly": [ + 732, + 1493, + 808, + 1493, + 808, + 1522, + 732, + 1522 + ], + "score": 0.85, + "latex": "3 2 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 907, + 1944, + 940, + 1944, + 940, + 1970, + 907, + 1970 + ], + "score": 0.75, + "latex": "M" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 250.0, + 1404.0, + 250.0, + 1404.0, + 283.0, + 296.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 280.0, + 1404.0, + 280.0, + 1404.0, + 313.0, + 296.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 312.0, + 1404.0, + 312.0, + 1404.0, + 345.0, + 295.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 340.0, + 1406.0, + 340.0, + 1406.0, + 377.0, + 295.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1310.0, + 369.0, + 1310.0, + 407.0, + 294.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1007.0, + 1401.0, + 1007.0, + 1401.0, + 1042.0, + 295.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1030.0, + 1274.0, + 1030.0, + 1274.0, + 1076.0, + 293.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 931.0, + 1114.0, + 931.0, + 1114.0, + 965.0, + 851.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 952.0, + 1125.0, + 952.0, + 1125.0, + 989.0, + 850.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1623.0, + 1010.0, + 1623.0, + 1010.0, + 1658.0, + 295.0, + 1658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1762.0, + 1404.0, + 1762.0, + 1404.0, + 1801.0, + 961.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1798.0, + 1403.0, + 1798.0, + 1403.0, + 1828.0, + 963.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1826.0, + 1404.0, + 1826.0, + 1404.0, + 1861.0, + 961.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1855.0, + 1405.0, + 1855.0, + 1405.0, + 1892.0, + 961.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1888.0, + 1205.0, + 1888.0, + 1205.0, + 1920.0, + 963.0, + 1920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 801.0, + 679.0, + 801.0, + 679.0, + 841.0, + 317.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 835.0, + 836.0, + 835.0, + 836.0, + 877.0, + 323.0, + 877.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 865.0, + 835.0, + 865.0, + 835.0, + 902.0, + 319.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 895.0, + 834.0, + 895.0, + 834.0, + 927.0, + 332.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 924.0, + 426.0, + 924.0, + 426.0, + 956.0, + 331.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 945.0, + 1756.0, + 945.0, + 1794.0, + 293.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1786.0, + 945.0, + 1786.0, + 945.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1820.0, + 943.0, + 1820.0, + 943.0, + 1854.0, + 295.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 945.0, + 1850.0, + 945.0, + 1883.0, + 294.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 945.0, + 1881.0, + 945.0, + 1914.0, + 294.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 945.0, + 1909.0, + 945.0, + 1949.0, + 293.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 906.0, + 1943.0, + 906.0, + 1975.0, + 294.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 946.0, + 1972.0, + 946.0, + 2005.0, + 293.0, + 2005.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2001.0, + 946.0, + 2001.0, + 946.0, + 2035.0, + 294.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1677.0, + 1405.0, + 1677.0, + 1405.0, + 1720.0, + 294.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1712.0, + 1147.0, + 1712.0, + 1147.0, + 1745.0, + 295.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1399.0, + 1405.0, + 1399.0, + 1405.0, + 1436.0, + 295.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1467.0, + 293.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1498.0, + 292.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1490.0, + 731.0, + 1490.0, + 731.0, + 1528.0, + 293.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1490.0, + 1404.0, + 1490.0, + 1404.0, + 1528.0, + 809.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1521.0, + 1405.0, + 1521.0, + 1405.0, + 1558.0, + 293.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1551.0, + 541.0, + 1551.0, + 541.0, + 1587.0, + 294.0, + 1587.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 341, + 379, + 1353, + 379, + 1353, + 721, + 341, + 721 + ], + "score": 0.984, + "html": "
Benchmark DARTS‡R-DARTS‡DARTSOursPC-DARTS+SDARTS+Ourst
DPL2ESADARSADV
C10S13.843.112.783.013.102.683.112.782.732.68
S24.853.483.313.263.352.633.022.752.652.63
S33.342.932.512.742.592.422.512.532.492.42
S47.203.583.563.714.842.863.022.932.872.86
C100S129.4625.9324.2528.3724.0322.4118.8717.0216.8816.92
S226.0522.3022.2423.2523.5221.6118.2317.5617.2416.14
S328.9022.3623.9923.7323.3721.1318.0517.7317.1215.86
S422.8522.1821.9421.2623.2021.5517.1617.1715.4617.48
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1650, + 982, + 1650, + 982, + 1957, + 299, + 1957 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 939, + 878, + 939, + 878, + 1215, + 298, + 1215 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1481, + 1403, + 1481, + 1403, + 1636, + 298, + 1636 + ], + "score": 0.978 + }, + { + "category_id": 6, + "poly": [ + 297, + 248, + 1404, + 248, + 1404, + 372, + 297, + 372 + ], + "score": 0.975 + }, + { + "category_id": 5, + "poly": [ + 1085, + 1780, + 1314, + 1780, + 1314, + 1999, + 1085, + 1999 + ], + "score": 0.964, + "html": "
βError (%)
12.65±0.04
0.7 0.42.76±0.16
0.13.04±0.19 3.11±0.16
04.58±1.3
" + }, + { + "category_id": 5, + "poly": [ + 900, + 1107, + 1406, + 1107, + 1406, + 1224, + 900, + 1224 + ], + "score": 0.963, + "html": "
MethodSettingAcc (%）Memory Cost
PC-DARTS K = 2 97.09±0.1419.9G 股银 3.75h
PC-DARTS- K = 2 97.35±0.0220.8G 3.41h
" + }, + { + "category_id": 1, + "poly": [ + 296, + 739, + 1404, + 739, + 1404, + 925, + 296, + 925 + ], + "score": 0.952 + }, + { + "category_id": 6, + "poly": [ + 899, + 946, + 1404, + 946, + 1404, + 1099, + 899, + 1099 + ], + "score": 0.936 + }, + { + "category_id": 6, + "poly": [ + 1002, + 1651, + 1403, + 1651, + 1403, + 1772, + 1002, + 1772 + ], + "score": 0.93 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 298, + 1229, + 1405, + 1229, + 1405, + 1385, + 298, + 1385 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 300, + 1424, + 580, + 1424, + 580, + 1455, + 300, + 1455 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.663 + }, + { + "category_id": 1, + "poly": [ + 293, + 2004, + 1400, + 2004, + 1400, + 2035, + 293, + 2035 + ], + "score": 0.647 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.341 + }, + { + "category_id": 13, + "poly": [ + 674, + 1835, + 714, + 1835, + 714, + 1864, + 674, + 1864 + ], + "score": 0.9, + "latex": "3 \\times" + }, + { + "category_id": 13, + "poly": [ + 299, + 1321, + 398, + 1321, + 398, + 1353, + 299, + 1353 + ], + "score": 0.9, + "latex": "( + 0 . 2 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 649, + 1894, + 738, + 1894, + 738, + 1924, + 649, + 1924 + ], + "score": 0.89, + "latex": "1 5 . 8 6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1092, + 1574, + 1181, + 1574, + 1181, + 1603, + 1092, + 1603 + ], + "score": 0.89, + "latex": "9 7 . 4 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 798, + 1574, + 957, + 1574, + 957, + 1604, + 798, + 1604 + ], + "score": 0.89, + "latex": "9 7 . 3 3 \\% \\pm 0 . 0 9" + }, + { + "category_id": 13, + "poly": [ + 887, + 893, + 984, + 893, + 984, + 924, + 887, + 924 + ], + "score": 0.88, + "latex": "( + \\mathbf { 0 . 8 \\% } )" + }, + { + "category_id": 13, + "poly": [ + 773, + 1865, + 806, + 1865, + 806, + 1895, + 773, + 1895 + ], + "score": 0.88, + "latex": "S _ { 3 }" + }, + { + "category_id": 13, + "poly": [ + 492, + 1604, + 582, + 1604, + 582, + 1635, + 492, + 1635 + ], + "score": 0.87, + "latex": "9 7 . 3 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 1174, + 1712, + 1205, + 1712, + 1205, + 1743, + 1174, + 1743 + ], + "score": 0.87, + "latex": "\\beta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 817, + 832, + 901, + 832, + 901, + 863, + 817, + 863 + ], + "score": 0.86, + "latex": "( 3 . 4 2 \\%" + }, + { + "category_id": 13, + "poly": [ + 945, + 1231, + 1004, + 1231, + 1004, + 1260, + 945, + 1260 + ], + "score": 0.85, + "latex": "K { = } 2" + }, + { + "category_id": 13, + "poly": [ + 662, + 1545, + 684, + 1545, + 684, + 1575, + 662, + 1575 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1322, + 1484, + 1342, + 1484, + 1342, + 1514, + 1322, + 1514 + ], + "score": 0.82, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 1007, + 2004, + 1027, + 2004, + 1027, + 2035, + 1007, + 2035 + ], + "score": 0.82, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 795, + 1123, + 869, + 1123, + 869, + 1153, + 795, + 1153 + ], + "score": 0.81, + "latex": "\\left( K { = } 4 \\right)" + }, + { + "category_id": 13, + "poly": [ + 548, + 1354, + 577, + 1354, + 577, + 1380, + 548, + 1380 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 571, + 1973, + 593, + 1973, + 593, + 2004, + 571, + 2004 + ], + "score": 0.8, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 417, + 1063, + 447, + 1063, + 447, + 1090, + 417, + 1090 + ], + "score": 0.78, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 528, + 1002, + 566, + 1002, + 566, + 1030, + 528, + 1030 + ], + "score": 0.43, + "latex": "\\mathrm { { X u } }" + }, + { + "category_id": 14, + "poly": [ + 900, + 1107, + 1414, + 1107, + 1414, + 1220, + 900, + 1220 + ], + "score": 0.34, + "latex": "\\begin{array} { r } { \\overline { { \\mathbf { M e t h o d } \\mathbf { \\Gamma } \\mathbf { S e t t i n g } \\mathbf { A c c } \\left( \\% \\right) \\mathbf { M e m o r y } \\mathbf { C o s t } } } } \\\\ { \\overline { { \\mathrm { P C - D A R T S } \\mathbf { \\Gamma } K = 2 \\ 9 7 . 0 9 \\pm 0 . 1 4 \\ 1 9 . 9 \\mathrm { G } \\ 3 . 7 5 \\mathrm { h } } } } \\\\ { \\overline { { \\mathrm { P C - D A R T S - } K } } = 2 \\ 9 7 . 3 5 \\pm 0 . 0 2 \\ 2 0 . 8 \\mathrm { G } \\ 3 . 4 1 \\mathrm { h } } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 246.0, + 1404.0, + 246.0, + 1404.0, + 283.0, + 292.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 277.0, + 1405.0, + 277.0, + 1405.0, + 314.0, + 294.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 308.0, + 1404.0, + 308.0, + 1404.0, + 344.0, + 294.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 340.0, + 1384.0, + 340.0, + 1384.0, + 373.0, + 296.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 946.0, + 1405.0, + 946.0, + 1405.0, + 980.0, + 900.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 978.0, + 1405.0, + 978.0, + 1405.0, + 1010.0, + 900.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1007.0, + 1405.0, + 1007.0, + 1405.0, + 1038.0, + 899.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 897.0, + 1038.0, + 1406.0, + 1038.0, + 1406.0, + 1070.0, + 897.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1071.0, + 1332.0, + 1071.0, + 1332.0, + 1100.0, + 899.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1647.0, + 1403.0, + 1647.0, + 1403.0, + 1684.0, + 1002.0, + 1684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1679.0, + 1403.0, + 1679.0, + 1403.0, + 1712.0, + 1002.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1002.0, + 1711.0, + 1173.0, + 1711.0, + 1173.0, + 1743.0, + 1002.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1711.0, + 1403.0, + 1711.0, + 1403.0, + 1743.0, + 1206.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 1740.0, + 1223.0, + 1740.0, + 1223.0, + 1774.0, + 1000.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1422.0, + 585.0, + 1422.0, + 585.0, + 1458.0, + 296.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1650.0, + 983.0, + 1650.0, + 983.0, + 1683.0, + 296.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1681.0, + 983.0, + 1681.0, + 983.0, + 1714.0, + 296.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1713.0, + 983.0, + 1713.0, + 983.0, + 1744.0, + 294.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1741.0, + 984.0, + 1741.0, + 984.0, + 1774.0, + 295.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1772.0, + 983.0, + 1772.0, + 983.0, + 1806.0, + 295.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1802.0, + 984.0, + 1802.0, + 984.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1834.0, + 673.0, + 1834.0, + 673.0, + 1867.0, + 295.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 1834.0, + 982.0, + 1834.0, + 982.0, + 1867.0, + 715.0, + 1867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1864.0, + 772.0, + 1864.0, + 772.0, + 1896.0, + 293.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1864.0, + 983.0, + 1864.0, + 983.0, + 1896.0, + 807.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1894.0, + 648.0, + 1894.0, + 648.0, + 1927.0, + 295.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1894.0, + 985.0, + 1894.0, + 985.0, + 1927.0, + 739.0, + 1927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1923.0, + 781.0, + 1923.0, + 781.0, + 1959.0, + 294.0, + 1959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 939.0, + 879.0, + 939.0, + 879.0, + 972.0, + 296.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 970.0, + 880.0, + 970.0, + 880.0, + 1004.0, + 296.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 999.0, + 527.0, + 999.0, + 527.0, + 1035.0, + 293.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 999.0, + 880.0, + 999.0, + 880.0, + 1035.0, + 567.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1030.0, + 880.0, + 1030.0, + 880.0, + 1065.0, + 294.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1060.0, + 416.0, + 1060.0, + 416.0, + 1095.0, + 291.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 1060.0, + 883.0, + 1060.0, + 883.0, + 1095.0, + 448.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1091.0, + 882.0, + 1091.0, + 882.0, + 1126.0, + 292.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1122.0, + 794.0, + 1122.0, + 794.0, + 1156.0, + 295.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1122.0, + 880.0, + 1122.0, + 880.0, + 1156.0, + 870.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1151.0, + 880.0, + 1151.0, + 880.0, + 1187.0, + 295.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 882.0, + 1179.0, + 882.0, + 1220.0, + 294.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1481.0, + 1321.0, + 1481.0, + 1321.0, + 1519.0, + 294.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1343.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1519.0, + 1343.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1512.0, + 1405.0, + 1512.0, + 1405.0, + 1548.0, + 293.0, + 1548.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1541.0, + 661.0, + 1541.0, + 661.0, + 1582.0, + 292.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1541.0, + 1404.0, + 1541.0, + 1404.0, + 1582.0, + 685.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1566.0, + 797.0, + 1566.0, + 797.0, + 1613.0, + 292.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1566.0, + 1091.0, + 1566.0, + 1091.0, + 1613.0, + 958.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1566.0, + 1408.0, + 1566.0, + 1408.0, + 1613.0, + 1182.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1602.0, + 491.0, + 1602.0, + 491.0, + 1643.0, + 293.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1602.0, + 1167.0, + 1602.0, + 1167.0, + 1643.0, + 583.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 775.0, + 294.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 769.0, + 1404.0, + 769.0, + 1404.0, + 808.0, + 293.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 800.0, + 1404.0, + 800.0, + 1404.0, + 837.0, + 294.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 832.0, + 816.0, + 832.0, + 816.0, + 868.0, + 295.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 832.0, + 1404.0, + 832.0, + 1404.0, + 868.0, + 902.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 860.0, + 1407.0, + 860.0, + 1407.0, + 898.0, + 291.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 892.0, + 886.0, + 892.0, + 886.0, + 928.0, + 294.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 892.0, + 1338.0, + 892.0, + 1338.0, + 928.0, + 985.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1231.0, + 944.0, + 1231.0, + 944.0, + 1264.0, + 297.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1231.0, + 1403.0, + 1231.0, + 1403.0, + 1264.0, + 1005.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1260.0, + 1406.0, + 1260.0, + 1406.0, + 1298.0, + 292.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1287.0, + 1405.0, + 1287.0, + 1405.0, + 1331.0, + 292.0, + 1331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1320.0, + 298.0, + 1320.0, + 298.0, + 1358.0, + 295.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1320.0, + 1406.0, + 1320.0, + 1406.0, + 1358.0, + 399.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1352.0, + 547.0, + 1352.0, + 547.0, + 1390.0, + 293.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 1352.0, + 1278.0, + 1352.0, + 1278.0, + 1390.0, + 578.0, + 1390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1998.0, + 1006.0, + 1998.0, + 1006.0, + 2039.0, + 293.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 1998.0, + 1405.0, + 1998.0, + 1405.0, + 2039.0, + 1028.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 536, + 1405, + 536, + 1405, + 1026, + 298, + 1026 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 297, + 1758, + 1404, + 1758, + 1404, + 2035, + 297, + 2035 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 338, + 1403, + 338, + 1403, + 522, + 298, + 522 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 297, + 1209, + 1412, + 1209, + 1412, + 1367, + 297, + 1367 + ], + "score": 0.975, + "html": "
EpochlS0S2S3
#PParams (M)Error (%)#PParams (M)Error (%)#PParams (M)Error (%)
1506.6±1.13.3±0.32.74±0.066.0±0.03.9±0.32.58±0.11|6.0±1.03.6±0.32.55±0.03
2007.3±0.63.2±0.32.71±0.118.0±0.04.3±0.12.65±0.217.6±0.54.3±0.22.66±0.09
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1431, + 1405, + 1431, + 1405, + 1523, + 298, + 1523 + ], + "score": 0.967 + }, + { + "category_id": 6, + "poly": [ + 300, + 1106, + 1398, + 1106, + 1398, + 1201, + 300, + 1201 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1398, + 229, + 1398, + 322, + 298, + 322 + ], + "score": 0.942 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.894 + }, + { + "category_id": 0, + "poly": [ + 301, + 1600, + 767, + 1600, + 767, + 1636, + 301, + 1636 + ], + "score": 0.887 + }, + { + "category_id": 0, + "poly": [ + 299, + 1688, + 786, + 1688, + 786, + 1720, + 299, + 1720 + ], + "score": 0.865 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.787 + }, + { + "category_id": 13, + "poly": [ + 859, + 261, + 1130, + 261, + 1130, + 295, + 859, + 295 + ], + "score": 0.91, + "latex": "\\beta \\in \\{ 1 , 0 . 7 , 0 . 4 , 0 . 1 , 0 \\}" + }, + { + "category_id": 13, + "poly": [ + 690, + 690, + 761, + 690, + 761, + 721, + 690, + 721 + ], + "score": 0.91, + "latex": "\\beta = 1" + }, + { + "category_id": 13, + "poly": [ + 298, + 262, + 373, + 262, + 373, + 293, + 298, + 293 + ], + "score": 0.9, + "latex": "\\beta = 0" + }, + { + "category_id": 13, + "poly": [ + 597, + 964, + 750, + 964, + 750, + 995, + 597, + 995 + ], + "score": 0.89, + "latex": "( 2 . 5 9 { \\pm } 0 . 0 8 \\% )" + }, + { + "category_id": 13, + "poly": [ + 478, + 430, + 636, + 430, + 636, + 460, + 478, + 460 + ], + "score": 0.89, + "latex": "9 7 . 2 5 \\% \\pm 0 . 0 9" + }, + { + "category_id": 13, + "poly": [ + 442, + 293, + 474, + 293, + 474, + 323, + 442, + 323 + ], + "score": 0.88, + "latex": "\\beta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 904, + 1290, + 904, + 1290, + 933, + 1214, + 933 + ], + "score": 0.87, + "latex": "2 . 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 993, + 782, + 1112, + 782, + 1112, + 813, + 993, + 813 + ], + "score": 0.87, + "latex": "( \\# P = 0 )" + }, + { + "category_id": 13, + "poly": [ + 432, + 934, + 511, + 934, + 511, + 964, + 432, + 964 + ], + "score": 0.87, + "latex": "2 . 6 3 \\%" + }, + { + "category_id": 13, + "poly": [ + 1082, + 369, + 1135, + 369, + 1135, + 398, + 1082, + 398 + ], + "score": 0.87, + "latex": "1 \\times 1" + }, + { + "category_id": 13, + "poly": [ + 981, + 934, + 1136, + 934, + 1136, + 963, + 981, + 963 + ], + "score": 0.87, + "latex": "( 2 . 7 1 { \\pm } 0 . 1 1 \\% )" + }, + { + "category_id": 13, + "poly": [ + 1254, + 691, + 1275, + 691, + 1275, + 721, + 1254, + 721 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 658, + 232, + 678, + 232, + 678, + 262, + 658, + 262 + ], + "score": 0.84, + "latex": "\\beta" + }, + { + "category_id": 13, + "poly": [ + 377, + 995, + 465, + 995, + 465, + 1026, + 377, + 1026 + ], + "score": 0.84, + "latex": "( 2 . 4 2 \\% )" + }, + { + "category_id": 13, + "poly": [ + 687, + 782, + 736, + 782, + 736, + 813, + 687, + 813 + ], + "score": 0.83, + "latex": "\\# P" + }, + { + "category_id": 13, + "poly": [ + 1078, + 965, + 1166, + 965, + 1166, + 994, + 1078, + 994 + ], + "score": 0.8, + "latex": "( 2 . 5 3 \\% )" + }, + { + "category_id": 13, + "poly": [ + 445, + 1139, + 494, + 1139, + 494, + 1171, + 445, + 1171 + ], + "score": 0.79, + "latex": "\\# P" + }, + { + "category_id": 13, + "poly": [ + 1123, + 1766, + 1145, + 1766, + 1145, + 1787, + 1123, + 1787 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1104.0, + 1404.0, + 1104.0, + 1404.0, + 1143.0, + 292.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1136.0, + 444.0, + 1136.0, + 444.0, + 1173.0, + 292.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1136.0, + 1403.0, + 1136.0, + 1403.0, + 1173.0, + 495.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1168.0, + 412.0, + 1168.0, + 412.0, + 1200.0, + 294.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1596.0, + 775.0, + 1596.0, + 775.0, + 1643.0, + 292.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1688.0, + 790.0, + 1688.0, + 790.0, + 1724.0, + 295.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 860.0, + 2085.0, + 860.0, + 2116.0, + 839.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 535.0, + 1406.0, + 535.0, + 1406.0, + 573.0, + 293.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 566.0, + 1407.0, + 566.0, + 1407.0, + 605.0, + 292.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 600.0, + 1405.0, + 600.0, + 1405.0, + 631.0, + 296.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 1403.0, + 630.0, + 1403.0, + 664.0, + 295.0, + 664.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 654.0, + 1405.0, + 654.0, + 1405.0, + 696.0, + 292.0, + 696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 690.0, + 689.0, + 690.0, + 689.0, + 724.0, + 295.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 690.0, + 1253.0, + 690.0, + 1253.0, + 724.0, + 762.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 690.0, + 1407.0, + 690.0, + 1407.0, + 724.0, + 1276.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 720.0, + 1405.0, + 720.0, + 1405.0, + 758.0, + 293.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 747.0, + 1405.0, + 747.0, + 1405.0, + 787.0, + 293.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 783.0, + 686.0, + 783.0, + 686.0, + 814.0, + 295.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 783.0, + 992.0, + 783.0, + 992.0, + 814.0, + 737.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 783.0, + 1402.0, + 783.0, + 1402.0, + 814.0, + 1113.0, + 814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 1405.0, + 812.0, + 1405.0, + 846.0, + 295.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 841.0, + 1405.0, + 841.0, + 1405.0, + 878.0, + 293.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 871.0, + 1405.0, + 871.0, + 1405.0, + 909.0, + 293.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 905.0, + 1213.0, + 905.0, + 1213.0, + 936.0, + 296.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 905.0, + 1405.0, + 905.0, + 1405.0, + 936.0, + 1291.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 932.0, + 431.0, + 932.0, + 431.0, + 968.0, + 292.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 512.0, + 932.0, + 980.0, + 932.0, + 980.0, + 968.0, + 512.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 932.0, + 1406.0, + 932.0, + 1406.0, + 968.0, + 1137.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 963.0, + 596.0, + 963.0, + 596.0, + 997.0, + 293.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 963.0, + 1077.0, + 963.0, + 1077.0, + 997.0, + 751.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 963.0, + 1405.0, + 963.0, + 1405.0, + 997.0, + 1167.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 994.0, + 376.0, + 994.0, + 376.0, + 1029.0, + 295.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 994.0, + 478.0, + 994.0, + 478.0, + 1029.0, + 466.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1759.0, + 1122.0, + 1759.0, + 1122.0, + 1792.0, + 297.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1759.0, + 1400.0, + 1759.0, + 1400.0, + 1792.0, + 1146.0, + 1792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1789.0, + 1402.0, + 1789.0, + 1402.0, + 1822.0, + 295.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1406.0, + 1821.0, + 1406.0, + 1857.0, + 294.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1406.0, + 1849.0, + 1406.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1879.0, + 1405.0, + 1879.0, + 1405.0, + 1917.0, + 292.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1908.0, + 1404.0, + 1908.0, + 1404.0, + 1948.0, + 292.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1940.0, + 1402.0, + 1940.0, + 1402.0, + 1979.0, + 292.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2001.0, + 1190.0, + 2001.0, + 1190.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 335.0, + 1403.0, + 335.0, + 1403.0, + 374.0, + 295.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 364.0, + 1081.0, + 364.0, + 1081.0, + 405.0, + 293.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 364.0, + 1405.0, + 364.0, + 1405.0, + 405.0, + 1136.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1407.0, + 397.0, + 1407.0, + 436.0, + 293.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 427.0, + 477.0, + 427.0, + 477.0, + 464.0, + 293.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 427.0, + 1405.0, + 427.0, + 1405.0, + 464.0, + 637.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 460.0, + 1405.0, + 460.0, + 1405.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 490.0, + 491.0, + 490.0, + 491.0, + 524.0, + 293.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 1405.0, + 1429.0, + 1405.0, + 1466.0, + 293.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1460.0, + 1405.0, + 1460.0, + 1405.0, + 1497.0, + 295.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1491.0, + 1401.0, + 1491.0, + 1401.0, + 1526.0, + 293.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 657.0, + 228.0, + 657.0, + 265.0, + 293.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 265.0, + 679.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 261.0, + 858.0, + 261.0, + 858.0, + 295.0, + 374.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 261.0, + 1403.0, + 261.0, + 1403.0, + 295.0, + 1131.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 292.0, + 441.0, + 292.0, + 441.0, + 326.0, + 295.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 292.0, + 951.0, + 292.0, + 951.0, + 326.0, + 475.0, + 326.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 293, + 1402, + 293, + 1402, + 542, + 298, + 542 + ], + "score": 0.985 + }, + { + "category_id": 1, + "poly": [ + 298, + 1349, + 1404, + 1349, + 1404, + 1625, + 298, + 1625 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1640, + 1403, + 1640, + 1403, + 1855, + 298, + 1855 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1047, + 1403, + 1047, + 1403, + 1202, + 298, + 1202 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 319, + 616, + 1381, + 616, + 1381, + 841, + 319, + 841 + ], + "score": 0.967 + }, + { + "category_id": 4, + "poly": [ + 296, + 874, + 1404, + 874, + 1404, + 998, + 296, + 998 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 301, + 2002, + 754, + 2002, + 754, + 2034, + 301, + 2034 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 301, + 1921, + 661, + 1921, + 661, + 1955, + 301, + 1955 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 299, + 1267, + 543, + 1267, + 543, + 1302, + 299, + 1302 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.879 + }, + { + "category_id": 0, + "poly": [ + 301, + 230, + 810, + 230, + 810, + 261, + 301, + 261 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2111, + 840, + 2111 + ], + "score": 0.787 + }, + { + "category_id": 13, + "poly": [ + 1049, + 936, + 1138, + 936, + 1138, + 965, + 1049, + 965 + ], + "score": 0.88, + "latex": "9 4 . 8 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1146, + 1534, + 1187, + 1534, + 1187, + 1562, + 1146, + 1562 + ], + "score": 0.86, + "latex": "3 \\times" + }, + { + "category_id": 13, + "poly": [ + 1245, + 936, + 1334, + 936, + 1334, + 966, + 1245, + 966 + ], + "score": 0.86, + "latex": "9 7 . 5 8 \\%" + }, + { + "category_id": 13, + "poly": [ + 1380, + 1116, + 1401, + 1116, + 1401, + 1137, + 1380, + 1137 + ], + "score": 0.73, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 391, + 1085, + 413, + 1085, + 413, + 1108, + 391, + 1108 + ], + "score": 0.73, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1379, + 881, + 1402, + 881, + 1402, + 903, + 1379, + 903 + ], + "score": 0.6, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 642.0, + 1153.0, + 642.0, + 1153.0, + 651.0, + 1139.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 662.0, + 1153.0, + 662.0, + 1153.0, + 671.0, + 1141.0, + 671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 683.0, + 1153.0, + 683.0, + 1153.0, + 692.0, + 1141.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 697.0, + 336.0, + 697.0, + 336.0, + 707.0, + 325.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 695.0, + 833.0, + 695.0, + 833.0, + 704.0, + 822.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 702.0, + 1155.0, + 702.0, + 1155.0, + 712.0, + 1139.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 707.0, + 336.0, + 707.0, + 336.0, + 718.0, + 325.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 709.0, + 557.0, + 709.0, + 557.0, + 716.0, + 551.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 712.0, + 833.0, + 712.0, + 833.0, + 722.0, + 821.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 716.0, + 561.0, + 716.0, + 561.0, + 734.0, + 547.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 722.0, + 1155.0, + 722.0, + 1155.0, + 731.0, + 1137.0, + 731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 730.0, + 337.0, + 730.0, + 337.0, + 741.0, + 327.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 731.0, + 559.0, + 731.0, + 559.0, + 741.0, + 549.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 731.0, + 833.0, + 731.0, + 833.0, + 740.0, + 821.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 742.0, + 1151.0, + 742.0, + 1151.0, + 751.0, + 1139.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 763.0, + 1151.0, + 763.0, + 1151.0, + 772.0, + 1138.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 787.0, + 925.0, + 787.0, + 925.0, + 800.0, + 900.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 788.0, + 949.0, + 788.0, + 949.0, + 797.0, + 926.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 787.0, + 1003.0, + 787.0, + 1003.0, + 800.0, + 953.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 787.0, + 1028.0, + 787.0, + 1028.0, + 800.0, + 1006.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 788.0, + 1053.0, + 788.0, + 1053.0, + 797.0, + 1036.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1062.0, + 788.0, + 1080.0, + 788.0, + 1080.0, + 797.0, + 1062.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 788.0, + 1171.0, + 788.0, + 1171.0, + 797.0, + 1153.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 787.0, + 1305.0, + 787.0, + 1305.0, + 800.0, + 1172.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 787.0, + 1333.0, + 787.0, + 1333.0, + 800.0, + 1306.0, + 800.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1337.0, + 788.0, + 1354.0, + 788.0, + 1354.0, + 797.0, + 1337.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 809.0, + 499.0, + 809.0, + 499.0, + 839.0, + 372.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 809.0, + 776.0, + 809.0, + 776.0, + 839.0, + 640.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 809.0, + 1046.0, + 809.0, + 1046.0, + 839.0, + 919.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1187.0, + 809.0, + 1325.0, + 809.0, + 1325.0, + 839.0, + 1187.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 717.5, + 339.0, + 717.5, + 339.0, + 730.0, + 323.0, + 730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 770.0, + 445.0, + 770.0, + 445.0, + 790.0, + 338.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 779.5, + 467.0, + 779.5, + 467.0, + 794.0, + 440.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 873.0, + 1378.0, + 873.0, + 1378.0, + 909.0, + 294.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 873.0, + 1407.0, + 873.0, + 1407.0, + 909.0, + 1403.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 904.0, + 1405.0, + 904.0, + 1405.0, + 940.0, + 293.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 935.0, + 1048.0, + 935.0, + 1048.0, + 971.0, + 294.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 935.0, + 1244.0, + 935.0, + 1244.0, + 971.0, + 1139.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 935.0, + 1405.0, + 935.0, + 1405.0, + 971.0, + 1335.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 965.0, + 1133.0, + 965.0, + 1133.0, + 1000.0, + 295.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1918.0, + 666.0, + 1918.0, + 666.0, + 1960.0, + 293.0, + 1960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1263.0, + 548.0, + 1263.0, + 548.0, + 1311.0, + 292.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 227.0, + 816.0, + 227.0, + 816.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 330.0, + 293.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 325.0, + 1404.0, + 325.0, + 1404.0, + 361.0, + 293.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 354.0, + 1406.0, + 354.0, + 1406.0, + 392.0, + 293.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 388.0, + 1404.0, + 388.0, + 1404.0, + 422.0, + 294.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 416.0, + 1406.0, + 416.0, + 1406.0, + 454.0, + 293.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 483.0, + 294.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1407.0, + 478.0, + 1407.0, + 516.0, + 294.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 511.0, + 845.0, + 511.0, + 845.0, + 545.0, + 296.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1351.0, + 1404.0, + 1351.0, + 1404.0, + 1384.0, + 297.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1381.0, + 1404.0, + 1381.0, + 1404.0, + 1414.0, + 296.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1409.0, + 1404.0, + 1409.0, + 1404.0, + 1445.0, + 294.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1439.0, + 1404.0, + 1439.0, + 1404.0, + 1478.0, + 293.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1471.0, + 1405.0, + 1471.0, + 1405.0, + 1508.0, + 293.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1504.0, + 1402.0, + 1504.0, + 1402.0, + 1537.0, + 296.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1533.0, + 1145.0, + 1533.0, + 1145.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1533.0, + 1404.0, + 1533.0, + 1404.0, + 1565.0, + 1188.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1563.0, + 1406.0, + 1563.0, + 1406.0, + 1599.0, + 294.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1590.0, + 1341.0, + 1590.0, + 1341.0, + 1633.0, + 293.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1640.0, + 1405.0, + 1640.0, + 1405.0, + 1675.0, + 296.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1672.0, + 1404.0, + 1672.0, + 1404.0, + 1704.0, + 294.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1701.0, + 1405.0, + 1701.0, + 1405.0, + 1735.0, + 293.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1732.0, + 1403.0, + 1732.0, + 1403.0, + 1766.0, + 294.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1760.0, + 1404.0, + 1760.0, + 1404.0, + 1797.0, + 293.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1792.0, + 1406.0, + 1792.0, + 1406.0, + 1829.0, + 293.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1824.0, + 1059.0, + 1824.0, + 1059.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1047.0, + 1404.0, + 1047.0, + 1404.0, + 1085.0, + 296.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1078.0, + 390.0, + 1078.0, + 390.0, + 1115.0, + 296.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1078.0, + 1403.0, + 1078.0, + 1403.0, + 1115.0, + 414.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1105.0, + 1379.0, + 1105.0, + 1379.0, + 1148.0, + 291.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1105.0, + 1407.0, + 1105.0, + 1407.0, + 1148.0, + 1402.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1136.0, + 1405.0, + 1136.0, + 1405.0, + 1176.0, + 293.0, + 1176.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1167.0, + 1017.0, + 1167.0, + 1017.0, + 1208.0, + 293.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1996.0, + 759.0, + 1996.0, + 759.0, + 2042.0, + 295.0, + 2042.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.887 + }, + { + "category_id": 1, + "poly": [ + 297, + 277, + 1400, + 277, + 1400, + 341, + 297, + 341 + ], + "score": 0.829 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.829 + }, + { + "category_id": 1, + "poly": [ + 295, + 469, + 1402, + 469, + 1402, + 533, + 295, + 533 + ], + "score": 0.779 + }, + { + "category_id": 1, + "poly": [ + 298, + 1179, + 1400, + 1179, + 1400, + 1244, + 298, + 1244 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 297, + 1261, + 1401, + 1261, + 1401, + 1325, + 297, + 1325 + ], + "score": 0.761 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 262, + 300, + 262 + ], + "score": 0.761 + }, + { + "category_id": 1, + "poly": [ + 296, + 936, + 1402, + 936, + 1402, + 1001, + 296, + 1001 + ], + "score": 0.759 + }, + { + "category_id": 1, + "poly": [ + 299, + 1615, + 1400, + 1615, + 1400, + 1681, + 299, + 1681 + ], + "score": 0.757 + }, + { + "category_id": 1, + "poly": [ + 296, + 743, + 1402, + 743, + 1402, + 808, + 296, + 808 + ], + "score": 0.755 + }, + { + "category_id": 1, + "poly": [ + 297, + 1098, + 1402, + 1098, + 1402, + 1163, + 297, + 1163 + ], + "score": 0.752 + }, + { + "category_id": 1, + "poly": [ + 297, + 662, + 1399, + 662, + 1399, + 727, + 297, + 727 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 296, + 1533, + 1401, + 1533, + 1401, + 1600, + 296, + 1600 + ], + "score": 0.743 + }, + { + "category_id": 1, + "poly": [ + 297, + 1017, + 1402, + 1017, + 1402, + 1081, + 297, + 1081 + ], + "score": 0.741 + }, + { + "category_id": 1, + "poly": [ + 299, + 1858, + 1397, + 1858, + 1397, + 1924, + 299, + 1924 + ], + "score": 0.738 + }, + { + "category_id": 1, + "poly": [ + 295, + 1341, + 1400, + 1341, + 1400, + 1407, + 295, + 1407 + ], + "score": 0.736 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1402, + 1941, + 1402, + 2033, + 299, + 2033 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 291, + 1808, + 1398, + 1808, + 1398, + 1843, + 291, + 1843 + ], + "score": 0.71 + }, + { + "category_id": 1, + "poly": [ + 297, + 550, + 1406, + 550, + 1406, + 645, + 297, + 645 + ], + "score": 0.696 + }, + { + "category_id": 1, + "poly": [ + 297, + 1423, + 1405, + 1423, + 1405, + 1518, + 297, + 1518 + ], + "score": 0.668 + }, + { + "category_id": 1, + "poly": [ + 298, + 1696, + 1405, + 1696, + 1405, + 1790, + 298, + 1790 + ], + "score": 0.667 + }, + { + "category_id": 1, + "poly": [ + 298, + 357, + 1404, + 357, + 1404, + 450, + 298, + 450 + ], + "score": 0.645 + }, + { + "category_id": 1, + "poly": [ + 297, + 824, + 1404, + 824, + 1404, + 918, + 297, + 918 + ], + "score": 0.52 + }, + { + "category_id": 1, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 262, + 300, + 262 + ], + "score": 0.112 + }, + { + "category_id": 13, + "poly": [ + 742, + 1490, + 760, + 1490, + 760, + 1512, + 742, + 1512 + ], + "score": 0.57, + "latex": "=" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 276.0, + 1402.0, + 276.0, + 1402.0, + 312.0, + 296.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 308.0, + 498.0, + 308.0, + 498.0, + 340.0, + 322.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 466.0, + 1405.0, + 466.0, + 1405.0, + 507.0, + 292.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 501.0, + 729.0, + 501.0, + 729.0, + 533.0, + 324.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1181.0, + 1402.0, + 1181.0, + 1402.0, + 1217.0, + 294.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1212.0, + 1003.0, + 1212.0, + 1003.0, + 1245.0, + 322.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 1401.0, + 1260.0, + 1401.0, + 1296.0, + 296.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1292.0, + 1405.0, + 1292.0, + 1405.0, + 1328.0, + 324.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 938.0, + 1406.0, + 938.0, + 1406.0, + 974.0, + 295.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 969.0, + 1324.0, + 969.0, + 1324.0, + 1002.0, + 323.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1612.0, + 1405.0, + 1612.0, + 1405.0, + 1656.0, + 293.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1645.0, + 813.0, + 1645.0, + 813.0, + 1682.0, + 323.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 1403.0, + 741.0, + 1403.0, + 783.0, + 293.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 776.0, + 1089.0, + 776.0, + 1089.0, + 809.0, + 321.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1096.0, + 1406.0, + 1096.0, + 1406.0, + 1138.0, + 292.0, + 1138.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1129.0, + 931.0, + 1129.0, + 931.0, + 1161.0, + 325.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 661.0, + 1403.0, + 661.0, + 1403.0, + 700.0, + 293.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 693.0, + 743.0, + 693.0, + 743.0, + 728.0, + 322.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1535.0, + 1403.0, + 1535.0, + 1403.0, + 1572.0, + 296.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1567.0, + 1002.0, + 1567.0, + 1002.0, + 1600.0, + 321.0, + 1600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1016.0, + 1407.0, + 1016.0, + 1407.0, + 1055.0, + 293.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1048.0, + 588.0, + 1048.0, + 588.0, + 1083.0, + 320.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1856.0, + 1402.0, + 1856.0, + 1402.0, + 1897.0, + 294.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1891.0, + 1006.0, + 1891.0, + 1006.0, + 1925.0, + 323.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1339.0, + 1406.0, + 1339.0, + 1406.0, + 1379.0, + 294.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1373.0, + 670.0, + 1373.0, + 670.0, + 1408.0, + 324.0, + 1408.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1402.0, + 1941.0, + 1402.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1973.0, + 1404.0, + 1973.0, + 1404.0, + 2006.0, + 322.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1996.0, + 410.0, + 1996.0, + 410.0, + 2037.0, + 320.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1807.0, + 1404.0, + 1807.0, + 1404.0, + 1847.0, + 294.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 552.0, + 1404.0, + 552.0, + 1404.0, + 587.0, + 296.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 585.0, + 1402.0, + 585.0, + 1402.0, + 616.0, + 324.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 613.0, + 788.0, + 613.0, + 788.0, + 647.0, + 323.0, + 647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1461.0, + 295.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1453.0, + 1403.0, + 1453.0, + 1403.0, + 1492.0, + 320.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1486.0, + 741.0, + 1486.0, + 741.0, + 1520.0, + 325.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1486.0, + 935.0, + 1486.0, + 935.0, + 1520.0, + 761.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1694.0, + 1403.0, + 1694.0, + 1403.0, + 1735.0, + 292.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1765.0, + 322.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1756.0, + 401.0, + 1756.0, + 401.0, + 1793.0, + 320.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 359.0, + 1404.0, + 359.0, + 1404.0, + 393.0, + 296.0, + 393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 390.0, + 1404.0, + 390.0, + 1404.0, + 424.0, + 323.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 419.0, + 396.0, + 419.0, + 396.0, + 453.0, + 321.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 822.0, + 1405.0, + 822.0, + 1405.0, + 864.0, + 294.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 856.0, + 1405.0, + 856.0, + 1405.0, + 890.0, + 322.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 883.0, + 411.0, + 883.0, + 411.0, + 922.0, + 320.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.83 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1399, + 228, + 1399, + 323, + 298, + 323 + ], + "score": 0.761 + }, + { + "category_id": 1, + "poly": [ + 297, + 933, + 1403, + 933, + 1403, + 998, + 297, + 998 + ], + "score": 0.726 + }, + { + "category_id": 1, + "poly": [ + 295, + 1017, + 1402, + 1017, + 1402, + 1082, + 295, + 1082 + ], + "score": 0.722 + }, + { + "category_id": 1, + "poly": [ + 295, + 766, + 1401, + 766, + 1401, + 831, + 295, + 831 + ], + "score": 0.713 + }, + { + "category_id": 1, + "poly": [ + 297, + 849, + 1400, + 849, + 1400, + 915, + 297, + 915 + ], + "score": 0.71 + }, + { + "category_id": 1, + "poly": [ + 297, + 1296, + 1364, + 1296, + 1364, + 1332, + 297, + 1332 + ], + "score": 0.698 + }, + { + "category_id": 1, + "poly": [ + 291, + 1659, + 1403, + 1659, + 1403, + 1725, + 291, + 1725 + ], + "score": 0.693 + }, + { + "category_id": 1, + "poly": [ + 291, + 1213, + 1398, + 1213, + 1398, + 1278, + 291, + 1278 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 295, + 1099, + 1400, + 1099, + 1400, + 1194, + 295, + 1194 + ], + "score": 0.666 + }, + { + "category_id": 1, + "poly": [ + 291, + 1576, + 1404, + 1576, + 1404, + 1642, + 291, + 1642 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 295, + 1349, + 1396, + 1349, + 1396, + 1443, + 295, + 1443 + ], + "score": 0.656 + }, + { + "category_id": 1, + "poly": [ + 296, + 455, + 1399, + 455, + 1399, + 521, + 296, + 521 + ], + "score": 0.627 + }, + { + "category_id": 1, + "poly": [ + 295, + 540, + 1404, + 540, + 1404, + 634, + 295, + 634 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 298, + 1462, + 1401, + 1462, + 1401, + 1557, + 298, + 1557 + ], + "score": 0.615 + }, + { + "category_id": 1, + "poly": [ + 291, + 1744, + 1401, + 1744, + 1401, + 1808, + 291, + 1808 + ], + "score": 0.59 + }, + { + "category_id": 1, + "poly": [ + 300, + 343, + 1400, + 343, + 1400, + 437, + 300, + 437 + ], + "score": 0.575 + }, + { + "category_id": 1, + "poly": [ + 301, + 1827, + 1401, + 1827, + 1401, + 1922, + 301, + 1922 + ], + "score": 0.547 + }, + { + "category_id": 1, + "poly": [ + 300, + 1942, + 1404, + 1942, + 1404, + 2034, + 300, + 2034 + ], + "score": 0.42 + }, + { + "category_id": 1, + "poly": [ + 296, + 653, + 1406, + 653, + 1406, + 748, + 296, + 748 + ], + "score": 0.136 + }, + { + "category_id": 13, + "poly": [ + 563, + 378, + 579, + 378, + 579, + 402, + 563, + 402 + ], + "score": 0.32, + "latex": "^ +" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 1405.0, + 224.0, + 1405.0, + 269.0, + 292.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 260.0, + 1404.0, + 260.0, + 1404.0, + 300.0, + 321.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 291.0, + 518.0, + 291.0, + 518.0, + 323.0, + 324.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 931.0, + 1405.0, + 931.0, + 1405.0, + 971.0, + 293.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 964.0, + 846.0, + 964.0, + 846.0, + 996.0, + 325.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1012.0, + 1406.0, + 1012.0, + 1406.0, + 1057.0, + 292.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1047.0, + 1139.0, + 1047.0, + 1139.0, + 1083.0, + 324.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 1405.0, + 766.0, + 1405.0, + 802.0, + 294.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 797.0, + 481.0, + 797.0, + 481.0, + 831.0, + 323.0, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 850.0, + 1404.0, + 850.0, + 1404.0, + 886.0, + 295.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 881.0, + 1238.0, + 881.0, + 1238.0, + 916.0, + 323.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1293.0, + 1364.0, + 1293.0, + 1364.0, + 1335.0, + 294.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1657.0, + 1406.0, + 1657.0, + 1406.0, + 1697.0, + 293.0, + 1697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1688.0, + 400.0, + 1688.0, + 400.0, + 1727.0, + 318.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1215.0, + 1403.0, + 1215.0, + 1403.0, + 1251.0, + 294.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1244.0, + 1305.0, + 1244.0, + 1305.0, + 1280.0, + 323.0, + 1280.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1099.0, + 1405.0, + 1099.0, + 1405.0, + 1137.0, + 295.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1128.0, + 1406.0, + 1128.0, + 1406.0, + 1170.0, + 320.0, + 1170.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1161.0, + 520.0, + 1161.0, + 520.0, + 1195.0, + 319.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1577.0, + 1403.0, + 1577.0, + 1403.0, + 1613.0, + 295.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1609.0, + 629.0, + 1609.0, + 629.0, + 1641.0, + 323.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1351.0, + 1401.0, + 1351.0, + 1401.0, + 1385.0, + 296.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1380.0, + 1403.0, + 1380.0, + 1403.0, + 1414.0, + 323.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1413.0, + 648.0, + 1413.0, + 648.0, + 1443.0, + 324.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 456.0, + 1404.0, + 456.0, + 1404.0, + 492.0, + 296.0, + 492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 487.0, + 888.0, + 487.0, + 888.0, + 522.0, + 322.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 540.0, + 1404.0, + 540.0, + 1404.0, + 574.0, + 297.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 571.0, + 1405.0, + 571.0, + 1405.0, + 605.0, + 324.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 602.0, + 520.0, + 602.0, + 520.0, + 635.0, + 319.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1462.0, + 1403.0, + 1462.0, + 1403.0, + 1500.0, + 297.0, + 1500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1494.0, + 1405.0, + 1494.0, + 1405.0, + 1530.0, + 321.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1523.0, + 466.0, + 1523.0, + 466.0, + 1558.0, + 322.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1744.0, + 1403.0, + 1744.0, + 1403.0, + 1780.0, + 295.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1775.0, + 1030.0, + 1775.0, + 1030.0, + 1807.0, + 323.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 343.0, + 1405.0, + 343.0, + 1405.0, + 380.0, + 294.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 373.0, + 562.0, + 373.0, + 562.0, + 409.0, + 321.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 373.0, + 1404.0, + 373.0, + 1404.0, + 409.0, + 580.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 407.0, + 707.0, + 407.0, + 707.0, + 438.0, + 320.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1827.0, + 1407.0, + 1827.0, + 1407.0, + 1865.0, + 297.0, + 1865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1857.0, + 1407.0, + 1857.0, + 1407.0, + 1895.0, + 322.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1889.0, + 887.0, + 1889.0, + 887.0, + 1922.0, + 322.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1941.0, + 1404.0, + 1941.0, + 1404.0, + 1978.0, + 295.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1969.0, + 1404.0, + 1969.0, + 1404.0, + 2010.0, + 320.0, + 2010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 2004.0, + 1029.0, + 2004.0, + 1029.0, + 2034.0, + 325.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 655.0, + 1401.0, + 655.0, + 1401.0, + 686.0, + 298.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 685.0, + 1403.0, + 685.0, + 1403.0, + 720.0, + 322.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 715.0, + 614.0, + 715.0, + 614.0, + 750.0, + 325.0, + 750.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.873 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 293, + 225, + 1406, + 225, + 1406, + 804, + 293, + 804 + ], + "score": 0.625 + }, + { + "category_id": 13, + "poly": [ + 742, + 606, + 760, + 606, + 760, + 628, + 742, + 628 + ], + "score": 0.57, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 904, + 378, + 922, + 378, + 922, + 401, + 904, + 401 + ], + "score": 0.54, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 525, + 492, + 544, + 492, + 544, + 513, + 525, + 513 + ], + "score": 0.4, + "latex": "{ . } = { }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 230.0, + 1404.0, + 230.0, + 1404.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 262.0, + 590.0, + 262.0, + 590.0, + 297.0, + 324.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 310.0, + 1403.0, + 310.0, + 1403.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 346.0, + 1403.0, + 346.0, + 1403.0, + 381.0, + 323.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 375.0, + 903.0, + 375.0, + 903.0, + 412.0, + 323.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 375.0, + 1098.0, + 375.0, + 1098.0, + 412.0, + 923.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 422.0, + 1406.0, + 422.0, + 1406.0, + 470.0, + 291.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 460.0, + 1402.0, + 460.0, + 1402.0, + 495.0, + 323.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 486.0, + 524.0, + 486.0, + 524.0, + 526.0, + 323.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 486.0, + 721.0, + 486.0, + 721.0, + 526.0, + 545.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 542.0, + 1404.0, + 542.0, + 1404.0, + 577.0, + 295.0, + 577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 568.0, + 1403.0, + 568.0, + 1403.0, + 611.0, + 321.0, + 611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 602.0, + 741.0, + 602.0, + 741.0, + 639.0, + 324.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 602.0, + 937.0, + 602.0, + 937.0, + 639.0, + 761.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 652.0, + 1405.0, + 652.0, + 1405.0, + 693.0, + 291.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 683.0, + 1226.0, + 683.0, + 1226.0, + 723.0, + 321.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 739.0, + 1403.0, + 739.0, + 1403.0, + 774.0, + 295.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 768.0, + 1053.0, + 768.0, + 1053.0, + 805.0, + 321.0, + 805.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 374, + 1321, + 1319, + 1321, + 1319, + 1619, + 374, + 1619 + ], + "score": 0.983, + "html": "
BenchmarkDARTSDARTS-ESDARTS-ADAoursours (EV)
C10S14.66±0.713.05±0.073.03±0.082.76±0.070.37±0.19
S24.42±0.403.41±0.143.59±0.312.79±0.040.41±0.08
S34.12±0.853.71±1.142.99±0.342.65±0.040.31±0.09
S46.95±0.184.17±0.213.89±0.672.91±0.040.30±0.11
C100S129.93±0.4128.90±0.8124.94±0.8123.26±0.590.60±0.15
S228.75±0.9224.68±1.4326.88±1.1122.31±0.650.54±0.15
S329.01±0.2426.99±1.7924.55±0.6321.47±0.400.40±0.03
S424.77±1.5123.90±2.0123.66±0.9021.75±0.260.93±0.15
" + }, + { + "category_id": 1, + "poly": [ + 298, + 352, + 1403, + 352, + 1403, + 481, + 298, + 481 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 605, + 1401, + 605, + 1401, + 698, + 298, + 698 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1651, + 1405, + 1651, + 1405, + 1807, + 297, + 1807 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 830, + 1405, + 830, + 1405, + 924, + 298, + 924 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 621, + 498, + 1076, + 498, + 1076, + 591, + 621, + 591 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 296, + 1820, + 1402, + 1820, + 1402, + 1884, + 296, + 1884 + ], + "score": 0.952 + }, + { + "category_id": 1, + "poly": [ + 297, + 1072, + 1400, + 1072, + 1400, + 1136, + 297, + 1136 + ], + "score": 0.951 + }, + { + "category_id": 1, + "poly": [ + 297, + 1971, + 1397, + 1971, + 1397, + 2035, + 297, + 2035 + ], + "score": 0.944 + }, + { + "category_id": 6, + "poly": [ + 298, + 1189, + 1404, + 1189, + 1404, + 1314, + 298, + 1314 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 643, + 715, + 905, + 715, + 905, + 761, + 643, + 761 + ], + "score": 0.915 + }, + { + "category_id": 0, + "poly": [ + 301, + 295, + 733, + 295, + 733, + 329, + 301, + 329 + ], + "score": 0.914 + }, + { + "category_id": 8, + "poly": [ + 639, + 768, + 1060, + 768, + 1060, + 813, + 639, + 813 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 300, + 961, + 532, + 961, + 532, + 993, + 300, + 993 + ], + "score": 0.907 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 301, + 1018, + 630, + 1018, + 630, + 1051, + 301, + 1051 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 301, + 226, + 509, + 226, + 509, + 261, + 301, + 261 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 298, + 1917, + 1036, + 1917, + 1036, + 1950, + 298, + 1950 + ], + "score": 0.897 + }, + { + "category_id": 9, + "poly": [ + 1366, + 526, + 1400, + 526, + 1400, + 556, + 1366, + 556 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 771, + 1400, + 771, + 1400, + 800, + 1366, + 800 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1366, + 719, + 1399, + 719, + 1399, + 748, + 1366, + 748 + ], + "score": 0.875 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 864, + 2088, + 864, + 2113, + 836, + 2113 + ], + "score": 0.854 + }, + { + "category_id": 14, + "poly": [ + 622, + 495, + 1079, + 495, + 1079, + 591, + 622, + 591 + ], + "score": 0.95, + "latex": "\\bar { o } ^ { ( i , j ) } ( x ) = \\sum _ { o \\in \\mathcal { O } } \\frac { \\exp ( \\alpha _ { o } ^ { ( i , j ) } ) } { \\sum _ { o ^ { \\prime } \\in \\mathcal { O } } \\exp ( \\alpha _ { o ^ { \\prime } } ^ { ( i , j ) } ) } o ( x )" + }, + { + "category_id": 14, + "poly": [ + 637, + 712, + 1061, + 712, + 1061, + 816, + 637, + 816 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { \\underset { \\alpha } { \\operatorname* { m i n } } } & { \\mathcal { L } _ { v a l } ( w ^ { * } ( \\alpha ) , \\alpha ) } \\\\ { s . t . } & { w ^ { * } ( \\alpha ) = \\arg \\underset { w } { \\operatorname* { m i n } } \\mathcal { L } _ { t r a i n } ( w , \\alpha ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 881, + 445, + 939, + 445, + 939, + 477, + 881, + 477 + ], + "score": 0.9, + "latex": "\\bar { o } ^ { ( i , j ) }" + }, + { + "category_id": 13, + "poly": [ + 1109, + 454, + 1127, + 454, + 1127, + 476, + 1109, + 476 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 468, + 641, + 488, + 641, + 488, + 663, + 468, + 663 + ], + "score": 0.74, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 298, + 450, + 324, + 450, + 324, + 476, + 298, + 476 + ], + "score": 0.61, + "latex": "\\mathcal { O }" + }, + { + "category_id": 13, + "poly": [ + 666, + 449, + 708, + 449, + 708, + 481, + 666, + 481 + ], + "score": 0.53, + "latex": "i , j" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1192.0, + 1404.0, + 1192.0, + 1404.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1222.0, + 1405.0, + 1222.0, + 1405.0, + 1255.0, + 293.0, + 1255.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1254.0, + 1406.0, + 1254.0, + 1406.0, + 1287.0, + 294.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1282.0, + 1078.0, + 1282.0, + 1078.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 296.0, + 734.0, + 296.0, + 734.0, + 332.0, + 296.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 962.0, + 534.0, + 962.0, + 534.0, + 995.0, + 297.0, + 995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1016.0, + 633.0, + 1016.0, + 633.0, + 1054.0, + 296.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 220.0, + 514.0, + 220.0, + 514.0, + 270.0, + 293.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1916.0, + 1038.0, + 1916.0, + 1038.0, + 1953.0, + 294.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 869.0, + 2085.0, + 869.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 353.0, + 1403.0, + 353.0, + 1403.0, + 387.0, + 293.0, + 387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 383.0, + 1407.0, + 383.0, + 1407.0, + 421.0, + 292.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 450.0, + 293.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 445.0, + 297.0, + 445.0, + 297.0, + 486.0, + 293.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 445.0, + 665.0, + 445.0, + 665.0, + 486.0, + 325.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 445.0, + 880.0, + 445.0, + 880.0, + 486.0, + 709.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 445.0, + 1108.0, + 445.0, + 1108.0, + 486.0, + 940.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 445.0, + 1284.0, + 445.0, + 1284.0, + 486.0, + 1128.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 606.0, + 1403.0, + 606.0, + 1403.0, + 639.0, + 294.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 636.0, + 467.0, + 636.0, + 467.0, + 668.0, + 293.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 636.0, + 1405.0, + 636.0, + 1405.0, + 668.0, + 489.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 664.0, + 1401.0, + 664.0, + 1401.0, + 703.0, + 293.0, + 703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1649.0, + 1405.0, + 1649.0, + 1405.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1678.0, + 1405.0, + 1678.0, + 1405.0, + 1718.0, + 294.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1750.0, + 294.0, + 1750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1777.0, + 296.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1772.0, + 1197.0, + 1772.0, + 1197.0, + 1812.0, + 292.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 826.0, + 1405.0, + 826.0, + 1405.0, + 867.0, + 293.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 861.0, + 1405.0, + 861.0, + 1405.0, + 895.0, + 295.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 893.0, + 1064.0, + 893.0, + 1064.0, + 927.0, + 296.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1819.0, + 1404.0, + 1819.0, + 1404.0, + 1856.0, + 295.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1339.0, + 1851.0, + 1339.0, + 1887.0, + 295.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1071.0, + 1405.0, + 1071.0, + 1405.0, + 1109.0, + 295.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1104.0, + 1296.0, + 1104.0, + 1296.0, + 1140.0, + 295.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1402.0, + 1970.0, + 1402.0, + 2007.0, + 293.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1402.0, + 2003.0, + 1402.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1065, + 1404, + 1065, + 1404, + 1312, + 296, + 1312 + ], + "score": 0.985 + }, + { + "category_id": 5, + "poly": [ + 343, + 769, + 1354, + 769, + 1354, + 1024, + 343, + 1024 + ], + "score": 0.983, + "html": "
BackbonesParams (M)AccAPAP50AP75APsAPMAPL
MobileNetV2 (2018)3.472.028.346.729.314.830.738.1
SingPath NAS (2019)4.375.030.749.832.215.433.941.6
MnasNet-A2 (2019)4.875.630.550.232.016.634.141.1
MobileNetV3 (2019)5.475.229.949.330.814.933.341.1
MixNet-M (2019)5.077.031.351.732.417.035.041.9
FairDARTS-C (2020b)5.377.231.951.933.017.435.343.0
DARTS-A (Ours)5.577.832.552.834.118.036.143.4
" + }, + { + "category_id": 6, + "poly": [ + 297, + 248, + 1397, + 248, + 1397, + 311, + 297, + 311 + ], + "score": 0.942 + }, + { + "category_id": 3, + "poly": [ + 318, + 1358, + 1389, + 1358, + 1389, + 1821, + 318, + 1821 + ], + "score": 0.932 + }, + { + "category_id": 6, + "poly": [ + 429, + 736, + 1266, + 736, + 1266, + 767, + 429, + 767 + ], + "score": 0.929 + }, + { + "category_id": 4, + "poly": [ + 295, + 1855, + 1407, + 1855, + 1407, + 2010, + 295, + 2010 + ], + "score": 0.905 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 295, + 1855, + 1407, + 1855, + 1407, + 2010, + 295, + 2010 + ], + "score": 0.101 + }, + { + "category_id": 13, + "poly": [ + 473, + 1127, + 550, + 1127, + 550, + 1157, + 473, + 1157 + ], + "score": 0.9, + "latex": "9 7 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 360, + 1127, + 437, + 1127, + 437, + 1158, + 360, + 1158 + ], + "score": 0.88, + "latex": "9 6 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 673, + 279, + 688, + 279, + 688, + 306, + 673, + 306 + ], + "score": 0.46, + "latex": "^ \\ddag" + }, + { + "category_id": 5, + "poly": [ + 590, + 312, + 1109, + 312, + 1109, + 680, + 590, + 680 + ], + "score": 0.924, + "html": "
ModelsParams Error (M) (%)Cost GPU Days
ResNet (2016)1.722.10-
AmoebaNet (2019) 3.118.933150
PNAS (2018)3.219.53150
ENAS (2018)4.619.430.45
DARTS (2019b)120.58±0.44*(*0.4
GDAS (2019b)3.418.380.2
P-DARTS (2019b） 3.617.49‡0.3
R-DARTS (2020)118.01±0.261.6
DARTS- (avg.)3.317.51±0.250.4
DARTS- (best)3.417.160.4
" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 244.0, + 1403.0, + 244.0, + 1403.0, + 287.0, + 292.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 276.0, + 672.0, + 276.0, + 672.0, + 315.0, + 293.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 276.0, + 891.0, + 276.0, + 891.0, + 315.0, + 689.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 314.0, + 1382.0, + 361.0, + 1382.0, + 361.0, + 1476.0, + 314.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 1368.0, + 412.0, + 1368.0, + 412.0, + 1411.0, + 378.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1418.0, + 446.0, + 1418.0, + 446.0, + 1456.0, + 433.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 1383.0, + 504.0, + 1383.0, + 504.0, + 1477.0, + 457.0, + 1477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1378.0, + 645.0, + 1378.0, + 645.0, + 1478.0, + 597.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1362.0, + 790.0, + 1362.0, + 790.0, + 1478.0, + 722.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 1426.0, + 844.0, + 1426.0, + 844.0, + 1441.0, + 828.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1367.0, + 929.0, + 1367.0, + 929.0, + 1478.0, + 883.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1362.0, + 1101.0, + 1362.0, + 1101.0, + 1468.0, + 1051.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1104.0, + 1373.0, + 1155.0, + 1373.0, + 1155.0, + 1406.0, + 1104.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1388.0, + 1334.0, + 1388.0, + 1334.0, + 1407.0, + 1307.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 1371.0, + 1378.0, + 1371.0, + 1378.0, + 1456.0, + 1338.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 1454.0, + 778.0, + 1454.0, + 778.0, + 1473.0, + 756.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1451.0, + 1150.0, + 1451.0, + 1150.0, + 1465.0, + 1133.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1338.0, + 1450.0, + 1365.0, + 1450.0, + 1365.0, + 1468.0, + 1338.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1470.0, + 351.0, + 1470.0, + 351.0, + 1485.0, + 334.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1480.0, + 494.0, + 1480.0, + 494.0, + 1495.0, + 477.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 1474.0, + 638.0, + 1474.0, + 638.0, + 1489.0, + 618.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1473.0, + 658.0, + 1473.0, + 658.0, + 1479.0, + 652.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 1479.0, + 779.0, + 1479.0, + 779.0, + 1497.0, + 755.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 1465.0, + 816.0, + 1465.0, + 816.0, + 1480.0, + 800.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1470.0, + 919.0, + 1470.0, + 919.0, + 1485.0, + 904.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1071.0, + 1470.0, + 1093.0, + 1470.0, + 1093.0, + 1489.0, + 1071.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1475.0, + 1160.0, + 1475.0, + 1160.0, + 1493.0, + 1132.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1470.0, + 1204.0, + 1470.0, + 1204.0, + 1489.0, + 1176.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 1469.0, + 1246.0, + 1469.0, + 1246.0, + 1483.0, + 1214.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1491.0, + 351.0, + 1491.0, + 351.0, + 1508.0, + 334.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1486.0, + 921.0, + 1486.0, + 921.0, + 1502.0, + 902.0, + 1502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 1489.0, + 1365.0, + 1489.0, + 1365.0, + 1507.0, + 1339.0, + 1507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1504.0, + 361.0, + 1504.0, + 361.0, + 1517.0, + 352.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1502.0, + 419.0, + 1502.0, + 419.0, + 1531.0, + 384.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1503.0, + 438.0, + 1503.0, + 438.0, + 1518.0, + 422.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 1508.0, + 502.0, + 1508.0, + 502.0, + 1514.0, + 496.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1503.0, + 562.0, + 1503.0, + 562.0, + 1533.0, + 527.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 1503.0, + 579.0, + 1503.0, + 579.0, + 1518.0, + 564.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1507.0, + 646.0, + 1507.0, + 646.0, + 1515.0, + 636.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1503.0, + 721.0, + 1503.0, + 721.0, + 1533.0, + 668.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1504.0, + 789.0, + 1504.0, + 789.0, + 1515.0, + 778.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1503.0, + 846.0, + 1503.0, + 846.0, + 1533.0, + 811.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1503.0, + 863.0, + 1503.0, + 863.0, + 1518.0, + 847.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1504.0, + 930.0, + 1504.0, + 930.0, + 1517.0, + 919.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1502.0, + 986.0, + 1502.0, + 986.0, + 1532.0, + 954.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 1503.0, + 1005.0, + 1503.0, + 1005.0, + 1518.0, + 989.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1502.0, + 1100.0, + 1502.0, + 1100.0, + 1513.0, + 1090.0, + 1513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1500.0, + 1147.0, + 1500.0, + 1147.0, + 1515.0, + 1130.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1500.0, + 1191.0, + 1500.0, + 1191.0, + 1517.0, + 1174.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1499.0, + 1238.0, + 1499.0, + 1238.0, + 1530.0, + 1197.0, + 1530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1499.0, + 1280.0, + 1499.0, + 1280.0, + 1515.0, + 1263.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1307.0, + 1499.0, + 1324.0, + 1499.0, + 1324.0, + 1515.0, + 1307.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1542.0, + 1012.0, + 1542.0, + 1012.0, + 1573.0, + 340.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 1543.0, + 1395.0, + 1543.0, + 1395.0, + 1574.0, + 1039.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 1606.0, + 530.0, + 1606.0, + 530.0, + 1625.0, + 503.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 1625.0, + 542.0, + 1625.0, + 542.0, + 1727.0, + 485.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1616.0, + 588.0, + 1616.0, + 588.0, + 1657.0, + 556.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 617.0, + 1656.0, + 625.0, + 1656.0, + 625.0, + 1667.0, + 617.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1616.0, + 689.0, + 1616.0, + 689.0, + 1721.0, + 633.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1635.0, + 821.0, + 1635.0, + 821.0, + 1720.0, + 773.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1631.0, + 905.0, + 1631.0, + 905.0, + 1640.0, + 893.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1618.0, + 952.0, + 1618.0, + 952.0, + 1712.0, + 920.0, + 1712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1616.0, + 1102.0, + 1616.0, + 1102.0, + 1723.0, + 1050.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1632.0, + 1191.0, + 1632.0, + 1191.0, + 1647.0, + 1182.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1701.0, + 671.0, + 1701.0, + 671.0, + 1720.0, + 650.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1716.0, + 530.0, + 1716.0, + 530.0, + 1752.0, + 505.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1720.0, + 672.0, + 1720.0, + 672.0, + 1754.0, + 650.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 1725.0, + 812.0, + 1725.0, + 812.0, + 1740.0, + 793.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1725.0, + 951.0, + 1725.0, + 951.0, + 1740.0, + 930.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1709.0, + 1093.0, + 1709.0, + 1093.0, + 1747.0, + 1069.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 1749.0, + 539.0, + 1749.0, + 539.0, + 1762.0, + 528.0, + 1762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 1747.0, + 578.0, + 1747.0, + 578.0, + 1763.0, + 561.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1748.0, + 613.0, + 1748.0, + 613.0, + 1763.0, + 596.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 1748.0, + 682.0, + 1748.0, + 682.0, + 1763.0, + 667.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 1748.0, + 719.0, + 1748.0, + 719.0, + 1763.0, + 702.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 1748.0, + 753.0, + 1748.0, + 753.0, + 1763.0, + 738.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1748.0, + 861.0, + 1748.0, + 861.0, + 1763.0, + 843.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1748.0, + 895.0, + 1748.0, + 895.0, + 1763.0, + 878.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1749.0, + 962.0, + 1749.0, + 962.0, + 1760.0, + 952.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1748.0, + 1002.0, + 1748.0, + 1002.0, + 1763.0, + 984.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1748.0, + 1036.0, + 1748.0, + 1036.0, + 1763.0, + 1019.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1094.0, + 1750.0, + 1102.0, + 1750.0, + 1102.0, + 1760.0, + 1094.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1745.0, + 1145.0, + 1745.0, + 1145.0, + 1764.0, + 1124.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1748.0, + 1178.0, + 1748.0, + 1178.0, + 1763.0, + 1161.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 562.0, + 1757.0, + 596.0, + 1757.0, + 596.0, + 1778.0, + 562.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1757.0, + 738.0, + 1757.0, + 738.0, + 1778.0, + 703.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1758.0, + 877.0, + 1758.0, + 877.0, + 1777.0, + 846.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 1757.0, + 1019.0, + 1757.0, + 1019.0, + 1778.0, + 985.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1757.0, + 1161.0, + 1757.0, + 1161.0, + 1778.0, + 1127.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1787.0, + 1188.0, + 1787.0, + 1188.0, + 1819.0, + 506.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1306.0, + 1373.0, + 1337.0, + 1373.0, + 1337.0, + 1391.0, + 1306.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1427.0, + 1203.0, + 1427.0, + 1203.0, + 1445.0, + 1174.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1443.0, + 1297.0, + 1443.0, + 1297.0, + 1463.0, + 1256.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1472.0, + 954.0, + 1472.0, + 954.0, + 1481.0, + 945.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.75, + 1640.0, + 953.75, + 1640.0, + 953.75, + 1658.5, + 926.75, + 1658.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 1709.0, + 551.0, + 1709.0, + 551.0, + 1719.5, + 543.0, + 1719.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1700.5, + 956.0, + 1700.5, + 956.0, + 1720.5, + 924.0, + 1720.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 730.0, + 1272.0, + 730.0, + 1272.0, + 775.0, + 427.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1857.0, + 1404.0, + 1857.0, + 1404.0, + 1890.0, + 296.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1921.0, + 293.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1917.0, + 1405.0, + 1917.0, + 1405.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1984.0, + 296.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1979.0, + 766.0, + 1979.0, + 766.0, + 2012.0, + 297.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1063.0, + 1405.0, + 1063.0, + 1405.0, + 1103.0, + 294.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1097.0, + 1406.0, + 1097.0, + 1406.0, + 1131.0, + 295.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1126.0, + 359.0, + 1126.0, + 359.0, + 1162.0, + 294.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 438.0, + 1126.0, + 472.0, + 1126.0, + 472.0, + 1162.0, + 438.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1126.0, + 1405.0, + 1126.0, + 1405.0, + 1162.0, + 551.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1158.0, + 1404.0, + 1158.0, + 1404.0, + 1192.0, + 295.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1185.0, + 1406.0, + 1185.0, + 1406.0, + 1227.0, + 291.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1217.0, + 1406.0, + 1217.0, + 1406.0, + 1254.0, + 294.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1246.0, + 1405.0, + 1246.0, + 1405.0, + 1286.0, + 290.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1280.0, + 863.0, + 1280.0, + 863.0, + 1314.0, + 294.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1857.0, + 1404.0, + 1857.0, + 1404.0, + 1890.0, + 296.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1886.0, + 1405.0, + 1886.0, + 1405.0, + 1921.0, + 293.0, + 1921.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1917.0, + 1405.0, + 1917.0, + 1405.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1947.0, + 1405.0, + 1947.0, + 1405.0, + 1984.0, + 296.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1979.0, + 766.0, + 1979.0, + 766.0, + 2012.0, + 297.0, + 2012.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1511, + 1404, + 1511, + 1404, + 1634, + 298, + 1634 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 599, + 238, + 1075, + 238, + 1075, + 600, + 599, + 600 + ], + "score": 0.968 + }, + { + "category_id": 3, + "poly": [ + 676, + 911, + 1019, + 911, + 1019, + 1264, + 676, + 1264 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 296, + 803, + 1403, + 803, + 1403, + 867, + 296, + 867 + ], + "score": 0.946 + }, + { + "category_id": 4, + "poly": [ + 296, + 631, + 1403, + 631, + 1403, + 695, + 296, + 695 + ], + "score": 0.945 + }, + { + "category_id": 4, + "poly": [ + 297, + 1321, + 1400, + 1321, + 1400, + 1386, + 297, + 1386 + ], + "score": 0.93 + }, + { + "category_id": 0, + "poly": [ + 300, + 750, + 715, + 750, + 715, + 783, + 300, + 783 + ], + "score": 0.916 + }, + { + "category_id": 0, + "poly": [ + 301, + 1454, + 590, + 1454, + 590, + 1486, + 301, + 1486 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 294, + 1729, + 1337, + 1729, + 1337, + 1763, + 294, + 1763 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 104, + 299, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.848 + }, + { + "category_id": 1, + "poly": [ + 300, + 1805, + 711, + 1805, + 711, + 1842, + 300, + 1842 + ], + "score": 0.815 + }, + { + "category_id": 0, + "poly": [ + 301, + 1672, + 646, + 1672, + 646, + 1704, + 301, + 1704 + ], + "score": 0.685 + }, + { + "category_id": 1, + "poly": [ + 301, + 1672, + 646, + 1672, + 646, + 1704, + 301, + 1704 + ], + "score": 0.233 + }, + { + "category_id": 13, + "poly": [ + 410, + 1354, + 440, + 1354, + 440, + 1386, + 410, + 1386 + ], + "score": 0.88, + "latex": "\\beta _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 424, + 837, + 445, + 837, + 445, + 867, + 424, + 867 + ], + "score": 0.83, + "latex": "\\beta" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 239.0, + 651.0, + 239.0, + 651.0, + 260.0, + 616.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 281.0, + 650.0, + 281.0, + 650.0, + 300.0, + 622.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 322.0, + 649.0, + 322.0, + 649.0, + 342.0, + 622.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 344.0, + 791.0, + 344.0, + 791.0, + 363.0, + 764.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 349.0, + 622.0, + 349.0, + 622.0, + 458.0, + 599.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 363.0, + 649.0, + 363.0, + 649.0, + 383.0, + 622.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 377.0, + 791.0, + 377.0, + 791.0, + 397.0, + 768.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 407.0, + 643.0, + 407.0, + 643.0, + 423.0, + 623.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 410.0, + 791.0, + 410.0, + 791.0, + 430.0, + 768.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 406.0, + 876.0, + 406.0, + 876.0, + 426.0, + 794.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 447.0, + 650.0, + 447.0, + 650.0, + 466.0, + 622.0, + 466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 443.0, + 809.0, + 443.0, + 809.0, + 489.0, + 685.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 448.0, + 864.0, + 448.0, + 864.0, + 481.0, + 829.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 456.0, + 913.0, + 456.0, + 913.0, + 479.0, + 883.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 448.0, + 966.0, + 448.0, + 966.0, + 481.0, + 931.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 454.0, + 1015.0, + 454.0, + 1015.0, + 479.0, + 985.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 456.0, + 1066.0, + 456.0, + 1066.0, + 479.0, + 1035.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 487.0, + 652.0, + 487.0, + 652.0, + 509.0, + 620.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 483.0, + 798.0, + 483.0, + 798.0, + 506.0, + 686.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 501.0, + 797.0, + 501.0, + 797.0, + 521.0, + 686.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 518.0, + 797.0, + 518.0, + 797.0, + 538.0, + 687.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 529.0, + 651.0, + 529.0, + 651.0, + 549.0, + 622.0, + 549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 534.0, + 797.0, + 534.0, + 797.0, + 554.0, + 687.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 566.0, + 675.0, + 566.0, + 675.0, + 584.0, + 659.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 563.0, + 747.0, + 563.0, + 747.0, + 586.0, + 717.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 563.0, + 811.0, + 563.0, + 811.0, + 587.0, + 780.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 563.0, + 874.0, + 563.0, + 874.0, + 586.0, + 843.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 565.0, + 937.0, + 565.0, + 937.0, + 585.0, + 909.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 565.0, + 1001.0, + 565.0, + 1001.0, + 585.0, + 972.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 565.0, + 1065.0, + 565.0, + 1065.0, + 585.0, + 1036.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 581.0, + 881.0, + 581.0, + 881.0, + 603.0, + 837.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 910.0, + 953.0, + 910.0, + 953.0, + 942.0, + 818.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 963.0, + 772.0, + 963.0, + 772.0, + 988.0, + 762.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 950.0, + 1008.0, + 950.0, + 1008.0, + 984.0, + 942.0, + 984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 986.0, + 756.0, + 986.0, + 756.0, + 1021.0, + 713.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 978.0, + 1008.0, + 978.0, + 1008.0, + 1014.0, + 943.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1008.0, + 1007.0, + 1008.0, + 1007.0, + 1041.0, + 939.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1048.0, + 756.0, + 1048.0, + 756.0, + 1088.0, + 674.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 1035.0, + 1008.0, + 1035.0, + 1008.0, + 1100.0, + 941.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1121.0, + 760.0, + 1121.0, + 760.0, + 1154.0, + 712.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1188.0, + 764.0, + 1188.0, + 764.0, + 1220.0, + 713.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 758.0, + 1215.0, + 777.0, + 1215.0, + 777.0, + 1240.0, + 758.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1212.0, + 882.0, + 1212.0, + 882.0, + 1243.0, + 846.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 1212.0, + 978.0, + 1212.0, + 978.0, + 1243.0, + 943.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 852.0, + 1238.0, + 914.0, + 1238.0, + 914.0, + 1266.0, + 852.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 669.0, + 294.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 663.0, + 788.0, + 663.0, + 788.0, + 699.0, + 295.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1320.0, + 1404.0, + 1320.0, + 1404.0, + 1356.0, + 293.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1351.0, + 409.0, + 1351.0, + 409.0, + 1388.0, + 293.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 441.0, + 1351.0, + 453.0, + 1351.0, + 453.0, + 1388.0, + 441.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 749.0, + 717.0, + 749.0, + 717.0, + 785.0, + 296.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1454.0, + 593.0, + 1454.0, + 593.0, + 1490.0, + 297.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1674.0, + 647.0, + 1674.0, + 647.0, + 1707.0, + 298.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1511.0, + 1404.0, + 1511.0, + 1404.0, + 1547.0, + 293.0, + 1547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1540.0, + 1405.0, + 1540.0, + 1405.0, + 1577.0, + 293.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1572.0, + 1405.0, + 1572.0, + 1405.0, + 1607.0, + 293.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 836.0, + 1605.0, + 836.0, + 1636.0, + 294.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 841.0, + 294.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 834.0, + 423.0, + 834.0, + 423.0, + 871.0, + 294.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 834.0, + 690.0, + 834.0, + 690.0, + 871.0, + 446.0, + 871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1729.0, + 1339.0, + 1729.0, + 1339.0, + 1765.0, + 296.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1801.0, + 714.0, + 1801.0, + 714.0, + 1848.0, + 292.0, + 1848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1674.0, + 647.0, + 1674.0, + 647.0, + 1707.0, + 298.0, + 1707.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 299, + 1574, + 1394, + 1574, + 1394, + 2025, + 299, + 2025 + ], + "score": 0.982, + "html": "
MethodSearch SpaceDatasetFiguresTables
Main textSupp.Main textSupp.
DARTSS0CIFAR-107
DARTSS3CIFAR-103
DARTS-S0CIFAR-104,5,9,71,2
DARTS-S1-S4CIFAR-1034,6,9,87,96
DARTS-S5ImageNet5,102
DARTS-S5MS COCO11
DARTS-S6CIFAR-103
DARTS-S1-S4CIFAR-1004,1179
DARTS-S0CIFAR-1004,111,10
P-DARTS w/o M=2S0CIFAR-10124
P-DARTS w/ auxiliary skipS0CIFAR-10134
PC-DARTS w/o channel shufflingS0CIFAR-10145
PC-DARTS w/ auxiliary skipS0CIFAR-10155
" + }, + { + "category_id": 3, + "poly": [ + 322, + 875, + 1384, + 875, + 1384, + 1360, + 322, + 1360 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 332, + 215, + 1383, + 215, + 1383, + 701, + 332, + 701 + ], + "score": 0.974 + }, + { + "category_id": 4, + "poly": [ + 297, + 1392, + 1404, + 1392, + 1404, + 1488, + 297, + 1488 + ], + "score": 0.953 + }, + { + "category_id": 4, + "poly": [ + 297, + 736, + 1404, + 736, + 1404, + 859, + 297, + 859 + ], + "score": 0.937 + }, + { + "category_id": 6, + "poly": [ + 551, + 1536, + 1150, + 1536, + 1150, + 1569, + 551, + 1569 + ], + "score": 0.918 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.853 + }, + { + "category_id": 13, + "poly": [ + 527, + 829, + 618, + 829, + 618, + 858, + 527, + 858 + ], + "score": 0.79, + "latex": "9 7 . 4 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 886, + 829, + 977, + 829, + 977, + 859, + 886, + 859 + ], + "score": 0.77, + "latex": "9 7 . 2 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 429, + 829, + 516, + 829, + 516, + 859, + 429, + 859 + ], + "score": 0.77, + "latex": "9 7 . 5 0 \\%" + }, + { + "category_id": 13, + "poly": [ + 784, + 829, + 876, + 829, + 876, + 859, + 784, + 859 + ], + "score": 0.71, + "latex": "( 9 7 . 1 9 \\%" + }, + { + "category_id": 13, + "poly": [ + 648, + 773, + 670, + 773, + 670, + 795, + 648, + 795 + ], + "score": 0.69, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 639, + 1430, + 661, + 1430, + 661, + 1452, + 639, + 1452 + ], + "score": 0.67, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 403, + 1456, + 457, + 1456, + 457, + 1487, + 403, + 1487 + ], + "score": 0.53, + "latex": "^ { ( \\mathrm { g , h } ) }" + }, + { + "category_id": 13, + "poly": [ + 439, + 800, + 493, + 800, + 493, + 829, + 439, + 829 + ], + "score": 0.26, + "latex": "^ { ( \\mathrm { g , h } ) }" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 898.0, + 884.0, + 898.0, + 884.0, + 908.0, + 870.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 897.0, + 1160.0, + 897.0, + 1160.0, + 909.0, + 1141.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 914.0, + 566.0, + 914.0, + 566.0, + 954.0, + 549.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 918.0, + 887.0, + 918.0, + 887.0, + 930.0, + 868.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 918.0, + 1160.0, + 918.0, + 1160.0, + 930.0, + 1141.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 942.0, + 339.0, + 942.0, + 339.0, + 952.0, + 328.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 937.0, + 887.0, + 937.0, + 887.0, + 949.0, + 868.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 937.0, + 1160.0, + 937.0, + 1160.0, + 949.0, + 1141.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 953.0, + 339.0, + 953.0, + 339.0, + 964.0, + 328.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 947.0, + 566.0, + 947.0, + 566.0, + 979.0, + 549.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 957.0, + 888.0, + 957.0, + 888.0, + 970.0, + 868.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1140.0, + 957.0, + 1162.0, + 957.0, + 1162.0, + 970.0, + 1140.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 964.0, + 340.0, + 964.0, + 340.0, + 976.0, + 329.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 965.0, + 837.0, + 965.0, + 837.0, + 975.0, + 825.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 975.0, + 340.0, + 975.0, + 340.0, + 987.0, + 329.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 971.0, + 566.0, + 971.0, + 566.0, + 1001.0, + 549.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 975.0, + 838.0, + 975.0, + 838.0, + 992.0, + 824.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 977.0, + 887.0, + 977.0, + 887.0, + 990.0, + 865.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 977.0, + 1160.0, + 977.0, + 1160.0, + 990.0, + 1137.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 984.0, + 342.0, + 984.0, + 342.0, + 1010.0, + 323.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 988.0, + 837.0, + 988.0, + 837.0, + 998.0, + 825.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 996.0, + 504.0, + 996.0, + 504.0, + 1010.0, + 487.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 997.0, + 888.0, + 997.0, + 888.0, + 1009.0, + 865.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 997.0, + 1162.0, + 997.0, + 1162.0, + 1010.0, + 1137.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1009.0, + 341.0, + 1009.0, + 341.0, + 1018.0, + 331.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1018.0, + 888.0, + 1018.0, + 888.0, + 1030.0, + 864.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1018.0, + 1160.0, + 1018.0, + 1160.0, + 1030.0, + 1137.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1037.0, + 887.0, + 1037.0, + 887.0, + 1049.0, + 865.0, + 1049.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1042.0, + 1060.0, + 1042.0, + 1060.0, + 1058.0, + 878.0, + 1058.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 1043.0, + 1087.0, + 1043.0, + 1087.0, + 1057.0, + 1064.0, + 1057.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1093.0, + 1046.0, + 1111.0, + 1046.0, + 1111.0, + 1054.0, + 1093.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1043.0, + 1336.0, + 1043.0, + 1336.0, + 1055.0, + 1281.0, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1342.0, + 1044.0, + 1359.0, + 1044.0, + 1359.0, + 1053.0, + 1342.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1067.0, + 500.0, + 1067.0, + 500.0, + 1094.0, + 377.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1065.0, + 776.0, + 1065.0, + 776.0, + 1094.0, + 648.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1064.0, + 1051.0, + 1064.0, + 1051.0, + 1095.0, + 920.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 1068.0, + 1324.0, + 1068.0, + 1324.0, + 1096.0, + 1198.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1160.0, + 882.0, + 1160.0, + 882.0, + 1169.0, + 865.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 1160.0, + 1157.0, + 1160.0, + 1157.0, + 1169.0, + 1139.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1180.0, + 881.0, + 1180.0, + 881.0, + 1189.0, + 865.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1178.0, + 1157.0, + 1178.0, + 1157.0, + 1191.0, + 1137.0, + 1191.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1194.0, + 337.0, + 1194.0, + 337.0, + 1203.0, + 326.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 1185.0, + 564.0, + 1185.0, + 564.0, + 1220.0, + 542.0, + 1220.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1186.0, + 612.0, + 1186.0, + 612.0, + 1194.0, + 599.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1189.0, + 832.0, + 1189.0, + 832.0, + 1198.0, + 820.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1199.0, + 612.0, + 1199.0, + 612.0, + 1208.0, + 599.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1203.0, + 834.0, + 1203.0, + 834.0, + 1211.0, + 820.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1198.0, + 884.0, + 1198.0, + 884.0, + 1211.0, + 864.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1198.0, + 1158.0, + 1198.0, + 1158.0, + 1211.0, + 1137.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1214.0, + 339.0, + 1214.0, + 339.0, + 1224.0, + 326.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1209.0, + 565.0, + 1209.0, + 565.0, + 1264.0, + 541.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1213.0, + 613.0, + 1213.0, + 613.0, + 1222.0, + 599.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1217.0, + 884.0, + 1217.0, + 884.0, + 1231.0, + 864.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1219.0, + 1158.0, + 1219.0, + 1158.0, + 1231.0, + 1137.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1224.0, + 615.0, + 1224.0, + 615.0, + 1237.0, + 598.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1224.0, + 1296.0, + 1224.0, + 1296.0, + 1233.0, + 1283.0, + 1233.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1228.0, + 834.0, + 1228.0, + 834.0, + 1237.0, + 822.0, + 1237.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 1234.0, + 339.0, + 1234.0, + 339.0, + 1244.0, + 328.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1237.0, + 615.0, + 1237.0, + 615.0, + 1250.0, + 599.0, + 1250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1242.0, + 832.0, + 1242.0, + 832.0, + 1252.0, + 819.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1238.0, + 884.0, + 1238.0, + 884.0, + 1252.0, + 860.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1238.0, + 1157.0, + 1238.0, + 1157.0, + 1252.0, + 1134.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 1252.0, + 613.0, + 1252.0, + 613.0, + 1261.0, + 601.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1255.0, + 834.0, + 1255.0, + 834.0, + 1264.0, + 822.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 1258.0, + 883.0, + 1258.0, + 883.0, + 1271.0, + 861.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1260.0, + 1156.0, + 1260.0, + 1156.0, + 1269.0, + 1136.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 1266.0, + 613.0, + 1266.0, + 613.0, + 1275.0, + 602.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1278.0, + 883.0, + 1278.0, + 883.0, + 1292.0, + 860.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1280.0, + 1156.0, + 1280.0, + 1156.0, + 1288.0, + 1136.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1298.0, + 1157.0, + 1298.0, + 1157.0, + 1310.0, + 1134.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 1304.0, + 1004.0, + 1304.0, + 1004.0, + 1316.0, + 927.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1304.0, + 1030.0, + 1304.0, + 1030.0, + 1316.0, + 1006.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 1304.0, + 1082.0, + 1304.0, + 1082.0, + 1316.0, + 1035.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 1306.0, + 1107.0, + 1306.0, + 1107.0, + 1315.0, + 1088.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1306.0, + 1170.0, + 1306.0, + 1170.0, + 1315.0, + 1156.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 1304.0, + 1198.0, + 1304.0, + 1198.0, + 1316.0, + 1174.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1304.0, + 1225.0, + 1304.0, + 1225.0, + 1316.0, + 1203.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1304.0, + 1251.0, + 1304.0, + 1251.0, + 1316.0, + 1227.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1256.0, + 1304.0, + 1277.0, + 1304.0, + 1277.0, + 1316.0, + 1256.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1304.0, + 1359.0, + 1304.0, + 1359.0, + 1316.0, + 1281.0, + 1316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 1306.0, + 1380.0, + 1306.0, + 1380.0, + 1315.0, + 1364.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1326.0, + 501.0, + 1326.0, + 501.0, + 1357.0, + 367.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1326.0, + 774.0, + 1326.0, + 774.0, + 1357.0, + 642.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1328.0, + 1052.0, + 1328.0, + 1052.0, + 1357.0, + 913.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1327.0, + 1326.0, + 1327.0, + 1326.0, + 1359.0, + 1189.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.75, + 1020.0, + 469.75, + 1020.0, + 469.75, + 1055.0, + 335.75, + 1055.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1027.5, + 716.0, + 1027.5, + 716.0, + 1047.5, + 611.0, + 1047.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.25, + 1035.0, + 1285.25, + 1035.0, + 1285.25, + 1063.0, + 1132.25, + 1063.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1042.0, + 736.0, + 1042.0, + 736.0, + 1054.0, + 713.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1287.0, + 451.0, + 1287.0, + 451.0, + 1304.0, + 341.0, + 1304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1287.0, + 725.0, + 1287.0, + 725.0, + 1305.0, + 614.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 445.0, + 1275.0, + 506.0, + 1275.0, + 506.0, + 1289.5, + 445.0, + 1289.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 1291.0, + 749.0, + 1291.0, + 749.0, + 1300.0, + 722.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.75, + 1297.5, + 926.75, + 1297.5, + 926.75, + 1320.5, + 857.75, + 1320.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 241.0, + 1157.0, + 241.0, + 1157.0, + 250.0, + 1143.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 301.0, + 563.0, + 301.0, + 563.0, + 309.0, + 552.0, + 309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 302.0, + 885.0, + 302.0, + 885.0, + 311.0, + 869.0, + 311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 324.0, + 483.0, + 324.0, + 483.0, + 338.0, + 468.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 325.0, + 564.0, + 325.0, + 564.0, + 335.0, + 552.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 320.0, + 885.0, + 320.0, + 885.0, + 332.0, + 863.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 330.0, + 836.0, + 330.0, + 836.0, + 340.0, + 825.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 340.0, + 887.0, + 340.0, + 887.0, + 353.0, + 863.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 360.0, + 887.0, + 360.0, + 887.0, + 373.0, + 863.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 363.0, + 1155.0, + 363.0, + 1155.0, + 371.0, + 1143.0, + 371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 380.0, + 739.0, + 380.0, + 739.0, + 398.0, + 712.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 381.0, + 983.0, + 381.0, + 983.0, + 399.0, + 863.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 387.0, + 1006.0, + 387.0, + 1006.0, + 399.0, + 985.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 387.0, + 1031.0, + 387.0, + 1031.0, + 399.0, + 1011.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1039.0, + 385.0, + 1112.0, + 385.0, + 1112.0, + 401.0, + 1039.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 382.0, + 1200.0, + 382.0, + 1200.0, + 399.0, + 1143.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 385.0, + 1282.0, + 385.0, + 1282.0, + 401.0, + 1204.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1286.0, + 387.0, + 1308.0, + 387.0, + 1308.0, + 399.0, + 1286.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1311.0, + 387.0, + 1335.0, + 387.0, + 1335.0, + 399.0, + 1311.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 388.0, + 1359.0, + 388.0, + 1359.0, + 397.0, + 1339.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 409.0, + 502.0, + 409.0, + 502.0, + 438.0, + 375.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 409.0, + 776.0, + 409.0, + 776.0, + 438.0, + 649.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 409.0, + 1049.0, + 409.0, + 1049.0, + 441.0, + 925.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1198.0, + 409.0, + 1324.0, + 409.0, + 1324.0, + 441.0, + 1198.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 485.0, + 880.0, + 485.0, + 880.0, + 493.0, + 865.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 485.0, + 1156.0, + 485.0, + 1156.0, + 493.0, + 1138.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 504.0, + 879.0, + 504.0, + 879.0, + 513.0, + 865.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 502.0, + 1160.0, + 502.0, + 1160.0, + 515.0, + 1137.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 523.0, + 884.0, + 523.0, + 884.0, + 535.0, + 863.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 523.0, + 1158.0, + 523.0, + 1158.0, + 535.0, + 1137.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 542.0, + 883.0, + 542.0, + 883.0, + 554.0, + 863.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 542.0, + 1160.0, + 542.0, + 1160.0, + 554.0, + 1138.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 563.0, + 884.0, + 563.0, + 884.0, + 575.0, + 863.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 564.0, + 1156.0, + 564.0, + 1156.0, + 573.0, + 1139.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 583.0, + 560.0, + 583.0, + 560.0, + 592.0, + 548.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 583.0, + 884.0, + 583.0, + 884.0, + 595.0, + 862.0, + 595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 584.0, + 1155.0, + 584.0, + 1155.0, + 592.0, + 1137.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 605.0, + 879.0, + 605.0, + 879.0, + 613.0, + 863.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 624.0, + 881.0, + 624.0, + 881.0, + 633.0, + 866.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 642.0, + 951.0, + 642.0, + 951.0, + 663.0, + 862.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 651.0, + 975.0, + 651.0, + 975.0, + 659.0, + 952.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 648.0, + 1001.0, + 648.0, + 1001.0, + 661.0, + 980.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 651.0, + 1054.0, + 651.0, + 1054.0, + 659.0, + 1037.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 650.0, + 1081.0, + 650.0, + 1081.0, + 659.0, + 1065.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 651.0, + 1106.0, + 651.0, + 1106.0, + 659.0, + 1091.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1138.0, + 644.0, + 1152.0, + 644.0, + 1152.0, + 652.0, + 1138.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 651.0, + 1196.0, + 651.0, + 1196.0, + 659.0, + 1176.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 648.0, + 1226.0, + 648.0, + 1226.0, + 661.0, + 1202.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 651.0, + 1248.0, + 651.0, + 1248.0, + 659.0, + 1229.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1258.0, + 651.0, + 1274.0, + 651.0, + 1274.0, + 659.0, + 1258.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 650.0, + 1300.0, + 650.0, + 1300.0, + 659.0, + 1284.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 648.0, + 1330.0, + 648.0, + 1330.0, + 661.0, + 1310.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1335.0, + 648.0, + 1356.0, + 648.0, + 1356.0, + 661.0, + 1335.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 651.0, + 1380.0, + 651.0, + 1380.0, + 661.0, + 1365.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 671.0, + 500.0, + 671.0, + 500.0, + 698.0, + 368.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 671.0, + 773.0, + 671.0, + 773.0, + 698.0, + 643.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 671.0, + 1051.0, + 671.0, + 1051.0, + 700.0, + 915.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 670.0, + 1324.0, + 670.0, + 1324.0, + 698.0, + 1189.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.75, + 371.0, + 714.75, + 371.0, + 714.75, + 390.0, + 613.75, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 462.0, + 356.5, + 500.0, + 356.5, + 500.0, + 368.5, + 462.0, + 368.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 335.25, + 632.0, + 437.25, + 632.0, + 437.25, + 651.0, + 335.25, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.25, + 632.0, + 710.25, + 632.0, + 710.25, + 651.0, + 608.25, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 641.5, + 465.0, + 641.5, + 465.0, + 659.0, + 432.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 616.0, + 506.0, + 616.0, + 506.0, + 630.0, + 452.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 624.5, + 761.0, + 624.5, + 761.0, + 633.5, + 733.0, + 633.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 645.0, + 733.0, + 645.0, + 733.0, + 658.0, + 708.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1391.0, + 1405.0, + 1391.0, + 1405.0, + 1429.0, + 295.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 638.0, + 1423.0, + 638.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1423.0, + 1404.0, + 1423.0, + 1404.0, + 1460.0, + 662.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1454.0, + 402.0, + 1454.0, + 402.0, + 1489.0, + 293.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 1454.0, + 823.0, + 1454.0, + 823.0, + 1489.0, + 458.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 738.0, + 1404.0, + 738.0, + 1404.0, + 771.0, + 295.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 767.0, + 647.0, + 767.0, + 647.0, + 801.0, + 294.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 767.0, + 1404.0, + 767.0, + 1404.0, + 801.0, + 671.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 800.0, + 438.0, + 800.0, + 438.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 800.0, + 1404.0, + 800.0, + 1404.0, + 832.0, + 494.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 828.0, + 428.0, + 828.0, + 428.0, + 863.0, + 292.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 828.0, + 526.0, + 828.0, + 526.0, + 863.0, + 517.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 828.0, + 783.0, + 828.0, + 783.0, + 863.0, + 619.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 828.0, + 885.0, + 828.0, + 885.0, + 863.0, + 877.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 828.0, + 996.0, + 828.0, + 996.0, + 863.0, + 978.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 549.0, + 1532.0, + 1150.0, + 1532.0, + 1150.0, + 1574.0, + 549.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 321, + 357, + 1382, + 357, + 1382, + 1205, + 321, + 1205 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 466, + 1854, + 1234, + 1854, + 1234, + 1889, + 466, + 1889 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 290, + 1239, + 1399, + 1239, + 1399, + 1274, + 290, + 1274 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2112, + 836, + 2112 + ], + "score": 0.849 + }, + { + "category_id": 3, + "poly": [ + 298, + 1330, + 1407, + 1330, + 1407, + 1782, + 298, + 1782 + ], + "score": 0.494 + }, + { + "category_id": 3, + "poly": [ + 302, + 1610, + 1405, + 1610, + 1405, + 1776, + 302, + 1776 + ], + "score": 0.229 + }, + { + "category_id": 13, + "poly": [ + 500, + 392, + 522, + 392, + 522, + 407, + 500, + 407 + ], + "score": 0.67, + "latex": "c ^ { \\kappa - 2 }" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 384.0, + 577.0, + 384.0, + 577.0, + 409.0, + 536.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 385.0, + 499.0, + 385.0, + 499.0, + 413.0, + 492.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 385.0, + 529.0, + 385.0, + 529.0, + 413.0, + 523.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 418.0, + 579.0, + 418.0, + 579.0, + 445.0, + 538.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 421.0, + 617.0, + 421.0, + 617.0, + 453.0, + 585.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 425.0, + 868.0, + 425.0, + 868.0, + 452.0, + 826.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 436.0, + 920.0, + 436.0, + 920.0, + 461.0, + 874.0, + 461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 420.0, + 975.0, + 420.0, + 975.0, + 442.0, + 926.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 431.0, + 1064.0, + 431.0, + 1064.0, + 463.0, + 1024.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 494.0, + 511.0, + 529.0, + 511.0, + 529.0, + 540.0, + 494.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 454.0, + 578.0, + 454.0, + 578.0, + 547.0, + 535.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 464.0, + 618.0, + 464.0, + 618.0, + 540.0, + 588.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 518.0, + 663.0, + 518.0, + 663.0, + 544.0, + 624.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 532.0, + 702.0, + 532.0, + 702.0, + 561.0, + 670.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 477.0, + 763.0, + 477.0, + 763.0, + 512.0, + 721.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 482.0, + 867.0, + 482.0, + 867.0, + 511.0, + 831.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 476.0, + 940.0, + 476.0, + 940.0, + 506.0, + 876.0, + 506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 484.0, + 965.0, + 484.0, + 965.0, + 511.0, + 941.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 470.0, + 1062.0, + 470.0, + 1062.0, + 543.0, + 979.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 448.0, + 1147.0, + 448.0, + 1147.0, + 559.0, + 1064.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 495.0, + 1207.0, + 495.0, + 1207.0, + 528.0, + 1165.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 606.0, + 887.0, + 606.0, + 887.0, + 636.0, + 811.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 693.0, + 353.0, + 693.0, + 353.0, + 776.0, + 328.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 690.0, + 382.0, + 690.0, + 382.0, + 760.0, + 356.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 707.0, + 406.0, + 707.0, + 406.0, + 758.0, + 386.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 740.0, + 431.0, + 740.0, + 431.0, + 755.0, + 411.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 484.0, + 764.0, + 512.0, + 764.0, + 512.0, + 788.0, + 484.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 733.0, + 547.0, + 733.0, + 547.0, + 759.0, + 519.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 701.0, + 621.0, + 701.0, + 621.0, + 723.0, + 591.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 682.0, + 762.0, + 682.0, + 762.0, + 782.0, + 622.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 764.0, + 735.0, + 806.0, + 735.0, + 806.0, + 767.0, + 764.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 732.0, + 885.0, + 732.0, + 885.0, + 750.0, + 862.0, + 750.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 709.0, + 916.0, + 709.0, + 916.0, + 723.0, + 893.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 702.0, + 945.0, + 702.0, + 945.0, + 744.0, + 922.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 732.0, + 1027.0, + 732.0, + 1027.0, + 776.0, + 947.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 753.0, + 1055.0, + 753.0, + 1055.0, + 777.0, + 1028.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 753.0, + 1145.0, + 753.0, + 1145.0, + 764.0, + 1133.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 766.0, + 1174.0, + 766.0, + 1174.0, + 776.0, + 1161.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 730.0, + 1228.0, + 730.0, + 1228.0, + 779.0, + 1206.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 751.0, + 1251.0, + 751.0, + 1251.0, + 772.0, + 1229.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 733.0, + 1318.0, + 733.0, + 1318.0, + 765.0, + 1254.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 759.0, + 1374.0, + 759.0, + 1374.0, + 780.0, + 1352.0, + 780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 781.0, + 406.0, + 781.0, + 406.0, + 795.0, + 383.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 779.0, + 487.0, + 779.0, + 487.0, + 813.0, + 433.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 782.0, + 619.0, + 782.0, + 619.0, + 803.0, + 589.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 777.0, + 696.0, + 777.0, + 696.0, + 799.0, + 622.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 770.0, + 758.0, + 770.0, + 758.0, + 803.0, + 702.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 776.0, + 884.0, + 776.0, + 884.0, + 796.0, + 861.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 771.0, + 914.0, + 771.0, + 914.0, + 785.0, + 894.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 770.0, + 945.0, + 770.0, + 945.0, + 796.0, + 919.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 785.0, + 1157.0, + 785.0, + 1157.0, + 807.0, + 1128.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 786.0, + 1173.0, + 786.0, + 1173.0, + 795.0, + 1159.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 783.0, + 1202.0, + 783.0, + 1202.0, + 803.0, + 1179.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 772.0, + 1274.0, + 772.0, + 1274.0, + 791.0, + 1252.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1318.0, + 774.0, + 1341.0, + 774.0, + 1341.0, + 793.0, + 1318.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 796.0, + 910.0, + 796.0, + 910.0, + 802.0, + 904.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 832.0, + 605.0, + 832.0, + 605.0, + 869.0, + 527.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 833.0, + 1154.0, + 833.0, + 1154.0, + 867.0, + 1076.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 920.0, + 437.0, + 920.0, + 437.0, + 942.0, + 400.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 944.0, + 434.0, + 944.0, + 434.0, + 962.0, + 406.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 943.0, + 1243.0, + 943.0, + 1243.0, + 964.0, + 1208.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 977.0, + 397.0, + 977.0, + 397.0, + 1012.0, + 358.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 992.0, + 438.0, + 992.0, + 438.0, + 1019.0, + 399.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 975.0, + 445.0, + 975.0, + 445.0, + 981.0, + 439.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 448.0, + 959.0, + 471.0, + 959.0, + 471.0, + 1019.0, + 448.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 991.0, + 912.0, + 991.0, + 912.0, + 1014.0, + 882.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 965.0, + 988.0, + 965.0, + 988.0, + 1014.0, + 918.0, + 1014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 988.0, + 1028.0, + 988.0, + 1028.0, + 1009.0, + 989.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 996.0, + 1062.0, + 996.0, + 1062.0, + 1025.0, + 1032.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 977.0, + 1243.0, + 977.0, + 1243.0, + 1017.0, + 1164.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 953.0, + 1282.0, + 953.0, + 1282.0, + 1029.0, + 1244.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1014.0, + 530.0, + 1014.0, + 530.0, + 1044.0, + 498.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1002.0, + 953.0, + 1002.0, + 953.0, + 1024.0, + 917.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 1009.0, + 1112.0, + 1009.0, + 1112.0, + 1043.0, + 1069.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1301.0, + 1014.0, + 1333.0, + 1014.0, + 1333.0, + 1044.0, + 1301.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1028.0, + 437.0, + 1028.0, + 437.0, + 1052.0, + 400.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 1035.0, + 471.0, + 1035.0, + 471.0, + 1062.0, + 446.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 1030.0, + 500.0, + 1030.0, + 500.0, + 1040.0, + 491.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1032.0, + 954.0, + 1032.0, + 954.0, + 1054.0, + 918.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1027.0, + 986.0, + 1027.0, + 986.0, + 1056.0, + 956.0, + 1056.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1027.0, + 1246.0, + 1027.0, + 1246.0, + 1053.0, + 1207.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1030.0, + 1280.0, + 1030.0, + 1280.0, + 1065.0, + 1251.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1030.0, + 1304.0, + 1030.0, + 1304.0, + 1041.0, + 1294.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1048.0, + 394.0, + 1048.0, + 394.0, + 1073.0, + 357.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 1059.0, + 437.0, + 1059.0, + 437.0, + 1083.0, + 397.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 1074.0, + 472.0, + 1074.0, + 472.0, + 1104.0, + 443.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 581.0, + 1055.0, + 604.0, + 1055.0, + 604.0, + 1076.0, + 581.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1062.0, + 639.0, + 1062.0, + 639.0, + 1081.0, + 610.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1051.0, + 668.0, + 1051.0, + 668.0, + 1069.0, + 641.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1061.0, + 723.0, + 1061.0, + 723.0, + 1078.0, + 698.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 1074.0, + 750.0, + 1074.0, + 750.0, + 1091.0, + 726.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1067.0, + 779.0, + 1067.0, + 779.0, + 1092.0, + 754.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1062.0, + 950.0, + 1062.0, + 950.0, + 1082.0, + 916.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1062.0, + 986.0, + 1062.0, + 986.0, + 1095.0, + 956.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 1048.0, + 1200.0, + 1048.0, + 1200.0, + 1077.0, + 1165.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1058.0, + 1245.0, + 1058.0, + 1245.0, + 1083.0, + 1206.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1077.0, + 1272.0, + 1077.0, + 1272.0, + 1097.0, + 1253.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 1092.0, + 436.0, + 1092.0, + 436.0, + 1130.0, + 400.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.0, + 1090.0, + 608.0, + 1090.0, + 608.0, + 1113.0, + 579.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1087.0, + 639.0, + 1087.0, + 639.0, + 1104.0, + 610.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1090.0, + 665.0, + 1090.0, + 665.0, + 1114.0, + 640.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1087.0, + 723.0, + 1087.0, + 723.0, + 1129.0, + 668.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 725.0, + 1108.0, + 751.0, + 1108.0, + 751.0, + 1125.0, + 725.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1117.0, + 779.0, + 1117.0, + 779.0, + 1141.0, + 754.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1098.0, + 818.0, + 1098.0, + 818.0, + 1123.0, + 792.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1096.0, + 910.0, + 1096.0, + 910.0, + 1116.0, + 884.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1122.0, + 951.0, + 1122.0, + 951.0, + 1139.0, + 919.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 1091.0, + 1242.0, + 1091.0, + 1242.0, + 1132.0, + 1207.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1174.0, + 622.0, + 1174.0, + 622.0, + 1204.0, + 545.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 1174.0, + 1144.0, + 1174.0, + 1144.0, + 1204.0, + 1070.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 356.0, + 580.0, + 356.0, + 580.0, + 383.0, + 533.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 483.5, + 589.0, + 483.5, + 589.0, + 518.0, + 537.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 490.5, + 1024.0, + 490.5, + 1024.0, + 501.0, + 1019.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 579.25, + 542.0, + 624.25, + 542.0, + 624.25, + 568.0, + 579.25, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 733.5, + 1095.0, + 733.5, + 1095.0, + 762.0, + 1053.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 766.0, + 424.0, + 766.0, + 424.0, + 775.0, + 418.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 918.5, + 1244.0, + 918.5, + 1244.0, + 942.5, + 1205.0, + 942.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 961.5, + 444.0, + 961.5, + 444.0, + 988.0, + 396.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 962.5, + 1248.0, + 962.5, + 1248.0, + 988.0, + 1203.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1088.0, + 955.0, + 1088.0, + 955.0, + 1111.0, + 915.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 1848.0, + 1236.0, + 1848.0, + 1236.0, + 1897.0, + 461.0, + 1897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1235.0, + 1405.0, + 1235.0, + 1405.0, + 1278.0, + 293.0, + 1278.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1615.0, + 331.0, + 1615.0, + 331.0, + 1696.0, + 306.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1667.0, + 363.0, + 1667.0, + 363.0, + 1720.0, + 337.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1609.0, + 397.0, + 1609.0, + 397.0, + 1692.0, + 373.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1643.0, + 433.0, + 1643.0, + 433.0, + 1740.0, + 404.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1609.0, + 465.0, + 1609.0, + 465.0, + 1693.0, + 439.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1643.0, + 502.0, + 1643.0, + 502.0, + 1740.0, + 473.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1624.0, + 535.0, + 1624.0, + 535.0, + 1731.0, + 505.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 1642.0, + 570.0, + 1642.0, + 570.0, + 1741.0, + 541.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 582.0, + 1624.0, + 607.0, + 1624.0, + 607.0, + 1696.0, + 582.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1632.0, + 642.0, + 1632.0, + 642.0, + 1747.0, + 610.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 643.0, + 1623.0, + 674.0, + 1623.0, + 674.0, + 1695.0, + 643.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1637.0, + 709.0, + 1637.0, + 709.0, + 1745.0, + 676.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1624.0, + 742.0, + 1624.0, + 742.0, + 1695.0, + 716.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 1636.0, + 778.0, + 1636.0, + 778.0, + 1745.0, + 745.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 1622.0, + 811.0, + 1622.0, + 811.0, + 1696.0, + 784.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1637.0, + 845.0, + 1637.0, + 845.0, + 1746.0, + 816.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1625.0, + 881.0, + 1625.0, + 881.0, + 1695.0, + 849.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 1643.0, + 913.0, + 1643.0, + 913.0, + 1741.0, + 884.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1623.0, + 946.0, + 1623.0, + 946.0, + 1695.0, + 921.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1640.0, + 981.0, + 1640.0, + 981.0, + 1743.0, + 952.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1623.0, + 1016.0, + 1623.0, + 1016.0, + 1694.0, + 991.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1636.0, + 1051.0, + 1636.0, + 1051.0, + 1742.0, + 1019.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 1624.0, + 1119.0, + 1624.0, + 1119.0, + 1744.0, + 1067.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 1626.0, + 1155.0, + 1626.0, + 1155.0, + 1695.0, + 1123.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1639.0, + 1187.0, + 1639.0, + 1187.0, + 1741.0, + 1159.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1630.0, + 1218.0, + 1630.0, + 1218.0, + 1693.0, + 1193.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1227.0, + 1646.0, + 1255.0, + 1646.0, + 1255.0, + 1741.0, + 1227.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1629.0, + 1293.0, + 1629.0, + 1293.0, + 1693.0, + 1268.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 1673.0, + 1326.0, + 1673.0, + 1326.0, + 1715.0, + 1300.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1331.0, + 1621.0, + 1360.0, + 1621.0, + 1360.0, + 1694.0, + 1331.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1365.0, + 1643.0, + 1390.0, + 1643.0, + 1390.0, + 1745.0, + 1365.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1615.0, + 331.0, + 1615.0, + 331.0, + 1695.0, + 306.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 1665.0, + 366.0, + 1665.0, + 366.0, + 1720.0, + 339.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1611.0, + 397.0, + 1611.0, + 397.0, + 1692.0, + 372.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 1645.0, + 432.0, + 1645.0, + 432.0, + 1739.0, + 404.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1609.0, + 465.0, + 1609.0, + 465.0, + 1693.0, + 440.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 1645.0, + 500.0, + 1645.0, + 500.0, + 1737.0, + 475.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1623.0, + 535.0, + 1623.0, + 535.0, + 1695.0, + 511.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 540.0, + 1637.0, + 573.0, + 1637.0, + 573.0, + 1739.0, + 540.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 1624.0, + 608.0, + 1624.0, + 608.0, + 1695.0, + 583.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1637.0, + 639.0, + 1637.0, + 639.0, + 1744.0, + 610.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 642.0, + 1622.0, + 674.0, + 1622.0, + 674.0, + 1694.0, + 642.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1636.0, + 707.0, + 1636.0, + 707.0, + 1740.0, + 678.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1622.0, + 739.0, + 1622.0, + 739.0, + 1694.0, + 714.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1638.0, + 776.0, + 1638.0, + 776.0, + 1745.0, + 747.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1623.0, + 812.0, + 1623.0, + 812.0, + 1693.0, + 783.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 1638.0, + 845.0, + 1638.0, + 845.0, + 1742.0, + 816.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1624.0, + 881.0, + 1624.0, + 881.0, + 1694.0, + 847.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 1641.0, + 914.0, + 1641.0, + 914.0, + 1739.0, + 885.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1623.0, + 946.0, + 1623.0, + 946.0, + 1695.0, + 921.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1644.0, + 981.0, + 1644.0, + 981.0, + 1740.0, + 952.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 1624.0, + 1015.0, + 1624.0, + 1015.0, + 1693.0, + 994.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1019.0, + 1637.0, + 1051.0, + 1637.0, + 1051.0, + 1740.0, + 1019.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1624.0, + 1088.0, + 1624.0, + 1088.0, + 1693.0, + 1068.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1640.0, + 1119.0, + 1640.0, + 1119.0, + 1752.0, + 1090.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 1628.0, + 1153.0, + 1628.0, + 1153.0, + 1694.0, + 1123.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1643.0, + 1187.0, + 1643.0, + 1187.0, + 1738.0, + 1158.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1193.0, + 1629.0, + 1218.0, + 1629.0, + 1218.0, + 1692.0, + 1193.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1645.0, + 1254.0, + 1645.0, + 1254.0, + 1738.0, + 1225.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 1628.0, + 1293.0, + 1628.0, + 1293.0, + 1692.0, + 1268.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 1673.0, + 1326.0, + 1673.0, + 1326.0, + 1714.0, + 1300.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1332.0, + 1623.0, + 1357.0, + 1623.0, + 1357.0, + 1692.0, + 1332.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 1641.0, + 1392.0, + 1641.0, + 1392.0, + 1748.0, + 1364.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1691.0, + 519.0, + 1691.0, + 519.0, + 1738.0, + 506.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1692.0, + 861.0, + 1692.0, + 861.0, + 1752.0, + 850.0, + 1752.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 321, + 249, + 1377, + 249, + 1377, + 1099, + 321, + 1099 + ], + "score": 0.972 + }, + { + "category_id": 3, + "poly": [ + 342, + 1324, + 1324, + 1324, + 1324, + 1892, + 342, + 1892 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 297, + 1130, + 1407, + 1130, + 1407, + 1195, + 297, + 1195 + ], + "score": 0.944 + }, + { + "category_id": 4, + "poly": [ + 295, + 1924, + 1401, + 1924, + 1401, + 1990, + 295, + 1990 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.893 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.859 + }, + { + "category_id": 13, + "poly": [ + 300, + 1957, + 370, + 1957, + 370, + 1987, + 300, + 1987 + ], + "score": 0.65, + "latex": "\\left( \\mathrm { M } { = } 2 \\right)" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 250.0, + 619.0, + 250.0, + 619.0, + 276.0, + 585.0, + 276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 275.0, + 619.0, + 275.0, + 619.0, + 298.0, + 588.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 290.0, + 656.0, + 290.0, + 656.0, + 328.0, + 629.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 296.0, + 623.0, + 296.0, + 623.0, + 322.0, + 586.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 538.0, + 311.0, + 623.0, + 311.0, + 623.0, + 389.0, + 538.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 314.0, + 661.0, + 314.0, + 661.0, + 402.0, + 626.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 349.0, + 712.0, + 349.0, + 712.0, + 378.0, + 673.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 325.0, + 815.0, + 325.0, + 815.0, + 356.0, + 776.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 334.0, + 872.0, + 334.0, + 872.0, + 363.0, + 820.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 315.0, + 918.0, + 315.0, + 918.0, + 345.0, + 874.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 336.0, + 1006.0, + 336.0, + 1006.0, + 362.0, + 965.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1008.0, + 346.0, + 1089.0, + 346.0, + 1089.0, + 383.0, + 1008.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 383.0, + 577.0, + 383.0, + 577.0, + 412.0, + 542.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 391.0, + 619.0, + 391.0, + 619.0, + 446.0, + 584.0, + 446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 406.0, + 655.0, + 406.0, + 655.0, + 437.0, + 627.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 382.0, + 813.0, + 382.0, + 813.0, + 414.0, + 776.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 379.0, + 868.0, + 379.0, + 868.0, + 401.0, + 824.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 871.0, + 395.0, + 882.0, + 395.0, + 882.0, + 404.0, + 871.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 384.0, + 914.0, + 384.0, + 914.0, + 412.0, + 883.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 922.0, + 375.0, + 1090.0, + 375.0, + 1090.0, + 459.0, + 922.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 395.0, + 1154.0, + 395.0, + 1154.0, + 431.0, + 1107.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 454.0, + 613.0, + 454.0, + 613.0, + 467.0, + 589.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 510.0, + 888.0, + 510.0, + 888.0, + 539.0, + 809.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 596.0, + 349.0, + 596.0, + 349.0, + 609.0, + 330.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 597.0, + 375.0, + 597.0, + 375.0, + 606.0, + 360.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 594.0, + 432.0, + 594.0, + 432.0, + 609.0, + 412.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 593.0, + 918.0, + 593.0, + 918.0, + 610.0, + 892.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 586.0, + 944.0, + 586.0, + 944.0, + 610.0, + 919.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 591.0, + 1191.0, + 591.0, + 1191.0, + 608.0, + 1167.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 585.0, + 1220.0, + 585.0, + 1220.0, + 610.0, + 1195.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 328.0, + 655.0, + 352.0, + 655.0, + 352.0, + 674.0, + 328.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 619.0, + 382.0, + 619.0, + 382.0, + 649.0, + 358.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 621.0, + 460.0, + 621.0, + 460.0, + 672.0, + 385.0, + 672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 638.0, + 548.0, + 638.0, + 548.0, + 674.0, + 464.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 648.0, + 602.0, + 648.0, + 602.0, + 661.0, + 585.0, + 661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 612.0, + 677.0, + 612.0, + 677.0, + 625.0, + 658.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 641.0, + 717.0, + 641.0, + 717.0, + 650.0, + 702.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 641.0, + 782.0, + 641.0, + 782.0, + 670.0, + 720.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 617.0, + 888.0, + 617.0, + 888.0, + 636.0, + 862.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 657.0, + 914.0, + 657.0, + 914.0, + 666.0, + 900.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 615.0, + 944.0, + 615.0, + 944.0, + 680.0, + 918.0, + 680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 621.0, + 1025.0, + 621.0, + 1025.0, + 644.0, + 974.0, + 644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 632.0, + 1055.0, + 632.0, + 1055.0, + 657.0, + 1027.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 617.0, + 1092.0, + 617.0, + 1092.0, + 643.0, + 1059.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 611.0, + 1160.0, + 611.0, + 1160.0, + 630.0, + 1134.0, + 630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 606.0, + 1249.0, + 606.0, + 1249.0, + 678.0, + 1195.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 608.0, + 1333.0, + 608.0, + 1333.0, + 643.0, + 1280.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 623.0, + 1370.0, + 623.0, + 1370.0, + 645.0, + 1350.0, + 645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 669.0, + 379.0, + 669.0, + 379.0, + 682.0, + 359.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 679.0, + 408.0, + 679.0, + 408.0, + 702.0, + 382.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 675.0, + 607.0, + 675.0, + 607.0, + 693.0, + 583.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 665.0, + 653.0, + 665.0, + 653.0, + 683.0, + 612.0, + 683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 676.0, + 674.0, + 676.0, + 674.0, + 691.0, + 656.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 661.0, + 719.0, + 661.0, + 719.0, + 675.0, + 702.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 667.0, + 810.0, + 667.0, + 810.0, + 686.0, + 785.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 679.0, + 887.0, + 679.0, + 887.0, + 698.0, + 858.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 678.0, + 915.0, + 678.0, + 915.0, + 691.0, + 894.0, + 691.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 672.0, + 971.0, + 672.0, + 971.0, + 685.0, + 948.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 676.0, + 1162.0, + 676.0, + 1162.0, + 700.0, + 1132.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 674.0, + 1217.0, + 674.0, + 1217.0, + 693.0, + 1164.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 665.0, + 1279.0, + 665.0, + 1279.0, + 689.0, + 1252.0, + 689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 687.0, + 697.0, + 687.0, + 697.0, + 706.0, + 677.0, + 706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 728.0, + 602.0, + 728.0, + 602.0, + 757.0, + 527.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 728.0, + 1153.0, + 728.0, + 1153.0, + 757.0, + 1077.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 811.0, + 437.0, + 811.0, + 437.0, + 832.0, + 405.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 811.0, + 1242.0, + 811.0, + 1242.0, + 832.0, + 1208.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 836.0, + 436.0, + 836.0, + 436.0, + 853.0, + 407.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 834.0, + 1242.0, + 834.0, + 1242.0, + 856.0, + 1210.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 858.0, + 438.0, + 858.0, + 438.0, + 879.0, + 405.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 446.0, + 849.0, + 472.0, + 849.0, + 472.0, + 875.0, + 446.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 854.0, + 1246.0, + 854.0, + 1246.0, + 881.0, + 1208.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 847.0, + 1279.0, + 847.0, + 1279.0, + 876.0, + 1248.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 359.0, + 870.0, + 396.0, + 870.0, + 396.0, + 903.0, + 359.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 877.0, + 951.0, + 877.0, + 951.0, + 897.0, + 918.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.0, + 871.0, + 1201.0, + 871.0, + 1201.0, + 900.0, + 1166.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 887.0, + 437.0, + 887.0, + 437.0, + 906.0, + 400.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 444.0, + 887.0, + 476.0, + 887.0, + 476.0, + 914.0, + 444.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 904.0, + 530.0, + 904.0, + 530.0, + 939.0, + 490.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 903.0, + 665.0, + 903.0, + 665.0, + 920.0, + 639.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 906.0, + 723.0, + 906.0, + 723.0, + 933.0, + 691.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 898.0, + 912.0, + 898.0, + 912.0, + 922.0, + 882.0, + 922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 897.0, + 951.0, + 897.0, + 951.0, + 917.0, + 916.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 887.0, + 1243.0, + 887.0, + 1243.0, + 907.0, + 1208.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 887.0, + 1280.0, + 887.0, + 1280.0, + 914.0, + 1248.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 904.0, + 1334.0, + 904.0, + 1334.0, + 938.0, + 1295.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 671.0, + 914.0, + 694.0, + 914.0, + 694.0, + 927.0, + 671.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 920.0, + 440.0, + 920.0, + 440.0, + 1003.0, + 356.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 929.0, + 470.0, + 929.0, + 470.0, + 992.0, + 443.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 578.0, + 958.0, + 605.0, + 958.0, + 605.0, + 983.0, + 578.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 612.0, + 955.0, + 664.0, + 955.0, + 664.0, + 996.0, + 612.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 954.0, + 724.0, + 954.0, + 724.0, + 1017.0, + 666.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 726.0, + 958.0, + 779.0, + 958.0, + 779.0, + 992.0, + 726.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 784.0, + 938.0, + 819.0, + 938.0, + 819.0, + 966.0, + 784.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 985.0, + 912.0, + 985.0, + 912.0, + 1009.0, + 882.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 921.0, + 986.0, + 921.0, + 986.0, + 1009.0, + 918.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 989.0, + 929.0, + 1061.0, + 929.0, + 1061.0, + 980.0, + 989.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1074.0, + 965.0, + 1109.0, + 965.0, + 1109.0, + 996.0, + 1074.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 938.0, + 1202.0, + 938.0, + 1202.0, + 970.0, + 1164.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1205.0, + 921.0, + 1245.0, + 921.0, + 1245.0, + 1003.0, + 1205.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 927.0, + 1276.0, + 927.0, + 1276.0, + 987.0, + 1250.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1009.0, + 436.0, + 1009.0, + 436.0, + 1024.0, + 403.0, + 1024.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1012.0, + 608.0, + 1012.0, + 608.0, + 1042.0, + 577.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 1004.0, + 950.0, + 1004.0, + 950.0, + 1021.0, + 921.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1005.0, + 986.0, + 1005.0, + 986.0, + 1034.0, + 955.0, + 1034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1066.0, + 622.0, + 1066.0, + 622.0, + 1095.0, + 546.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1065.0, + 1145.0, + 1065.0, + 1145.0, + 1097.0, + 1068.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1334.75, + 621.0, + 1357.75, + 621.0, + 1357.75, + 639.5, + 1334.75, + 639.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 631.0, + 976.0, + 631.0, + 976.0, + 649.0, + 946.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1008.5, + 1241.0, + 1008.5, + 1241.0, + 1025.0, + 1208.0, + 1025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 1333.0, + 1223.0, + 1333.0, + 1223.0, + 1354.0, + 1190.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 1355.0, + 1225.0, + 1355.0, + 1225.0, + 1380.0, + 1189.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1375.0, + 1262.0, + 1375.0, + 1262.0, + 1398.0, + 1236.0, + 1398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1380.0, + 1230.0, + 1380.0, + 1230.0, + 1404.0, + 1182.0, + 1404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 1418.0, + 399.0, + 1418.0, + 399.0, + 1431.0, + 381.0, + 1431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1410.0, + 424.0, + 1410.0, + 424.0, + 1432.0, + 402.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1407.0, + 634.0, + 1407.0, + 634.0, + 1425.0, + 611.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 1395.0, + 671.0, + 1395.0, + 671.0, + 1428.0, + 639.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1396.0, + 700.0, + 1396.0, + 700.0, + 1440.0, + 673.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1410.0, + 733.0, + 1410.0, + 733.0, + 1427.0, + 707.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1421.0, + 764.0, + 1421.0, + 764.0, + 1448.0, + 735.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1396.0, + 1181.0, + 1396.0, + 1181.0, + 1428.0, + 1144.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 1411.0, + 1226.0, + 1411.0, + 1226.0, + 1436.0, + 1186.0, + 1436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1414.0, + 1262.0, + 1414.0, + 1262.0, + 1439.0, + 1236.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1439.0, + 370.0, + 1439.0, + 370.0, + 1451.0, + 352.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1436.0, + 419.0, + 1436.0, + 419.0, + 1446.0, + 406.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 1438.0, + 469.0, + 1438.0, + 469.0, + 1457.0, + 449.0, + 1457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1440.0, + 552.0, + 1440.0, + 552.0, + 1463.0, + 524.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.0, + 1435.0, + 803.0, + 1435.0, + 803.0, + 1461.0, + 772.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1433.0, + 1323.0, + 1433.0, + 1323.0, + 1465.0, + 1283.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 349.0, + 1477.0, + 370.0, + 1477.0, + 370.0, + 1496.0, + 349.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1450.0, + 519.0, + 1450.0, + 519.0, + 1561.0, + 384.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1454.0, + 780.0, + 1454.0, + 780.0, + 1564.0, + 608.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1479.0, + 902.0, + 1479.0, + 902.0, + 1544.0, + 876.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 906.0, + 1480.0, + 922.0, + 1480.0, + 922.0, + 1508.0, + 906.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1480.0, + 948.0, + 1480.0, + 948.0, + 1566.0, + 928.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1485.0, + 972.0, + 1485.0, + 972.0, + 1538.0, + 952.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 1509.0, + 998.0, + 1509.0, + 998.0, + 1531.0, + 973.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1521.0, + 1046.0, + 1521.0, + 1046.0, + 1549.0, + 1001.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1509.0, + 1079.0, + 1509.0, + 1079.0, + 1533.0, + 1048.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1468.0, + 1180.0, + 1468.0, + 1180.0, + 1499.0, + 1143.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 1446.0, + 1227.0, + 1446.0, + 1227.0, + 1564.0, + 1185.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1456.0, + 1264.0, + 1456.0, + 1264.0, + 1523.0, + 1233.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 1556.0, + 999.0, + 1556.0, + 999.0, + 1579.0, + 974.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1604.0, + 1037.0, + 1604.0, + 1037.0, + 1631.0, + 916.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1602.0, + 1308.0, + 1602.0, + 1308.0, + 1636.0, + 1158.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1678.0, + 966.0, + 1678.0, + 966.0, + 1698.0, + 936.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1715.0, + 730.0, + 1715.0, + 730.0, + 1738.0, + 707.0, + 1738.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 1731.0, + 750.0, + 1731.0, + 750.0, + 1740.0, + 735.0, + 1740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1701.0, + 895.0, + 1701.0, + 895.0, + 1723.0, + 868.0, + 1723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1697.0, + 933.0, + 1697.0, + 933.0, + 1755.0, + 899.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 939.0, + 1701.0, + 963.0, + 1701.0, + 963.0, + 1756.0, + 939.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 1701.0, + 1022.0, + 1701.0, + 1022.0, + 1733.0, + 964.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1714.0, + 1065.0, + 1714.0, + 1065.0, + 1746.0, + 1027.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1745.0, + 781.0, + 1745.0, + 781.0, + 1767.0, + 753.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1752.0, + 726.0, + 1752.0, + 726.0, + 1761.0, + 712.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 1767.0, + 631.0, + 1767.0, + 631.0, + 1784.0, + 610.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1768.0, + 655.0, + 1768.0, + 655.0, + 1776.0, + 641.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1767.0, + 682.0, + 1767.0, + 682.0, + 1787.0, + 662.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 1759.0, + 934.0, + 1759.0, + 934.0, + 1780.0, + 898.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1761.0, + 964.0, + 1761.0, + 964.0, + 1790.0, + 937.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.0, + 1781.0, + 680.0, + 1781.0, + 680.0, + 1801.0, + 638.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1780.0, + 732.0, + 1780.0, + 732.0, + 1810.0, + 683.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 1787.0, + 898.0, + 1787.0, + 898.0, + 1813.0, + 867.0, + 1813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 1783.0, + 932.0, + 1783.0, + 932.0, + 1804.0, + 903.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1813.0, + 634.0, + 1813.0, + 634.0, + 1836.0, + 608.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1812.0, + 929.0, + 1812.0, + 929.0, + 1828.0, + 905.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1859.0, + 770.0, + 1859.0, + 770.0, + 1889.0, + 648.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1861.0, + 1038.0, + 1861.0, + 1038.0, + 1891.0, + 894.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1470.5, + 394.0, + 1470.5, + 394.0, + 1483.5, + 384.0, + 1483.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 772.75, + 1748.5, + 810.75, + 1748.5, + 810.75, + 1777.5, + 772.75, + 1777.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1129.0, + 1404.0, + 1129.0, + 1404.0, + 1165.0, + 295.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1157.0, + 391.0, + 1157.0, + 391.0, + 1199.0, + 292.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1925.0, + 1405.0, + 1925.0, + 1405.0, + 1961.0, + 294.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1954.0, + 299.0, + 1954.0, + 299.0, + 1992.0, + 295.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 1954.0, + 1044.0, + 1954.0, + 1044.0, + 1992.0, + 371.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 348, + 1173, + 1333, + 1173, + 1333, + 1827, + 348, + 1827 + ], + "score": 0.947 + }, + { + "category_id": 4, + "poly": [ + 296, + 855, + 1403, + 855, + 1403, + 920, + 296, + 920 + ], + "score": 0.942 + }, + { + "category_id": 4, + "poly": [ + 294, + 1862, + 1403, + 1862, + 1403, + 1927, + 294, + 1927 + ], + "score": 0.933 + }, + { + "category_id": 3, + "poly": [ + 347, + 365, + 1332, + 365, + 1332, + 821, + 347, + 821 + ], + "score": 0.916 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 813, + 76, + 813, + 105, + 299, + 105 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.847 + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1185.0, + 703.0, + 1185.0, + 703.0, + 1208.0, + 657.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1179.0, + 970.0, + 1179.0, + 970.0, + 1203.0, + 930.0, + 1203.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1210.0, + 696.0, + 1210.0, + 696.0, + 1232.0, + 661.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1203.0, + 968.0, + 1203.0, + 968.0, + 1224.0, + 937.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1250.0, + 654.0, + 1250.0, + 654.0, + 1285.0, + 614.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1235.0, + 702.0, + 1235.0, + 702.0, + 1259.0, + 658.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 1226.0, + 740.0, + 1226.0, + 740.0, + 1264.0, + 708.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1241.0, + 929.0, + 1241.0, + 929.0, + 1277.0, + 887.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 978.0, + 1219.0, + 1008.0, + 1219.0, + 1008.0, + 1260.0, + 978.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1304.0, + 378.0, + 1304.0, + 378.0, + 1322.0, + 356.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1287.0, + 511.0, + 1287.0, + 511.0, + 1388.0, + 382.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 516.0, + 1337.0, + 549.0, + 1337.0, + 549.0, + 1365.0, + 516.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1320.0, + 654.0, + 1320.0, + 654.0, + 1356.0, + 613.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1263.0, + 705.0, + 1263.0, + 705.0, + 1364.0, + 659.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1267.0, + 739.0, + 1267.0, + 739.0, + 1375.0, + 709.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1287.0, + 796.0, + 1287.0, + 796.0, + 1319.0, + 763.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1315.0, + 926.0, + 1315.0, + 926.0, + 1346.0, + 883.0, + 1346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1257.0, + 973.0, + 1257.0, + 973.0, + 1355.0, + 930.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1260.0, + 1011.0, + 1260.0, + 1011.0, + 1368.0, + 976.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1281.0, + 1065.0, + 1281.0, + 1065.0, + 1310.0, + 1036.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 1292.0, + 1160.0, + 1292.0, + 1160.0, + 1315.0, + 1134.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1304.0, + 1197.0, + 1304.0, + 1197.0, + 1364.0, + 1171.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1200.0, + 1262.0, + 1293.0, + 1262.0, + 1293.0, + 1382.0, + 1200.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1303.0, + 1334.0, + 1303.0, + 1334.0, + 1341.0, + 1295.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1370.0, + 414.0, + 1370.0, + 414.0, + 1407.0, + 351.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 1359.0, + 448.0, + 1359.0, + 448.0, + 1427.0, + 418.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1364.0, + 699.0, + 1364.0, + 699.0, + 1416.0, + 659.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 1361.0, + 973.0, + 1361.0, + 973.0, + 1407.0, + 932.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 1377.0, + 1165.0, + 1377.0, + 1165.0, + 1402.0, + 1133.0, + 1402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1375.0, + 1197.0, + 1375.0, + 1197.0, + 1412.0, + 1169.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1455.0, + 511.0, + 1455.0, + 511.0, + 1486.0, + 388.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1456.0, + 779.0, + 1456.0, + 779.0, + 1485.0, + 631.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1449.0, + 1039.0, + 1449.0, + 1039.0, + 1478.0, + 914.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1451.0, + 1306.0, + 1451.0, + 1306.0, + 1480.0, + 1158.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1526.0, + 704.0, + 1526.0, + 704.0, + 1550.0, + 665.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 670.0, + 1553.0, + 701.0, + 1553.0, + 701.0, + 1573.0, + 670.0, + 1573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1570.0, + 738.0, + 1570.0, + 738.0, + 1592.0, + 714.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1573.0, + 706.0, + 1573.0, + 706.0, + 1602.0, + 666.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1589.0, + 663.0, + 1589.0, + 663.0, + 1625.0, + 619.0, + 1625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1599.0, + 721.0, + 1599.0, + 721.0, + 1609.0, + 711.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1602.0, + 704.0, + 1602.0, + 704.0, + 1632.0, + 664.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1607.0, + 744.0, + 1607.0, + 744.0, + 1638.0, + 711.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1627.0, + 801.0, + 1627.0, + 801.0, + 1660.0, + 767.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1632.0, + 770.0, + 1632.0, + 770.0, + 1641.0, + 761.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1641.0, + 706.0, + 1641.0, + 706.0, + 1672.0, + 665.0, + 1672.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 870.0, + 1641.0, + 896.0, + 1641.0, + 896.0, + 1663.0, + 870.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1642.0, + 924.0, + 1642.0, + 924.0, + 1656.0, + 905.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1650.0, + 739.0, + 1650.0, + 739.0, + 1678.0, + 712.0, + 1678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1660.0, + 964.0, + 1660.0, + 964.0, + 1752.0, + 930.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1662.0, + 661.0, + 1662.0, + 661.0, + 1696.0, + 619.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1673.0, + 706.0, + 1673.0, + 706.0, + 1706.0, + 662.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1680.0, + 721.0, + 1680.0, + 721.0, + 1690.0, + 712.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1691.0, + 739.0, + 1691.0, + 739.0, + 1720.0, + 713.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1694.0, + 926.0, + 1694.0, + 926.0, + 1703.0, + 915.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1708.0, + 705.0, + 1708.0, + 705.0, + 1737.0, + 663.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 1703.0, + 1065.0, + 1703.0, + 1065.0, + 1733.0, + 1027.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1733.0, + 704.0, + 1733.0, + 704.0, + 1758.0, + 665.0, + 1758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 1728.0, + 896.0, + 1728.0, + 896.0, + 1749.0, + 868.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 1738.0, + 929.0, + 1738.0, + 929.0, + 1754.0, + 905.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1751.0, + 959.0, + 1751.0, + 959.0, + 1768.0, + 934.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 1735.0, + 1021.0, + 1735.0, + 1021.0, + 1768.0, + 963.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1797.0, + 771.0, + 1797.0, + 771.0, + 1826.0, + 648.0, + 1826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 1798.0, + 1037.0, + 1798.0, + 1037.0, + 1828.0, + 894.0, + 1828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 1227.0, + 976.0, + 1227.0, + 976.0, + 1250.5, + 930.0, + 1250.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.25, + 1288.0, + 1200.25, + 1288.0, + 1200.25, + 1303.5, + 1164.25, + 1303.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1311.0, + 483.0, + 1311.0, + 483.0, + 1331.5, + 451.0, + 1331.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1166.75, + 1328.5, + 1202.75, + 1328.5, + 1202.75, + 1342.5, + 1166.75, + 1342.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1621.0, + 929.0, + 1621.0, + 929.0, + 1639.0, + 901.0, + 1639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 855.0, + 1405.0, + 855.0, + 1405.0, + 891.0, + 295.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 885.0, + 1371.0, + 885.0, + 1371.0, + 924.0, + 292.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1862.0, + 1404.0, + 1862.0, + 1404.0, + 1898.0, + 294.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1893.0, + 1070.0, + 1893.0, + 1070.0, + 1929.0, + 296.0, + 1929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 361.0, + 1233.0, + 361.0, + 1233.0, + 381.0, + 1199.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 393.0, + 415.0, + 393.0, + 415.0, + 430.0, + 355.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 403.0, + 446.0, + 403.0, + 446.0, + 430.0, + 417.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 408.0, + 476.0, + 408.0, + 476.0, + 425.0, + 451.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 416.0, + 506.0, + 416.0, + 506.0, + 441.0, + 482.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 422.0, + 548.0, + 422.0, + 548.0, + 447.0, + 509.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 401.0, + 624.0, + 401.0, + 624.0, + 413.0, + 606.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 420.0, + 677.0, + 420.0, + 677.0, + 440.0, + 659.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 397.0, + 707.0, + 397.0, + 707.0, + 433.0, + 687.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 425.0, + 776.0, + 425.0, + 776.0, + 444.0, + 756.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 390.0, + 1158.0, + 390.0, + 1158.0, + 409.0, + 1137.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 385.0, + 1202.0, + 385.0, + 1202.0, + 441.0, + 1163.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 388.0, + 1231.0, + 388.0, + 1231.0, + 441.0, + 1203.0, + 441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 374.0, + 1289.0, + 374.0, + 1289.0, + 398.0, + 1235.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 408.0, + 1331.0, + 408.0, + 1331.0, + 434.0, + 1296.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 435.0, + 472.0, + 435.0, + 472.0, + 445.0, + 457.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 788.0, + 434.0, + 808.0, + 434.0, + 808.0, + 453.0, + 788.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 455.0, + 382.0, + 455.0, + 382.0, + 476.0, + 355.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 464.0, + 415.0, + 464.0, + 415.0, + 482.0, + 387.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 457.0, + 445.0, + 457.0, + 445.0, + 483.0, + 420.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 447.0, + 506.0, + 447.0, + 506.0, + 469.0, + 485.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 454.0, + 625.0, + 454.0, + 625.0, + 473.0, + 603.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 451.0, + 679.0, + 451.0, + 679.0, + 475.0, + 658.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 442.0, + 702.0, + 442.0, + 702.0, + 451.0, + 687.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 445.0, + 729.0, + 445.0, + 729.0, + 464.0, + 710.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 461.0, + 891.0, + 461.0, + 891.0, + 474.0, + 876.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 467.0, + 955.0, + 467.0, + 955.0, + 476.0, + 941.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 456.0, + 975.0, + 456.0, + 975.0, + 486.0, + 956.0, + 486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 469.0, + 991.0, + 469.0, + 991.0, + 478.0, + 977.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 475.0, + 1017.0, + 475.0, + 1017.0, + 493.0, + 995.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 474.0, + 1084.0, + 474.0, + 1084.0, + 496.0, + 1058.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1134.0, + 474.0, + 1162.0, + 474.0, + 1162.0, + 496.0, + 1134.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 445.0, + 1201.0, + 445.0, + 1201.0, + 490.0, + 1163.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 448.0, + 1232.0, + 448.0, + 1232.0, + 475.0, + 1202.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 873.0, + 484.0, + 896.0, + 484.0, + 896.0, + 500.0, + 873.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 482.0, + 934.0, + 482.0, + 934.0, + 500.0, + 915.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 489.0, + 1054.0, + 489.0, + 1054.0, + 507.0, + 1033.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 496.0, + 1197.0, + 496.0, + 1197.0, + 512.0, + 1164.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 512.0, + 511.0, + 512.0, + 511.0, + 544.0, + 388.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 513.0, + 780.0, + 513.0, + 780.0, + 543.0, + 632.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 542.0, + 1038.0, + 542.0, + 1038.0, + 576.0, + 915.0, + 576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 542.0, + 1308.0, + 542.0, + 1308.0, + 573.0, + 1155.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 903.0, + 623.0, + 924.0, + 623.0, + 924.0, + 635.0, + 903.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 867.0, + 638.0, + 895.0, + 638.0, + 895.0, + 659.0, + 867.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 640.0, + 925.0, + 640.0, + 925.0, + 653.0, + 901.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 658.0, + 967.0, + 658.0, + 967.0, + 747.0, + 933.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 664.0, + 673.0, + 664.0, + 673.0, + 677.0, + 646.0, + 677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 664.0, + 699.0, + 664.0, + 699.0, + 682.0, + 681.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 664.0, + 929.0, + 664.0, + 929.0, + 678.0, + 899.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 667.0, + 645.0, + 667.0, + 645.0, + 693.0, + 613.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 668.0, + 736.0, + 668.0, + 736.0, + 686.0, + 708.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 679.0, + 672.0, + 679.0, + 672.0, + 695.0, + 647.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 710.0, + 689.0, + 737.0, + 689.0, + 737.0, + 710.0, + 710.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 685.0, + 766.0, + 685.0, + 766.0, + 710.0, + 741.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 687.0, + 929.0, + 687.0, + 929.0, + 700.0, + 904.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 705.0, + 679.0, + 705.0, + 679.0, + 726.0, + 646.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 711.0, + 705.0, + 711.0, + 705.0, + 734.0, + 680.0, + 734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 698.0, + 809.0, + 698.0, + 809.0, + 724.0, + 775.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 709.0, + 930.0, + 709.0, + 930.0, + 724.0, + 901.0, + 724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 698.0, + 1066.0, + 698.0, + 1066.0, + 729.0, + 1028.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 732.0, + 642.0, + 732.0, + 642.0, + 754.0, + 613.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 730.0, + 676.0, + 730.0, + 676.0, + 746.0, + 648.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 741.0, + 703.0, + 741.0, + 703.0, + 754.0, + 681.0, + 754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 729.0, + 736.0, + 729.0, + 736.0, + 746.0, + 708.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 737.0, + 765.0, + 737.0, + 765.0, + 762.0, + 740.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 723.0, + 895.0, + 723.0, + 895.0, + 744.0, + 868.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 730.0, + 1025.0, + 730.0, + 1025.0, + 762.0, + 968.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 789.0, + 771.0, + 789.0, + 771.0, + 820.0, + 647.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 789.0, + 1038.0, + 789.0, + 1038.0, + 820.0, + 891.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 380.5, + 446.0, + 380.5, + 446.0, + 398.0, + 417.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 699.75, + 468.0, + 734.75, + 468.0, + 734.75, + 490.0, + 699.75, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 744.5, + 964.0, + 744.5, + 964.0, + 761.0, + 933.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 294.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 347, + 1045, + 1342, + 1045, + 1342, + 1447, + 347, + 1447 + ], + "score": 0.97 + }, + { + "category_id": 3, + "poly": [ + 328, + 282, + 1356, + 282, + 1356, + 842, + 328, + 842 + ], + "score": 0.954 + }, + { + "category_id": 4, + "poly": [ + 296, + 877, + 1401, + 877, + 1401, + 942, + 296, + 942 + ], + "score": 0.948 + }, + { + "category_id": 4, + "poly": [ + 295, + 1481, + 1402, + 1481, + 1402, + 1546, + 295, + 1546 + ], + "score": 0.947 + }, + { + "category_id": 3, + "poly": [ + 344, + 1641, + 1333, + 1641, + 1333, + 1897, + 344, + 1897 + ], + "score": 0.937 + }, + { + "category_id": 4, + "poly": [ + 294, + 1930, + 1405, + 1930, + 1405, + 1995, + 294, + 1995 + ], + "score": 0.923 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 866, + 2088, + 866, + 2113, + 834, + 2113 + ], + "score": 0.862 + }, + { + "category_id": 13, + "poly": [ + 484, + 1483, + 598, + 1483, + 598, + 1517, + 484, + 1517 + ], + "score": 0.92, + "latex": "\\beta _ { s k i p } = 1" + }, + { + "category_id": 13, + "poly": [ + 775, + 1932, + 836, + 1932, + 836, + 1964, + 775, + 1964 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1051.0, + 413.0, + 1051.0, + 413.0, + 1069.0, + 389.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1053.0, + 941.0, + 1053.0, + 941.0, + 1070.0, + 915.0, + 1070.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1071.0, + 383.0, + 1071.0, + 383.0, + 1092.0, + 354.0, + 1092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1073.0, + 409.0, + 1073.0, + 409.0, + 1087.0, + 389.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1069.0, + 911.0, + 1069.0, + 911.0, + 1096.0, + 879.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1068.0, + 941.0, + 1068.0, + 941.0, + 1088.0, + 913.0, + 1088.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1094.0, + 448.0, + 1094.0, + 448.0, + 1131.0, + 387.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1091.0, + 978.0, + 1091.0, + 978.0, + 1157.0, + 912.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1124.0, + 413.0, + 1124.0, + 413.0, + 1137.0, + 395.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1118.0, + 446.0, + 1118.0, + 446.0, + 1154.0, + 420.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1128.0, + 677.0, + 1128.0, + 677.0, + 1136.0, + 664.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1121.0, + 940.0, + 1121.0, + 940.0, + 1137.0, + 917.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1142.0, + 416.0, + 1142.0, + 416.0, + 1162.0, + 388.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 1136.0, + 550.0, + 1136.0, + 550.0, + 1165.0, + 513.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1135.0, + 725.0, + 1135.0, + 725.0, + 1149.0, + 709.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1141.0, + 774.0, + 1141.0, + 774.0, + 1160.0, + 732.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1143.0, + 940.0, + 1143.0, + 940.0, + 1161.0, + 917.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1135.0, + 1077.0, + 1135.0, + 1077.0, + 1165.0, + 1038.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1150.0, + 1143.0, + 1150.0, + 1143.0, + 1159.0, + 1131.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1146.0, + 1231.0, + 1146.0, + 1231.0, + 1160.0, + 1215.0, + 1160.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1161.0, + 383.0, + 1161.0, + 383.0, + 1182.0, + 354.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1170.0, + 416.0, + 1170.0, + 416.0, + 1187.0, + 388.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1160.0, + 476.0, + 1160.0, + 476.0, + 1202.0, + 419.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 1173.0, + 506.0, + 1173.0, + 506.0, + 1202.0, + 477.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 611.0, + 1163.0, + 622.0, + 1163.0, + 622.0, + 1173.0, + 611.0, + 1173.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 1160.0, + 680.0, + 1160.0, + 680.0, + 1179.0, + 659.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1160.0, + 727.0, + 1160.0, + 727.0, + 1190.0, + 706.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1182.0, + 774.0, + 1182.0, + 774.0, + 1201.0, + 753.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 1165.0, + 808.0, + 1165.0, + 808.0, + 1188.0, + 782.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1159.0, + 911.0, + 1159.0, + 911.0, + 1183.0, + 879.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1169.0, + 944.0, + 1169.0, + 944.0, + 1186.0, + 914.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1160.0, + 1002.0, + 1160.0, + 1002.0, + 1202.0, + 945.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1174.0, + 1034.0, + 1174.0, + 1034.0, + 1200.0, + 1004.0, + 1200.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1172.0, + 1150.0, + 1172.0, + 1150.0, + 1189.0, + 1127.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1172.0, + 1171.0, + 1192.0, + 1171.0, + 1192.0, + 1189.0, + 1172.0, + 1189.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1237.0, + 1162.0, + 1289.0, + 1162.0, + 1289.0, + 1185.0, + 1237.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1177.0, + 1311.0, + 1177.0, + 1311.0, + 1196.0, + 1290.0, + 1196.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1312.0, + 1163.0, + 1338.0, + 1163.0, + 1338.0, + 1184.0, + 1312.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1229.0, + 512.0, + 1229.0, + 512.0, + 1263.0, + 387.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 1233.0, + 779.0, + 1233.0, + 779.0, + 1262.0, + 631.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 1229.0, + 1039.0, + 1229.0, + 1039.0, + 1263.0, + 914.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1230.0, + 1308.0, + 1230.0, + 1308.0, + 1263.0, + 1157.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1296.0, + 623.0, + 1296.0, + 623.0, + 1305.0, + 614.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1313.0, + 937.0, + 1313.0, + 937.0, + 1321.0, + 925.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 1318.0, + 684.0, + 1318.0, + 684.0, + 1337.0, + 662.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 1320.0, + 986.0, + 1320.0, + 986.0, + 1333.0, + 969.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1326.0, + 708.0, + 1326.0, + 708.0, + 1342.0, + 683.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1328.0, + 783.0, + 1328.0, + 783.0, + 1359.0, + 730.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1327.0, + 1034.0, + 1327.0, + 1034.0, + 1345.0, + 990.0, + 1345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1351.0, + 626.0, + 1351.0, + 626.0, + 1360.0, + 614.0, + 1360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1342.0, + 681.0, + 1342.0, + 681.0, + 1356.0, + 663.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1341.0, + 734.0, + 1341.0, + 734.0, + 1367.0, + 704.0, + 1367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1346.0, + 886.0, + 1346.0, + 886.0, + 1363.0, + 866.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1343.0, + 940.0, + 1343.0, + 940.0, + 1366.0, + 916.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1345.0, + 986.0, + 1345.0, + 986.0, + 1364.0, + 965.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 1350.0, + 1068.0, + 1350.0, + 1068.0, + 1373.0, + 1044.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 1365.0, + 654.0, + 1365.0, + 654.0, + 1373.0, + 641.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1370.0, + 685.0, + 1370.0, + 685.0, + 1393.0, + 660.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 1364.0, + 986.0, + 1364.0, + 986.0, + 1377.0, + 970.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1012.0, + 1367.0, + 1034.0, + 1367.0, + 1034.0, + 1386.0, + 1012.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1414.0, + 773.0, + 1414.0, + 773.0, + 1450.0, + 648.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 1417.0, + 1036.0, + 1417.0, + 1036.0, + 1447.0, + 892.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1157.5, + 701.0, + 1157.5, + 701.0, + 1169.0, + 683.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.75, + 1173.5, + 751.75, + 1173.5, + 751.75, + 1190.0, + 728.75, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1337.0, + 809.0, + 1337.0, + 809.0, + 1365.0, + 778.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 638.25, + 1328.5, + 665.25, + 1328.5, + 665.25, + 1342.0, + 638.25, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 1358.5, + 1011.0, + 1358.5, + 1011.0, + 1375.5, + 986.0, + 1375.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 289.0, + 974.0, + 289.0, + 974.0, + 314.0, + 937.0, + 314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 318.0, + 975.0, + 318.0, + 975.0, + 339.0, + 937.0, + 339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 344.0, + 432.0, + 344.0, + 432.0, + 365.0, + 400.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 334.0, + 1014.0, + 334.0, + 1014.0, + 363.0, + 985.0, + 363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 364.0, + 431.0, + 364.0, + 431.0, + 384.0, + 400.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 358.0, + 928.0, + 358.0, + 928.0, + 390.0, + 889.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 376.0, + 361.0, + 376.0, + 361.0, + 398.0, + 334.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 386.0, + 395.0, + 386.0, + 395.0, + 402.0, + 372.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 374.0, + 492.0, + 374.0, + 492.0, + 398.0, + 468.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 372.0, + 977.0, + 372.0, + 977.0, + 400.0, + 931.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 374.0, + 1014.0, + 374.0, + 1014.0, + 404.0, + 985.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 398.0, + 1075.0, + 398.0, + 1075.0, + 427.0, + 1042.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 400.0, + 1044.0, + 400.0, + 1044.0, + 411.0, + 1032.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 405.0, + 492.0, + 405.0, + 492.0, + 428.0, + 467.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 437.0, + 411.0, + 461.0, + 411.0, + 461.0, + 423.0, + 437.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 412.0, + 431.0, + 412.0, + 431.0, + 439.0, + 401.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 413.0, + 395.0, + 413.0, + 395.0, + 429.0, + 372.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 421.0, + 1012.0, + 421.0, + 1012.0, + 445.0, + 987.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 435.0, + 397.0, + 435.0, + 397.0, + 452.0, + 371.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 435.0, + 693.0, + 435.0, + 693.0, + 450.0, + 676.0, + 450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 433.0, + 928.0, + 433.0, + 928.0, + 465.0, + 889.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 450.0, + 365.0, + 450.0, + 365.0, + 477.0, + 333.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 449.0, + 629.0, + 449.0, + 629.0, + 457.0, + 619.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 444.0, + 676.0, + 444.0, + 676.0, + 458.0, + 656.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 695.0, + 445.0, + 715.0, + 445.0, + 715.0, + 459.0, + 695.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 450.0, + 755.0, + 450.0, + 755.0, + 459.0, + 741.0, + 459.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 443.0, + 780.0, + 443.0, + 780.0, + 460.0, + 757.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 447.0, + 975.0, + 447.0, + 975.0, + 472.0, + 931.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 462.0, + 395.0, + 462.0, + 395.0, + 478.0, + 372.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 454.0, + 430.0, + 454.0, + 430.0, + 477.0, + 403.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 462.0, + 653.0, + 462.0, + 653.0, + 477.0, + 636.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 715.0, + 457.0, + 736.0, + 457.0, + 736.0, + 475.0, + 715.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 482.0, + 974.0, + 482.0, + 974.0, + 504.0, + 932.0, + 504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 467.0, + 1005.0, + 467.0, + 1005.0, + 482.0, + 990.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 471.0, + 1160.0, + 471.0, + 1160.0, + 484.0, + 1144.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 480.0, + 1187.0, + 480.0, + 1187.0, + 491.0, + 1167.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 467.0, + 1212.0, + 467.0, + 1212.0, + 480.0, + 1196.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 475.0, + 1266.0, + 475.0, + 1266.0, + 488.0, + 1246.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 485.0, + 1292.0, + 485.0, + 1292.0, + 499.0, + 1273.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 481.0, + 1317.0, + 481.0, + 1317.0, + 499.0, + 1296.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 510.0, + 497.0, + 510.0, + 497.0, + 541.0, + 372.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 509.0, + 775.0, + 509.0, + 775.0, + 543.0, + 623.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 500.0, + 1165.0, + 500.0, + 1165.0, + 518.0, + 1142.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1190.0, + 498.0, + 1214.0, + 498.0, + 1214.0, + 522.0, + 1190.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 500.0, + 1236.0, + 500.0, + 1236.0, + 509.0, + 1220.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1244.0, + 500.0, + 1266.0, + 500.0, + 1266.0, + 520.0, + 1244.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 516.0, + 1292.0, + 516.0, + 1292.0, + 529.0, + 1271.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 506.0, + 1353.0, + 506.0, + 1353.0, + 529.0, + 1328.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 524.0, + 1316.0, + 524.0, + 1316.0, + 543.0, + 1297.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 571.0, + 1044.0, + 571.0, + 1044.0, + 602.0, + 918.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 573.0, + 1321.0, + 573.0, + 1321.0, + 603.0, + 1171.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 636.0, + 675.0, + 636.0, + 675.0, + 654.0, + 647.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 656.0, + 638.0, + 656.0, + 638.0, + 674.0, + 616.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 661.0, + 666.0, + 661.0, + 666.0, + 669.0, + 652.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 675.0, + 708.0, + 675.0, + 708.0, + 768.0, + 676.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 683.0, + 670.0, + 683.0, + 670.0, + 693.0, + 656.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 706.0, + 670.0, + 706.0, + 670.0, + 717.0, + 659.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 718.0, + 810.0, + 718.0, + 810.0, + 749.0, + 773.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 744.0, + 639.0, + 744.0, + 639.0, + 764.0, + 616.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 751.0, + 673.0, + 751.0, + 673.0, + 768.0, + 649.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 750.0, + 736.0, + 750.0, + 736.0, + 767.0, + 709.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 743.0, + 878.0, + 743.0, + 878.0, + 757.0, + 862.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 927.0, + 751.0, + 938.0, + 751.0, + 938.0, + 760.0, + 927.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 752.0, + 979.0, + 752.0, + 979.0, + 761.0, + 965.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 745.0, + 1043.0, + 745.0, + 1043.0, + 764.0, + 1023.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 765.0, + 705.0, + 765.0, + 705.0, + 782.0, + 680.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 758.0, + 767.0, + 758.0, + 767.0, + 782.0, + 740.0, + 782.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 767.0, + 884.0, + 767.0, + 884.0, + 785.0, + 861.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 764.0, + 922.0, + 764.0, + 922.0, + 781.0, + 902.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 761.0, + 1003.0, + 761.0, + 1003.0, + 779.0, + 982.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 756.0, + 1070.0, + 756.0, + 1070.0, + 774.0, + 1051.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 811.0, + 772.0, + 811.0, + 772.0, + 841.0, + 647.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 894.0, + 813.0, + 1038.0, + 813.0, + 1038.0, + 843.0, + 894.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.75, + 342.5, + 976.75, + 342.5, + 976.75, + 364.5, + 933.75, + 364.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 432.0, + 383.5, + 466.0, + 383.5, + 466.0, + 399.5, + 432.0, + 399.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 500.0, + 415.0, + 535.0, + 415.0, + 535.0, + 452.0, + 500.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 412.5, + 978.0, + 412.5, + 978.0, + 436.5, + 933.0, + 436.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 779.0, + 447.5, + 803.0, + 447.5, + 803.0, + 468.5, + 779.0, + 468.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 932.0, + 507.5, + 977.0, + 507.5, + 977.0, + 531.5, + 932.0, + 531.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 737.0, + 957.0, + 737.0, + 957.0, + 753.0, + 942.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 876.0, + 1403.0, + 876.0, + 1403.0, + 915.0, + 294.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 907.0, + 1334.0, + 907.0, + 1334.0, + 946.0, + 292.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1477.0, + 483.0, + 1477.0, + 483.0, + 1521.0, + 292.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 1477.0, + 1407.0, + 1477.0, + 1407.0, + 1521.0, + 599.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1513.0, + 545.0, + 1513.0, + 545.0, + 1546.0, + 295.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 1653.0, + 411.0, + 1653.0, + 411.0, + 1666.0, + 390.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1688.0, + 417.0, + 1688.0, + 417.0, + 1708.0, + 388.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1683.0, + 445.0, + 1683.0, + 445.0, + 1708.0, + 422.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1688.0, + 944.0, + 1688.0, + 944.0, + 1708.0, + 915.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1683.0, + 973.0, + 1683.0, + 973.0, + 1708.0, + 949.0, + 1708.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1700.0, + 382.0, + 1700.0, + 382.0, + 1727.0, + 351.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1710.0, + 414.0, + 1710.0, + 414.0, + 1730.0, + 385.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1714.0, + 446.0, + 1714.0, + 446.0, + 1739.0, + 421.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1701.0, + 909.0, + 1701.0, + 909.0, + 1727.0, + 878.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1710.0, + 942.0, + 1710.0, + 942.0, + 1731.0, + 911.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1714.0, + 972.0, + 1714.0, + 972.0, + 1739.0, + 949.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 351.0, + 1758.0, + 381.0, + 1758.0, + 381.0, + 1785.0, + 351.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1742.0, + 413.0, + 1742.0, + 413.0, + 1759.0, + 389.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1748.0, + 445.0, + 1748.0, + 445.0, + 1770.0, + 421.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1731.0, + 490.0, + 1731.0, + 490.0, + 1753.0, + 466.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1743.0, + 576.0, + 1743.0, + 576.0, + 1760.0, + 556.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 1752.0, + 609.0, + 1752.0, + 609.0, + 1763.0, + 588.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1739.0, + 643.0, + 1739.0, + 643.0, + 1755.0, + 616.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1748.0, + 702.0, + 1748.0, + 702.0, + 1765.0, + 675.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1753.0, + 760.0, + 1753.0, + 760.0, + 1780.0, + 732.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1758.0, + 909.0, + 1758.0, + 909.0, + 1785.0, + 878.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1743.0, + 942.0, + 1743.0, + 942.0, + 1759.0, + 916.0, + 1759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1749.0, + 971.0, + 1749.0, + 971.0, + 1770.0, + 950.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 1731.0, + 1018.0, + 1731.0, + 1018.0, + 1755.0, + 993.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 1743.0, + 1104.0, + 1743.0, + 1104.0, + 1760.0, + 1084.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1113.0, + 1752.0, + 1138.0, + 1752.0, + 1138.0, + 1765.0, + 1113.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1739.0, + 1172.0, + 1739.0, + 1172.0, + 1755.0, + 1148.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1747.0, + 1231.0, + 1747.0, + 1231.0, + 1768.0, + 1204.0, + 1768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1761.0, + 1260.0, + 1761.0, + 1260.0, + 1778.0, + 1233.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1757.0, + 1287.0, + 1757.0, + 1287.0, + 1779.0, + 1263.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1797.0, + 413.0, + 1797.0, + 413.0, + 1830.0, + 389.0, + 1830.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 1777.0, + 446.0, + 1777.0, + 446.0, + 1806.0, + 419.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1776.0, + 584.0, + 1776.0, + 584.0, + 1802.0, + 552.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 586.0, + 1778.0, + 610.0, + 1778.0, + 610.0, + 1790.0, + 586.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1779.0, + 643.0, + 1779.0, + 643.0, + 1801.0, + 615.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.0, + 1777.0, + 760.0, + 1777.0, + 760.0, + 1831.0, + 674.0, + 1831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 1783.0, + 802.0, + 1783.0, + 802.0, + 1818.0, + 761.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 945.0, + 1778.0, + 973.0, + 1778.0, + 973.0, + 1806.0, + 945.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1777.0, + 1137.0, + 1777.0, + 1137.0, + 1802.0, + 1079.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 1779.0, + 1171.0, + 1779.0, + 1171.0, + 1803.0, + 1146.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1778.0, + 1230.0, + 1778.0, + 1230.0, + 1820.0, + 1178.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 1798.0, + 1258.0, + 1798.0, + 1258.0, + 1815.0, + 1232.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1260.0, + 1805.0, + 1288.0, + 1805.0, + 1288.0, + 1832.0, + 1260.0, + 1832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1786.0, + 1329.0, + 1786.0, + 1329.0, + 1815.0, + 1291.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1864.0, + 484.0, + 1864.0, + 484.0, + 1895.0, + 356.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1862.0, + 751.0, + 1862.0, + 751.0, + 1898.0, + 598.0, + 1898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1864.0, + 1010.0, + 1864.0, + 1010.0, + 1895.0, + 883.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 1863.0, + 1278.0, + 1863.0, + 1278.0, + 1896.0, + 1125.0, + 1896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 1758.5, + 738.0, + 1758.5, + 738.0, + 1778.5, + 705.0, + 1778.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1771.5, + 413.0, + 1771.5, + 413.0, + 1784.5, + 388.0, + 1784.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 649.0, + 1777.5, + 677.0, + 1777.5, + 677.0, + 1789.5, + 649.0, + 1789.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1771.5, + 940.0, + 1771.5, + 940.0, + 1784.5, + 915.0, + 1784.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1929.0, + 774.0, + 1929.0, + 774.0, + 1969.0, + 294.0, + 1969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1929.0, + 1406.0, + 1929.0, + 1406.0, + 1969.0, + 837.0, + 1969.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1961.0, + 1179.0, + 1961.0, + 1179.0, + 1995.0, + 294.0, + 1995.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2123.0, + 828.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 352, + 1356, + 1333, + 1356, + 1333, + 1887, + 352, + 1887 + ], + "score": 0.957 + }, + { + "category_id": 4, + "poly": [ + 295, + 632, + 1402, + 632, + 1402, + 697, + 295, + 697 + ], + "score": 0.945 + }, + { + "category_id": 3, + "poly": [ + 339, + 286, + 1338, + 286, + 1338, + 599, + 339, + 599 + ], + "score": 0.942 + }, + { + "category_id": 3, + "poly": [ + 339, + 815, + 1340, + 815, + 1340, + 1132, + 339, + 1132 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 291, + 1165, + 1405, + 1165, + 1405, + 1231, + 291, + 1231 + ], + "score": 0.931 + }, + { + "category_id": 4, + "poly": [ + 292, + 1921, + 1403, + 1921, + 1403, + 1985, + 292, + 1985 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 813, + 76, + 813, + 105, + 299, + 105 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.759 + }, + { + "category_id": 3, + "poly": [ + 352, + 1352, + 1333, + 1352, + 1333, + 1887, + 352, + 1887 + ], + "score": 0.182 + }, + { + "category_id": 13, + "poly": [ + 533, + 1169, + 595, + 1169, + 595, + 1202, + 533, + 1202 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 533, + 1924, + 595, + 1924, + 595, + 1957, + 533, + 1957 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 13, + "poly": [ + 533, + 634, + 595, + 634, + 595, + 668, + 533, + 668 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1397.0, + 450.0, + 1397.0, + 450.0, + 1421.0, + 426.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 1408.0, + 943.0, + 1408.0, + 943.0, + 1419.0, + 925.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1397.0, + 976.0, + 1397.0, + 976.0, + 1421.0, + 952.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1414.0, + 384.0, + 1414.0, + 384.0, + 1440.0, + 354.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1417.0, + 910.0, + 1417.0, + 910.0, + 1439.0, + 883.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1428.0, + 451.0, + 1428.0, + 451.0, + 1452.0, + 426.0, + 1452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1431.0, + 415.0, + 1431.0, + 415.0, + 1441.0, + 403.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1473.0, + 382.0, + 1473.0, + 382.0, + 1494.0, + 356.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1457.0, + 418.0, + 1457.0, + 418.0, + 1474.0, + 392.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1460.0, + 450.0, + 1460.0, + 450.0, + 1482.0, + 426.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1444.0, + 496.0, + 1444.0, + 496.0, + 1470.0, + 468.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1456.0, + 582.0, + 1456.0, + 582.0, + 1474.0, + 559.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1451.0, + 644.0, + 1451.0, + 644.0, + 1467.0, + 621.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1459.0, + 761.0, + 1459.0, + 761.0, + 1491.0, + 676.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1473.0, + 910.0, + 1473.0, + 910.0, + 1494.0, + 883.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1460.0, + 942.0, + 1460.0, + 942.0, + 1473.0, + 919.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1463.0, + 973.0, + 1463.0, + 973.0, + 1481.0, + 954.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1445.0, + 1020.0, + 1445.0, + 1020.0, + 1469.0, + 996.0, + 1469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 1454.0, + 1109.0, + 1454.0, + 1109.0, + 1475.0, + 1085.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 1462.0, + 1143.0, + 1462.0, + 1143.0, + 1481.0, + 1115.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1451.0, + 1171.0, + 1451.0, + 1171.0, + 1467.0, + 1147.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1459.0, + 1232.0, + 1459.0, + 1232.0, + 1480.0, + 1203.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 1467.0, + 1291.0, + 1467.0, + 1291.0, + 1492.0, + 1233.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1488.0, + 413.0, + 1488.0, + 413.0, + 1496.0, + 399.0, + 1496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 1492.0, + 450.0, + 1492.0, + 450.0, + 1519.0, + 420.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1489.0, + 587.0, + 1489.0, + 587.0, + 1514.0, + 555.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1491.0, + 644.0, + 1491.0, + 644.0, + 1515.0, + 621.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1492.0, + 704.0, + 1492.0, + 704.0, + 1514.0, + 678.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1485.0, + 943.0, + 1485.0, + 943.0, + 1497.0, + 920.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 1492.0, + 976.0, + 1492.0, + 976.0, + 1519.0, + 948.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1489.0, + 1172.0, + 1489.0, + 1172.0, + 1515.0, + 1082.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 1490.0, + 1231.0, + 1490.0, + 1231.0, + 1514.0, + 1176.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1512.0, + 417.0, + 1512.0, + 417.0, + 1523.0, + 398.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1514.0, + 705.0, + 1514.0, + 705.0, + 1534.0, + 677.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1511.0, + 732.0, + 1511.0, + 732.0, + 1526.0, + 707.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1499.0, + 804.0, + 1499.0, + 804.0, + 1526.0, + 773.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1514.0, + 942.0, + 1514.0, + 942.0, + 1522.0, + 926.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1300.0, + 1501.0, + 1330.0, + 1501.0, + 1330.0, + 1524.0, + 1300.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 1520.0, + 763.0, + 1520.0, + 763.0, + 1543.0, + 737.0, + 1543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 1519.0, + 1291.0, + 1519.0, + 1291.0, + 1545.0, + 1263.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1576.0, + 487.0, + 1576.0, + 487.0, + 1608.0, + 360.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 1577.0, + 755.0, + 1577.0, + 755.0, + 1612.0, + 603.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1577.0, + 1014.0, + 1577.0, + 1014.0, + 1609.0, + 887.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1577.0, + 1283.0, + 1577.0, + 1283.0, + 1612.0, + 1131.0, + 1612.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 1645.0, + 678.0, + 1645.0, + 678.0, + 1661.0, + 652.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 1669.0, + 671.0, + 1669.0, + 671.0, + 1674.0, + 666.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1672.0, + 711.0, + 1672.0, + 711.0, + 1734.0, + 680.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 1685.0, + 678.0, + 1685.0, + 678.0, + 1693.0, + 669.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1691.0, + 648.0, + 1691.0, + 648.0, + 1720.0, + 613.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1711.0, + 675.0, + 1711.0, + 675.0, + 1719.0, + 660.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1749.0, + 645.0, + 1749.0, + 645.0, + 1775.0, + 614.0, + 1775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1739.0, + 710.0, + 1739.0, + 710.0, + 1764.0, + 687.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 1722.0, + 754.0, + 1722.0, + 754.0, + 1745.0, + 729.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 1738.0, + 836.0, + 1738.0, + 836.0, + 1746.0, + 822.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1741.0, + 874.0, + 1741.0, + 874.0, + 1757.0, + 847.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1728.0, + 906.0, + 1728.0, + 906.0, + 1744.0, + 882.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1738.0, + 965.0, + 1738.0, + 965.0, + 1757.0, + 937.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1745.0, + 1022.0, + 1745.0, + 1022.0, + 1770.0, + 967.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1763.0, + 679.0, + 1763.0, + 679.0, + 1801.0, + 654.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1768.0, + 711.0, + 1768.0, + 711.0, + 1797.0, + 683.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1769.0, + 846.0, + 1769.0, + 846.0, + 1790.0, + 817.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 853.0, + 1768.0, + 876.0, + 1768.0, + 876.0, + 1780.0, + 853.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1767.0, + 907.0, + 1767.0, + 907.0, + 1795.0, + 877.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1762.0, + 994.0, + 1762.0, + 994.0, + 1811.0, + 909.0, + 1811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1795.0, + 1024.0, + 1795.0, + 1024.0, + 1823.0, + 995.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1031.0, + 1777.0, + 1064.0, + 1777.0, + 1064.0, + 1802.0, + 1031.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1854.0, + 746.0, + 1854.0, + 746.0, + 1886.0, + 620.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 1854.0, + 1014.0, + 1854.0, + 1014.0, + 1889.0, + 865.0, + 1889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.25, + 1405.5, + 420.25, + 1405.5, + 420.25, + 1418.5, + 396.25, + 1418.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.25, + 1427.0, + 979.25, + 1427.0, + 979.25, + 1451.0, + 949.25, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1462.5, + 613.0, + 1462.5, + 613.0, + 1479.0, + 591.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 1489.0, + 618.0, + 1489.0, + 618.0, + 1504.5, + 584.0, + 1504.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1487.5, + 684.0, + 1487.5, + 684.0, + 1506.0, + 648.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1509.5, + 1257.0, + 1509.5, + 1257.0, + 1525.5, + 1235.0, + 1525.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.75, + 1514.5, + 1226.75, + 1514.5, + 1226.75, + 1531.5, + 1207.75, + 1531.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 630.0, + 532.0, + 630.0, + 532.0, + 668.0, + 291.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 668.0, + 596.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 662.0, + 546.0, + 662.0, + 546.0, + 698.0, + 294.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 291.0, + 1000.0, + 291.0, + 1000.0, + 315.0, + 975.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 296.0, + 948.0, + 296.0, + 948.0, + 310.0, + 929.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 305.0, + 386.0, + 305.0, + 386.0, + 323.0, + 345.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 312.0, + 406.0, + 312.0, + 406.0, + 328.0, + 390.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 316.0, + 893.0, + 316.0, + 893.0, + 329.0, + 876.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 318.0, + 975.0, + 318.0, + 975.0, + 331.0, + 954.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 330.0, + 448.0, + 330.0, + 448.0, + 349.0, + 428.0, + 349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 534.0, + 331.0, + 556.0, + 331.0, + 556.0, + 350.0, + 534.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 602.0, + 330.0, + 622.0, + 330.0, + 622.0, + 347.0, + 602.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 325.0, + 702.0, + 325.0, + 702.0, + 344.0, + 684.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 331.0, + 922.0, + 331.0, + 922.0, + 340.0, + 909.0, + 340.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 326.0, + 1048.0, + 326.0, + 1048.0, + 350.0, + 1023.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 328.0, + 1270.0, + 328.0, + 1270.0, + 346.0, + 1249.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 340.0, + 638.0, + 340.0, + 638.0, + 350.0, + 626.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 336.0, + 722.0, + 336.0, + 722.0, + 353.0, + 700.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 340.0, + 760.0, + 340.0, + 760.0, + 350.0, + 746.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 761.0, + 335.0, + 781.0, + 335.0, + 781.0, + 352.0, + 761.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 335.0, + 969.0, + 335.0, + 969.0, + 344.0, + 955.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 338.0, + 1079.0, + 338.0, + 1079.0, + 358.0, + 1051.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1233.0, + 338.0, + 1252.0, + 338.0, + 1252.0, + 351.0, + 1233.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 337.0, + 1286.0, + 337.0, + 1286.0, + 351.0, + 1270.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 347.0, + 346.0, + 360.0, + 346.0, + 360.0, + 356.0, + 347.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 346.0, + 488.0, + 346.0, + 488.0, + 365.0, + 466.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 353.0, + 527.0, + 353.0, + 527.0, + 374.0, + 492.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 352.0, + 661.0, + 352.0, + 661.0, + 373.0, + 600.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 352.0, + 694.0, + 352.0, + 694.0, + 358.0, + 689.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 350.0, + 742.0, + 350.0, + 742.0, + 369.0, + 721.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 349.0, + 948.0, + 349.0, + 948.0, + 370.0, + 904.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 362.0, + 974.0, + 362.0, + 974.0, + 378.0, + 950.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 352.0, + 1142.0, + 352.0, + 1142.0, + 360.0, + 1132.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 352.0, + 1229.0, + 352.0, + 1229.0, + 374.0, + 1195.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.0, + 351.0, + 1309.0, + 351.0, + 1309.0, + 370.0, + 1287.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 369.0, + 448.0, + 369.0, + 448.0, + 380.0, + 430.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 370.0, + 899.0, + 370.0, + 899.0, + 392.0, + 874.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 371.0, + 947.0, + 371.0, + 947.0, + 385.0, + 929.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 974.0, + 372.0, + 999.0, + 372.0, + 999.0, + 396.0, + 974.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1123.0, + 373.0, + 1190.0, + 373.0, + 1190.0, + 396.0, + 1123.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 399.0, + 511.0, + 399.0, + 511.0, + 431.0, + 387.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 399.0, + 780.0, + 399.0, + 780.0, + 431.0, + 629.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 416.0, + 1038.0, + 416.0, + 1038.0, + 448.0, + 915.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 417.0, + 1307.0, + 417.0, + 1307.0, + 448.0, + 1155.0, + 448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 478.0, + 749.0, + 478.0, + 749.0, + 499.0, + 709.0, + 499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 480.0, + 684.0, + 480.0, + 684.0, + 488.0, + 673.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 496.0, + 645.0, + 496.0, + 645.0, + 541.0, + 605.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 521.0, + 668.0, + 521.0, + 668.0, + 540.0, + 646.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 669.0, + 495.0, + 708.0, + 495.0, + 708.0, + 521.0, + 669.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 516.0, + 722.0, + 516.0, + 722.0, + 525.0, + 713.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 491.0, + 787.0, + 491.0, + 787.0, + 532.0, + 747.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 794.0, + 501.0, + 816.0, + 501.0, + 816.0, + 519.0, + 794.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 498.0, + 882.0, + 498.0, + 882.0, + 539.0, + 863.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 519.0, + 922.0, + 519.0, + 922.0, + 536.0, + 883.0, + 536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 492.0, + 962.0, + 492.0, + 962.0, + 509.0, + 943.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 963.0, + 504.0, + 1001.0, + 504.0, + 1001.0, + 534.0, + 963.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 500.0, + 1041.0, + 500.0, + 1041.0, + 519.0, + 1004.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1046.0, + 509.0, + 1070.0, + 509.0, + 1070.0, + 528.0, + 1046.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 566.0, + 771.0, + 566.0, + 771.0, + 597.0, + 648.0, + 597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 566.0, + 1039.0, + 566.0, + 1039.0, + 600.0, + 892.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 320.5, + 427.0, + 320.5, + 427.0, + 334.5, + 405.0, + 334.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 337.0, + 462.0, + 337.0, + 462.0, + 350.0, + 449.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 337.0, + 681.0, + 337.0, + 681.0, + 351.0, + 666.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.25, + 339.0, + 807.25, + 339.0, + 807.25, + 361.0, + 780.25, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.75, + 360.0, + 1163.75, + 360.0, + 1163.75, + 374.0, + 1150.75, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 355.0, + 1334.0, + 355.0, + 1334.0, + 380.5, + 1308.0, + 380.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 825.0, + 1000.0, + 825.0, + 1000.0, + 849.0, + 976.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 828.0, + 950.0, + 828.0, + 950.0, + 846.0, + 928.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 838.0, + 367.0, + 838.0, + 367.0, + 856.0, + 345.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 846.0, + 405.0, + 846.0, + 405.0, + 860.0, + 391.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 853.0, + 429.0, + 853.0, + 429.0, + 870.0, + 407.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 850.0, + 893.0, + 850.0, + 893.0, + 863.0, + 877.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 853.0, + 974.0, + 853.0, + 974.0, + 865.0, + 954.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 869.0, + 404.0, + 869.0, + 404.0, + 879.0, + 392.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 863.0, + 447.0, + 863.0, + 447.0, + 881.0, + 427.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 875.0, + 463.0, + 875.0, + 463.0, + 883.0, + 451.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 531.0, + 866.0, + 555.0, + 866.0, + 555.0, + 885.0, + 531.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 868.0, + 617.0, + 868.0, + 617.0, + 878.0, + 603.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 873.0, + 677.0, + 873.0, + 677.0, + 882.0, + 667.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 861.0, + 701.0, + 861.0, + 701.0, + 876.0, + 685.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 872.0, + 721.0, + 872.0, + 721.0, + 886.0, + 703.0, + 886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 862.0, + 739.0, + 862.0, + 739.0, + 870.0, + 727.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 868.0, + 782.0, + 868.0, + 782.0, + 887.0, + 744.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 867.0, + 970.0, + 867.0, + 970.0, + 876.0, + 955.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1022.0, + 860.0, + 1047.0, + 860.0, + 1047.0, + 882.0, + 1022.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 868.0, + 1204.0, + 868.0, + 1204.0, + 876.0, + 1195.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.0, + 861.0, + 1289.0, + 861.0, + 1289.0, + 887.0, + 1232.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 879.0, + 361.0, + 879.0, + 361.0, + 892.0, + 345.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 880.0, + 490.0, + 880.0, + 490.0, + 899.0, + 467.0, + 899.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 508.0, + 891.0, + 529.0, + 891.0, + 529.0, + 908.0, + 508.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 601.0, + 889.0, + 624.0, + 889.0, + 624.0, + 907.0, + 601.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 641.0, + 886.0, + 662.0, + 886.0, + 662.0, + 904.0, + 641.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 882.0, + 742.0, + 882.0, + 742.0, + 901.0, + 720.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 881.0, + 948.0, + 881.0, + 948.0, + 902.0, + 905.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 895.0, + 974.0, + 895.0, + 974.0, + 912.0, + 950.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 885.0, + 1143.0, + 885.0, + 1143.0, + 894.0, + 1131.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1154.0, + 895.0, + 1165.0, + 895.0, + 1165.0, + 905.0, + 1154.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 886.0, + 1229.0, + 886.0, + 1229.0, + 905.0, + 1195.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1254.0, + 886.0, + 1264.0, + 886.0, + 1264.0, + 895.0, + 1254.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 901.0, + 447.0, + 901.0, + 447.0, + 915.0, + 430.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 904.0, + 901.0, + 904.0, + 901.0, + 927.0, + 874.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 929.0, + 905.0, + 948.0, + 905.0, + 948.0, + 919.0, + 929.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 906.0, + 1000.0, + 906.0, + 1000.0, + 930.0, + 975.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1126.0, + 905.0, + 1193.0, + 905.0, + 1193.0, + 932.0, + 1126.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 933.0, + 513.0, + 933.0, + 513.0, + 968.0, + 389.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 632.0, + 936.0, + 780.0, + 936.0, + 780.0, + 966.0, + 632.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 914.0, + 950.0, + 1039.0, + 950.0, + 1039.0, + 982.0, + 914.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 950.0, + 1307.0, + 950.0, + 1307.0, + 980.0, + 1156.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 1014.0, + 686.0, + 1014.0, + 686.0, + 1022.0, + 672.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1015.0, + 748.0, + 1015.0, + 748.0, + 1029.0, + 712.0, + 1029.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1060.0, + 626.0, + 1060.0, + 626.0, + 1073.0, + 607.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1055.0, + 667.0, + 1055.0, + 667.0, + 1073.0, + 647.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 1042.0, + 686.0, + 1042.0, + 686.0, + 1052.0, + 673.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1028.0, + 708.0, + 1028.0, + 708.0, + 1047.0, + 687.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1051.0, + 724.0, + 1051.0, + 724.0, + 1060.0, + 712.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1026.0, + 788.0, + 1026.0, + 788.0, + 1066.0, + 747.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1034.0, + 815.0, + 1034.0, + 815.0, + 1053.0, + 795.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1032.0, + 882.0, + 1032.0, + 882.0, + 1045.0, + 864.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1052.0, + 923.0, + 1052.0, + 923.0, + 1069.0, + 904.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 1026.0, + 960.0, + 1026.0, + 960.0, + 1051.0, + 926.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 1035.0, + 1002.0, + 1035.0, + 1002.0, + 1068.0, + 961.0, + 1068.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1006.0, + 1034.0, + 1042.0, + 1034.0, + 1042.0, + 1052.0, + 1006.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1042.0, + 1070.0, + 1042.0, + 1070.0, + 1062.0, + 1049.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1099.0, + 771.0, + 1099.0, + 771.0, + 1131.0, + 648.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 1101.0, + 1037.0, + 1101.0, + 1037.0, + 1132.0, + 893.0, + 1132.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 872.0, + 805.0, + 872.0, + 805.0, + 893.0, + 783.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.75, + 863.0, + 922.75, + 863.0, + 922.75, + 874.5, + 907.75, + 874.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 866.5, + 1077.0, + 866.5, + 1077.0, + 892.5, + 1044.0, + 892.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1287.75, + 883.0, + 1334.75, + 883.0, + 1334.75, + 911.0, + 1287.75, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.25, + 1053.5, + 901.25, + 1053.5, + 901.25, + 1073.5, + 858.25, + 1073.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1164.0, + 532.0, + 1164.0, + 532.0, + 1204.0, + 293.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1164.0, + 1404.0, + 1164.0, + 1404.0, + 1204.0, + 596.0, + 1204.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1195.0, + 548.0, + 1195.0, + 548.0, + 1232.0, + 293.0, + 1232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1918.0, + 532.0, + 1918.0, + 532.0, + 1958.0, + 292.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1918.0, + 1405.0, + 1918.0, + 1405.0, + 1958.0, + 596.0, + 1958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1950.0, + 547.0, + 1950.0, + 547.0, + 1986.0, + 294.0, + 1986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 294.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1366.0, + 419.0, + 1366.0, + 419.0, + 1382.0, + 392.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 425.0, + 1397.0, + 450.0, + 1397.0, + 450.0, + 1421.0, + 425.0, + 1421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1396.0, + 978.0, + 1396.0, + 978.0, + 1423.0, + 949.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1414.0, + 385.0, + 1414.0, + 385.0, + 1441.0, + 355.0, + 1441.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 1417.0, + 910.0, + 1417.0, + 910.0, + 1439.0, + 883.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 1428.0, + 452.0, + 1428.0, + 452.0, + 1454.0, + 424.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1431.0, + 943.0, + 1431.0, + 943.0, + 1442.0, + 919.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1428.0, + 978.0, + 1428.0, + 978.0, + 1454.0, + 949.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1472.0, + 384.0, + 1472.0, + 384.0, + 1498.0, + 354.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1455.0, + 420.0, + 1455.0, + 420.0, + 1475.0, + 391.0, + 1475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 1461.0, + 451.0, + 1461.0, + 451.0, + 1484.0, + 426.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 459.0, + 1443.0, + 496.0, + 1443.0, + 496.0, + 1471.0, + 459.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 1456.0, + 581.0, + 1456.0, + 581.0, + 1472.0, + 559.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 1450.0, + 644.0, + 1450.0, + 644.0, + 1467.0, + 621.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 1457.0, + 706.0, + 1457.0, + 706.0, + 1480.0, + 678.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1466.0, + 763.0, + 1466.0, + 763.0, + 1494.0, + 707.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1473.0, + 910.0, + 1473.0, + 910.0, + 1495.0, + 882.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1459.0, + 942.0, + 1459.0, + 942.0, + 1474.0, + 918.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 1461.0, + 975.0, + 1461.0, + 975.0, + 1484.0, + 952.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 988.0, + 1443.0, + 1022.0, + 1443.0, + 1022.0, + 1472.0, + 988.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1456.0, + 1108.0, + 1456.0, + 1108.0, + 1473.0, + 1086.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1463.0, + 1142.0, + 1463.0, + 1142.0, + 1479.0, + 1117.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1148.0, + 1450.0, + 1171.0, + 1450.0, + 1171.0, + 1467.0, + 1148.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1204.0, + 1460.0, + 1231.0, + 1460.0, + 1231.0, + 1479.0, + 1204.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 1467.0, + 1291.0, + 1467.0, + 1291.0, + 1494.0, + 1235.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1508.0, + 420.0, + 1508.0, + 420.0, + 1546.0, + 386.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 1491.0, + 451.0, + 1491.0, + 451.0, + 1519.0, + 421.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1492.0, + 585.0, + 1492.0, + 585.0, + 1514.0, + 556.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1485.0, + 705.0, + 1485.0, + 705.0, + 1531.0, + 590.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1510.0, + 763.0, + 1510.0, + 763.0, + 1546.0, + 707.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1494.0, + 806.0, + 1494.0, + 806.0, + 1529.0, + 767.0, + 1529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1482.0, + 946.0, + 1482.0, + 946.0, + 1526.0, + 916.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1492.0, + 976.0, + 1492.0, + 976.0, + 1520.0, + 947.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 1489.0, + 1113.0, + 1489.0, + 1113.0, + 1515.0, + 1082.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 1489.0, + 1174.0, + 1489.0, + 1174.0, + 1517.0, + 1117.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1203.0, + 1491.0, + 1291.0, + 1491.0, + 1291.0, + 1546.0, + 1203.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 1497.0, + 1332.0, + 1497.0, + 1332.0, + 1528.0, + 1294.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1576.0, + 487.0, + 1576.0, + 487.0, + 1608.0, + 360.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1577.0, + 752.0, + 1577.0, + 752.0, + 1608.0, + 604.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1576.0, + 1014.0, + 1576.0, + 1014.0, + 1608.0, + 887.0, + 1608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1578.0, + 1282.0, + 1578.0, + 1282.0, + 1610.0, + 1132.0, + 1610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1648.0, + 675.0, + 1648.0, + 675.0, + 1657.0, + 660.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1671.0, + 713.0, + 1671.0, + 713.0, + 1734.0, + 681.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1691.0, + 649.0, + 1691.0, + 649.0, + 1722.0, + 613.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 1710.0, + 676.0, + 1710.0, + 676.0, + 1719.0, + 665.0, + 1719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1747.0, + 648.0, + 1747.0, + 648.0, + 1777.0, + 613.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1736.0, + 711.0, + 1736.0, + 711.0, + 1766.0, + 685.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1723.0, + 755.0, + 1723.0, + 755.0, + 1745.0, + 730.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1734.0, + 841.0, + 1734.0, + 841.0, + 1752.0, + 818.0, + 1752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 1741.0, + 875.0, + 1741.0, + 875.0, + 1757.0, + 847.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1726.0, + 907.0, + 1726.0, + 907.0, + 1746.0, + 880.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1737.0, + 965.0, + 1737.0, + 965.0, + 1757.0, + 937.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1744.0, + 1023.0, + 1744.0, + 1023.0, + 1772.0, + 967.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1761.0, + 680.0, + 1761.0, + 680.0, + 1820.0, + 647.0, + 1820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1767.0, + 711.0, + 1767.0, + 711.0, + 1797.0, + 683.0, + 1797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 1767.0, + 849.0, + 1767.0, + 849.0, + 1794.0, + 815.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 1767.0, + 907.0, + 1767.0, + 907.0, + 1794.0, + 876.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1762.0, + 1023.0, + 1762.0, + 1023.0, + 1823.0, + 909.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1776.0, + 1064.0, + 1776.0, + 1064.0, + 1804.0, + 1032.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 1855.0, + 746.0, + 1855.0, + 746.0, + 1888.0, + 620.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1855.0, + 1013.0, + 1855.0, + 1013.0, + 1887.0, + 864.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.75, + 1366.0, + 945.75, + 1366.0, + 945.75, + 1382.5, + 915.75, + 1382.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.75, + 1405.5, + 418.75, + 1405.5, + 418.75, + 1419.0, + 396.75, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 1404.0, + 950.0, + 1404.0, + 950.0, + 1420.0, + 920.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1429.5, + 416.0, + 1429.5, + 416.0, + 1443.0, + 395.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 1463.0, + 613.0, + 1463.0, + 613.0, + 1478.5, + 591.0, + 1478.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1483.5, + 420.0, + 1483.5, + 420.0, + 1499.5, + 389.0, + 1499.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1177.0, + 1487.5, + 1202.0, + 1487.5, + 1202.0, + 1504.5, + 1177.0, + 1504.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1530.5, + 942.0, + 1530.5, + 942.0, + 1542.5, + 917.0, + 1542.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1736.5, + 679.0, + 1736.5, + 679.0, + 1749.0, + 653.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.25, + 1766.5, + 880.25, + 1766.5, + 880.25, + 1782.5, + 848.25, + 1782.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 4, + "poly": [ + 290, + 1387, + 1404, + 1387, + 1404, + 1451, + 290, + 1451 + ], + "score": 0.929 + }, + { + "category_id": 3, + "poly": [ + 353, + 826, + 1332, + 826, + 1332, + 1353, + 353, + 1353 + ], + "score": 0.894 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.834 + }, + { + "category_id": 3, + "poly": [ + 346, + 143, + 1341, + 143, + 1341, + 1384, + 346, + 1384 + ], + "score": 0.134 + }, + { + "category_id": 13, + "poly": [ + 533, + 1389, + 595, + 1389, + 595, + 1422, + 533, + 1422 + ], + "score": 0.91, + "latex": "\\beta _ { s k i p }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1386.0, + 532.0, + 1386.0, + 532.0, + 1423.0, + 292.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1386.0, + 1405.0, + 1386.0, + 1405.0, + 1423.0, + 596.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1419.0, + 546.0, + 1419.0, + 546.0, + 1451.0, + 294.0, + 1451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 832.0, + 418.0, + 832.0, + 418.0, + 849.0, + 389.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 832.0, + 944.0, + 832.0, + 944.0, + 849.0, + 915.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 854.0, + 420.0, + 854.0, + 420.0, + 887.0, + 393.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 860.0, + 454.0, + 860.0, + 454.0, + 924.0, + 422.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 857.0, + 945.0, + 857.0, + 945.0, + 885.0, + 921.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 865.0, + 975.0, + 865.0, + 975.0, + 887.0, + 954.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 881.0, + 384.0, + 881.0, + 384.0, + 907.0, + 355.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 884.0, + 883.0, + 910.0, + 883.0, + 910.0, + 906.0, + 884.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 404.0, + 898.0, + 415.0, + 898.0, + 415.0, + 907.0, + 404.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 930.0, + 898.0, + 940.0, + 898.0, + 940.0, + 906.0, + 930.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 896.0, + 975.0, + 896.0, + 975.0, + 915.0, + 954.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 939.0, + 384.0, + 939.0, + 384.0, + 962.0, + 356.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 421.0, + 924.0, + 452.0, + 924.0, + 452.0, + 987.0, + 421.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 910.0, + 495.0, + 910.0, + 495.0, + 937.0, + 461.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 920.0, + 581.0, + 920.0, + 581.0, + 939.0, + 559.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 927.0, + 614.0, + 927.0, + 614.0, + 948.0, + 585.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 914.0, + 646.0, + 914.0, + 646.0, + 935.0, + 618.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 925.0, + 704.0, + 925.0, + 704.0, + 945.0, + 676.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 708.0, + 933.0, + 763.0, + 933.0, + 763.0, + 959.0, + 708.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 939.0, + 910.0, + 939.0, + 910.0, + 962.0, + 883.0, + 962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 926.0, + 943.0, + 926.0, + 943.0, + 939.0, + 920.0, + 939.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 927.0, + 974.0, + 927.0, + 974.0, + 952.0, + 953.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 910.0, + 1021.0, + 910.0, + 1021.0, + 937.0, + 986.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1084.0, + 920.0, + 1110.0, + 920.0, + 1110.0, + 942.0, + 1084.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1117.0, + 926.0, + 1143.0, + 926.0, + 1143.0, + 947.0, + 1117.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 915.0, + 1174.0, + 915.0, + 1174.0, + 936.0, + 1146.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 927.0, + 1230.0, + 927.0, + 1230.0, + 944.0, + 1206.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1264.0, + 934.0, + 1290.0, + 934.0, + 1290.0, + 958.0, + 1264.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 952.0, + 417.0, + 952.0, + 417.0, + 964.0, + 394.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 956.0, + 586.0, + 956.0, + 586.0, + 981.0, + 555.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 952.0, + 642.0, + 952.0, + 642.0, + 981.0, + 587.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 957.0, + 703.0, + 957.0, + 703.0, + 980.0, + 678.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 948.0, + 957.0, + 976.0, + 957.0, + 976.0, + 986.0, + 948.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1082.0, + 956.0, + 1113.0, + 956.0, + 1113.0, + 981.0, + 1082.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 957.0, + 1172.0, + 957.0, + 1172.0, + 981.0, + 1118.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1175.0, + 956.0, + 1202.0, + 956.0, + 1202.0, + 972.0, + 1175.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 957.0, + 1230.0, + 957.0, + 1230.0, + 981.0, + 1206.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 975.0, + 418.0, + 975.0, + 418.0, + 1010.0, + 389.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 974.0, + 764.0, + 974.0, + 764.0, + 1012.0, + 677.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 963.0, + 805.0, + 963.0, + 805.0, + 994.0, + 767.0, + 994.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 975.0, + 944.0, + 975.0, + 944.0, + 1009.0, + 919.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 980.0, + 1230.0, + 980.0, + 1230.0, + 998.0, + 1207.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 983.0, + 1290.0, + 983.0, + 1290.0, + 1012.0, + 1263.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1294.0, + 966.0, + 1331.0, + 966.0, + 1331.0, + 992.0, + 1294.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 1044.0, + 485.0, + 1044.0, + 485.0, + 1074.0, + 361.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 605.0, + 1046.0, + 753.0, + 1046.0, + 753.0, + 1076.0, + 605.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 1044.0, + 1013.0, + 1044.0, + 1013.0, + 1075.0, + 888.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 1046.0, + 1279.0, + 1046.0, + 1279.0, + 1076.0, + 1131.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1110.0, + 678.0, + 1110.0, + 678.0, + 1127.0, + 648.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1133.0, + 675.0, + 1133.0, + 675.0, + 1141.0, + 657.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1140.0, + 710.0, + 1140.0, + 710.0, + 1164.0, + 685.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 656.0, + 1149.0, + 680.0, + 1149.0, + 680.0, + 1162.0, + 656.0, + 1162.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1160.0, + 644.0, + 1160.0, + 644.0, + 1183.0, + 615.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 614.0, + 1215.0, + 646.0, + 1215.0, + 646.0, + 1240.0, + 614.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 1201.0, + 681.0, + 1201.0, + 681.0, + 1244.0, + 651.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1171.0, + 711.0, + 1171.0, + 711.0, + 1230.0, + 686.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 723.0, + 1187.0, + 756.0, + 1187.0, + 756.0, + 1215.0, + 723.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1183.0, + 994.0, + 1183.0, + 994.0, + 1242.0, + 817.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1215.0, + 1021.0, + 1215.0, + 1021.0, + 1238.0, + 995.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 1202.0, + 1063.0, + 1202.0, + 1063.0, + 1228.0, + 1032.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1251.0, + 682.0, + 1251.0, + 682.0, + 1270.0, + 648.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1235.0, + 711.0, + 1235.0, + 711.0, + 1262.0, + 683.0, + 1262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 1239.0, + 844.0, + 1239.0, + 844.0, + 1266.0, + 814.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1247.0, + 937.0, + 1247.0, + 937.0, + 1269.0, + 908.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1245.0, + 1021.0, + 1245.0, + 1021.0, + 1269.0, + 996.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 1274.0, + 675.0, + 1274.0, + 675.0, + 1285.0, + 655.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1273.0, + 936.0, + 1273.0, + 936.0, + 1295.0, + 909.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 622.0, + 1321.0, + 745.0, + 1321.0, + 745.0, + 1353.0, + 622.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1321.0, + 1012.0, + 1321.0, + 1012.0, + 1355.0, + 866.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1232.75, + 937.5, + 1271.75, + 937.5, + 1271.75, + 959.0, + 1232.75, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 951.0, + 680.0, + 951.0, + 680.0, + 971.5, + 648.0, + 971.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 950.0, + 945.0, + 950.0, + 945.0, + 966.0, + 916.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 974.5, + 1260.0, + 974.5, + 1260.0, + 994.5, + 1229.0, + 994.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2125.0, + 829.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 838.0, + 414.0, + 838.0, + 414.0, + 847.0, + 399.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 839.0, + 942.0, + 839.0, + 942.0, + 849.0, + 926.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 889.0, + 373.0, + 889.0, + 373.0, + 900.0, + 363.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 858.0, + 456.0, + 858.0, + 456.0, + 987.0, + 422.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 861.0, + 978.0, + 861.0, + 978.0, + 923.0, + 949.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 945.0, + 377.0, + 945.0, + 377.0, + 956.0, + 363.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 910.0, + 496.0, + 910.0, + 496.0, + 938.0, + 466.0, + 938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 563.0, + 924.0, + 580.0, + 924.0, + 580.0, + 934.0, + 563.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 931.0, + 613.0, + 931.0, + 613.0, + 945.0, + 588.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 914.0, + 644.0, + 914.0, + 644.0, + 935.0, + 618.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 927.0, + 702.0, + 927.0, + 702.0, + 944.0, + 680.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 941.0, + 734.0, + 941.0, + 734.0, + 956.0, + 709.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 934.0, + 759.0, + 934.0, + 759.0, + 955.0, + 735.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 885.0, + 942.0, + 905.0, + 942.0, + 905.0, + 958.0, + 885.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 934.0, + 970.0, + 934.0, + 970.0, + 945.0, + 958.0, + 945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 912.0, + 1020.0, + 912.0, + 1020.0, + 937.0, + 996.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 919.0, + 1171.0, + 919.0, + 1171.0, + 934.0, + 1147.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1207.0, + 927.0, + 1230.0, + 927.0, + 1230.0, + 944.0, + 1207.0, + 944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 945.0, + 1258.0, + 945.0, + 1258.0, + 956.0, + 1240.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1263.0, + 934.0, + 1288.0, + 934.0, + 1288.0, + 956.0, + 1263.0, + 956.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 951.0, + 419.0, + 951.0, + 419.0, + 988.0, + 393.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 956.0, + 587.0, + 956.0, + 587.0, + 979.0, + 557.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 955.0, + 645.0, + 955.0, + 645.0, + 983.0, + 615.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 652.0, + 958.0, + 672.0, + 958.0, + 672.0, + 967.0, + 652.0, + 967.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 958.0, + 731.0, + 958.0, + 731.0, + 997.0, + 677.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 986.0, + 759.0, + 986.0, + 759.0, + 1009.0, + 735.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 966.0, + 804.0, + 966.0, + 804.0, + 990.0, + 776.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 958.0, + 1171.0, + 958.0, + 1171.0, + 981.0, + 1146.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1174.0, + 953.0, + 1204.0, + 953.0, + 1204.0, + 973.0, + 1174.0, + 973.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 959.0, + 1260.0, + 959.0, + 1260.0, + 997.0, + 1206.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 988.0, + 1285.0, + 988.0, + 1285.0, + 1007.0, + 1266.0, + 1007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 967.0, + 1328.0, + 967.0, + 1328.0, + 990.0, + 1304.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 1043.0, + 488.0, + 1043.0, + 488.0, + 1076.0, + 360.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1046.0, + 755.0, + 1046.0, + 755.0, + 1076.0, + 604.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 1043.0, + 1014.0, + 1043.0, + 1014.0, + 1076.0, + 887.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 1046.0, + 1283.0, + 1046.0, + 1283.0, + 1076.0, + 1132.0, + 1076.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 654.0, + 1111.0, + 677.0, + 1111.0, + 677.0, + 1127.0, + 654.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1142.0, + 709.0, + 1142.0, + 709.0, + 1166.0, + 687.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 660.0, + 1152.0, + 677.0, + 1152.0, + 677.0, + 1163.0, + 660.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1160.0, + 642.0, + 1160.0, + 642.0, + 1183.0, + 615.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 1171.0, + 709.0, + 1171.0, + 709.0, + 1199.0, + 687.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 1188.0, + 754.0, + 1188.0, + 754.0, + 1213.0, + 721.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1185.0, + 846.0, + 1185.0, + 846.0, + 1206.0, + 818.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1185.0, + 875.0, + 1185.0, + 875.0, + 1201.0, + 851.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1190.0, + 909.0, + 1190.0, + 909.0, + 1217.0, + 879.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1217.0, + 641.0, + 1217.0, + 641.0, + 1238.0, + 615.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 1206.0, + 678.0, + 1206.0, + 678.0, + 1244.0, + 650.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1202.0, + 711.0, + 1202.0, + 711.0, + 1265.0, + 681.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 1202.0, + 878.0, + 1202.0, + 878.0, + 1222.0, + 849.0, + 1222.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 879.0, + 1217.0, + 908.0, + 1217.0, + 908.0, + 1238.0, + 879.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1209.0, + 993.0, + 1209.0, + 993.0, + 1244.0, + 933.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 997.0, + 1215.0, + 1023.0, + 1215.0, + 1023.0, + 1248.0, + 997.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1202.0, + 1064.0, + 1202.0, + 1064.0, + 1229.0, + 1037.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1244.0, + 838.0, + 1244.0, + 838.0, + 1259.0, + 818.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1247.0, + 936.0, + 1247.0, + 936.0, + 1269.0, + 908.0, + 1269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 1237.0, + 1023.0, + 1237.0, + 1023.0, + 1271.0, + 996.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1272.0, + 936.0, + 1272.0, + 936.0, + 1294.0, + 908.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 1320.0, + 747.0, + 1320.0, + 747.0, + 1353.0, + 619.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1322.0, + 1013.0, + 1322.0, + 1013.0, + 1354.0, + 866.0, + 1354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 887.5, + 905.0, + 887.5, + 905.0, + 898.0, + 886.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 904.0, + 1193.5, + 933.0, + 1193.5, + 933.0, + 1215.0, + 904.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.25, + 1254.0, + 683.25, + 1254.0, + 683.25, + 1269.5, + 651.25, + 1269.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/QmxFsofRvW9/QmxFsofRvW9.md b/parse/train/QmxFsofRvW9/QmxFsofRvW9.md new file mode 100644 index 0000000000000000000000000000000000000000..d35ed05945a603d699748cad0ae04aec58733c1b --- /dev/null +++ b/parse/train/QmxFsofRvW9/QmxFsofRvW9.md @@ -0,0 +1,287 @@ +# Unsupervised Speech Recognition + +Alexei Baevski4 Wei-Ning Hsu4 Alexis Conneau∗ Michael Auli4 + +4 Facebook AI  Google AI + +# Abstract + +Despite rapid progress in the recent past, current speech recognition systems still require labeled training data which limits this technology to a small fraction of the languages spoken around the globe. This paper describes wav2vec-U, short for wav2vec Unsupervised, a method to train speech recognition models without any labeled data. We leverage self-supervised speech representations to segment unlabeled audio and learn a mapping from these representations to phonemes via adversarial training. The right representations are key to the success of our method. Compared to the best previous unsupervised work, wav2vec-U reduces the phone error rate on the TIMIT benchmark from 26.1 to 11.3. On the larger English Librispeech benchmark, wav2vec-U achieves a word error rate of 5.9 on test-other, rivaling some of the best published systems trained on 960 hours of labeled data from only two years ago. We also experiment on nine other languages, including low-resource languages such as Kyrgyz, Swahili and Tatar. The code is available at https://github.com/pytorch/fairseq/tree/ master/examples/wav2vec/unsupervised + +# 1 Introduction + +Speech recognition performance on the much studied English Librispeech benchmark [Panayotov et al., 2015] has seen rapid improvement over the last few years due to advances in model architectures [Dong et al., 2018, Synnaeve et al., 2020, Gulati et al., 2020], semi-supervised learning [Xu et al., 2020b, Park et al., 2020] and self-supervised learning [van den Oord et al., 2018, Chung and Glass, 2018, Chung et al., 2019b, Baevski et al., 2020c]. However, all of these techniques require transcribed speech data which is not available for the vast majority of the nearly 7,000 languages of the world [Lewis et al., 2016]. As a result, speech recognition technology is only available for about 125 different languages [Google, 2021]. On the other hand, humans learn a lot about speech simply by listening to others around them and without explicit supervision [Werker and Tees, 1984, Hirsh-Pasek et al., 1987, Polka and Werker, 1994, Jusczyk et al., 1999, Johnson and Jusczyk, 2001]. + +Unsupervised learning has been very successful in machine translation resulting in systems that obtain remarkable accuracy given no labeled training data at all [Conneau et al., 2018, Lample et al., 2018, Artetxe et al., 2018]. Inspired by this, there has been some work on unsupervised speech recognition based on learning to align unlabeled text and audio [Yeh et al., 2019] or adversarial learning [Liu et al., 2018, Chen et al., 2019]. These approaches showed promising initial results but their error rates are still high, with evaluation being limited to the small-scale and clean TIMIT benchmark. + +In this work, we introduce a framework for unsupervised learning of speech recognition models. Wav2vec-U, or wav2vec Unsupervised, leverages self-supervised representations from wav2vec 2.0 [Baevski et al., 2020c] to embed the speech audio and to segment the audio into units with a simple $\mathrm { k }$ -means clustering method (see Figure 1 for an illustration of our approach). We find that the quality of the audio representations is key to the success of unsupervised speech recognition. Similar to Liu et al. [2018] and Chen et al. [2019], we learn a mapping between segments and phonemes using adversarial training but different to their work, we also enable the algorithm to label segments as silences. We also introduce an unsupervised cross-validation metric to enable model development without labeled development data. Our unsupervised speech recognition model, the generator, is very lightweight: it consists of a single temporal convolution comprising only about 90k parameters to which we input frozen wav2vec 2.0 representations. + +![](images/0b5b4f3465570676296b5d72eebf6d66830b97659f2ec44658268dac36522d53.jpg) +Figure 1: Illustration of wav2vec Unsupervised: we learn self-supervised representations with wav2vec 2.0 on unlabeled speech audio (Step 1), identify clusters in the representations with $\mathbf { k }$ -means (Step 2) to segment the audio (Step 3). Next, we build segment representations by mean pooling the wav2vec 2.0 representations, performing PCA and a second mean pooling step between adjacent segments (Step 4). This is input to the generator which outputs a phoneme sequence (Step 5) fed to the discriminator, similar to phonemized unlabeled text (Step 6), for adversarial training (Step 7). + +Experimental results demonstrate the viability of the framework for a variety of settings and languages. wav2vec-U improves the phone error rate (PER) on the small-scale TIMIT benchmark from 26.1 to 11.3 compared to the next best known unsupervised approach. To get a better sense of the performance compared to the best supervised methods, we measure performance on the larger Librispeech benchmark where our method achieves word error rate (WER) 5.9 on test-other. We also evaluate on six other European languages of the multilingual Librispeech benchmark [Pratap et al., 2020] and on three non-European low-resource languages. + +# 2 Speech and Text Representations + +Next, we describe how we build suitable speech and text representations for unsupervised learning. +Good representations are essential to learning a mapping from speech to text without supervision. + +# 2.1 Self-supervised Learning of Speech Audio Representations + +In the first step, we learn representations of the speech audio signal using self-supervised learning. There has been a lot of recent work in this direction which has shown strong performance in extremely low-labeled data setups across a range of languages [Conneau et al., 2020] and tasks [Fan et al., 2021, Pepino et al., 2021, Wang et al., 2021]. + +Wav2vec 2.0 consists of a convolutional feature encoder $f : \mathcal X \mapsto \mathcal Z$ that maps a raw audio sequence $X$ to latent speech representations $z _ { 1 } , \dots , z _ { T }$ , which a Transformer $g : { \mathcal { Z } } \mapsto { \mathcal { C } }$ then turns into context representations $c _ { 1 } , \ldots , c _ { T }$ [Baevski et al., 2020b,a]. Each $z _ { t }$ represents about $2 5 \mathrm { m s }$ of audio strided by $2 0 \mathrm { m s }$ and the Transformer architecture follows BERT [Vaswani et al., 2017, Devlin et al., 2019]. During training, latent representations are discretized to $q _ { 1 } , \ldots , q _ { T }$ with a quantization module ${ \mathcal { Z } } \mapsto { \mathcal { Q } }$ to represent the targets in the objective. Quantization uses a Gumbel softmax to choose entries from two codebooks [Jegou et al., 2011, Jang et al., 2016, Baevski et al., 2020b]. + +In our experiments, we use the publicly available English model pre-trained on $5 3 \mathrm { k }$ hours of LibriLight [Kahn et al., 2020b] as well as XLSR-53 which was pre-trained on nearly 60k hours of speech audio in 53 languages [Conneau et al., 2020]. + +# 2.2 Pre-processing and Embedding the Audio Data + +Removing Silences. Most datasets we use for our experiments have audio data with silences. However, these parts of the audio do not correspond to any transcription and we therefore remove silences as much as possible. We apply rVAD, an unsupervised voice activity detection (VAD) model which determines the segments in the audio data corresponding to silences, and we remove these sections [Tan et al., 2020]. We ablate this choice in Appendix C. + +Speech Audio Representations. After silence removal, we embed the unlabeled speech audio with wav2vec 2.0 to obtain speech representations. Specifically, we use the representations of the context Transformer network $c _ { 1 } , \ldots , c _ { T }$ $( \ S 2 . 1 )$ . The context network contains 24 Transformer blocks and we denote the output of block $l$ at time-step $t$ as $c _ { t } ^ { l }$ . Our goal is to learn a model which can map from audio representations $c _ { t } ^ { l }$ to phonemes using no supervision. However, the representations of the uppermost block of wav2vec 2.0 may not be well suited for this task. These features are trained to directly predict masked latent representations spanning $2 5 \mathrm { m s }$ of speech audio which is much shorter than the typical duration of a phoneme. + +To get a better sense of this, we train supervised phoneme recognizers with a CTC loss [Graves et al., 2006] on top of the frozen representations of each of the 24 blocks of the English wav2vec 2.0 LARGE model pre-trained on Libri-Light. We then evaluate phone error rate (PER) with respect to the phonemized transcriptions of Librispeech dev-other. The classifier takes as input $c _ { t } ^ { l }$ and contains a single softmax-normalized linear layer mapping to the phoneme inventory. Figure 2 shows that most of the first ten blocks as well as the final blocks provide very poor performance, while blocks 15-19 provide error rates below $9 \%$ PER. Block 15 achieves the best error rate of $7 . 5 \%$ PER. A similar insight has been used in the concurrent work of Hsu et al. [2021b]. Appendix A shows that this choice generalizes to other languages. For brevity we drop the superscript $l$ and refer to block 15 representations simply as $c _ { 1 } , \ldots , c _ { T }$ . + +![](images/4893afedc781e5fe6ccd0b0db23368e07e04f56d069fce1ae7128f54dac18113.jpg) +Figure 2: Supervised phoneme recognition using representations from different wav2vec 2.0 blocks on dev-other of English Librispeech. Low and high blocks do not provide good features, while as blocks 14-19 do. Block 15 performs best. + +# 2.3 Segmenting the Audio Signal + +Once the speech signal is embedded, we identify segments corresponding to meaningful units that can be mapped to phonemes. Segmentation has been shown to be crucial in prior work [Chung et al., 2018] since the right boundaries in the input representations make it more aligned to phonetic sequences. There has been a lot of prior work in unsupervised speech segmentation [Kamper et al., 2017a,b, Rasanen et al., 2015, Kreuk et al., 2020] but here we simply use a method based on clustering the wav2vec 2.0 speech representations $c _ { 1 } , \ldots , c _ { T }$ . In a first step, we collect all the speech representations for the unlabeled speech data and perform k-means clustering to identify $K = 1 2 8$ clusters. We use the FAISS library to do fast clustering on GPUs [Johnson et al., 2019]. Next, each $c _ { t }$ is labeled with the corresponding cluster ID $i _ { t } \in \{ 1 , \ldots , K \}$ and we introduce speech segment boundaries whenever the cluster ID changes. + +Once the speech audio representations are segmented, we compute a 512-dimensional PCA over all speech representations output by wav2vec 2.0 for the training set. Next, we mean-pool the PCA representations for a particular segment to obtain an average representation of the segment. The PCA retains only the most important features and we found this to be effective. Segment boundaries are noisy due to the lack of supervision and we therefore found it useful to also mean-pool pairs of adjacent segment representations to increase robustness. This results in sequences of speech segment representation $S = s _ { 1 } , \ldots , s _ { T } , S \sim { \mathcal { S } }$ for a given utterance. Appendix B shows an illustration of the segmentation strategy on an actual example as well as a quantitative evaluation of the strategy compared to human segmented data. + +# 2.4 Pre-processing the Text Data + +Similar to how we segment the unlabeled speech audio data into suitable units for unsupervised learning, we do the same for the unlabeled text data. We apply two pre-processing steps to the text data: phonemization and silence token insertion. + +Phonemes characterize the different sounds which distinguish words from each other, e.g., for the word cat there are three phonemes corresponding to the three distinct sounds in the pronunciation of the word: /K/, /AE/, $/ \mathrm { T } / .$ . We phonemize the text data because we found it easier to learn a mapping between speech audio and the different sounds of a word rather than between audio and words or letters. Phonemization converts a sequence of words $Y$ into a sequence of phonemes $P = [ p _ { 1 } , \cdots , p _ { M } ]$ , where $p _ { m } \in O$ and $O$ is the phoneme inventory. We use off-the-shelf tools for this step which we detail in Appendix $\ S \operatorname { E } . 2$ . + +The unlabeled speech audio data is pre-processed by applying unsupervised silence removal. However, this process is not always accurate and many silences in the speech audio remain. To deal with this, we enable the unsupervised model to label some segments with a phonemic silence token (SIL; $\ S \ 3 . 1 \AA .$ ). However, the phonemized unlabeled text data does not contain any silence tokens and this may pose difficulties for adversarial learning $( \ S 3 )$ . We remedy this by inserting silence markers at the beginning and end of the phonemized unlabeled text data; we also randomly insert SIL between words, or groups of phonemes corresponding to words at a rate of $2 5 \%$ . Appendix C evaluates these choices. + +# 3 Unsupervised Learning + +We use adversarial training to train an unsupervised speech recognition model using the representations of the unlabeled speech audio data and the unlabeled phonemized text data [Liu et al., 2018, Chen et al., 2019]. In the following, we detail the model architecture, the training objective as well as the unsupervised cross-validation metric we developed. + +# 3.1 Model Architecture + +Generative adversarial networks (GAN; Goodfellow et al. 2014) train a generator network $\mathcal { G }$ and a discriminator/critic network $\mathcal { C }$ where the generator produces samples which are then judged by the discriminator. The discriminator is trained to classify whether samples are from the generator or from the real data distribution. The objective of the generator is to produce samples that are indistinguishable by the discriminator. + +Concretely, $\mathcal { G }$ takes as input a sequence of $T$ segment representations $S = [ s _ { 1 } , \dotsc , s _ { T } ]$ (§ 2.3) which are then mapped to a sequence of $M$ phonemes $\mathcal { G } ( S ) \bar { = } [ p _ { 1 } , \dotsc , p _ { M } ]$ . The generator predicts a distribution over the phoneme set $O$ for each segment and outputs the phoneme with the highest probability. If the argmax prediction of consecutive segments result in the same phoneme, then we sample one of these segments, therefore $M \leq T$ . + +The phoneme set $O$ includes a silence label SIL to enable labeling silences in the speech audio as such. Without a silence label, we noticed that the model was repurposing a particular phoneme to label silences which resulted in much lower performance since it interfered with subsequent language model (LM) decoding. In the backward pass, we back-propagate through segments sampled at the generator output. We do not modify the segment representations $S$ during unsupervised training. The generator is parameterized as a single layer convolutional neural network (CNN). + +The discriminator takes as input either a sequence $P ^ { r } \sim \mathcal { P } ^ { r }$ of one-hot vectors denoting phonemized text from the real data distribution ${ \mathcal { P } } ^ { r }$ or a sequence of output distributions from the generator $\mathcal { G } ( S )$ . Each input vector has $| O |$ dimensions to represent the distribution over phonemes for each segment. The discriminator is also a CNN which outputs a probability indicating how likely the sample is to be from the data distribution. + +# 3.2 Objective + +In our setup we use the original GAN objective with a gradient penalty [Goodfellow et al., 2014, Arjovsky et al., 2017], a segment smoothness penalty and a phoneme diversity penalty: + +$$ +\operatorname* { m i n } _ { \mathcal { G } } \operatorname* { m a x } _ { \mathcal { C } } \quad \mathbb { E } _ { \mathcal { P } ^ { r } \sim \mathcal { P } ^ { r } } \left[ \log \mathcal { C } ( \boldsymbol { P } ^ { r } ) \right] - \underset { S \sim \mathcal { S } } { \mathbb { E } } \left[ \log \left( 1 - \mathcal { C } ( \mathcal { G } ( S ) ) \right) \right] - \lambda \mathcal { L } _ { g p } + \gamma \mathcal { L } _ { s p } + \eta \mathcal { L } _ { p d } +$$ + +where $P ^ { r } \sim \mathcal { P } ^ { r }$ is phonemized unlabeled text, $\mathcal { G } ( S )$ is the transcription output by the generator of input segment representations $S$ for some unlabeled speech audio. The first term trains the discriminator to assign high probability to real transcriptions, the second term encourages the discriminator to assign low probability to generator outputs, $\mathcal { L } _ { g p }$ is a gradient penalty, $\mathcal { L } _ { s p }$ is a smoothness penalty and $\mathcal { L } _ { p d }$ is a phoneme diversity loss which we detail next. During training we alternate updates for the discriminator and the generator. We also alternate batches of predicted transcriptions from the generator and phonemized unlabeled text. + +Gradient penalty. To stabilize training, we penalize the gradient norm of the discriminator with respect to the input [Gulrajani et al., 2017]. The penalty is computed for random samples $\tilde { P } \sim \tilde { \mathcal { P } }$ which are a linear combination of the activations of pairs of real and fake samples.2 + +$$ +\mathcal { L } _ { g p } = \underset { \tilde { P } \sim \tilde { \mathcal { P } } } { \mathbb { E } } \left[ \left( \| \nabla \mathcal { C } ( \tilde { P } ) \| - 1 \right) ^ { 2 } \right] +$$ + +Segment smoothness penalty. The $\mathbf { k }$ -means segmentation of the speech audio is more granular than a typical phonemized transcription and neighboring representations are highly correlated. We therefore found it useful to add a penalty which encourages the generator to produce similar outputs for adjacent segments where $p _ { t } \in \mathbb { R } ^ { | O | }$ : + +$$ +\mathcal { L } _ { s p } = \sum _ { ( p _ { t } , p _ { t + 1 } ) \in \mathcal { G } ( S ) } \| p _ { t } - p _ { t + 1 } \| ^ { 2 } +$$ + +Phoneme diversity loss. We also found it helpful to penalize low usage of the phoneme vocabulary by the generator on the batch level. In particular, we maximize the entropy of the averaged softmax distribution $H _ { \mathcal { G } } ( \mathcal { G } ( S ) )$ of the generator over the phoneme vocabulary across a batch $B$ of utterances: + +$$ +\mathcal { L } _ { p d } = \frac { 1 } { | B | } \sum _ { S \in B } - H _ { \mathcal { G } } ( \mathcal { G } ( S ) ) +$$ + +# 3.3 Unsupervised Cross-Validation Metric + +Our goal is to build speech recognition models without any supervision. To this end, we developed a cross-validation metric which does not require labeled data. We use the metric for early stopping, selecting a random seed, and hyper-parameter selection $( \lambda , \gamma , \eta )$ . + +We consider two quantities in our metric: LM negative log-likelihood (NLL) and vocabulary usage. LM-NLL serves as an indicator of fluency for a given transcription and it is measured with a language model $p _ { L M }$ trained on phonemized text data $( \ S ~ 2 . 4 )$ . Vocabulary usage is the proportion of the phoneme vocabulary being output by the model via Viterbi decoding. Measuring vocabulary usage identifies degenerate models which output fluent but trivial transcriptions. + +We deaudio rbias criptions for a given generator configuration . LM-NLL is measured in the standard way o $\mathcal { G }$ and unlabeled speechr the phonemized tran$\{ X _ { j } \} _ { j = 1 } ^ { N _ { s } }$ $\mathbf { \bar { \mathcal { P } } } = \{ P _ { j } \} _ { j = 1 } ^ { N _ { s } }$ scriptions: $\begin{array} { r } { N L L _ { L M } ( \mathcal { P } ) = \frac { 1 } { N _ { s } } \sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \end{array}$ where $\begin{array} { r } { N L L _ { L M } ( P ) = - \frac { 1 } { M } \sum _ { t = 1 } ^ { M } \log p _ { L M } ( p _ { t } ) } \end{array}$ using $p _ { L M } ( p _ { t } )$ as shorthand for $p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )$ .3 On the other hand, we use $U ( { \mathcal { P } } ) =$ $\begin{array} { r } { \frac { 1 } { | O | } \bar { \sum _ { o \in O } } [ o \in \mathcal { P } ] \in [ 0 , 1 ] } \end{array}$ to denote the vocabulary usage of $\mathcal { P }$ . + +In a first step, we generate phoneme transcriptions for different training checkpoints or hyperparameter settings and denote the transcriptions of the configuration with the lowest vocabulary-usage adjusted NLL as $\begin{array} { r } { \hat { \mathcal { P } } = \arg \operatorname* { m i n } _ { \mathcal { P } } N L L _ { L M } ( \mathcal { P } ) - \log U ( \mathcal { P } ) } \end{array}$ .4 Next, we discard model configurations which do not satisfy the following using $\hat { \mathcal { P } }$ as the anchor: + +$$ +N L L _ { L M } ( \mathcal { P } ) < N L L _ { L M } ( \hat { \mathcal { P } } ) + \log \left( \frac { U ( \mathcal { P } ) } { U ( \hat { \mathcal { P } } ) } \right) + \log 1 . 2 +$$ + +The second term on the right hand side introduces a margin over the NLL of the anchor transcription $N L L _ { L M } ( \hat { \mathcal { P } } )$ based on the vocabulary usage of $\mathcal { P }$ and $\hat { \mathcal { P } }$ : If $U ( \hat { \mathcal { P } } )$ is much lower compared to $U ( \mathcal { P } )$ , then we allow model configurations which produce transcriptions with higher NLL compared to $\hat { \mathcal { P } }$ . However, if $U ( \hat { \mathcal { P } } )$ is a lot higher than $U ( \mathcal { P } )$ , then the model configuration will not satisfy the constraint. The $\log 1 . 2$ factor serves as another margin allowing checkpoints with slightly worse vocabulary-usage adjusted NLL to be included. + +In a final step, we take into account the length of the transcriptions: out of the configurations ${ \mathcal { P } } ^ { \prime }$ which satisfy the above constraint, we select the one which has the highest sum of log probability without normalizing the length: + +$$ +\mathcal { P } ^ { * } = \arg \operatorname* { m a x } _ { \mathcal { P } ^ { \prime } } \sum _ { j = 1 } ^ { N _ { s } } \sum _ { t = 1 } ^ { M } \log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \dots , p _ { M } ^ { j } ] +$$ + +This selects model configurations which produce phoneme sequences that score high under the language model but are not too long. Appendix D compares accuracy when developing with this metric compared to a labeled development set. + +# 4 Results + +# 4.1 Comparison to Supervised Speech Recognition on Librispeech + +We first test our approach on Librispeech to get a sense of how unsupervised speech recognition compares to the best supervised systems trained on a large amount of labeled data. Librispeech is a standard benchmark in the speech recognition community which provides about 960 hours of transcribed read audiobooks. We use the language modeling data of Librispeech as unlabeled text data for unsupervised training. In Appendix G we show that far less unlabeled text and speech audio are sufficient to reach a similar level of performance. We experiment with the frozen representations of a wav2vec 2.0 LARGE model trained on the $5 3 . 2 \mathrm { k }$ hours of Libri-Light (LL-60k) which we denote as wav2vec-U LARGE. We also consider self-training over three iterations by first training an HMM on the labels generated by the GANm then fine-tuning the original wav2vec 2.0 model on the labels of the HMM for Librispeech followed by then fine-tuning on Libri-Light; Appendix F investigates alternatives. + +wav2vec-U LARGE with self-training (wav2vec- $\mathbf { \partial } . \mathbf { U } + \mathbf { S } \mathbf { T } ,$ ) and a Transformer language model achieves WER 5.9 on test-other, the noisy test set. This shows that unsupervised speech recognition can perform remarkably well compared to the best supervised systems of the recent past on this much studied benchmark. Also, self-training is effective even when the teacher model is unsupervised as per the improvement over GAN training (wav2vec-U). Interestingly, self-training on just Librispeech, or 960 hours of unlabeled speech audio, achieves already very good performance of WER 6.4 on dev-other compared to self-training on all of Libri-Light (53.2k hours) which compares at 6.0 WER. We note that the number of parameters trained during adversarial training is very small: the generator contains only about 90k parameters for a single temporal convolution mapping to the phoneme set from frozen wav2vec 2.0 representations. + +Table 1: WER on Librispeech dev/test sets when using 960 hours of unlabeled audio from Librispeech (LS-960) or $5 3 . 2 \mathrm { k }$ hours from Libri-Light (LL-60k) using representations from wav2vec 2.0 LARGE. Librispeech provides clean dev/test sets which are less challenging than the other sets. We report results for GAN training only (wav2vec-U) and with subsequent self-training (wav2vec- $\mathbf { \partial } . \mathbf { U } + \mathbf { S } \mathbf { T } ,$ . + +
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
+ +# 4.2 Comparison to Prior Unsupervised Work + +Prior work on unsupervised speech recognition focused on the TIMIT benchmark. In order to perform a direct comparison to these approaches, we report results on this benchmark as well. We consider two setups to compare to previous work: in the matched setting, the unlabeled text data is simply the transcriptions of the unlabeled audio data but unpaired. In the unmatched setup, the unlabeled text data does not contain the transcriptions for the audio data which is a more realistic setting. + +We measure performance on the standard Kaldi dev and test sets (core-dev/core-test) as well as a slightly larger version of the test set (all-test) to be able to compare to Liu et al. [2018] and Chen et al. [2019]. Further details of the two setups can be found in Appendix $\ S \operatorname { E } . 1$ . We report performance for wav2vec-U with a 4-gram language model trained on the language modeling data of TIMIT and we also consider self-training (wav2vec- $\mathbf { \partial } . \mathbf { U } + \mathbf { S } \mathbf { T }$ ). + +Table 2 shows that wav2vec-U outperforms prior unsupervised work in both the matched and unmatched settings, reducing PER on all-test in the matched setup by $57 \%$ relative compared to Chen et al. [2019]. Our method has lower performance than the best supervised methods but it performs still very well at PER 12 on core-test in the matched setup compared to PER 8.3 for the state of the art [Baevski et al., 2020c]. + +# 4.3 Performance on non-English languages + +To get a sense of how well the method works on non-English data, we experiment on six languages of the multilingual Librispeech corpus (MLS; Pratap et al. 2020). As baseline we consider the supervised systems of Pratap et al. [2020] trained on between $2 \mathrm { k }$ and 161 hours of labeled data, depending on the language. For adversarial learning we use 100 hours of unlabeled audio data from MLS for every language as well as the MLS language modeling data. As input to wav2vec-U we use the representations from XLSR-53 [Conneau et al., 2020], a wav2vec 2.0 model pre-trained on 53 languages. Table 3 shows that wav2vec-U generalizes across a range of languages. Performance is lower than supervised systems but it shows the viability for other languages. + +Next, we turn to three low-resource languages, Swahili, Kyrgyz, and Tatar. Swahili is an African language, Kyrgyz and Tatar are Turkic languages with only about $4 . 3 \mathrm { m }$ and $5 . 2 \mathrm { m }$ speakers, respectively.5 We use between 1.8 hours (Kyrgyz) and 9.2 hours of unlabeled audio (Swahili), see Appendix $\ S \operatorname { E } . 1$ . + +Table 2: TIMIT Phoneme Error Rate (PER) in comparison to previous work for the matched and unmatched training data setups (Appendix $\ S \operatorname { E . 1 }$ ). PER is measured on the Kaldi dev and test sets (core-dev/core-test) as well as a slightly larger version of the test set (all-test) as used by some of the prior work. $( ^ { * } )$ indicates experiments that do not use the standard split excluding SA utterances. + +
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
+ +Table 3: WER on the Multilingual Librispeech (MLS) dataset using representations from the wav2vec 2.0 XLSR-53 model. We consider German (de), Dutch (nl), French (fr), Spanish (es), Italian (it), Portuguese (pt). + +
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
+ +To compare to prior work, we measure WER for Swahili and PER for Kyrgyz and Tatar. For Tatar and Kyrgyz we opted to use a reduced self-training regime for faster experimental turn-around where we only perform HMM self-training and we expect better performance with the full self-training setup (Appendix F). Table 4 and Table 5 show that wav2vec-U achieves good performance on these low-resource languages compared to previous work that utilized labeled data. We note that for Tatar and Kyrgyz we use a much smaller amount of speech audio than prior work: compared to XLSR-53 we use $1 . 8 \mathrm { h }$ unlabeled data vs 17h of labeled data for Kyrgyz and 4.6h vs. 17h for Tatar. + +# 5 Related Work + +This paper builds on a large body of prior work which includes semi-supervised speech recognition such as self-training [Kahn et al., 2020a, Xu et al., 2020b, Park et al., 2020]. Some of the earliest work in self-supervised learning of speech representations was was done by van den Oord et al. [2018] for phoneme recognition which was simplified in Schneider et al. [2019] who applied it to full speech recognition. Other work includes language model-style pre-training [Chung et al., 2019a] and learning fixed size representations of audio segments [Chung and Glass, 2018]. There is also work on quantization of the continuous speech data [Baevski et al., 2020b,a, Liu et al., 2019, van Niekerk et al., 2020, Baevski et al., 2020c, Hsu et al., 2021b] and on robustness to domain shift [Hsu et al., 2021a], multilingual pre-training [Kawakami et al., 2020, Conneau et al., 2020] as well as combining speech and vision [Harwath et al., 2020]. + +Table 4: PER for low-resource languages, Tatar (tt) and Kyrgyz (ky). + +
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
+ +Table 5: WER for Swahili from the ALFFA corpus. We compare to the supervised baseline of the ALFFA project. + +
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
+ +Learning to map speech to phonemes without supervision using adversarial learning has been explored by Liu et al. [2018] who learn a mapping matrix between segment identifiers and phonemes. However, their work still relied on data segmented into phonemes by human annotators. This has been later extended to use an automatic segmentation [Chen et al., 2019] which is iteratively refined with HMMs. However, cross validation is still performed using labeled data (personal communication with authors). We also explored HMMs to refine segmentation boundaries (Table A2) but did not find it as effective as self-training. Our work is in part inspired by aligning word embedding spaces of different languages [Mikolov et al., 2013, Artetxe et al., 2017, Conneau et al., 2018] and full unsupervised machine translation [Lample et al., 2018, Artetxe et al., 2018, Conneau and Lample, 2019]. + +# 6 Conclusion and Future Work + +wav2vec-U is a framework which enables building speech recognition models without labeled data. It embeds and segments the speech audio with self-supervised representations from wav2vec 2.0, learns a mapping to phonemes with adversarial learning, and cross-validates hyper-parameter choices as well as early stopping with an unsupervised metric. Experiments on the standard Librispeech benchmark show performance close to the state of the art models from only a few years ago, even though these models relied on nearly 1,000 hours of labeled data. + +Compared to the previous best unsupervised speech recognition approach, wav2vec-U reduces TIMIT phone error rate from 26.1 to 11.3. We also demonstrate the viability of our approach on several languages other than English, some of which are low-resource. The ability to build speech recognition models solely from unlabeled speech audio and unlabeled text drastically lowers the effort to build speech technology for many more languages of the world. + +Our approach requires phonemization of the text for the language of interest. Moreover, phonemizers are not available for all languages and this presents a bottleneck. To address this, future work may develop phonemizers for more languages, explore phonemization approaches that generalize across languages, or unsupervised training with graphemic text units such as letters. + +We explored a simple segmentation technique based on self-supervised representations, however, there is a large body of research on segmentation and some of these techniques may lead to improvements over our simple approach [Varadarajan et al., 2008, Zhang and Glass, 2009, Gish et al., 2009, Lee and Glass, 2012, Lee et al., 2015, Ondel et al., 2016, Kamper et al., 2017a,b, Kreuk et al., 2020]. Also, wav2vec 2.0 learns representations for fixed size units with a fixed stride, however, phonemic units are of variable size. Another direction is to learn variable sized representations during pre-training. + +# Acknowledgments and Disclosure of Funding + +We thank Zhouhan Lin for helping with initial explorations in this project, Tatiana Likhomanenko for helpful discussions about self-training, Da-Rong Liu for sharing details to reproduce the setup of Chen et al. [2019], Marc’Aurelio Ranzato for general helpful discussions, and Ruth Kipng’eno, Ruth Ndila Ndeto as well as Mark Mutitu for error analysis of our Swahili model. + +# References + +S. T. Abate, W. Menzel, and B. Tafila. An amharic speech corpus for large vocabulary continuous speech recognition. In Proc. of Interspeech, 2005. +D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro, Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In Proc. of ICML, 2016. +R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. Proc. of LREC, 2020. +M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. Proc. of ICML, 2017. +M. Artetxe, G. Labaka, and E. Agirre. Learning bilingual word embeddings with (almost) no bilingual data. In Proc. of ACL, 2017. +M. Artetxe, G. Labaka, E. Agirre, and K. Cho. Unsupervised neural machine translation. Proc. of ICLR, 2018. +A. Baevski and M. Auli. Adaptive input representations for neural language modeling. In Proc. of ICLR, 2018. +A. Baevski, M. Auli, and A. Mohamed. Effectiveness of self-supervised pre-training for speech recognition. Proc. of ICASSP, 2020a. +A. Baevski, S. Schneider, and M. Auli. vq-wav2vec: Self-supervised learning of discrete speech representations. In Proc. of ICLR, 2020b. +A. Baevski, Y. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Proc. of NeurIPS, 2020c. +L. Besacier, E. Gauthier, M. Mangeot, P. Bretier, P. Bagshaw, O. Rosec, T. Moudenc, F. Pellegrino, S. Voisin, E. Marsico, and P. Nocera. Speech technologies for african languages: example of a multilingual calculator for education. In Proc. of Interspeech, 2015. +D. Can, V. R. Martinez, P. Papadopoulos, and S. S. Narayanan. Pykaldi: A python wrapper for kaldi. In Proc. of ICASSP, 2018. +K.-Y. Chen, C.-P. Tsai, D.-R. Liu, H.-Y. Lee, and L. shan Lee. Completely unsupervised speech recognition by a generative adversarial network harmonized with iteratively refined hidden markov models. In Proc. of Interspeech, 2019. +Y. Chung, W. Weng, S. Tong, and J. R. Glass. Unsupervised cross-modal alignment of speech and text embedding spaces. Proc. of NIPS, 2018. +Y. Chung, W. Hsu, H. Tang, and J. R. Glass. An unsupervised autoregressive model for speech representation learning. Proc. of Interspeech, 2019a. +Y.-A. Chung and J. Glass. Speech2vec: A sequence-to-sequence framework for learning word embeddings from speech. Proc. of Interspeech, 2018. +Y.-A. Chung, W.-N. Hsu, H. Tang, and J. Glass. An unsupervised autoregressive model for speech representation learning. Proc. of Interspeech, 2019b. +A. Conneau and G. Lample. Cross-lingual language model pretraining. Proc. of NeurIPS, 2019. +A. Conneau, G. Lample, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel data. Proc. of ICLR, 2018. +A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli. Unsupervised cross-lingual representation learning for speech recognition. arXiv, abs/2006.13979, 2020. +J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. Proc. of NAACL, 2019. +L. Dong, S. Xu, and B. Xu. Speech-transformer: A no-recurrence sequence-to-sequence model for speech recognition. In Proc. of ICASSP, 2018. +Z. Fan, M. Li, S. Zhou, and B. Xu. Exploring wav2vec 2.0 on speaker verification and language identification. arXiv, 2021. +R. Fer, P. Matejka, F. Grézl, O. Plchot, K. Vesel ˇ y, and J. H. \` Cernock ˇ y. Multilingually trained \` bottleneck features in spoken language recognition. Computer Speech & Language, 46, 2017. +J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium, 1993. +H. Gelas, L. Besacier, and F. Pellegrino. Developments of Swahili resources for an automatic speech recognition system. In Proc. of SLTU, 2012. +H. Gish, M. Siu, A. Chan, and W. Belfield. Unsupervised training of an hmm-based speech recognizer for topic classification. In Proc. of Interspeech, 2009. +I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks. Proc. of NIPS, 2014. +Google. Google cloud: Speech-to-text. https://cloud.google.com/speech-to-text, 2021. Accessed: 2021-05-13. +A. Graves, S. Fernández, and F. Gomez. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. In Proc. of ICML, 2006. +A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wu, and R. Pang. Conformer: Convolution-augmented transformer for speech recognition. Proc. of Interspeech, 2020. +I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein gans. Proc. of NIPS, 2017. +W. Han, Z. Zhang, Y. Zhang, J. Yu, C.-C. Chiu, J. Qin, A. Gulati, R. Pang, and Y. Wu. Contextnet: Improving convolutional neural networks for automatic speech recognition with global context. Proc. of Interspeech, 2020. +D. Harwath and J. Glass. Towards visually grounded sub-word speech unit discovery. In Proc. of ICASSP, pages 3017–3021. IEEE, 2019. +D. Harwath, W.-N. Hsu, and J. Glass. Learning hierarchical discrete linguistic units from visuallygrounded speech. In Proc. of ICLR, 2020. +K. Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July 2011. Association for Computational Linguistics. +K. Hirsh-Pasek, D. G. Kemler Nelson, P. W. Jusczyk, K. W. Cassidy, B. Druss, and L. Kennedy. Clauses are perceptual units for young infants. Cognition, 26(3):269–286, 1987. +W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V. Pratap, J. Kahn, A. Lee, R. Collobert, G. Synnaeve, et al. Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training. arXiv preprint arXiv:2104.01027, 2021a. +W.-N. Hsu, Y.-H. H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mohamed. Hubert: How much can a bad teacher benefit ASR pre-training? In Proc. of ICASSP, 2021b. +E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. Proc. of ICLR, 2016. +H. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Trans. Pattern Anal. Mach. Intell., 33(1):117–128, Jan. 2011. +E. K. Johnson and P. W. Jusczyk. Word segmentation by 8-month-olds: When speech cues count more than statistics. Journal of Memory and Language, 44(4):548–567, 2001. +J. Johnson, M. Douze, and H. Jégou. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 2019. +P. W. Jusczyk, D. M. Houston, and M. Newsome. The beginnings of word segmentation in englishlearning infants. Cognitive Psychology, 39(3):159–207, 1999. +J. Kahn, A. Lee, and A. Hannun. Self-training for end-to-end speech recognition. In Proc. of ICASSP, 2020a. +J. Kahn et al. Libri-light: A benchmark for asr with limited or no supervision. In Proc. of ICASSP, 2020b. +H. Kamper, A. Jansen, and S. Goldwater. A segmental framework for fully-unsupervised largevocabulary speech recognition. Comput. Speech Lang., 46(C), Nov. 2017a. +H. Kamper, K. Livescu, and S. Goldwater. An embedded segmental k-means model for unsupervised segmentation and clustering of speech. Proc. of ASRU, 2017b. +K. Kawakami, L. Wang, C. Dyer, P. Blunsom, and A. van den Oord. Learning robust and multilingual speech representations. Proc. of EMNLP, 2020. +D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of ICLR, 2015. +F. Kreuk, J. Keshet, and Y. Adi. Self-supervised contrastive learning for unsupervised phoneme segmentation. Proc. of Interspeech, 2020. +G. Lample, L. Denoyer, and M. Ranzato. Unsupervised machine translation using monolingual corpora only. In Proc. of ICLR, 2018. +C. Lee and J. R. Glass. A nonparametric bayesian approach to acoustic model discovery. In Proc. of ACL, 2012. +C. Lee, T. J. O’Donnell, and J. R. Glass. Unsupervised lexicon discovery from acoustic input. TACL, 2015. +M. P. Lewis, G. F. Simon, and C. D. Fennig. Ethnologue: Languages of the world, nineteenth edition. Online version: http://www.ethnologue.com, 2016. +T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert. slimipl: Language-model-free iterative pseudo-labeling. arXiv, 2021. +A. H. Liu, T. Tu, H. yi Lee, and L. shan Lee. Towards unsupervised speech recognition and synthesis with quantized speech representation learning. Proc. of ICASSP, 2019. +D.-R. Liu, K.-Y. Chen, H.-Y. Lee, and L. shan Lee. Completely unsupervised phoneme recognition by adversarially learning mapping relationships from audio embeddings. Proc. of Interspeech, 2018. +T. Mikolov, Q. V. Le, and I. Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. +M. Mohri, F. Pereira, and M. Riley. Weighted finite-state transducers in speech recognition. Computer Speech & Language, 16(1):69–88, 2002. +L. Ondel, L. Burget, and J. Cernocký. Variational inference for acoustic unit discovery. In Proc. of SLTU, 2016. +M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, 2019. +V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public domain audio books. In Proc. of ICASSP, pages 5206–5210. IEEE, 2015. +D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le. Specaugment: A simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, 2019. +D. S. Park, Y. Zhang, Y. Jia, W. Han, C.-C. Chiu, B. Li, Y. Wu, and Q. V. Le. Improved noisy student training for automatic speech recognition. Proc. of Interspeech, 2020. +K. Park and J. Kim. g2pe. https://github.com/Kyubyong/g2p, 2019. +L. Pepino, P. Riera, and L. Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings. arXiv, 2021. +L. Polka and J. F. Werker. Developmental changes in perception of nonnative vowel contrasts. Journal of Experimental Psychology: Human perception and performance, 20(2):421, 1994. +D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech recognition toolkit. In Proc. of ASRU, 2011. +V. Pratap, A. Hannun, Q. Xu, J. Cai, J. Kahn, G. Synnaeve, V. Liptchinsky, and R. Collobert. Wav2letter++: A fast open-source speech recognition system. In Proc. of ICASSP, 2019. +V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert. Mls: A large-scale multilingual dataset for speech research. In Proc. of Interspeech, 2020. +O. Rasanen, G. Doyle, and M. C. Frank. Unsupervised word discovery from speech using automatic segmentation into syllable-like units. In Proc. of Interspeech, 2015. +M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio. Light gated recurrent units for speech recognition. IEEE Trans. on Emerging Topics in Comp. Intel., 2, 2018. +M. Ravanelli, T. Parcollet, and Y. Bengio. The pytorch-kaldi speech recognition toolkit. Proc. of ICASSP, 2019. +M. Rivière, A. Joulin, P.-E. Mazaré, and E. Dupoux. Unsupervised pretraining transfers well across languages. In Proc. of ICASSP, 2020. +S. Schneider, A. Baevski, R. Collobert, and M. Auli. wav2vec: Unsupervised pre-training for speech recognition. In Proc. of Interspeech, 2019. +N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. +G. Synnaeve, Q. Xu, J. Kahn, T. Likhomanenko, E. Grave, V. Pratap, A. Sriram, V. Liptchinsky, and R. Collobert. End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern Architectures. Proc. of ICML workshop on Self-supervision in Audio and Speech (SAS), 2020. +M. Tachbelie, S. T. Abate, and L. Besacier. Using different acoustic, lexical and language modeling units for asr of an under-resourced language - amharic. Speech Communication, 56, 2014. +Z. Tan, A. K. Sarkar, and N. Dehak. rvad: An unsupervised segment-based robust voice activity detection method. Computer speech & language, 59:1–21, 2020. +A. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding. Proc. of NIPS, 2018. +B. van Niekerk, L. Nortje, and H. Kamper. Vector-quantized neural networks for acoustic unit discovery in the zerospeech 2020 challenge. Proc. of Interspeech, 2020. +B. Varadarajan, S. Khudanpur, and E. Dupoux. Unsupervised learning of acoustic sub-word units. In Proc. of ACL, 2008. +A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In Proc. of NIPS, 2017. +C. Wang, A. Wu, J. Pino, A. Baevski, M. Auli, and A. Conneau. Large-scale self- and semi-supervised learning for speech translation. arXiv, 2021. +J. F. Werker and R. C. Tees. Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant behavior and development, 7(1):49–63, 1984. +H. Xu, K. Li, Y. Wang, J. Wang, S. Kang, X. Chen, D. Povey, and S. Khudanpur. Neural network language modeling with letter-based features and importance sampling. In Proc. of ICASSP, 2018. +Q. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and M. Auli. Self-training and pre-training are complementary for speech recognition. In Proc. of ICASSP, 2020a. +Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert. Iterative pseudolabeling for speech recognition. Proc. of Interspeech, 2020b. +C.-K. Yeh, J. Chen, C. Yu, and D. Yu. Unsupervised speech recognition via segmental empirical output distribution matching. In Proc. of ICLR, 2019. +N. Zeghidour, Q. Xu, V. Liptchinsky, N. Usunier, G. Synnaeve, and R. Collobert. Fully convolutional speech recognition. arXiv, abs/1812.06864, 2018. +F. Zhang, Y. Wang, X. Zhang, C. Liu, Y. Saraf, and G. Zweig. Faster, simpler and more accurate hybrid asr systems using wordpieces. Proc. of Interspeech, 2020a. +Y. Zhang and J. R. Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian posteriorgrams. IEEE Workshop on Automatic Speech Recognition & Understanding, 2009. +Y. Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V. Le, and Y. Wu. Pushing the limits of semi-supervised learning for automatic speech recognition. Proc. of NeurIPS SAS Workshop, 2020b. \ No newline at end of file diff --git a/parse/train/QmxFsofRvW9/QmxFsofRvW9_content_list.json b/parse/train/QmxFsofRvW9/QmxFsofRvW9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2742a9f4a400c5e61b5d6372fbe5fc06e7362cd7 --- /dev/null +++ b/parse/train/QmxFsofRvW9/QmxFsofRvW9_content_list.json @@ -0,0 +1,1090 @@ +[ + { + "type": "text", + "text": "Unsupervised Speech Recognition ", + "text_level": 1, + "bbox": [ + 292, + 122, + 705, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexei Baevski4 Wei-Ning Hsu4 Alexis Conneau\u0003∗ Michael Auli4 ", + "bbox": [ + 243, + 199, + 750, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Facebook AI \u0003 Google AI ", + "bbox": [ + 400, + 234, + 598, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract ", + "text_level": 1, + "bbox": [ + 462, + 286, + 535, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite rapid progress in the recent past, current speech recognition systems still require labeled training data which limits this technology to a small fraction of the languages spoken around the globe. This paper describes wav2vec-U, short for wav2vec Unsupervised, a method to train speech recognition models without any labeled data. We leverage self-supervised speech representations to segment unlabeled audio and learn a mapping from these representations to phonemes via adversarial training. The right representations are key to the success of our method. Compared to the best previous unsupervised work, wav2vec-U reduces the phone error rate on the TIMIT benchmark from 26.1 to 11.3. On the larger English Librispeech benchmark, wav2vec-U achieves a word error rate of 5.9 on test-other, rivaling some of the best published systems trained on 960 hours of labeled data from only two years ago. We also experiment on nine other languages, including low-resource languages such as Kyrgyz, Swahili and Tatar. The code is available at https://github.com/pytorch/fairseq/tree/ master/examples/wav2vec/unsupervised ", + "bbox": [ + 233, + 319, + 766, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Introduction ", + "text_level": 1, + "bbox": [ + 174, + 553, + 310, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speech recognition performance on the much studied English Librispeech benchmark [Panayotov et al., 2015] has seen rapid improvement over the last few years due to advances in model architectures [Dong et al., 2018, Synnaeve et al., 2020, Gulati et al., 2020], semi-supervised learning [Xu et al., 2020b, Park et al., 2020] and self-supervised learning [van den Oord et al., 2018, Chung and Glass, 2018, Chung et al., 2019b, Baevski et al., 2020c]. However, all of these techniques require transcribed speech data which is not available for the vast majority of the nearly 7,000 languages of the world [Lewis et al., 2016]. As a result, speech recognition technology is only available for about 125 different languages [Google, 2021]. On the other hand, humans learn a lot about speech simply by listening to others around them and without explicit supervision [Werker and Tees, 1984, Hirsh-Pasek et al., 1987, Polka and Werker, 1994, Jusczyk et al., 1999, Johnson and Jusczyk, 2001]. ", + "bbox": [ + 174, + 584, + 825, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised learning has been very successful in machine translation resulting in systems that obtain remarkable accuracy given no labeled training data at all [Conneau et al., 2018, Lample et al., 2018, Artetxe et al., 2018]. Inspired by this, there has been some work on unsupervised speech recognition based on learning to align unlabeled text and audio [Yeh et al., 2019] or adversarial learning [Liu et al., 2018, Chen et al., 2019]. These approaches showed promising initial results but their error rates are still high, with evaluation being limited to the small-scale and clean TIMIT benchmark. ", + "bbox": [ + 174, + 729, + 825, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce a framework for unsupervised learning of speech recognition models. Wav2vec-U, or wav2vec Unsupervised, leverages self-supervised representations from wav2vec 2.0 [Baevski et al., 2020c] to embed the speech audio and to segment the audio into units with a simple $\\mathrm { k }$ -means clustering method (see Figure 1 for an illustration of our approach). We find that the quality of the audio representations is key to the success of unsupervised speech recognition. Similar to Liu et al. [2018] and Chen et al. [2019], we learn a mapping between segments and phonemes using adversarial training but different to their work, we also enable the algorithm to label segments as silences. We also introduce an unsupervised cross-validation metric to enable model development without labeled development data. Our unsupervised speech recognition model, the generator, is very lightweight: it consists of a single temporal convolution comprising only about 90k parameters to which we input frozen wav2vec 2.0 representations. ", + "bbox": [ + 174, + 819, + 825, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b5b4f3465570676296b5d72eebf6d66830b97659f2ec44658268dac36522d53.jpg", + "image_caption": [ + "Figure 1: Illustration of wav2vec Unsupervised: we learn self-supervised representations with wav2vec 2.0 on unlabeled speech audio (Step 1), identify clusters in the representations with $\\mathbf { k }$ -means (Step 2) to segment the audio (Step 3). Next, we build segment representations by mean pooling the wav2vec 2.0 representations, performing PCA and a second mean pooling step between adjacent segments (Step 4). This is input to the generator which outputs a phoneme sequence (Step 5) fed to the discriminator, similar to phonemized unlabeled text (Step 6), for adversarial training (Step 7). " + ], + "image_footnote": [], + "bbox": [ + 187, + 90, + 807, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 443, + 825, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimental results demonstrate the viability of the framework for a variety of settings and languages. wav2vec-U improves the phone error rate (PER) on the small-scale TIMIT benchmark from 26.1 to 11.3 compared to the next best known unsupervised approach. To get a better sense of the performance compared to the best supervised methods, we measure performance on the larger Librispeech benchmark where our method achieves word error rate (WER) 5.9 on test-other. We also evaluate on six other European languages of the multilingual Librispeech benchmark [Pratap et al., 2020] and on three non-European low-resource languages. ", + "bbox": [ + 174, + 546, + 826, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 Speech and Text Representations ", + "text_level": 1, + "bbox": [ + 174, + 667, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next, we describe how we build suitable speech and text representations for unsupervised learning. \nGood representations are essential to learning a mapping from speech to text without supervision. ", + "bbox": [ + 173, + 703, + 825, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 Self-supervised Learning of Speech Audio Representations ", + "text_level": 1, + "bbox": [ + 173, + 751, + 619, + 767 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first step, we learn representations of the speech audio signal using self-supervised learning. There has been a lot of recent work in this direction which has shown strong performance in extremely low-labeled data setups across a range of languages [Conneau et al., 2020] and tasks [Fan et al., 2021, Pepino et al., 2021, Wang et al., 2021]. ", + "bbox": [ + 174, + 779, + 825, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Wav2vec 2.0 consists of a convolutional feature encoder $f : \\mathcal X \\mapsto \\mathcal Z$ that maps a raw audio sequence $X$ to latent speech representations $z _ { 1 } , \\dots , z _ { T }$ , which a Transformer $g : { \\mathcal { Z } } \\mapsto { \\mathcal { C } }$ then turns into context representations $c _ { 1 } , \\ldots , c _ { T }$ [Baevski et al., 2020b,a]. Each $z _ { t }$ represents about $2 5 \\mathrm { m s }$ of audio strided by $2 0 \\mathrm { m s }$ and the Transformer architecture follows BERT [Vaswani et al., 2017, Devlin et al., 2019]. During training, latent representations are discretized to $q _ { 1 } , \\ldots , q _ { T }$ with a quantization module ${ \\mathcal { Z } } \\mapsto { \\mathcal { Q } }$ to represent the targets in the objective. Quantization uses a Gumbel softmax to choose entries from two codebooks [Jegou et al., 2011, Jang et al., 2016, Baevski et al., 2020b]. ", + "bbox": [ + 174, + 842, + 825, + 911 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 92, + 823, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our experiments, we use the publicly available English model pre-trained on $5 3 \\mathrm { k }$ hours of LibriLight [Kahn et al., 2020b] as well as XLSR-53 which was pre-trained on nearly 60k hours of speech audio in 53 languages [Conneau et al., 2020]. ", + "bbox": [ + 176, + 126, + 823, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 Pre-processing and Embedding the Audio Data ", + "text_level": 1, + "bbox": [ + 174, + 184, + 540, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Removing Silences. Most datasets we use for our experiments have audio data with silences. However, these parts of the audio do not correspond to any transcription and we therefore remove silences as much as possible. We apply rVAD, an unsupervised voice activity detection (VAD) model which determines the segments in the audio data corresponding to silences, and we remove these sections [Tan et al., 2020]. We ablate this choice in Appendix C. ", + "bbox": [ + 174, + 209, + 825, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Speech Audio Representations. After silence removal, we embed the unlabeled speech audio with wav2vec 2.0 to obtain speech representations. Specifically, we use the representations of the context Transformer network $c _ { 1 } , \\ldots , c _ { T }$ $( \\ S 2 . 1 )$ . The context network contains 24 Transformer blocks and we denote the output of block $l$ at time-step $t$ as $c _ { t } ^ { l }$ . Our goal is to learn a model which can map from audio representations $c _ { t } ^ { l }$ to phonemes using no supervision. However, the representations of the uppermost block of wav2vec 2.0 may not be well suited for this task. These features are trained to directly predict masked latent representations spanning $2 5 \\mathrm { m s }$ of speech audio which is much shorter than the typical duration of a phoneme. ", + "bbox": [ + 173, + 292, + 825, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To get a better sense of this, we train supervised phoneme recognizers with a CTC loss [Graves et al., 2006] on top of the frozen representations of each of the 24 blocks of the English wav2vec 2.0 LARGE model pre-trained on Libri-Light. We then evaluate phone error rate (PER) with respect to the phonemized transcriptions of Librispeech dev-other. The classifier takes as input $c _ { t } ^ { l }$ and contains a single softmax-normalized linear layer mapping to the phoneme inventory. Figure 2 shows that most of the first ten blocks as well as the final blocks provide very poor performance, while blocks 15-19 provide error rates below $9 \\%$ PER. Block 15 achieves the best error rate of $7 . 5 \\%$ PER. A similar insight has been used in the concurrent work of Hsu et al. [2021b]. Appendix A shows that this choice generalizes to other languages. For brevity we drop the superscript $l$ and refer to block 15 representations simply as $c _ { 1 } , \\ldots , c _ { T }$ . ", + "bbox": [ + 173, + 411, + 825, + 550 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4893afedc781e5fe6ccd0b0db23368e07e04f56d069fce1ae7128f54dac18113.jpg", + "image_caption": [ + "Figure 2: Supervised phoneme recognition using representations from different wav2vec 2.0 blocks on dev-other of English Librispeech. Low and high blocks do not provide good features, while as blocks 14-19 do. Block 15 performs best. " + ], + "image_footnote": [], + "bbox": [ + 176, + 565, + 803, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3 Segmenting the Audio Signal ", + "text_level": 1, + "bbox": [ + 176, + 747, + 413, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the speech signal is embedded, we identify segments corresponding to meaningful units that can be mapped to phonemes. Segmentation has been shown to be crucial in prior work [Chung et al., 2018] since the right boundaries in the input representations make it more aligned to phonetic sequences. There has been a lot of prior work in unsupervised speech segmentation [Kamper et al., 2017a,b, Rasanen et al., 2015, Kreuk et al., 2020] but here we simply use a method based on clustering the wav2vec 2.0 speech representations $c _ { 1 } , \\ldots , c _ { T }$ . In a first step, we collect all the speech representations for the unlabeled speech data and perform k-means clustering to identify $K = 1 2 8$ clusters. We use the FAISS library to do fast clustering on GPUs [Johnson et al., 2019]. Next, each $c _ { t }$ is labeled with the corresponding cluster ID $i _ { t } \\in \\{ 1 , \\ldots , K \\}$ and we introduce speech segment boundaries whenever the cluster ID changes. ", + "bbox": [ + 174, + 772, + 825, + 911 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the speech audio representations are segmented, we compute a 512-dimensional PCA over all speech representations output by wav2vec 2.0 for the training set. Next, we mean-pool the PCA representations for a particular segment to obtain an average representation of the segment. The PCA retains only the most important features and we found this to be effective. Segment boundaries are noisy due to the lack of supervision and we therefore found it useful to also mean-pool pairs of adjacent segment representations to increase robustness. This results in sequences of speech segment representation $S = s _ { 1 } , \\ldots , s _ { T } , S \\sim { \\mathcal { S } }$ for a given utterance. Appendix B shows an illustration of the segmentation strategy on an actual example as well as a quantitative evaluation of the strategy compared to human segmented data. ", + "bbox": [ + 173, + 90, + 825, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4 Pre-processing the Text Data ", + "text_level": 1, + "bbox": [ + 176, + 237, + 411, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to how we segment the unlabeled speech audio data into suitable units for unsupervised learning, we do the same for the unlabeled text data. We apply two pre-processing steps to the text data: phonemization and silence token insertion. ", + "bbox": [ + 176, + 263, + 823, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Phonemes characterize the different sounds which distinguish words from each other, e.g., for the word cat there are three phonemes corresponding to the three distinct sounds in the pronunciation of the word: /K/, /AE/, $/ \\mathrm { T } / .$ . We phonemize the text data because we found it easier to learn a mapping between speech audio and the different sounds of a word rather than between audio and words or letters. Phonemization converts a sequence of words $Y$ into a sequence of phonemes $P = [ p _ { 1 } , \\cdots , p _ { M } ]$ , where $p _ { m } \\in O$ and $O$ is the phoneme inventory. We use off-the-shelf tools for this step which we detail in Appendix $\\ S \\operatorname { E } . 2$ . ", + "bbox": [ + 174, + 313, + 825, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The unlabeled speech audio data is pre-processed by applying unsupervised silence removal. However, this process is not always accurate and many silences in the speech audio remain. To deal with this, we enable the unsupervised model to label some segments with a phonemic silence token (SIL; $\\ S \\ 3 . 1 \\AA .$ ). However, the phonemized unlabeled text data does not contain any silence tokens and this may pose difficulties for adversarial learning $( \\ S 3 )$ . We remedy this by inserting silence markers at the beginning and end of the phonemized unlabeled text data; we also randomly insert SIL between words, or groups of phonemes corresponding to words at a rate of $2 5 \\%$ . Appendix C evaluates these choices. ", + "bbox": [ + 174, + 416, + 826, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 Unsupervised Learning ", + "text_level": 1, + "bbox": [ + 176, + 536, + 401, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use adversarial training to train an unsupervised speech recognition model using the representations of the unlabeled speech audio data and the unlabeled phonemized text data [Liu et al., 2018, Chen et al., 2019]. In the following, we detail the model architecture, the training objective as well as the unsupervised cross-validation metric we developed. ", + "bbox": [ + 174, + 571, + 826, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 Model Architecture ", + "text_level": 1, + "bbox": [ + 174, + 648, + 349, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generative adversarial networks (GAN; Goodfellow et al. 2014) train a generator network $\\mathcal { G }$ and a discriminator/critic network $\\mathcal { C }$ where the generator produces samples which are then judged by the discriminator. The discriminator is trained to classify whether samples are from the generator or from the real data distribution. The objective of the generator is to produce samples that are indistinguishable by the discriminator. ", + "bbox": [ + 174, + 676, + 825, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, $\\mathcal { G }$ takes as input a sequence of $T$ segment representations $S = [ s _ { 1 } , \\dotsc , s _ { T } ]$ (§ 2.3) which are then mapped to a sequence of $M$ phonemes $\\mathcal { G } ( S ) \\bar { = } [ p _ { 1 } , \\dotsc , p _ { M } ]$ . The generator predicts a distribution over the phoneme set $O$ for each segment and outputs the phoneme with the highest probability. If the argmax prediction of consecutive segments result in the same phoneme, then we sample one of these segments, therefore $M \\leq T$ . ", + "bbox": [ + 174, + 752, + 825, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The phoneme set $O$ includes a silence label SIL to enable labeling silences in the speech audio as such. Without a silence label, we noticed that the model was repurposing a particular phoneme to label silences which resulted in much lower performance since it interfered with subsequent language model (LM) decoding. In the backward pass, we back-propagate through segments sampled at the generator output. We do not modify the segment representations $S$ during unsupervised training. The generator is parameterized as a single layer convolutional neural network (CNN). ", + "bbox": [ + 174, + 827, + 823, + 911 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The discriminator takes as input either a sequence $P ^ { r } \\sim \\mathcal { P } ^ { r }$ of one-hot vectors denoting phonemized text from the real data distribution ${ \\mathcal { P } } ^ { r }$ or a sequence of output distributions from the generator $\\mathcal { G } ( S )$ . Each input vector has $| O |$ dimensions to represent the distribution over phonemes for each segment. The discriminator is also a CNN which outputs a probability indicating how likely the sample is to be from the data distribution. ", + "bbox": [ + 173, + 90, + 826, + 161 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2 Objective ", + "text_level": 1, + "bbox": [ + 174, + 176, + 281, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our setup we use the original GAN objective with a gradient penalty [Goodfellow et al., 2014, Arjovsky et al., 2017], a segment smoothness penalty and a phoneme diversity penalty: ", + "bbox": [ + 176, + 202, + 825, + 231 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a15593f8fdca0cdc2e5a2bf415cdd52fd286427c617f23aeff5dbf4f23d44a11.jpg", + "text": "$$\n\\operatorname* { m i n } _ { \\mathcal { G } } \\operatorname* { m a x } _ { \\mathcal { C } } \\quad \\mathbb { E } _ { \\mathcal { P } ^ { r } \\sim \\mathcal { P } ^ { r } } \\left[ \\log \\mathcal { C } ( \\boldsymbol { P } ^ { r } ) \\right] - \\underset { S \\sim \\mathcal { S } } { \\mathbb { E } } \\left[ \\log \\left( 1 - \\mathcal { C } ( \\mathcal { G } ( S ) ) \\right) \\right] - \\lambda \\mathcal { L } _ { g p } + \\gamma \\mathcal { L } _ { s p } + \\eta \\mathcal { L } _ { p d }\n$$", + "text_format": "latex", + "bbox": [ + 222, + 236, + 774, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P ^ { r } \\sim \\mathcal { P } ^ { r }$ is phonemized unlabeled text, $\\mathcal { G } ( S )$ is the transcription output by the generator of input segment representations $S$ for some unlabeled speech audio. The first term trains the discriminator to assign high probability to real transcriptions, the second term encourages the discriminator to assign low probability to generator outputs, $\\mathcal { L } _ { g p }$ is a gradient penalty, $\\mathcal { L } _ { s p }$ is a smoothness penalty and $\\mathcal { L } _ { p d }$ is a phoneme diversity loss which we detail next. During training we alternate updates for the discriminator and the generator. We also alternate batches of predicted transcriptions from the generator and phonemized unlabeled text. ", + "bbox": [ + 173, + 266, + 825, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gradient penalty. To stabilize training, we penalize the gradient norm of the discriminator with respect to the input [Gulrajani et al., 2017]. The penalty is computed for random samples $\\tilde { P } \\sim \\tilde { \\mathcal { P } }$ which are a linear combination of the activations of pairs of real and fake samples.2 ", + "bbox": [ + 173, + 377, + 825, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/0a8f6af23cdb1998f6ef4e92c36819c82eb870dd5b721a5d19fc4f0384f93881.jpg", + "text": "$$\n\\mathcal { L } _ { g p } = \\underset { \\tilde { P } \\sim \\tilde { \\mathcal { P } } } { \\mathbb { E } } \\left[ \\left( \\| \\nabla \\mathcal { C } ( \\tilde { P } ) \\| - 1 \\right) ^ { 2 } \\right]\n$$", + "text_format": "latex", + "bbox": [ + 387, + 428, + 609, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Segment smoothness penalty. The $\\mathbf { k }$ -means segmentation of the speech audio is more granular than a typical phonemized transcription and neighboring representations are highly correlated. We therefore found it useful to add a penalty which encourages the generator to produce similar outputs for adjacent segments where $p _ { t } \\in \\mathbb { R } ^ { | O | }$ : ", + "bbox": [ + 173, + 474, + 825, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/dc18aa35eab1ad24f09a6f15551a94289a62fd2eada4a74b608bff6bec0d7c91.jpg", + "text": "$$\n\\mathcal { L } _ { s p } = \\sum _ { ( p _ { t } , p _ { t + 1 } ) \\in \\mathcal { G } ( S ) } \\| p _ { t } - p _ { t + 1 } \\| ^ { 2 }\n$$", + "text_format": "latex", + "bbox": [ + 383, + 537, + 614, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Phoneme diversity loss. We also found it helpful to penalize low usage of the phoneme vocabulary by the generator on the batch level. In particular, we maximize the entropy of the averaged softmax distribution $H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )$ of the generator over the phoneme vocabulary across a batch $B$ of utterances: ", + "bbox": [ + 173, + 585, + 825, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/feda12d555ac0640b6d3cc1aeab0e1b03d7947fe4dd41ee8f6f0dbf28f5d27c6.jpg", + "text": "$$\n\\mathcal { L } _ { p d } = \\frac { 1 } { | B | } \\sum _ { S \\in B } - H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )\n$$", + "text_format": "latex", + "bbox": [ + 401, + 633, + 596, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3 Unsupervised Cross-Validation Metric ", + "text_level": 1, + "bbox": [ + 174, + 686, + 480, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to build speech recognition models without any supervision. To this end, we developed a cross-validation metric which does not require labeled data. We use the metric for early stopping, selecting a random seed, and hyper-parameter selection $( \\lambda , \\gamma , \\eta )$ . ", + "bbox": [ + 173, + 712, + 826, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider two quantities in our metric: LM negative log-likelihood (NLL) and vocabulary usage. LM-NLL serves as an indicator of fluency for a given transcription and it is measured with a language model $p _ { L M }$ trained on phonemized text data $( \\ S ~ 2 . 4 )$ . Vocabulary usage is the proportion of the phoneme vocabulary being output by the model via Viterbi decoding. Measuring vocabulary usage identifies degenerate models which output fluent but trivial transcriptions. ", + "bbox": [ + 173, + 760, + 825, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We deaudio rbias criptions for a given generator configuration . LM-NLL is measured in the standard way o $\\mathcal { G }$ and unlabeled speechr the phonemized tran$\\{ X _ { j } \\} _ { j = 1 } ^ { N _ { s } }$ $\\mathbf { \\bar { \\mathcal { P } } } = \\{ P _ { j } \\} _ { j = 1 } ^ { N _ { s } }$ scriptions: $\\begin{array} { r } { N L L _ { L M } ( \\mathcal { P } ) = \\frac { 1 } { N _ { s } } \\sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \\end{array}$ where $\\begin{array} { r } { N L L _ { L M } ( P ) = - \\frac { 1 } { M } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ) } \\end{array}$ using $p _ { L M } ( p _ { t } )$ as shorthand for $p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )$ .3 On the other hand, we use $U ( { \\mathcal { P } } ) =$ $\\begin{array} { r } { \\frac { 1 } { | O | } \\bar { \\sum _ { o \\in O } } [ o \\in \\mathcal { P } ] \\in [ 0 , 1 ] } \\end{array}$ to denote the vocabulary usage of $\\mathcal { P }$ . ", + "bbox": [ + 174, + 835, + 826, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 90, + 823, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In a first step, we generate phoneme transcriptions for different training checkpoints or hyperparameter settings and denote the transcriptions of the configuration with the lowest vocabulary-usage adjusted NLL as $\\begin{array} { r } { \\hat { \\mathcal { P } } = \\arg \\operatorname* { m i n } _ { \\mathcal { P } } N L L _ { L M } ( \\mathcal { P } ) - \\log U ( \\mathcal { P } ) } \\end{array}$ .4 Next, we discard model configurations which do not satisfy the following using $\\hat { \\mathcal { P } }$ as the anchor: ", + "bbox": [ + 174, + 127, + 825, + 189 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/b07976d07eb1264788bb86ec8c503a84fc3c183351681c48efba85a9d06b98dd.jpg", + "text": "$$\nN L L _ { L M } ( \\mathcal { P } ) < N L L _ { L M } ( \\hat { \\mathcal { P } } ) + \\log \\left( \\frac { U ( \\mathcal { P } ) } { U ( \\hat { \\mathcal { P } } ) } \\right) + \\log 1 . 2\n$$", + "text_format": "latex", + "bbox": [ + 312, + 196, + 686, + 239 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second term on the right hand side introduces a margin over the NLL of the anchor transcription $N L L _ { L M } ( \\hat { \\mathcal { P } } )$ based on the vocabulary usage of $\\mathcal { P }$ and $\\hat { \\mathcal { P } }$ : If $U ( \\hat { \\mathcal { P } } )$ is much lower compared to $U ( \\mathcal { P } )$ , then we allow model configurations which produce transcriptions with higher NLL compared to $\\hat { \\mathcal { P } }$ . However, if $U ( \\hat { \\mathcal { P } } )$ is a lot higher than $U ( \\mathcal { P } )$ , then the model configuration will not satisfy the constraint. The $\\log 1 . 2$ factor serves as another margin allowing checkpoints with slightly worse vocabulary-usage adjusted NLL to be included. ", + "bbox": [ + 173, + 247, + 825, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In a final step, we take into account the length of the transcriptions: out of the configurations ${ \\mathcal { P } } ^ { \\prime }$ which satisfy the above constraint, we select the one which has the highest sum of log probability without normalizing the length: ", + "bbox": [ + 174, + 342, + 825, + 385 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/2930d2c4de9bfc9242467fc6ca39076d0e51025e6eb1f2125a1e21c64638989c.jpg", + "text": "$$\n\\mathcal { P } ^ { * } = \\arg \\operatorname* { m a x } _ { \\mathcal { P } ^ { \\prime } } \\sum _ { j = 1 } ^ { N _ { s } } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \\dots , p _ { M } ^ { j } ]\n$$", + "text_format": "latex", + "bbox": [ + 276, + 393, + 723, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This selects model configurations which produce phoneme sequences that score high under the language model but are not too long. Appendix D compares accuracy when developing with this metric compared to a labeled development set. ", + "bbox": [ + 174, + 446, + 826, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4 Results ", + "text_level": 1, + "bbox": [ + 174, + 512, + 266, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1 Comparison to Supervised Speech Recognition on Librispeech ", + "text_level": 1, + "bbox": [ + 174, + 544, + 645, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first test our approach on Librispeech to get a sense of how unsupervised speech recognition compares to the best supervised systems trained on a large amount of labeled data. Librispeech is a standard benchmark in the speech recognition community which provides about 960 hours of transcribed read audiobooks. We use the language modeling data of Librispeech as unlabeled text data for unsupervised training. In Appendix G we show that far less unlabeled text and speech audio are sufficient to reach a similar level of performance. We experiment with the frozen representations of a wav2vec 2.0 LARGE model trained on the $5 3 . 2 \\mathrm { k }$ hours of Libri-Light (LL-60k) which we denote as wav2vec-U LARGE. We also consider self-training over three iterations by first training an HMM on the labels generated by the GANm then fine-tuning the original wav2vec 2.0 model on the labels of the HMM for Librispeech followed by then fine-tuning on Libri-Light; Appendix F investigates alternatives. ", + "bbox": [ + 173, + 570, + 825, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "wav2vec-U LARGE with self-training (wav2vec- $\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,$ ) and a Transformer language model achieves WER 5.9 on test-other, the noisy test set. This shows that unsupervised speech recognition can perform remarkably well compared to the best supervised systems of the recent past on this much studied benchmark. Also, self-training is effective even when the teacher model is unsupervised as per the improvement over GAN training (wav2vec-U). Interestingly, self-training on just Librispeech, or 960 hours of unlabeled speech audio, achieves already very good performance of WER 6.4 on dev-other compared to self-training on all of Libri-Light (53.2k hours) which compares at 6.0 WER. We note that the number of parameters trained during adversarial training is very small: the generator contains only about 90k parameters for a single temporal convolution mapping to the phoneme set from frozen wav2vec 2.0 representations. ", + "bbox": [ + 173, + 729, + 826, + 867 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d5f6a9816458b5d9d90dd620c97a44d70ecc200d9e2e077d3e3e3e887499080.jpg", + "table_caption": [ + "Table 1: WER on Librispeech dev/test sets when using 960 hours of unlabeled audio from Librispeech (LS-960) or $5 3 . 2 \\mathrm { k }$ hours from Libri-Light (LL-60k) using representations from wav2vec 2.0 LARGE. Librispeech provides clean dev/test sets which are less challenging than the other sets. We report results for GAN training only (wav2vec-U) and with subsequent self-training (wav2vec- $\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,$ . " + ], + "table_footnote": [], + "table_body": "
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
", + "bbox": [ + 174, + 154, + 823, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2 Comparison to Prior Unsupervised Work ", + "text_level": 1, + "bbox": [ + 173, + 435, + 500, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prior work on unsupervised speech recognition focused on the TIMIT benchmark. In order to perform a direct comparison to these approaches, we report results on this benchmark as well. We consider two setups to compare to previous work: in the matched setting, the unlabeled text data is simply the transcriptions of the unlabeled audio data but unpaired. In the unmatched setup, the unlabeled text data does not contain the transcriptions for the audio data which is a more realistic setting. ", + "bbox": [ + 174, + 462, + 825, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We measure performance on the standard Kaldi dev and test sets (core-dev/core-test) as well as a slightly larger version of the test set (all-test) to be able to compare to Liu et al. [2018] and Chen et al. [2019]. Further details of the two setups can be found in Appendix $\\ S \\operatorname { E } . 1$ . We report performance for wav2vec-U with a 4-gram language model trained on the language modeling data of TIMIT and we also consider self-training (wav2vec- $\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T }$ ). ", + "bbox": [ + 174, + 537, + 825, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows that wav2vec-U outperforms prior unsupervised work in both the matched and unmatched settings, reducing PER on all-test in the matched setup by $57 \\%$ relative compared to Chen et al. [2019]. Our method has lower performance than the best supervised methods but it performs still very well at PER 12 on core-test in the matched setup compared to PER 8.3 for the state of the art [Baevski et al., 2020c]. ", + "bbox": [ + 174, + 613, + 825, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 Performance on non-English languages ", + "text_level": 1, + "bbox": [ + 176, + 700, + 485, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To get a sense of how well the method works on non-English data, we experiment on six languages of the multilingual Librispeech corpus (MLS; Pratap et al. 2020). As baseline we consider the supervised systems of Pratap et al. [2020] trained on between $2 \\mathrm { k }$ and 161 hours of labeled data, depending on the language. For adversarial learning we use 100 hours of unlabeled audio data from MLS for every language as well as the MLS language modeling data. As input to wav2vec-U we use the representations from XLSR-53 [Conneau et al., 2020], a wav2vec 2.0 model pre-trained on 53 languages. Table 3 shows that wav2vec-U generalizes across a range of languages. Performance is lower than supervised systems but it shows the viability for other languages. ", + "bbox": [ + 174, + 726, + 825, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we turn to three low-resource languages, Swahili, Kyrgyz, and Tatar. Swahili is an African language, Kyrgyz and Tatar are Turkic languages with only about $4 . 3 \\mathrm { m }$ and $5 . 2 \\mathrm { m }$ speakers, respectively.5 We use between 1.8 hours (Kyrgyz) and 9.2 hours of unlabeled audio (Swahili), see Appendix $\\ S \\operatorname { E } . 1$ . ", + "bbox": [ + 176, + 843, + 826, + 886 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/77fd4860fa4ec19b6701be89b6868bc4fe53c80cb80e10678edaf6050869d0ac.jpg", + "table_caption": [ + "Table 2: TIMIT Phoneme Error Rate (PER) in comparison to previous work for the matched and unmatched training data setups (Appendix $\\ S \\operatorname { E . 1 }$ ). PER is measured on the Kaldi dev and test sets (core-dev/core-test) as well as a slightly larger version of the test set (all-test) as used by some of the prior work. $( ^ { * } )$ indicates experiments that do not use the standard split excluding SA utterances. " + ], + "table_footnote": [], + "table_body": "
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
", + "bbox": [ + 240, + 152, + 758, + 458 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b406f6c833d0575a5f43e3163dc4e9bbbbefc546de30a4f3e6945916da6d56ac.jpg", + "table_caption": [ + "Table 3: WER on the Multilingual Librispeech (MLS) dataset using representations from the wav2vec 2.0 XLSR-53 model. We consider German (de), Dutch (nl), French (fr), Spanish (es), Italian (it), Portuguese (pt). " + ], + "table_footnote": [], + "table_body": "
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
", + "bbox": [ + 176, + 516, + 821, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To compare to prior work, we measure WER for Swahili and PER for Kyrgyz and Tatar. For Tatar and Kyrgyz we opted to use a reduced self-training regime for faster experimental turn-around where we only perform HMM self-training and we expect better performance with the full self-training setup (Appendix F). Table 4 and Table 5 show that wav2vec-U achieves good performance on these low-resource languages compared to previous work that utilized labeled data. We note that for Tatar and Kyrgyz we use a much smaller amount of speech audio than prior work: compared to XLSR-53 we use $1 . 8 \\mathrm { h }$ unlabeled data vs 17h of labeled data for Kyrgyz and 4.6h vs. 17h for Tatar. ", + "bbox": [ + 173, + 675, + 825, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 Related Work ", + "text_level": 1, + "bbox": [ + 174, + 795, + 321, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper builds on a large body of prior work which includes semi-supervised speech recognition such as self-training [Kahn et al., 2020a, Xu et al., 2020b, Park et al., 2020]. Some of the earliest work in self-supervised learning of speech representations was was done by van den Oord et al. [2018] for phoneme recognition which was simplified in Schneider et al. [2019] who applied it to full speech recognition. Other work includes language model-style pre-training [Chung et al., 2019a] and learning fixed size representations of audio segments [Chung and Glass, 2018]. There is also work on quantization of the continuous speech data [Baevski et al., 2020b,a, Liu et al., 2019, van Niekerk et al., 2020, Baevski et al., 2020c, Hsu et al., 2021b] and on robustness to domain shift [Hsu et al., 2021a], multilingual pre-training [Kawakami et al., 2020, Conneau et al., 2020] as well as combining speech and vision [Harwath et al., 2020]. ", + "bbox": [ + 174, + 827, + 825, + 911 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2681bd270ee947c4735ec6105ffa578b603c0a8f14a89fb155bf16d2679be651.jpg", + "table_caption": [ + "Table 4: PER for low-resource languages, Tatar (tt) and Kyrgyz (ky). " + ], + "table_footnote": [], + "table_body": "
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
", + "bbox": [ + 176, + 137, + 501, + 285 + ], + "page_idx": 8 + }, + { + "type": "table", + "img_path": "images/75c82e8aef7073d98bca3ee508ceb7960d3a4e7ae6963489d1130fb03b1695b7.jpg", + "table_caption": [ + "Table 5: WER for Swahili from the ALFFA corpus. We compare to the supervised baseline of the ALFFA project. " + ], + "table_footnote": [], + "table_body": "
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
", + "bbox": [ + 565, + 151, + 779, + 273 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 329, + 825, + 385 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Learning to map speech to phonemes without supervision using adversarial learning has been explored by Liu et al. [2018] who learn a mapping matrix between segment identifiers and phonemes. However, their work still relied on data segmented into phonemes by human annotators. This has been later extended to use an automatic segmentation [Chen et al., 2019] which is iteratively refined with HMMs. However, cross validation is still performed using labeled data (personal communication with authors). We also explored HMMs to refine segmentation boundaries (Table A2) but did not find it as effective as self-training. Our work is in part inspired by aligning word embedding spaces of different languages [Mikolov et al., 2013, Artetxe et al., 2017, Conneau et al., 2018] and full unsupervised machine translation [Lample et al., 2018, Artetxe et al., 2018, Conneau and Lample, 2019]. ", + "bbox": [ + 173, + 391, + 825, + 516 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 Conclusion and Future Work ", + "text_level": 1, + "bbox": [ + 174, + 556, + 449, + 574 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "wav2vec-U is a framework which enables building speech recognition models without labeled data. It embeds and segments the speech audio with self-supervised representations from wav2vec 2.0, learns a mapping to phonemes with adversarial learning, and cross-validates hyper-parameter choices as well as early stopping with an unsupervised metric. Experiments on the standard Librispeech benchmark show performance close to the state of the art models from only a few years ago, even though these models relied on nearly 1,000 hours of labeled data. ", + "bbox": [ + 174, + 599, + 825, + 684 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Compared to the previous best unsupervised speech recognition approach, wav2vec-U reduces TIMIT phone error rate from 26.1 to 11.3. We also demonstrate the viability of our approach on several languages other than English, some of which are low-resource. The ability to build speech recognition models solely from unlabeled speech audio and unlabeled text drastically lowers the effort to build speech technology for many more languages of the world. ", + "bbox": [ + 174, + 690, + 823, + 760 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our approach requires phonemization of the text for the language of interest. Moreover, phonemizers are not available for all languages and this presents a bottleneck. To address this, future work may develop phonemizers for more languages, explore phonemization approaches that generalize across languages, or unsupervised training with graphemic text units such as letters. ", + "bbox": [ + 174, + 766, + 825, + 821 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We explored a simple segmentation technique based on self-supervised representations, however, there is a large body of research on segmentation and some of these techniques may lead to improvements over our simple approach [Varadarajan et al., 2008, Zhang and Glass, 2009, Gish et al., 2009, Lee and Glass, 2012, Lee et al., 2015, Ondel et al., 2016, Kamper et al., 2017a,b, Kreuk et al., 2020]. Also, wav2vec 2.0 learns representations for fixed size units with a fixed stride, however, phonemic units are of variable size. Another direction is to learn variable sized representations during pre-training. ", + "bbox": [ + 174, + 828, + 825, + 911 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Acknowledgments and Disclosure of Funding ", + "text_level": 1, + "bbox": [ + 174, + 88, + 553, + 107 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We thank Zhouhan Lin for helping with initial explorations in this project, Tatiana Likhomanenko for helpful discussions about self-training, Da-Rong Liu for sharing details to reproduce the setup of Chen et al. [2019], Marc’Aurelio Ranzato for general helpful discussions, and Ruth Kipng’eno, Ruth Ndila Ndeto as well as Mark Mutitu for error analysis of our Swahili model. ", + "bbox": [ + 174, + 121, + 826, + 176 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "References ", + "text_level": 1, + "bbox": [ + 174, + 195, + 266, + 212 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "S. T. Abate, W. Menzel, and B. Tafila. An amharic speech corpus for large vocabulary continuous speech recognition. In Proc. of Interspeech, 2005. \nD. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro, Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin. In Proc. of ICML, 2016. \nR. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. Proc. of LREC, 2020. \nM. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. Proc. of ICML, 2017. \nM. Artetxe, G. Labaka, and E. Agirre. Learning bilingual word embeddings with (almost) no bilingual data. In Proc. of ACL, 2017. \nM. Artetxe, G. Labaka, E. Agirre, and K. Cho. Unsupervised neural machine translation. Proc. of ICLR, 2018. \nA. Baevski and M. Auli. Adaptive input representations for neural language modeling. In Proc. of ICLR, 2018. \nA. Baevski, M. Auli, and A. Mohamed. Effectiveness of self-supervised pre-training for speech recognition. Proc. of ICASSP, 2020a. \nA. Baevski, S. Schneider, and M. Auli. vq-wav2vec: Self-supervised learning of discrete speech representations. In Proc. of ICLR, 2020b. \nA. Baevski, Y. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. In Proc. of NeurIPS, 2020c. \nL. Besacier, E. Gauthier, M. Mangeot, P. Bretier, P. Bagshaw, O. Rosec, T. Moudenc, F. Pellegrino, S. Voisin, E. Marsico, and P. Nocera. Speech technologies for african languages: example of a multilingual calculator for education. In Proc. of Interspeech, 2015. \nD. Can, V. R. Martinez, P. Papadopoulos, and S. S. Narayanan. Pykaldi: A python wrapper for kaldi. In Proc. of ICASSP, 2018. \nK.-Y. Chen, C.-P. Tsai, D.-R. Liu, H.-Y. Lee, and L. shan Lee. Completely unsupervised speech recognition by a generative adversarial network harmonized with iteratively refined hidden markov models. In Proc. of Interspeech, 2019. \nY. Chung, W. Weng, S. Tong, and J. R. Glass. Unsupervised cross-modal alignment of speech and text embedding spaces. Proc. of NIPS, 2018. \nY. Chung, W. Hsu, H. Tang, and J. R. Glass. An unsupervised autoregressive model for speech representation learning. Proc. of Interspeech, 2019a. \nY.-A. Chung and J. Glass. Speech2vec: A sequence-to-sequence framework for learning word embeddings from speech. Proc. of Interspeech, 2018. \nY.-A. Chung, W.-N. Hsu, H. Tang, and J. Glass. An unsupervised autoregressive model for speech representation learning. Proc. of Interspeech, 2019b. \nA. Conneau and G. Lample. Cross-lingual language model pretraining. Proc. of NeurIPS, 2019. \nA. Conneau, G. Lample, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel data. Proc. of ICLR, 2018. \nA. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli. Unsupervised cross-lingual representation learning for speech recognition. arXiv, abs/2006.13979, 2020. \nJ. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. Proc. of NAACL, 2019. \nL. Dong, S. Xu, and B. Xu. Speech-transformer: A no-recurrence sequence-to-sequence model for speech recognition. In Proc. of ICASSP, 2018. \nZ. Fan, M. Li, S. Zhou, and B. Xu. Exploring wav2vec 2.0 on speaker verification and language identification. arXiv, 2021. \nR. Fer, P. Matejka, F. Grézl, O. Plchot, K. Vesel ˇ y, and J. H. \\` Cernock ˇ y. Multilingually trained \\` bottleneck features in spoken language recognition. Computer Speech & Language, 46, 2017. \nJ. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. The DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium, 1993. \nH. Gelas, L. Besacier, and F. Pellegrino. Developments of Swahili resources for an automatic speech recognition system. In Proc. of SLTU, 2012. \nH. Gish, M. Siu, A. Chan, and W. Belfield. Unsupervised training of an hmm-based speech recognizer for topic classification. In Proc. of Interspeech, 2009. \nI. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial networks. Proc. of NIPS, 2014. \nGoogle. Google cloud: Speech-to-text. https://cloud.google.com/speech-to-text, 2021. Accessed: 2021-05-13. \nA. Graves, S. Fernández, and F. Gomez. Connectionist temporal classification: Labelling unsegmented sequence data with recurrent neural networks. In Proc. of ICML, 2006. \nA. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wu, and R. Pang. Conformer: Convolution-augmented transformer for speech recognition. Proc. of Interspeech, 2020. \nI. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein gans. Proc. of NIPS, 2017. \nW. Han, Z. Zhang, Y. Zhang, J. Yu, C.-C. Chiu, J. Qin, A. Gulati, R. Pang, and Y. Wu. Contextnet: Improving convolutional neural networks for automatic speech recognition with global context. Proc. of Interspeech, 2020. \nD. Harwath and J. Glass. Towards visually grounded sub-word speech unit discovery. In Proc. of ICASSP, pages 3017–3021. IEEE, 2019. \nD. Harwath, W.-N. Hsu, and J. Glass. Learning hierarchical discrete linguistic units from visuallygrounded speech. In Proc. of ICLR, 2020. \nK. Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July 2011. Association for Computational Linguistics. \nK. Hirsh-Pasek, D. G. Kemler Nelson, P. W. Jusczyk, K. W. Cassidy, B. Druss, and L. Kennedy. Clauses are perceptual units for young infants. Cognition, 26(3):269–286, 1987. \nW.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V. Pratap, J. Kahn, A. Lee, R. Collobert, G. Synnaeve, et al. Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training. arXiv preprint arXiv:2104.01027, 2021a. \nW.-N. Hsu, Y.-H. H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mohamed. Hubert: How much can a bad teacher benefit ASR pre-training? In Proc. of ICASSP, 2021b. \nE. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. Proc. of ICLR, 2016. \nH. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Trans. Pattern Anal. Mach. Intell., 33(1):117–128, Jan. 2011. \nE. K. Johnson and P. W. Jusczyk. Word segmentation by 8-month-olds: When speech cues count more than statistics. Journal of Memory and Language, 44(4):548–567, 2001. \nJ. Johnson, M. Douze, and H. Jégou. Billion-scale similarity search with gpus. IEEE Transactions on Big Data, 2019. \nP. W. Jusczyk, D. M. Houston, and M. Newsome. The beginnings of word segmentation in englishlearning infants. Cognitive Psychology, 39(3):159–207, 1999. \nJ. Kahn, A. Lee, and A. Hannun. Self-training for end-to-end speech recognition. In Proc. of ICASSP, 2020a. \nJ. Kahn et al. Libri-light: A benchmark for asr with limited or no supervision. In Proc. of ICASSP, 2020b. \nH. Kamper, A. Jansen, and S. Goldwater. A segmental framework for fully-unsupervised largevocabulary speech recognition. Comput. Speech Lang., 46(C), Nov. 2017a. \nH. Kamper, K. Livescu, and S. Goldwater. An embedded segmental k-means model for unsupervised segmentation and clustering of speech. Proc. of ASRU, 2017b. \nK. Kawakami, L. Wang, C. Dyer, P. Blunsom, and A. van den Oord. Learning robust and multilingual speech representations. Proc. of EMNLP, 2020. \nD. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of ICLR, 2015. \nF. Kreuk, J. Keshet, and Y. Adi. Self-supervised contrastive learning for unsupervised phoneme segmentation. Proc. of Interspeech, 2020. \nG. Lample, L. Denoyer, and M. Ranzato. Unsupervised machine translation using monolingual corpora only. In Proc. of ICLR, 2018. \nC. Lee and J. R. Glass. A nonparametric bayesian approach to acoustic model discovery. In Proc. of ACL, 2012. \nC. Lee, T. J. O’Donnell, and J. R. Glass. Unsupervised lexicon discovery from acoustic input. TACL, 2015. \nM. P. Lewis, G. F. Simon, and C. D. Fennig. Ethnologue: Languages of the world, nineteenth edition. Online version: http://www.ethnologue.com, 2016. \nT. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert. slimipl: Language-model-free iterative pseudo-labeling. arXiv, 2021. \nA. H. Liu, T. Tu, H. yi Lee, and L. shan Lee. Towards unsupervised speech recognition and synthesis with quantized speech representation learning. Proc. of ICASSP, 2019. \nD.-R. Liu, K.-Y. Chen, H.-Y. Lee, and L. shan Lee. Completely unsupervised phoneme recognition by adversarially learning mapping relationships from audio embeddings. Proc. of Interspeech, 2018. \nT. Mikolov, Q. V. Le, and I. Sutskever. Exploiting similarities among languages for machine translation. arXiv preprint arXiv:1309.4168, 2013. \nM. Mohri, F. Pereira, and M. Riley. Weighted finite-state transducers in speech recognition. Computer Speech & Language, 16(1):69–88, 2002. \nL. Ondel, L. Burget, and J. Cernocký. Variational inference for acoustic unit discovery. In Proc. of SLTU, 2016. \nM. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli. fairseq: A fast, extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, 2019. \nV. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public domain audio books. In Proc. of ICASSP, pages 5206–5210. IEEE, 2015. \nD. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le. Specaugment: A simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, 2019. \nD. S. Park, Y. Zhang, Y. Jia, W. Han, C.-C. Chiu, B. Li, Y. Wu, and Q. V. Le. Improved noisy student training for automatic speech recognition. Proc. of Interspeech, 2020. \nK. Park and J. Kim. g2pe. https://github.com/Kyubyong/g2p, 2019. \nL. Pepino, P. Riera, and L. Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings. arXiv, 2021. \nL. Polka and J. F. Werker. Developmental changes in perception of nonnative vowel contrasts. Journal of Experimental Psychology: Human perception and performance, 20(2):421, 1994. \nD. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek, Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech recognition toolkit. In Proc. of ASRU, 2011. \nV. Pratap, A. Hannun, Q. Xu, J. Cai, J. Kahn, G. Synnaeve, V. Liptchinsky, and R. Collobert. Wav2letter++: A fast open-source speech recognition system. In Proc. of ICASSP, 2019. \nV. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert. Mls: A large-scale multilingual dataset for speech research. In Proc. of Interspeech, 2020. \nO. Rasanen, G. Doyle, and M. C. Frank. Unsupervised word discovery from speech using automatic segmentation into syllable-like units. In Proc. of Interspeech, 2015. \nM. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio. Light gated recurrent units for speech recognition. IEEE Trans. on Emerging Topics in Comp. Intel., 2, 2018. \nM. Ravanelli, T. Parcollet, and Y. Bengio. The pytorch-kaldi speech recognition toolkit. Proc. of ICASSP, 2019. \nM. Rivière, A. Joulin, P.-E. Mazaré, and E. Dupoux. Unsupervised pretraining transfers well across languages. In Proc. of ICASSP, 2020. \nS. Schneider, A. Baevski, R. Collobert, and M. Auli. wav2vec: Unsupervised pre-training for speech recognition. In Proc. of Interspeech, 2019. \nN. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, 2014. \nG. Synnaeve, Q. Xu, J. Kahn, T. Likhomanenko, E. Grave, V. Pratap, A. Sriram, V. Liptchinsky, and R. Collobert. End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern Architectures. Proc. of ICML workshop on Self-supervision in Audio and Speech (SAS), 2020. \nM. Tachbelie, S. T. Abate, and L. Besacier. Using different acoustic, lexical and language modeling units for asr of an under-resourced language - amharic. Speech Communication, 56, 2014. \nZ. Tan, A. K. Sarkar, and N. Dehak. rvad: An unsupervised segment-based robust voice activity detection method. Computer speech & language, 59:1–21, 2020. \nA. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding. Proc. of NIPS, 2018. \nB. van Niekerk, L. Nortje, and H. Kamper. Vector-quantized neural networks for acoustic unit discovery in the zerospeech 2020 challenge. Proc. of Interspeech, 2020. \nB. Varadarajan, S. Khudanpur, and E. Dupoux. Unsupervised learning of acoustic sub-word units. In Proc. of ACL, 2008. \nA. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. In Proc. of NIPS, 2017. \nC. Wang, A. Wu, J. Pino, A. Baevski, M. Auli, and A. Conneau. Large-scale self- and semi-supervised learning for speech translation. arXiv, 2021. \nJ. F. Werker and R. C. Tees. Cross-language speech perception: Evidence for perceptual reorganization during the first year of life. Infant behavior and development, 7(1):49–63, 1984. \nH. Xu, K. Li, Y. Wang, J. Wang, S. Kang, X. Chen, D. Povey, and S. Khudanpur. Neural network language modeling with letter-based features and importance sampling. In Proc. of ICASSP, 2018. \nQ. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and M. Auli. Self-training and pre-training are complementary for speech recognition. In Proc. of ICASSP, 2020a. \nQ. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert. Iterative pseudolabeling for speech recognition. Proc. of Interspeech, 2020b. \nC.-K. Yeh, J. Chen, C. Yu, and D. Yu. Unsupervised speech recognition via segmental empirical output distribution matching. In Proc. of ICLR, 2019. \nN. Zeghidour, Q. Xu, V. Liptchinsky, N. Usunier, G. Synnaeve, and R. Collobert. Fully convolutional speech recognition. arXiv, abs/1812.06864, 2018. \nF. Zhang, Y. Wang, X. Zhang, C. Liu, Y. Saraf, and G. Zweig. Faster, simpler and more accurate hybrid asr systems using wordpieces. Proc. of Interspeech, 2020a. \nY. Zhang and J. R. Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian posteriorgrams. IEEE Workshop on Automatic Speech Recognition & Understanding, 2009. \nY. Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V. Le, and Y. Wu. Pushing the limits of semi-supervised learning for automatic speech recognition. Proc. of NeurIPS SAS Workshop, 2020b. ", + "bbox": [ + 171, + 217, + 828, + 906 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 54, + 828, + 914 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 87, + 830, + 920 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 73, + 828, + 921 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "", + "bbox": [ + 169, + 85, + 828, + 565 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/QmxFsofRvW9/QmxFsofRvW9_middle.json b/parse/train/QmxFsofRvW9/QmxFsofRvW9_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..a79db16b08449274f4236c5253a74beb093d8615 --- /dev/null +++ b/parse/train/QmxFsofRvW9/QmxFsofRvW9_middle.json @@ -0,0 +1,33614 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 179, + 97, + 432, + 117 + ], + "lines": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "spans": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "score": 1.0, + "content": "Unsupervised Speech Recognition", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 149, + 158, + 459, + 171 + ], + "lines": [ + { + "bbox": [ + 149, + 156, + 461, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 225, + 172 + ], + "score": 1.0, + "content": "Alexei Baevski4", + "type": "text" + }, + { + "bbox": [ + 230, + 156, + 461, + 173 + ], + "score": 1.0, + "content": "Wei-Ning Hsu4 Alexis Conneau\u0003∗ Michael Auli4", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 245, + 186, + 366, + 199 + ], + "lines": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "spans": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "score": 1.0, + "content": "4 Facebook AI \u0003 Google AI", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 283, + 227, + 328, + 240 + ], + "lines": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "spans": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 253, + 469, + 416 + ], + "lines": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "score": 1.0, + "content": "Despite rapid progress in the recent past, current speech recognition systems", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "score": 1.0, + "content": "still require labeled training data which limits this technology to a small fraction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "score": 1.0, + "content": "of the languages spoken around the globe. This paper describes wav2vec-U,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "score": 1.0, + "content": "short for wav2vec Unsupervised, a method to train speech recognition models", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "score": 1.0, + "content": "without any labeled data. We leverage self-supervised speech representations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "to segment unlabeled audio and learn a mapping from these representations to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "score": 1.0, + "content": "phonemes via adversarial training. The right representations are key to the success", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "score": 1.0, + "content": "of our method. Compared to the best previous unsupervised work, wav2vec-U", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "score": 1.0, + "content": "reduces the phone error rate on the TIMIT benchmark from 26.1 to 11.3. On", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "score": 1.0, + "content": "the larger English Librispeech benchmark, wav2vec-U achieves a word error rate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "score": 1.0, + "content": "of 5.9 on test-other, rivaling some of the best published systems trained on 960", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "score": 1.0, + "content": "hours of labeled data from only two years ago. We also experiment on nine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "score": 1.0, + "content": "other languages, including low-resource languages such as Kyrgyz, Swahili and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "score": 1.0, + "content": "Tatar. The code is available at https://github.com/pytorch/fairseq/tree/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "score": 1.0, + "content": "master/examples/wav2vec/unsupervised", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11 + }, + { + "type": "title", + "bbox": [ + 107, + 438, + 190, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "Speech recognition performance on the much studied English Librispeech benchmark [Panayotov", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "score": 1.0, + "content": "et al., 2015] has seen rapid improvement over the last few years due to advances in model architec-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "tures [Dong et al., 2018, Synnaeve et al., 2020, Gulati et al., 2020], semi-supervised learning [Xu", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "et al., 2020b, Park et al., 2020] and self-supervised learning [van den Oord et al., 2018, Chung and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "Glass, 2018, Chung et al., 2019b, Baevski et al., 2020c]. However, all of these techniques require", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "transcribed speech data which is not available for the vast majority of the nearly 7,000 languages", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "of the world [Lewis et al., 2016]. As a result, speech recognition technology is only available for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "about 125 different languages [Google, 2021]. On the other hand, humans learn a lot about speech", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "simply by listening to others around them and without explicit supervision [Werker and Tees, 1984,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "score": 1.0, + "content": "Hirsh-Pasek et al., 1987, Polka and Werker, 1994, Jusczyk et al., 1999, Johnson and Jusczyk, 2001].", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "Unsupervised learning has been very successful in machine translation resulting in systems that obtain", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "remarkable accuracy given no labeled training data at all [Conneau et al., 2018, Lample et al., 2018,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "Artetxe et al., 2018]. Inspired by this, there has been some work on unsupervised speech recognition", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "score": 1.0, + "content": "based on learning to align unlabeled text and audio [Yeh et al., 2019] or adversarial learning [Liu", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "score": 1.0, + "content": "et al., 2018, Chen et al., 2019]. These approaches showed promising initial results but their error", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "score": 1.0, + "content": "rates are still high, with evaluation being limited to the small-scale and clean TIMIT benchmark.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "score": 1.0, + "content": "In this work, we introduce a framework for unsupervised learning of speech recognition models.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "Wav2vec-U, or wav2vec Unsupervised, leverages self-supervised representations from wav2vec", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "2.0 [Baevski et al., 2020c] to embed the speech audio and to segment the audio into units with a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 135, + 695 + ], + "score": 1.0, + "content": "simple", + "type": "text" + }, + { + "bbox": [ + 135, + 682, + 141, + 692 + ], + "score": 0.46, + "content": "\\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 682, + 505, + 695 + ], + "score": 1.0, + "content": "-means clustering method (see Figure 1 for an illustration of our approach). We find that the", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 120, + 702, + 245, + 712 + ], + "lines": [ + { + "bbox": [ + 119, + 701, + 246, + 712 + ], + "spans": [ + { + "bbox": [ + 119, + 701, + 246, + 712 + ], + "score": 1.0, + "content": "∗Work done while at Facebook AI.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 731, + 385, + 742 + ], + "lines": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "spans": [ + { + "bbox": [ + 106, + 730, + 386, + 743 + ], + "score": 1.0, + "content": "35th Conference on Neural Information Processing Systems (NeurIPS 2021).", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 179, + 97, + 432, + 117 + ], + "lines": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "spans": [ + { + "bbox": [ + 178, + 96, + 433, + 119 + ], + "score": 1.0, + "content": "Unsupervised Speech Recognition", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 149, + 158, + 459, + 171 + ], + "lines": [ + { + "bbox": [ + 149, + 156, + 461, + 173 + ], + "spans": [ + { + "bbox": [ + 149, + 157, + 225, + 172 + ], + "score": 1.0, + "content": "Alexei Baevski4", + "type": "text" + }, + { + "bbox": [ + 230, + 156, + 461, + 173 + ], + "score": 1.0, + "content": "Wei-Ning Hsu4 Alexis Conneau\u0003∗ Michael Auli4", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1, + "bbox_fs": [ + 149, + 156, + 461, + 173 + ] + }, + { + "type": "text", + "bbox": [ + 245, + 186, + 366, + 199 + ], + "lines": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "spans": [ + { + "bbox": [ + 243, + 184, + 368, + 201 + ], + "score": 1.0, + "content": "4 Facebook AI \u0003 Google AI", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2, + "bbox_fs": [ + 243, + 184, + 368, + 201 + ] + }, + { + "type": "title", + "bbox": [ + 283, + 227, + 328, + 240 + ], + "lines": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "spans": [ + { + "bbox": [ + 281, + 226, + 330, + 242 + ], + "score": 1.0, + "content": "Abstract", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 143, + 253, + 469, + 416 + ], + "lines": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "spans": [ + { + "bbox": [ + 141, + 253, + 469, + 266 + ], + "score": 1.0, + "content": "Despite rapid progress in the recent past, current speech recognition systems", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "spans": [ + { + "bbox": [ + 141, + 264, + 469, + 276 + ], + "score": 1.0, + "content": "still require labeled training data which limits this technology to a small fraction", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "spans": [ + { + "bbox": [ + 141, + 274, + 470, + 288 + ], + "score": 1.0, + "content": "of the languages spoken around the globe. This paper describes wav2vec-U,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "spans": [ + { + "bbox": [ + 141, + 285, + 469, + 298 + ], + "score": 1.0, + "content": "short for wav2vec Unsupervised, a method to train speech recognition models", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 469, + 309 + ], + "score": 1.0, + "content": "without any labeled data. We leverage self-supervised speech representations", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "spans": [ + { + "bbox": [ + 141, + 308, + 469, + 320 + ], + "score": 1.0, + "content": "to segment unlabeled audio and learn a mapping from these representations to", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "spans": [ + { + "bbox": [ + 142, + 319, + 470, + 330 + ], + "score": 1.0, + "content": "phonemes via adversarial training. The right representations are key to the success", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "spans": [ + { + "bbox": [ + 141, + 329, + 470, + 342 + ], + "score": 1.0, + "content": "of our method. Compared to the best previous unsupervised work, wav2vec-U", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "spans": [ + { + "bbox": [ + 141, + 341, + 469, + 352 + ], + "score": 1.0, + "content": "reduces the phone error rate on the TIMIT benchmark from 26.1 to 11.3. On", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "spans": [ + { + "bbox": [ + 142, + 352, + 469, + 363 + ], + "score": 1.0, + "content": "the larger English Librispeech benchmark, wav2vec-U achieves a word error rate", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "spans": [ + { + "bbox": [ + 141, + 361, + 469, + 374 + ], + "score": 1.0, + "content": "of 5.9 on test-other, rivaling some of the best published systems trained on 960", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "spans": [ + { + "bbox": [ + 141, + 372, + 469, + 386 + ], + "score": 1.0, + "content": "hours of labeled data from only two years ago. We also experiment on nine", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "spans": [ + { + "bbox": [ + 141, + 383, + 470, + 397 + ], + "score": 1.0, + "content": "other languages, including low-resource languages such as Kyrgyz, Swahili and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "spans": [ + { + "bbox": [ + 141, + 394, + 469, + 407 + ], + "score": 1.0, + "content": "Tatar. The code is available at https://github.com/pytorch/fairseq/tree/", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "spans": [ + { + "bbox": [ + 141, + 406, + 333, + 417 + ], + "score": 1.0, + "content": "master/examples/wav2vec/unsupervised", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 11, + "bbox_fs": [ + 141, + 253, + 470, + 417 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 438, + 190, + 452 + ], + "lines": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 192, + 454 + ], + "score": 1.0, + "content": "1 Introduction", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "Speech recognition performance on the much studied English Librispeech benchmark [Panayotov", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 506, + 486 + ], + "score": 1.0, + "content": "et al., 2015] has seen rapid improvement over the last few years due to advances in model architec-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 505, + 498 + ], + "score": 1.0, + "content": "tures [Dong et al., 2018, Synnaeve et al., 2020, Gulati et al., 2020], semi-supervised learning [Xu", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "et al., 2020b, Park et al., 2020] and self-supervised learning [van den Oord et al., 2018, Chung and", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 506, + 520 + ], + "score": 1.0, + "content": "Glass, 2018, Chung et al., 2019b, Baevski et al., 2020c]. However, all of these techniques require", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 532 + ], + "score": 1.0, + "content": "transcribed speech data which is not available for the vast majority of the nearly 7,000 languages", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 542 + ], + "score": 1.0, + "content": "of the world [Lewis et al., 2016]. As a result, speech recognition technology is only available for", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 106, + 540, + 505, + 552 + ], + "score": 1.0, + "content": "about 125 different languages [Google, 2021]. On the other hand, humans learn a lot about speech", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 551, + 506, + 564 + ], + "score": 1.0, + "content": "simply by listening to others around them and without explicit supervision [Werker and Tees, 1984,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 507, + 574 + ], + "score": 1.0, + "content": "Hirsh-Pasek et al., 1987, Polka and Werker, 1994, Jusczyk et al., 1999, Johnson and Jusczyk, 2001].", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 464, + 507, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 578, + 505, + 644 + ], + "lines": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 578, + 505, + 590 + ], + "score": 1.0, + "content": "Unsupervised learning has been very successful in machine translation resulting in systems that obtain", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "remarkable accuracy given no labeled training data at all [Conneau et al., 2018, Lample et al., 2018,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "Artetxe et al., 2018]. Inspired by this, there has been some work on unsupervised speech recognition", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 504, + 623 + ], + "score": 1.0, + "content": "based on learning to align unlabeled text and audio [Yeh et al., 2019] or adversarial learning [Liu", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 634 + ], + "score": 1.0, + "content": "et al., 2018, Chen et al., 2019]. These approaches showed promising initial results but their error", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 493, + 645 + ], + "score": 1.0, + "content": "rates are still high, with evaluation being limited to the small-scale and clean TIMIT benchmark.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 578, + 506, + 645 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 649, + 505, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 507, + 662 + ], + "score": 1.0, + "content": "In this work, we introduce a framework for unsupervised learning of speech recognition models.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "spans": [ + { + "bbox": [ + 105, + 659, + 506, + 673 + ], + "score": 1.0, + "content": "Wav2vec-U, or wav2vec Unsupervised, leverages self-supervised representations from wav2vec", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 671, + 506, + 684 + ], + "score": 1.0, + "content": "2.0 [Baevski et al., 2020c] to embed the speech audio and to segment the audio into units with a", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 682, + 505, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 682, + 135, + 695 + ], + "score": 1.0, + "content": "simple", + "type": "text" + }, + { + "bbox": [ + 135, + 682, + 141, + 692 + ], + "score": 0.46, + "content": "\\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 682, + 505, + 695 + ], + "score": 1.0, + "content": "-means clustering method (see Figure 1 for an illustration of our approach). We find that the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "quality of the audio representations is key to the success of unsupervised speech recognition. Similar", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "to Liu et al. [2018] and Chen et al. [2019], we learn a mapping between segments and phonemes", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "using adversarial training but different to their work, we also enable the algorithm to label segments", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "as silences. We also introduce an unsupervised cross-validation metric to enable model development", + "type": "text", + "cross_page": true + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "without labeled development data. Our unsupervised speech recognition model, the generator, is very", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "lightweight: it consists of a single temporal convolution comprising only about 90k parameters to", + "type": "text", + "cross_page": true + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "score": 1.0, + "content": "which we input frozen wav2vec 2.0 representations.", + "type": "text", + "cross_page": true + } + ], + "index": 15 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 649, + 507, + 695 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "score": 0.973, + "type": "image", + "image_path": "0b5b4f3465570676296b5d72eebf6d66830b97659f2ec44658268dac36522d53.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 72, + 494, + 132.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 132.33333333333334, + 494, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 192.66666666666669, + 494, + 253.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 259, + 505, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "Figure 1: Illustration of wav2vec Unsupervised: we learn self-supervised representations with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 469, + 283 + ], + "score": 1.0, + "content": "wav2vec 2.0 on unlabeled speech audio (Step 1), identify clusters in the representations with", + "type": "text" + }, + { + "bbox": [ + 469, + 271, + 476, + 281 + ], + "score": 0.4, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "-means", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "(Step 2) to segment the audio (Step 3). Next, we build segment representations by mean pooling the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "wav2vec 2.0 representations, performing PCA and a second mean pooling step between adjacent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "segments (Step 4). This is input to the generator which outputs a phoneme sequence (Step 5) fed to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "score": 1.0, + "content": "the discriminator, similar to phonemized unlabeled text (Step 6), for adversarial training (Step 7).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 506, + 364 + ], + "score": 1.0, + "content": "quality of the audio representations is key to the success of unsupervised speech recognition. Similar", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 505, + 375 + ], + "score": 1.0, + "content": "to Liu et al. [2018] and Chen et al. [2019], we learn a mapping between segments and phonemes", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "using adversarial training but different to their work, we also enable the algorithm to label segments", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 506, + 397 + ], + "score": 1.0, + "content": "as silences. We also introduce an unsupervised cross-validation metric to enable model development", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 394, + 505, + 408 + ], + "score": 1.0, + "content": "without labeled development data. Our unsupervised speech recognition model, the generator, is very", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "spans": [ + { + "bbox": [ + 105, + 406, + 505, + 419 + ], + "score": 1.0, + "content": "lightweight: it consists of a single temporal convolution comprising only about 90k parameters to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 316, + 429 + ], + "score": 1.0, + "content": "which we input frozen wav2vec 2.0 representations.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "Experimental results demonstrate the viability of the framework for a variety of settings and languages.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "wav2vec-U improves the phone error rate (PER) on the small-scale TIMIT benchmark from 26.1", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "to 11.3 compared to the next best known unsupervised approach. To get a better sense of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "performance compared to the best supervised methods, we measure performance on the larger", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "Librispeech benchmark where our method achieves word error rate (WER) 5.9 on test-other. We also", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "score": 1.0, + "content": "evaluate on six other European languages of the multilingual Librispeech benchmark [Pratap et al.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "score": 1.0, + "content": "2020] and on three non-European low-resource languages.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19 + }, + { + "type": "title", + "bbox": [ + 107, + 529, + 295, + 543 + ], + "lines": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "score": 1.0, + "content": "2 Speech and Text Representations", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 106, + 557, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "score": 1.0, + "content": "Next, we describe how we build suitable speech and text representations for unsupervised learning.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "score": 1.0, + "content": "Good representations are essential to learning a mapping from speech to text without supervision.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 106, + 595, + 379, + 608 + ], + "lines": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "score": 1.0, + "content": "2.1 Self-supervised Learning of Speech Audio Representations", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "score": 1.0, + "content": "In the first step, we learn representations of the speech audio signal using self-supervised learning.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "There has been a lot of recent work in this direction which has shown strong performance in extremely", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "score": 1.0, + "content": "low-labeled data setups across a range of languages [Conneau et al., 2020] and tasks [Fan et al., 2021,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "score": 1.0, + "content": "Pepino et al., 2021, Wang et al., 2021].", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 348, + 679 + ], + "score": 1.0, + "content": "Wav2vec 2.0 consists of a convolutional feature encoder", + "type": "text" + }, + { + "bbox": [ + 348, + 667, + 406, + 678 + ], + "score": 0.91, + "content": "f : \\mathcal X \\mapsto \\mathcal Z", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "that maps a raw audio", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 145, + 691 + ], + "score": 1.0, + "content": "sequence", + "type": "text" + }, + { + "bbox": [ + 146, + 678, + 156, + 688 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 677, + 285, + 691 + ], + "score": 1.0, + "content": "to latent speech representations", + "type": "text" + }, + { + "bbox": [ + 285, + 680, + 328, + 689 + ], + "score": 0.89, + "content": "z _ { 1 } , \\dots , z _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 677, + 417, + 691 + ], + "score": 1.0, + "content": ", which a Transformer", + "type": "text" + }, + { + "bbox": [ + 418, + 678, + 462, + 689 + ], + "score": 0.92, + "content": "g : { \\mathcal { Z } } \\mapsto { \\mathcal { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "then turns", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 220, + 701 + ], + "score": 1.0, + "content": "into context representations", + "type": "text" + }, + { + "bbox": [ + 221, + 690, + 263, + 700 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 689, + 390, + 701 + ], + "score": 1.0, + "content": "[Baevski et al., 2020b,a]. Each", + "type": "text" + }, + { + "bbox": [ + 390, + 690, + 400, + 700 + ], + "score": 0.84, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 689, + 469, + 701 + ], + "score": 1.0, + "content": "represents about", + "type": "text" + }, + { + "bbox": [ + 470, + 689, + 493, + 699 + ], + "score": 0.47, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 172, + 712 + ], + "score": 1.0, + "content": "audio strided by", + "type": "text" + }, + { + "bbox": [ + 173, + 700, + 196, + 711 + ], + "score": 0.31, + "content": "2 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "and the Transformer architecture follows BERT [Vaswani et al., 2017, Devlin", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 382, + 724 + ], + "score": 1.0, + "content": "et al., 2019]. During training, latent representations are discretized to", + "type": "text" + }, + { + "bbox": [ + 383, + 712, + 425, + 722 + ], + "score": 0.9, + "content": "q _ { 1 } , \\ldots , q _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "with a quantization", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 33 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 753 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 115, + 72, + 494, + 253 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "spans": [ + { + "bbox": [ + 115, + 72, + 494, + 253 + ], + "score": 0.973, + "type": "image", + "image_path": "0b5b4f3465570676296b5d72eebf6d66830b97659f2ec44658268dac36522d53.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 115, + 72, + 494, + 132.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 115, + 132.33333333333334, + 494, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 115, + 192.66666666666669, + 494, + 253.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 259, + 505, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 259, + 505, + 272 + ], + "score": 1.0, + "content": "Figure 1: Illustration of wav2vec Unsupervised: we learn self-supervised representations with", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 270, + 506, + 283 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 469, + 283 + ], + "score": 1.0, + "content": "wav2vec 2.0 on unlabeled speech audio (Step 1), identify clusters in the representations with", + "type": "text" + }, + { + "bbox": [ + 469, + 271, + 476, + 281 + ], + "score": 0.4, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 270, + 506, + 283 + ], + "score": 1.0, + "content": "-means", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 505, + 295 + ], + "score": 1.0, + "content": "(Step 2) to segment the audio (Step 3). Next, we build segment representations by mean pooling the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "wav2vec 2.0 representations, performing PCA and a second mean pooling step between adjacent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "segments (Step 4). This is input to the generator which outputs a phoneme sequence (Step 5) fed to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 496, + 327 + ], + "score": 1.0, + "content": "the discriminator, similar to phonemized unlabeled text (Step 6), for adversarial training (Step 7).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 5.5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 106, + 351, + 505, + 428 + ], + "lines": [], + "index": 12, + "bbox_fs": [ + 105, + 351, + 506, + 429 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 433, + 506, + 510 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 507, + 447 + ], + "score": 1.0, + "content": "Experimental results demonstrate the viability of the framework for a variety of settings and languages.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 444, + 506, + 457 + ], + "score": 1.0, + "content": "wav2vec-U improves the phone error rate (PER) on the small-scale TIMIT benchmark from 26.1", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 455, + 505, + 468 + ], + "score": 1.0, + "content": "to 11.3 compared to the next best known unsupervised approach. To get a better sense of the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 506, + 479 + ], + "score": 1.0, + "content": "performance compared to the best supervised methods, we measure performance on the larger", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 477, + 505, + 488 + ], + "score": 1.0, + "content": "Librispeech benchmark where our method achieves word error rate (WER) 5.9 on test-other. We also", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 104, + 487, + 507, + 501 + ], + "score": 1.0, + "content": "evaluate on six other European languages of the multilingual Librispeech benchmark [Pratap et al.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 342, + 512 + ], + "score": 1.0, + "content": "2020] and on three non-European low-resource languages.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 19, + "bbox_fs": [ + 104, + 432, + 507, + 512 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 529, + 295, + 543 + ], + "lines": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 104, + 527, + 296, + 547 + ], + "score": 1.0, + "content": "2 Speech and Text Representations", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "list", + "bbox": [ + 106, + 557, + 505, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 555, + 506, + 571 + ], + "score": 1.0, + "content": "Next, we describe how we build suitable speech and text representations for unsupervised learning.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 498, + 581 + ], + "score": 1.0, + "content": "Good representations are essential to learning a mapping from speech to text without supervision.", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 555, + 506, + 581 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 595, + 379, + 608 + ], + "lines": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "spans": [ + { + "bbox": [ + 104, + 594, + 380, + 612 + ], + "score": 1.0, + "content": "2.1 Self-supervised Learning of Speech Audio Representations", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 662 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 507, + 631 + ], + "score": 1.0, + "content": "In the first step, we learn representations of the speech audio signal using self-supervised learning.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 505, + 641 + ], + "score": 1.0, + "content": "There has been a lot of recent work in this direction which has shown strong performance in extremely", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 507, + 653 + ], + "score": 1.0, + "content": "low-labeled data setups across a range of languages [Conneau et al., 2020] and tasks [Fan et al., 2021,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 264, + 663 + ], + "score": 1.0, + "content": "Pepino et al., 2021, Wang et al., 2021].", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 617, + 507, + 663 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 667, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 667, + 505, + 679 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 348, + 679 + ], + "score": 1.0, + "content": "Wav2vec 2.0 consists of a convolutional feature encoder", + "type": "text" + }, + { + "bbox": [ + 348, + 667, + 406, + 678 + ], + "score": 0.91, + "content": "f : \\mathcal X \\mapsto \\mathcal Z", + "type": "inline_equation" + }, + { + "bbox": [ + 406, + 667, + 505, + 679 + ], + "score": 1.0, + "content": "that maps a raw audio", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 677, + 506, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 145, + 691 + ], + "score": 1.0, + "content": "sequence", + "type": "text" + }, + { + "bbox": [ + 146, + 678, + 156, + 688 + ], + "score": 0.81, + "content": "X", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 677, + 285, + 691 + ], + "score": 1.0, + "content": "to latent speech representations", + "type": "text" + }, + { + "bbox": [ + 285, + 680, + 328, + 689 + ], + "score": 0.89, + "content": "z _ { 1 } , \\dots , z _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 677, + 417, + 691 + ], + "score": 1.0, + "content": ", which a Transformer", + "type": "text" + }, + { + "bbox": [ + 418, + 678, + 462, + 689 + ], + "score": 0.92, + "content": "g : { \\mathcal { Z } } \\mapsto { \\mathcal { C } }", + "type": "inline_equation" + }, + { + "bbox": [ + 462, + 677, + 506, + 691 + ], + "score": 1.0, + "content": "then turns", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 220, + 701 + ], + "score": 1.0, + "content": "into context representations", + "type": "text" + }, + { + "bbox": [ + 221, + 690, + 263, + 700 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 689, + 390, + 701 + ], + "score": 1.0, + "content": "[Baevski et al., 2020b,a]. Each", + "type": "text" + }, + { + "bbox": [ + 390, + 690, + 400, + 700 + ], + "score": 0.84, + "content": "z _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 689, + 469, + 701 + ], + "score": 1.0, + "content": "represents about", + "type": "text" + }, + { + "bbox": [ + 470, + 689, + 493, + 699 + ], + "score": 0.47, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 700, + 504, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 172, + 712 + ], + "score": 1.0, + "content": "audio strided by", + "type": "text" + }, + { + "bbox": [ + 173, + 700, + 196, + 711 + ], + "score": 0.31, + "content": "2 0 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 700, + 504, + 712 + ], + "score": 1.0, + "content": "and the Transformer architecture follows BERT [Vaswani et al., 2017, Devlin", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 711, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 382, + 724 + ], + "score": 1.0, + "content": "et al., 2019]. During training, latent representations are discretized to", + "type": "text" + }, + { + "bbox": [ + 383, + 712, + 425, + 722 + ], + "score": 0.9, + "content": "q _ { 1 } , \\ldots , q _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 711, + 505, + 724 + ], + "score": 1.0, + "content": "with a quantization", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 140, + 85 + ], + "score": 1.0, + "content": "module", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 140, + 73, + 174, + 83 + ], + "score": 0.91, + "content": "{ \\mathcal { Z } } \\mapsto { \\mathcal { Q } }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 175, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "to represent the targets in the objective. Quantization uses a Gumbel softmax to", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "score": 1.0, + "content": "choose entries from two codebooks [Jegou et al., 2011, Jang et al., 2016, Baevski et al., 2020b].", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 667, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 504, + 95 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 140, + 85 + ], + "score": 1.0, + "content": "module", + "type": "text" + }, + { + "bbox": [ + 140, + 73, + 174, + 83 + ], + "score": 0.91, + "content": "{ \\mathcal { Z } } \\mapsto { \\mathcal { Q } }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 72, + 505, + 85 + ], + "score": 1.0, + "content": "to represent the targets in the objective. Quantization uses a Gumbel softmax to", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 491, + 96 + ], + "score": 1.0, + "content": "choose entries from two codebooks [Jegou et al., 2011, Jang et al., 2016, Baevski et al., 2020b].", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 426, + 112 + ], + "score": 1.0, + "content": "In our experiments, we use the publicly available English model pre-trained on", + "type": "text" + }, + { + "bbox": [ + 427, + 100, + 444, + 110 + ], + "score": 0.38, + "content": "5 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "hours of Libri-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Light [Kahn et al., 2020b] as well as XLSR-53 which was pre-trained on nearly 60k hours of speech", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "score": 1.0, + "content": "audio in 53 languages [Conneau et al., 2020].", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3 + }, + { + "type": "title", + "bbox": [ + 107, + 146, + 331, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "score": 1.0, + "content": "2.2 Pre-processing and Embedding the Audio Data", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "Removing Silences. Most datasets we use for our experiments have audio data with silences.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "However, these parts of the audio do not correspond to any transcription and we therefore remove", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "silences as much as possible. We apply rVAD, an unsupervised voice activity detection (VAD) model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "which determines the segments in the audio data corresponding to silences, and we remove these", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "score": 1.0, + "content": "sections [Tan et al., 2020]. We ablate this choice in Appendix C.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "Speech Audio Representations. After silence removal, we embed the unlabeled speech audio with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "score": 1.0, + "content": "wav2vec 2.0 to obtain speech representations. Specifically, we use the representations of the context", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 194, + 267 + ], + "score": 1.0, + "content": "Transformer network", + "type": "text" + }, + { + "bbox": [ + 195, + 256, + 237, + 266 + ], + "score": 0.86, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 255, + 266, + 266 + ], + "score": 0.66, + "content": "( \\ S 2 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 253, + 506, + 267 + ], + "score": 1.0, + "content": ". The context network contains 24 Transformer blocks and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 232, + 279 + ], + "score": 1.0, + "content": "we denote the output of block", + "type": "text" + }, + { + "bbox": [ + 232, + 266, + 236, + 275 + ], + "score": 0.69, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 264, + 289, + 279 + ], + "score": 1.0, + "content": "at time-step", + "type": "text" + }, + { + "bbox": [ + 289, + 266, + 294, + 275 + ], + "score": 0.63, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 264, + 307, + 279 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 307, + 265, + 316, + 277 + ], + "score": 0.87, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 264, + 505, + 279 + ], + "score": 1.0, + "content": ". Our goal is to learn a model which can map", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 214, + 290 + ], + "score": 1.0, + "content": "from audio representations", + "type": "text" + }, + { + "bbox": [ + 214, + 276, + 223, + 289 + ], + "score": 0.89, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "to phonemes using no supervision. However, the representations of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "uppermost block of wav2vec 2.0 may not be well suited for this task. These features are trained to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 326, + 312 + ], + "score": 1.0, + "content": "directly predict masked latent representations spanning", + "type": "text" + }, + { + "bbox": [ + 326, + 299, + 349, + 310 + ], + "score": 0.55, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "of speech audio which is much shorter", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "score": 1.0, + "content": "than the typical duration of a phoneme.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "To get a better sense of this, we train supervised phoneme recognizers with a CTC loss [Graves", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "et al., 2006] on top of the frozen representations of each of the 24 blocks of the English wav2vec 2.0", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "LARGE model pre-trained on Libri-Light. We then evaluate phone error rate (PER) with respect to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 437, + 372 + ], + "score": 1.0, + "content": "the phonemized transcriptions of Librispeech dev-other. The classifier takes as input", + "type": "text" + }, + { + "bbox": [ + 437, + 359, + 446, + 371 + ], + "score": 0.88, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 358, + 506, + 372 + ], + "score": 1.0, + "content": "and contains a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "single softmax-normalized linear layer mapping to the phoneme inventory. Figure 2 shows that most", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "of the first ten blocks as well as the final blocks provide very poor performance, while blocks 15-19", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 212, + 405 + ], + "score": 1.0, + "content": "provide error rates below", + "type": "text" + }, + { + "bbox": [ + 212, + 392, + 227, + 402 + ], + "score": 0.86, + "content": "9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 391, + 416, + 405 + ], + "score": 1.0, + "content": "PER. Block 15 achieves the best error rate of", + "type": "text" + }, + { + "bbox": [ + 416, + 392, + 439, + 402 + ], + "score": 0.87, + "content": "7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "PER. A similar", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "insight has been used in the concurrent work of Hsu et al. [2021b]. Appendix A shows that this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 411, + 426 + ], + "score": 1.0, + "content": "choice generalizes to other languages. For brevity we drop the superscript", + "type": "text" + }, + { + "bbox": [ + 411, + 414, + 416, + 424 + ], + "score": 0.49, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "and refer to block 15", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 257, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 209, + 438 + ], + "score": 1.0, + "content": "representations simply as", + "type": "text" + }, + { + "bbox": [ + 210, + 426, + 252, + 436 + ], + "score": 0.89, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 424, + 257, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "spans": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "score": 0.965, + "type": "image", + "image_path": "4893afedc781e5fe6ccd0b0db23368e07e04f56d069fce1ae7128f54dac18113.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 108, + 448, + 492, + 476.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 108, + 476.0, + 492, + 504.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 504.0, + 492, + 532.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 538, + 504, + 572 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "Figure 2: Supervised phoneme recognition using representations from different wav2vec 2.0 blocks", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "on dev-other of English Librispeech. Low and high blocks do not provide good features, while as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "score": 1.0, + "content": "blocks 14-19 do. Block 15 performs best.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 592, + 253, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "score": 1.0, + "content": "2.3 Segmenting the Audio Signal", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "Once the speech signal is embedded, we identify segments corresponding to meaningful units that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "score": 1.0, + "content": "can be mapped to phonemes. Segmentation has been shown to be crucial in prior work [Chung", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "et al., 2018] since the right boundaries in the input representations make it more aligned to phonetic", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "sequences. There has been a lot of prior work in unsupervised speech segmentation [Kamper et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "2017a,b, Rasanen et al., 2015, Kreuk et al., 2020] but here we simply use a method based on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 304, + 681 + ], + "score": 1.0, + "content": "clustering the wav2vec 2.0 speech representations", + "type": "text" + }, + { + "bbox": [ + 304, + 669, + 347, + 678 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 667, + 506, + 681 + ], + "score": 1.0, + "content": ". In a first step, we collect all the speech", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 465, + 690 + ], + "score": 1.0, + "content": "representations for the unlabeled speech data and perform k-means clustering to identify", + "type": "text" + }, + { + "bbox": [ + 466, + 678, + 504, + 688 + ], + "score": 0.89, + "content": "K = 1 2 8", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "clusters. We use the FAISS library to do fast clustering on GPUs [Johnson et al., 2019]. Next, each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 115, + 711 + ], + "score": 0.84, + "content": "c _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 700, + 297, + 713 + ], + "score": 1.0, + "content": "is labeled with the corresponding cluster ID", + "type": "text" + }, + { + "bbox": [ + 298, + 700, + 364, + 712 + ], + "score": 0.84, + "content": "i _ { t } \\in \\{ 1 , \\ldots , K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "and we introduce speech segment", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "score": 1.0, + "content": "boundaries whenever the cluster ID changes.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 73, + 504, + 95 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 106, + 72, + 505, + 96 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 426, + 112 + ], + "score": 1.0, + "content": "In our experiments, we use the publicly available English model pre-trained on", + "type": "text" + }, + { + "bbox": [ + 427, + 100, + 444, + 110 + ], + "score": 0.38, + "content": "5 3 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 100, + 505, + 112 + ], + "score": 1.0, + "content": "hours of Libri-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 111, + 505, + 123 + ], + "score": 1.0, + "content": "Light [Kahn et al., 2020b] as well as XLSR-53 which was pre-trained on nearly 60k hours of speech", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 290, + 134 + ], + "score": 1.0, + "content": "audio in 53 languages [Conneau et al., 2020].", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3, + "bbox_fs": [ + 105, + 100, + 505, + 134 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 146, + 331, + 158 + ], + "lines": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "spans": [ + { + "bbox": [ + 105, + 145, + 332, + 160 + ], + "score": 1.0, + "content": "2.2 Pre-processing and Embedding the Audio Data", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 166, + 505, + 221 + ], + "lines": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 166, + 506, + 178 + ], + "score": 1.0, + "content": "Removing Silences. Most datasets we use for our experiments have audio data with silences.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 505, + 189 + ], + "score": 1.0, + "content": "However, these parts of the audio do not correspond to any transcription and we therefore remove", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 188, + 505, + 200 + ], + "score": 1.0, + "content": "silences as much as possible. We apply rVAD, an unsupervised voice activity detection (VAD) model", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 211 + ], + "score": 1.0, + "content": "which determines the segments in the audio data corresponding to silences, and we remove these", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 366, + 222 + ], + "score": 1.0, + "content": "sections [Tan et al., 2020]. We ablate this choice in Appendix C.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 166, + 506, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 505, + 245 + ], + "score": 1.0, + "content": "Speech Audio Representations. After silence removal, we embed the unlabeled speech audio with", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 256 + ], + "score": 1.0, + "content": "wav2vec 2.0 to obtain speech representations. Specifically, we use the representations of the context", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 194, + 267 + ], + "score": 1.0, + "content": "Transformer network", + "type": "text" + }, + { + "bbox": [ + 195, + 256, + 237, + 266 + ], + "score": 0.86, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 255, + 266, + 266 + ], + "score": 0.66, + "content": "( \\ S 2 . 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 253, + 506, + 267 + ], + "score": 1.0, + "content": ". The context network contains 24 Transformer blocks and", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 264, + 505, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 232, + 279 + ], + "score": 1.0, + "content": "we denote the output of block", + "type": "text" + }, + { + "bbox": [ + 232, + 266, + 236, + 275 + ], + "score": 0.69, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 237, + 264, + 289, + 279 + ], + "score": 1.0, + "content": "at time-step", + "type": "text" + }, + { + "bbox": [ + 289, + 266, + 294, + 275 + ], + "score": 0.63, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 264, + 307, + 279 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 307, + 265, + 316, + 277 + ], + "score": 0.87, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 316, + 264, + 505, + 279 + ], + "score": 1.0, + "content": ". Our goal is to learn a model which can map", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 276, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 214, + 290 + ], + "score": 1.0, + "content": "from audio representations", + "type": "text" + }, + { + "bbox": [ + 214, + 276, + 223, + 289 + ], + "score": 0.89, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 277, + 505, + 290 + ], + "score": 1.0, + "content": "to phonemes using no supervision. However, the representations of the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 505, + 301 + ], + "score": 1.0, + "content": "uppermost block of wav2vec 2.0 may not be well suited for this task. These features are trained to", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 326, + 312 + ], + "score": 1.0, + "content": "directly predict masked latent representations spanning", + "type": "text" + }, + { + "bbox": [ + 326, + 299, + 349, + 310 + ], + "score": 0.55, + "content": "2 5 \\mathrm { m s }", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "of speech audio which is much shorter", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 310, + 266, + 323 + ], + "score": 1.0, + "content": "than the typical duration of a phoneme.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 233, + 506, + 323 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 326, + 505, + 436 + ], + "lines": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 327, + 505, + 339 + ], + "score": 1.0, + "content": "To get a better sense of this, we train supervised phoneme recognizers with a CTC loss [Graves", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 506, + 349 + ], + "score": 1.0, + "content": "et al., 2006] on top of the frozen representations of each of the 24 blocks of the English wav2vec 2.0", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 506, + 361 + ], + "score": 1.0, + "content": "LARGE model pre-trained on Libri-Light. We then evaluate phone error rate (PER) with respect to", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 358, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 358, + 437, + 372 + ], + "score": 1.0, + "content": "the phonemized transcriptions of Librispeech dev-other. The classifier takes as input", + "type": "text" + }, + { + "bbox": [ + 437, + 359, + 446, + 371 + ], + "score": 0.88, + "content": "c _ { t } ^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 358, + 506, + 372 + ], + "score": 1.0, + "content": "and contains a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "single softmax-normalized linear layer mapping to the phoneme inventory. Figure 2 shows that most", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 394 + ], + "score": 1.0, + "content": "of the first ten blocks as well as the final blocks provide very poor performance, while blocks 15-19", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 391, + 506, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 212, + 405 + ], + "score": 1.0, + "content": "provide error rates below", + "type": "text" + }, + { + "bbox": [ + 212, + 392, + 227, + 402 + ], + "score": 0.86, + "content": "9 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 391, + 416, + 405 + ], + "score": 1.0, + "content": "PER. Block 15 achieves the best error rate of", + "type": "text" + }, + { + "bbox": [ + 416, + 392, + 439, + 402 + ], + "score": 0.87, + "content": "7 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 391, + 506, + 405 + ], + "score": 1.0, + "content": "PER. A similar", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 505, + 415 + ], + "score": 1.0, + "content": "insight has been used in the concurrent work of Hsu et al. [2021b]. Appendix A shows that this", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 414, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 411, + 426 + ], + "score": 1.0, + "content": "choice generalizes to other languages. For brevity we drop the superscript", + "type": "text" + }, + { + "bbox": [ + 411, + 414, + 416, + 424 + ], + "score": 0.49, + "content": "l", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 414, + 505, + 426 + ], + "score": 1.0, + "content": "and refer to block 15", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 424, + 257, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 209, + 438 + ], + "score": 1.0, + "content": "representations simply as", + "type": "text" + }, + { + "bbox": [ + 210, + 426, + 252, + 436 + ], + "score": 0.89, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 424, + 257, + 438 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 327, + 506, + 438 + ] + }, + { + "type": "image", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 448, + 492, + 532 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "spans": [ + { + "bbox": [ + 108, + 448, + 492, + 532 + ], + "score": 0.965, + "type": "image", + "image_path": "4893afedc781e5fe6ccd0b0db23368e07e04f56d069fce1ae7128f54dac18113.jpg" + } + ] + } + ], + "index": 30, + "virtual_lines": [ + { + "bbox": [ + 108, + 448, + 492, + 476.0 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 108, + 476.0, + 492, + 504.0 + ], + "spans": [], + "index": 30 + }, + { + "bbox": [ + 108, + 504.0, + 492, + 532.0 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 538, + 504, + 572 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 505, + 552 + ], + "score": 1.0, + "content": "Figure 2: Supervised phoneme recognition using representations from different wav2vec 2.0 blocks", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 562 + ], + "score": 1.0, + "content": "on dev-other of English Librispeech. Low and high blocks do not provide good features, while as", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 560, + 274, + 573 + ], + "score": 1.0, + "content": "blocks 14-19 do. Block 15 performs best.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 108, + 592, + 253, + 604 + ], + "lines": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 254, + 606 + ], + "score": 1.0, + "content": "2.3 Segmenting the Audio Signal", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 612, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 625 + ], + "score": 1.0, + "content": "Once the speech signal is embedded, we identify segments corresponding to meaningful units that", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 637 + ], + "score": 1.0, + "content": "can be mapped to phonemes. Segmentation has been shown to be crucial in prior work [Chung", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "spans": [ + { + "bbox": [ + 105, + 634, + 505, + 647 + ], + "score": 1.0, + "content": "et al., 2018] since the right boundaries in the input representations make it more aligned to phonetic", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 506, + 658 + ], + "score": 1.0, + "content": "sequences. There has been a lot of prior work in unsupervised speech segmentation [Kamper et al.,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 506, + 668 + ], + "score": 1.0, + "content": "2017a,b, Rasanen et al., 2015, Kreuk et al., 2020] but here we simply use a method based on", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 304, + 681 + ], + "score": 1.0, + "content": "clustering the wav2vec 2.0 speech representations", + "type": "text" + }, + { + "bbox": [ + 304, + 669, + 347, + 678 + ], + "score": 0.87, + "content": "c _ { 1 } , \\ldots , c _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 667, + 506, + 681 + ], + "score": 1.0, + "content": ". In a first step, we collect all the speech", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 678, + 504, + 690 + ], + "spans": [ + { + "bbox": [ + 104, + 678, + 465, + 690 + ], + "score": 1.0, + "content": "representations for the unlabeled speech data and perform k-means clustering to identify", + "type": "text" + }, + { + "bbox": [ + 466, + 678, + 504, + 688 + ], + "score": 0.89, + "content": "K = 1 2 8", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 505, + 702 + ], + "score": 1.0, + "content": "clusters. We use the FAISS library to do fast clustering on GPUs [Johnson et al., 2019]. Next, each", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 700, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 106, + 701, + 115, + 711 + ], + "score": 0.84, + "content": "c _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 700, + 297, + 713 + ], + "score": 1.0, + "content": "is labeled with the corresponding cluster ID", + "type": "text" + }, + { + "bbox": [ + 298, + 700, + 364, + 712 + ], + "score": 0.84, + "content": "i _ { t } \\in \\{ 1 , \\ldots , K \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 700, + 506, + 713 + ], + "score": 1.0, + "content": "and we introduce speech segment", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 287, + 724 + ], + "score": 1.0, + "content": "boundaries whenever the cluster ID changes.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5, + "bbox_fs": [ + 104, + 612, + 506, + 724 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Once the speech audio representations are segmented, we compute a 512-dimensional PCA over", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "all speech representations output by wav2vec 2.0 for the training set. Next, we mean-pool the PCA", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "representations for a particular segment to obtain an average representation of the segment. The", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "PCA retains only the most important features and we found this to be effective. Segment boundaries", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "are noisy due to the lack of supervision and we therefore found it useful to also mean-pool pairs of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "adjacent segment representations to increase robustness. This results in sequences of speech segment", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 167, + 151 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 167, + 138, + 263, + 149 + ], + "score": 0.91, + "content": "S = s _ { 1 } , \\ldots , s _ { T } , S \\sim { \\mathcal { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "for a given utterance. Appendix B shows an illustration of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "score": 1.0, + "content": "the segmentation strategy on an actual example as well as a quantitative evaluation of the strategy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "score": 1.0, + "content": "compared to human segmented data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "title", + "bbox": [ + 108, + 188, + 252, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "score": 1.0, + "content": "2.4 Pre-processing the Text Data", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 209, + 504, + 243 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "Similar to how we segment the unlabeled speech audio data into suitable units for unsupervised", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "learning, we do the same for the unlabeled text data. We apply two pre-processing steps to the text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "spans": [ + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "score": 1.0, + "content": "data: phonemization and silence token insertion.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 248, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "score": 1.0, + "content": "Phonemes characterize the different sounds which distinguish words from each other, e.g., for the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "score": 1.0, + "content": "word cat there are three phonemes corresponding to the three distinct sounds in the pronunciation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 207, + 282 + ], + "score": 1.0, + "content": "of the word: /K/, /AE/,", + "type": "text" + }, + { + "bbox": [ + 207, + 270, + 222, + 281 + ], + "score": 0.31, + "content": "/ \\mathrm { T } / .", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 270, + 506, + 282 + ], + "score": 1.0, + "content": ". We phonemize the text data because we found it easier to learn a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "mapping between speech audio and the different sounds of a word rather than between audio and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 370, + 303 + ], + "score": 1.0, + "content": "words or letters. Phonemization converts a sequence of words", + "type": "text" + }, + { + "bbox": [ + 370, + 292, + 380, + 302 + ], + "score": 0.73, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "into a sequence of phonemes", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 181, + 315 + ], + "score": 0.91, + "content": "P = [ p _ { 1 } , \\cdots , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 302, + 212, + 316 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 213, + 303, + 247, + 314 + ], + "score": 0.92, + "content": "p _ { m } \\in O", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 302, + 265, + 316 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 266, + 303, + 275, + 313 + ], + "score": 0.82, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "is the phoneme inventory. We use off-the-shelf tools for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 314, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 259, + 326 + ], + "score": 1.0, + "content": "this step which we detail in Appendix", + "type": "text" + }, + { + "bbox": [ + 259, + 314, + 282, + 325 + ], + "score": 0.73, + "content": "\\ S \\operatorname { E } . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 314, + 286, + 326 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 506, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "score": 1.0, + "content": "The unlabeled speech audio data is pre-processed by applying unsupervised silence removal. However,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "this process is not always accurate and many silences in the speech audio remain. To deal with this,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 479, + 364 + ], + "score": 1.0, + "content": "we enable the unsupervised model to label some segments with a phonemic silence token (SIL;", + "type": "text" + }, + { + "bbox": [ + 479, + 352, + 501, + 363 + ], + "score": 0.7, + "content": "\\ S \\ 3 . 1 \\AA .", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 352, + 506, + 364 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "However, the phonemized unlabeled text data does not contain any silence tokens and this may pose", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 243, + 387 + ], + "score": 1.0, + "content": "difficulties for adversarial learning", + "type": "text" + }, + { + "bbox": [ + 244, + 374, + 262, + 385 + ], + "score": 0.42, + "content": "( \\ S 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 373, + 506, + 387 + ], + "score": 1.0, + "content": ". We remedy this by inserting silence markers at the beginning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "and end of the phonemized unlabeled text data; we also randomly insert SIL between words, or", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 396, + 504, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 331, + 408 + ], + "score": 1.0, + "content": "groups of phonemes corresponding to words at a rate of", + "type": "text" + }, + { + "bbox": [ + 331, + 396, + 351, + 406 + ], + "score": 0.87, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 396, + 504, + 408 + ], + "score": 1.0, + "content": ". Appendix C evaluates these choices.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 108, + 425, + 246, + 440 + ], + "lines": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "spans": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "score": 1.0, + "content": "3 Unsupervised Learning", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 506, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "score": 1.0, + "content": "We use adversarial training to train an unsupervised speech recognition model using the representa-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "tions of the unlabeled speech audio data and the unlabeled phonemized text data [Liu et al., 2018,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "Chen et al., 2019]. In the following, we detail the model architecture, the training objective as well as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "score": 1.0, + "content": "the unsupervised cross-validation metric we developed.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 214, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "score": 1.0, + "content": "3.1 Model Architecture", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 478, + 549 + ], + "score": 1.0, + "content": "Generative adversarial networks (GAN; Goodfellow et al. 2014) train a generator network", + "type": "text" + }, + { + "bbox": [ + 478, + 537, + 486, + 547 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 230, + 560 + ], + "score": 1.0, + "content": "a discriminator/critic network", + "type": "text" + }, + { + "bbox": [ + 231, + 547, + 238, + 557 + ], + "score": 0.79, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "where the generator produces samples which are then judged by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "score": 1.0, + "content": "the discriminator. The discriminator is trained to classify whether samples are from the generator", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "or from the real data distribution. The objective of the generator is to produce samples that are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "score": 1.0, + "content": "indistinguishable by the discriminator.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 153, + 609 + ], + "score": 1.0, + "content": "Concretely,", + "type": "text" + }, + { + "bbox": [ + 154, + 597, + 162, + 607 + ], + "score": 0.8, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 595, + 273, + 609 + ], + "score": 1.0, + "content": "takes as input a sequence of", + "type": "text" + }, + { + "bbox": [ + 274, + 597, + 282, + 606 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 595, + 380, + 609 + ], + "score": 1.0, + "content": "segment representations", + "type": "text" + }, + { + "bbox": [ + 380, + 596, + 449, + 608 + ], + "score": 0.83, + "content": "S = [ s _ { 1 } , \\dotsc , s _ { T } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "(§ 2.3) which", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 249, + 619 + ], + "score": 1.0, + "content": "are then mapped to a sequence of", + "type": "text" + }, + { + "bbox": [ + 249, + 608, + 261, + 617 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 606, + 307, + 619 + ], + "score": 1.0, + "content": "phonemes", + "type": "text" + }, + { + "bbox": [ + 307, + 607, + 396, + 619 + ], + "score": 0.93, + "content": "\\mathcal { G } ( S ) \\bar { = } [ p _ { 1 } , \\dotsc , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 606, + 505, + 619 + ], + "score": 1.0, + "content": ". The generator predicts a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 246, + 631 + ], + "score": 1.0, + "content": "distribution over the phoneme set", + "type": "text" + }, + { + "bbox": [ + 247, + 618, + 256, + 628 + ], + "score": 0.8, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "for each segment and outputs the phoneme with the highest", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "probability. If the argmax prediction of consecutive segments result in the same phoneme, then we", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 305, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 268, + 653 + ], + "score": 1.0, + "content": "sample one of these segments, therefore", + "type": "text" + }, + { + "bbox": [ + 268, + 640, + 300, + 651 + ], + "score": 0.91, + "content": "M \\leq T", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 639, + 305, + 653 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 178, + 669 + ], + "score": 1.0, + "content": "The phoneme set", + "type": "text" + }, + { + "bbox": [ + 178, + 657, + 187, + 666 + ], + "score": 0.74, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "includes a silence label SIL to enable labeling silences in the speech audio as", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "such. Without a silence label, we noticed that the model was repurposing a particular phoneme to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "score": 1.0, + "content": "label silences which resulted in much lower performance since it interfered with subsequent language", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "model (LM) decoding. In the backward pass, we back-propagate through segments sampled at the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 360, + 713 + ], + "score": 1.0, + "content": "generator output. We do not modify the segment representations", + "type": "text" + }, + { + "bbox": [ + 361, + 701, + 368, + 710 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "during unsupervised training. The", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "score": 1.0, + "content": "generator is parameterized as a single layer convolutional neural network (CNN).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 505, + 171 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 506, + 86 + ], + "score": 1.0, + "content": "Once the speech audio representations are segmented, we compute a 512-dimensional PCA over", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "score": 1.0, + "content": "all speech representations output by wav2vec 2.0 for the training set. Next, we mean-pool the PCA", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "representations for a particular segment to obtain an average representation of the segment. The", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "PCA retains only the most important features and we found this to be effective. Segment boundaries", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 506, + 129 + ], + "score": 1.0, + "content": "are noisy due to the lack of supervision and we therefore found it useful to also mean-pool pairs of", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 106, + 127, + 506, + 141 + ], + "score": 1.0, + "content": "adjacent segment representations to increase robustness. This results in sequences of speech segment", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 138, + 506, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 167, + 151 + ], + "score": 1.0, + "content": "representation", + "type": "text" + }, + { + "bbox": [ + 167, + 138, + 263, + 149 + ], + "score": 0.91, + "content": "S = s _ { 1 } , \\ldots , s _ { T } , S \\sim { \\mathcal { S } }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 138, + 506, + 151 + ], + "score": 1.0, + "content": "for a given utterance. Appendix B shows an illustration of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 506, + 163 + ], + "score": 1.0, + "content": "the segmentation strategy on an actual example as well as a quantitative evaluation of the strategy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "spans": [ + { + "bbox": [ + 106, + 160, + 255, + 173 + ], + "score": 1.0, + "content": "compared to human segmented data.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 73, + 506, + 173 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 188, + 252, + 200 + ], + "lines": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 187, + 254, + 202 + ], + "score": 1.0, + "content": "2.4 Pre-processing the Text Data", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "text", + "bbox": [ + 108, + 209, + 504, + 243 + ], + "lines": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 505, + 222 + ], + "score": 1.0, + "content": "Similar to how we segment the unlabeled speech audio data into suitable units for unsupervised", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 221, + 505, + 233 + ], + "score": 1.0, + "content": "learning, we do the same for the unlabeled text data. We apply two pre-processing steps to the text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "spans": [ + { + "bbox": [ + 107, + 233, + 302, + 244 + ], + "score": 1.0, + "content": "data: phonemization and silence token insertion.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 210, + 505, + 244 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 248, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "spans": [ + { + "bbox": [ + 106, + 249, + 504, + 260 + ], + "score": 1.0, + "content": "Phonemes characterize the different sounds which distinguish words from each other, e.g., for the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 260, + 505, + 271 + ], + "score": 1.0, + "content": "word cat there are three phonemes corresponding to the three distinct sounds in the pronunciation", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 270, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 207, + 282 + ], + "score": 1.0, + "content": "of the word: /K/, /AE/,", + "type": "text" + }, + { + "bbox": [ + 207, + 270, + 222, + 281 + ], + "score": 0.31, + "content": "/ \\mathrm { T } / .", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 270, + 506, + 282 + ], + "score": 1.0, + "content": ". We phonemize the text data because we found it easier to learn a", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 281, + 506, + 293 + ], + "score": 1.0, + "content": "mapping between speech audio and the different sounds of a word rather than between audio and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 291, + 505, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 370, + 303 + ], + "score": 1.0, + "content": "words or letters. Phonemization converts a sequence of words", + "type": "text" + }, + { + "bbox": [ + 370, + 292, + 380, + 302 + ], + "score": 0.73, + "content": "Y", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 291, + 505, + 303 + ], + "score": 1.0, + "content": "into a sequence of phonemes", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 302, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 107, + 303, + 181, + 315 + ], + "score": 0.91, + "content": "P = [ p _ { 1 } , \\cdots , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 302, + 212, + 316 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 213, + 303, + 247, + 314 + ], + "score": 0.92, + "content": "p _ { m } \\in O", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 302, + 265, + 316 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 266, + 303, + 275, + 313 + ], + "score": 0.82, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 302, + 506, + 316 + ], + "score": 1.0, + "content": "is the phoneme inventory. We use off-the-shelf tools for", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 314, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 314, + 259, + 326 + ], + "score": 1.0, + "content": "this step which we detail in Appendix", + "type": "text" + }, + { + "bbox": [ + 259, + 314, + 282, + 325 + ], + "score": 0.73, + "content": "\\ S \\operatorname { E } . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 314, + 286, + 326 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 249, + 506, + 326 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 330, + 506, + 407 + ], + "lines": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 507, + 343 + ], + "score": 1.0, + "content": "The unlabeled speech audio data is pre-processed by applying unsupervised silence removal. However,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 341, + 506, + 353 + ], + "score": 1.0, + "content": "this process is not always accurate and many silences in the speech audio remain. To deal with this,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 352, + 506, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 352, + 479, + 364 + ], + "score": 1.0, + "content": "we enable the unsupervised model to label some segments with a phonemic silence token (SIL;", + "type": "text" + }, + { + "bbox": [ + 479, + 352, + 501, + 363 + ], + "score": 0.7, + "content": "\\ S \\ 3 . 1 \\AA .", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 352, + 506, + 364 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 362, + 506, + 376 + ], + "score": 1.0, + "content": "However, the phonemized unlabeled text data does not contain any silence tokens and this may pose", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 373, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 243, + 387 + ], + "score": 1.0, + "content": "difficulties for adversarial learning", + "type": "text" + }, + { + "bbox": [ + 244, + 374, + 262, + 385 + ], + "score": 0.42, + "content": "( \\ S 3 )", + "type": "inline_equation" + }, + { + "bbox": [ + 262, + 373, + 506, + 387 + ], + "score": 1.0, + "content": ". We remedy this by inserting silence markers at the beginning", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 385, + 506, + 397 + ], + "score": 1.0, + "content": "and end of the phonemized unlabeled text data; we also randomly insert SIL between words, or", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 396, + 504, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 331, + 408 + ], + "score": 1.0, + "content": "groups of phonemes corresponding to words at a rate of", + "type": "text" + }, + { + "bbox": [ + 331, + 396, + 351, + 406 + ], + "score": 0.87, + "content": "2 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 351, + 396, + 504, + 408 + ], + "score": 1.0, + "content": ". Appendix C evaluates these choices.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 329, + 507, + 408 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 425, + 246, + 440 + ], + "lines": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "spans": [ + { + "bbox": [ + 103, + 423, + 248, + 444 + ], + "score": 1.0, + "content": "3 Unsupervised Learning", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 453, + 506, + 497 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 507, + 466 + ], + "score": 1.0, + "content": "We use adversarial training to train an unsupervised speech recognition model using the representa-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 106, + 464, + 506, + 476 + ], + "score": 1.0, + "content": "tions of the unlabeled speech audio data and the unlabeled phonemized text data [Liu et al., 2018,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 488 + ], + "score": 1.0, + "content": "Chen et al., 2019]. In the following, we detail the model architecture, the training objective as well as", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 487, + 329, + 498 + ], + "score": 1.0, + "content": "the unsupervised cross-validation metric we developed.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 452, + 507, + 498 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 514, + 214, + 526 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 215, + 527 + ], + "score": 1.0, + "content": "3.1 Model Architecture", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 536, + 505, + 591 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 478, + 549 + ], + "score": 1.0, + "content": "Generative adversarial networks (GAN; Goodfellow et al. 2014) train a generator network", + "type": "text" + }, + { + "bbox": [ + 478, + 537, + 486, + 547 + ], + "score": 0.83, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 487, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 230, + 560 + ], + "score": 1.0, + "content": "a discriminator/critic network", + "type": "text" + }, + { + "bbox": [ + 231, + 547, + 238, + 557 + ], + "score": 0.79, + "content": "\\mathcal { C }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 546, + 505, + 560 + ], + "score": 1.0, + "content": "where the generator produces samples which are then judged by", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 506, + 571 + ], + "score": 1.0, + "content": "the discriminator. The discriminator is trained to classify whether samples are from the generator", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 505, + 582 + ], + "score": 1.0, + "content": "or from the real data distribution. The objective of the generator is to produce samples that are", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "spans": [ + { + "bbox": [ + 106, + 579, + 262, + 593 + ], + "score": 1.0, + "content": "indistinguishable by the discriminator.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 536, + 506, + 593 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 105, + 595, + 506, + 609 + ], + "spans": [ + { + "bbox": [ + 105, + 595, + 153, + 609 + ], + "score": 1.0, + "content": "Concretely,", + "type": "text" + }, + { + "bbox": [ + 154, + 597, + 162, + 607 + ], + "score": 0.8, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 162, + 595, + 273, + 609 + ], + "score": 1.0, + "content": "takes as input a sequence of", + "type": "text" + }, + { + "bbox": [ + 274, + 597, + 282, + 606 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 283, + 595, + 380, + 609 + ], + "score": 1.0, + "content": "segment representations", + "type": "text" + }, + { + "bbox": [ + 380, + 596, + 449, + 608 + ], + "score": 0.83, + "content": "S = [ s _ { 1 } , \\dotsc , s _ { T } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 449, + 595, + 506, + 609 + ], + "score": 1.0, + "content": "(§ 2.3) which", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 249, + 619 + ], + "score": 1.0, + "content": "are then mapped to a sequence of", + "type": "text" + }, + { + "bbox": [ + 249, + 608, + 261, + 617 + ], + "score": 0.82, + "content": "M", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 606, + 307, + 619 + ], + "score": 1.0, + "content": "phonemes", + "type": "text" + }, + { + "bbox": [ + 307, + 607, + 396, + 619 + ], + "score": 0.93, + "content": "\\mathcal { G } ( S ) \\bar { = } [ p _ { 1 } , \\dotsc , p _ { M } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 396, + 606, + 505, + 619 + ], + "score": 1.0, + "content": ". The generator predicts a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 246, + 631 + ], + "score": 1.0, + "content": "distribution over the phoneme set", + "type": "text" + }, + { + "bbox": [ + 247, + 618, + 256, + 628 + ], + "score": 0.8, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "for each segment and outputs the phoneme with the highest", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "probability. If the argmax prediction of consecutive segments result in the same phoneme, then we", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 305, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 268, + 653 + ], + "score": 1.0, + "content": "sample one of these segments, therefore", + "type": "text" + }, + { + "bbox": [ + 268, + 640, + 300, + 651 + ], + "score": 0.91, + "content": "M \\leq T", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 639, + 305, + 653 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 595, + 506, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 504, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 506, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 178, + 669 + ], + "score": 1.0, + "content": "The phoneme set", + "type": "text" + }, + { + "bbox": [ + 178, + 657, + 187, + 666 + ], + "score": 0.74, + "content": "O", + "type": "inline_equation" + }, + { + "bbox": [ + 188, + 655, + 506, + 669 + ], + "score": 1.0, + "content": "includes a silence label SIL to enable labeling silences in the speech audio as", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 680 + ], + "score": 1.0, + "content": "such. Without a silence label, we noticed that the model was repurposing a particular phoneme to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 506, + 692 + ], + "score": 1.0, + "content": "label silences which resulted in much lower performance since it interfered with subsequent language", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 702 + ], + "score": 1.0, + "content": "model (LM) decoding. In the backward pass, we back-propagate through segments sampled at the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 360, + 713 + ], + "score": 1.0, + "content": "generator output. We do not modify the segment representations", + "type": "text" + }, + { + "bbox": [ + 361, + 701, + 368, + 710 + ], + "score": 0.82, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "during unsupervised training. The", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 433, + 723 + ], + "score": 1.0, + "content": "generator is parameterized as a single layer convolutional neural network (CNN).", + "type": "text" + } + ], + "index": 48 + } + ], + "index": 45.5, + "bbox_fs": [ + 105, + 655, + 506, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 304, + 86 + ], + "score": 1.0, + "content": "The discriminator takes as input either a sequence", + "type": "text" + }, + { + "bbox": [ + 304, + 73, + 343, + 83 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "of one-hot vectors denoting phonemized", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 244, + 96 + ], + "score": 1.0, + "content": "text from the real data distribution", + "type": "text" + }, + { + "bbox": [ + 244, + 84, + 257, + 94 + ], + "score": 0.85, + "content": "{ \\mathcal { P } } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 83, + 481, + 96 + ], + "score": 1.0, + "content": "or a sequence of output distributions from the generator", + "type": "text" + }, + { + "bbox": [ + 481, + 83, + 502, + 96 + ], + "score": 0.92, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 83, + 506, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 195, + 108 + ], + "score": 1.0, + "content": "Each input vector has", + "type": "text" + }, + { + "bbox": [ + 195, + 95, + 209, + 106 + ], + "score": 0.91, + "content": "| O |", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "dimensions to represent the distribution over phonemes for each segment.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "score": 1.0, + "content": "The discriminator is also a CNN which outputs a probability indicating how likely the sample is to be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "score": 1.0, + "content": "from the data distribution.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 172, + 152 + ], + "lines": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "score": 1.0, + "content": "3.2 Objective", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 160, + 505, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "score": 1.0, + "content": "In our setup we use the original GAN objective with a gradient penalty [Goodfellow et al., 2014,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "score": 1.0, + "content": "Arjovsky et al., 2017], a segment smoothness penalty and a phoneme diversity penalty:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "interline_equation", + "bbox": [ + 136, + 187, + 474, + 207 + ], + "lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "score": 0.91, + "content": "\\operatorname* { m i n } _ { \\mathcal { G } } \\operatorname* { m a x } _ { \\mathcal { C } } \\quad \\mathbb { E } _ { \\mathcal { P } ^ { r } \\sim \\mathcal { P } ^ { r } } \\left[ \\log \\mathcal { C } ( \\boldsymbol { P } ^ { r } ) \\right] - \\underset { S \\sim \\mathcal { S } } { \\mathbb { E } } \\left[ \\log \\left( 1 - \\mathcal { C } ( \\mathcal { G } ( S ) ) \\right) \\right] - \\lambda \\mathcal { L } _ { g p } + \\gamma \\mathcal { L } _ { s p } + \\eta \\mathcal { L } _ { p d }", + "type": "interline_equation", + "image_path": "a15593f8fdca0cdc2e5a2bf415cdd52fd286427c617f23aeff5dbf4f23d44a11.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 134, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 212, + 175, + 222 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 210, + 302, + 225 + ], + "score": 1.0, + "content": "is phonemized unlabeled text,", + "type": "text" + }, + { + "bbox": [ + 303, + 212, + 325, + 223 + ], + "score": 0.94, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 210, + 505, + 225 + ], + "score": 1.0, + "content": "is the transcription output by the generator", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 245, + 235 + ], + "score": 1.0, + "content": "of input segment representations", + "type": "text" + }, + { + "bbox": [ + 246, + 223, + 254, + 233 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "for some unlabeled speech audio. The first term trains the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "discriminator to assign high probability to real transcriptions, the second term encourages the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 360, + 258 + ], + "score": 1.0, + "content": "discriminator to assign low probability to generator outputs,", + "type": "text" + }, + { + "bbox": [ + 360, + 245, + 377, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { g p }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "is a gradient penalty,", + "type": "text" + }, + { + "bbox": [ + 469, + 245, + 486, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { s p }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 205, + 269 + ], + "score": 1.0, + "content": "smoothness penalty and", + "type": "text" + }, + { + "bbox": [ + 206, + 256, + 223, + 267 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { p d }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is a phoneme diversity loss which we detail next. During training we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "alternate updates for the discriminator and the generator. We also alternate batches of predicted", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "score": 1.0, + "content": "transcriptions from the generator and phonemized unlabeled text.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12 + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "score": 1.0, + "content": "Gradient penalty. To stabilize training, we penalize the gradient norm of the discriminator with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 473, + 324 + ], + "score": 1.0, + "content": "respect to the input [Gulrajani et al., 2017]. The penalty is computed for random samples", + "type": "text" + }, + { + "bbox": [ + 473, + 311, + 504, + 322 + ], + "score": 0.93, + "content": "\\tilde { P } \\sim \\tilde { \\mathcal { P } }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "score": 1.0, + "content": "which are a linear combination of the activations of pairs of real and fake samples.2", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 339, + 373, + 367 + ], + "lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { g p } = \\underset { \\tilde { P } \\sim \\tilde { \\mathcal { P } } } { \\mathbb { E } } \\left[ \\left( \\| \\nabla \\mathcal { C } ( \\tilde { P } ) \\| - 1 \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "0a8f6af23cdb1998f6ef4e92c36819c82eb870dd5b721a5d19fc4f0384f93881.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 376, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 258, + 389 + ], + "score": 1.0, + "content": "Segment smoothness penalty. The", + "type": "text" + }, + { + "bbox": [ + 259, + 378, + 266, + 387 + ], + "score": 0.65, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "-means segmentation of the speech audio is more granular", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "score": 1.0, + "content": "than a typical phonemized transcription and neighboring representations are highly correlated. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "therefore found it useful to add a penalty which encourages the generator to produce similar outputs", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 409, + 268, + 423 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 222, + 423 + ], + "score": 1.0, + "content": "for adjacent segments where", + "type": "text" + }, + { + "bbox": [ + 223, + 409, + 263, + 422 + ], + "score": 0.93, + "content": "p _ { t } \\in \\mathbb { R } ^ { | O | }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 410, + 268, + 423 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5 + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 426, + 376, + 455 + ], + "lines": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "spans": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { s p } = \\sum _ { ( p _ { t } , p _ { t + 1 } ) \\in \\mathcal { G } ( S ) } \\| p _ { t } - p _ { t + 1 } \\| ^ { 2 }", + "type": "interline_equation", + "image_path": "dc18aa35eab1ad24f09a6f15551a94289a62fd2eada4a74b608bff6bec0d7c91.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 426, + 376, + 440.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 235, + 440.5, + 376, + 455.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 464, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "Phoneme diversity loss. We also found it helpful to penalize low usage of the phoneme vocabulary", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "by the generator on the batch level. In particular, we maximize the entropy of the averaged softmax", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 503, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 154, + 500 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 154, + 487, + 197, + 499 + ], + "score": 0.93, + "content": "H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 486, + 436, + 500 + ], + "score": 1.0, + "content": "of the generator over the phoneme vocabulary across a batch", + "type": "text" + }, + { + "bbox": [ + 437, + 487, + 446, + 497 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 486, + 503, + 500 + ], + "score": 1.0, + "content": "of utterances:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27 + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 502, + 365, + 533 + ], + "lines": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "spans": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { p d } = \\frac { 1 } { | B | } \\sum _ { S \\in B } - H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "interline_equation", + "image_path": "feda12d555ac0640b6d3cc1aeab0e1b03d7947fe4dd41ee8f6f0dbf28f5d27c6.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 502, + 365, + 517.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 246, + 517.5, + 365, + 533.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 544, + 294, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "score": 1.0, + "content": "3.3 Unsupervised Cross-Validation Metric", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 506, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "Our goal is to build speech recognition models without any supervision. To this end, we developed a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "score": 1.0, + "content": "cross-validation metric which does not require labeled data. We use the metric for early stopping,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 585, + 369, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 330, + 600 + ], + "score": 1.0, + "content": "selecting a random seed, and hyper-parameter selection", + "type": "text" + }, + { + "bbox": [ + 330, + 587, + 364, + 598 + ], + "score": 0.89, + "content": "( \\lambda , \\gamma , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 585, + 369, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "We consider two quantities in our metric: LM negative log-likelihood (NLL) and vocabulary usage.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "LM-NLL serves as an indicator of fluency for a given transcription and it is measured with a language", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 135, + 637 + ], + "score": 1.0, + "content": "model", + "type": "text" + }, + { + "bbox": [ + 135, + 626, + 156, + 636 + ], + "score": 0.87, + "content": "p _ { L M }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 624, + 295, + 637 + ], + "score": 1.0, + "content": "trained on phonemized text data", + "type": "text" + }, + { + "bbox": [ + 295, + 625, + 324, + 636 + ], + "score": 0.81, + "content": "( \\ S ~ 2 . 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 624, + 505, + 637 + ], + "score": 1.0, + "content": ". Vocabulary usage is the proportion of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "score": 1.0, + "content": "phoneme vocabulary being output by the model via Viterbi decoding. Measuring vocabulary usage", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "score": 1.0, + "content": "identifies degenerate models which output fluent but trivial transcriptions.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 662, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 101, + 662, + 511, + 694 + ], + "spans": [ + { + "bbox": [ + 101, + 662, + 130, + 694 + ], + "score": 1.0, + "content": "We deaudio", + "type": "text" + }, + { + "bbox": [ + 168, + 662, + 179, + 694 + ], + "score": 1.0, + "content": "rbias", + "type": "text" + }, + { + "bbox": [ + 236, + 662, + 410, + 694 + ], + "score": 1.0, + "content": "criptions for a given generator configuration . LM-NLL is measured in the standard way o", + "type": "text" + }, + { + "bbox": [ + 410, + 663, + 418, + 673 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 662, + 511, + 694 + ], + "score": 1.0, + "content": "and unlabeled speechr the phonemized tran-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 131, + 673, + 236, + 689 + ], + "spans": [ + { + "bbox": [ + 131, + 673, + 168, + 689 + ], + "score": 0.93, + "content": "\\{ X _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 673, + 236, + 689 + ], + "score": 0.93, + "content": "\\mathbf { \\bar { \\mathcal { P } } } = \\{ P _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 686, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 151, + 703 + ], + "score": 1.0, + "content": "scriptions:", + "type": "text" + }, + { + "bbox": [ + 151, + 688, + 314, + 704 + ], + "score": 0.89, + "content": "\\begin{array} { r } { N L L _ { L M } ( \\mathcal { P } ) = \\frac { 1 } { N _ { s } } \\sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 686, + 342, + 706 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 342, + 687, + 505, + 703 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N L L _ { L M } ( P ) = - \\frac { 1 } { M } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 41 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 711, + 324, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 325, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 325, + 725 + ], + "score": 1.0, + "content": "2We simply shorten longer sequence if the lengths differ.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 753 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 72, + 506, + 128 + ], + "lines": [ + { + "bbox": [ + 106, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 106, + 72, + 304, + 86 + ], + "score": 1.0, + "content": "The discriminator takes as input either a sequence", + "type": "text" + }, + { + "bbox": [ + 304, + 73, + 343, + 83 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "of one-hot vectors denoting phonemized", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 83, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 83, + 244, + 96 + ], + "score": 1.0, + "content": "text from the real data distribution", + "type": "text" + }, + { + "bbox": [ + 244, + 84, + 257, + 94 + ], + "score": 0.85, + "content": "{ \\mathcal { P } } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 83, + 481, + 96 + ], + "score": 1.0, + "content": "or a sequence of output distributions from the generator", + "type": "text" + }, + { + "bbox": [ + 481, + 83, + 502, + 96 + ], + "score": 0.92, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 83, + 506, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 93, + 507, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 195, + 108 + ], + "score": 1.0, + "content": "Each input vector has", + "type": "text" + }, + { + "bbox": [ + 195, + 95, + 209, + 106 + ], + "score": 0.91, + "content": "| O |", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 93, + 507, + 108 + ], + "score": 1.0, + "content": "dimensions to represent the distribution over phonemes for each segment.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 119 + ], + "score": 1.0, + "content": "The discriminator is also a CNN which outputs a probability indicating how likely the sample is to be", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 213, + 128 + ], + "score": 1.0, + "content": "from the data distribution.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 72, + 507, + 128 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 140, + 172, + 152 + ], + "lines": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "spans": [ + { + "bbox": [ + 104, + 137, + 173, + 155 + ], + "score": 1.0, + "content": "3.2 Objective", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 160, + 505, + 183 + ], + "lines": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 507, + 173 + ], + "score": 1.0, + "content": "In our setup we use the original GAN objective with a gradient penalty [Goodfellow et al., 2014,", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 171, + 455, + 185 + ], + "score": 1.0, + "content": "Arjovsky et al., 2017], a segment smoothness penalty and a phoneme diversity penalty:", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 159, + 507, + 185 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 136, + 187, + 474, + 207 + ], + "lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "score": 0.91, + "content": "\\operatorname* { m i n } _ { \\mathcal { G } } \\operatorname* { m a x } _ { \\mathcal { C } } \\quad \\mathbb { E } _ { \\mathcal { P } ^ { r } \\sim \\mathcal { P } ^ { r } } \\left[ \\log \\mathcal { C } ( \\boldsymbol { P } ^ { r } ) \\right] - \\underset { S \\sim \\mathcal { S } } { \\mathbb { E } } \\left[ \\log \\left( 1 - \\mathcal { C } ( \\mathcal { G } ( S ) ) \\right) \\right] - \\lambda \\mathcal { L } _ { g p } + \\gamma \\mathcal { L } _ { s p } + \\eta \\mathcal { L } _ { p d }", + "type": "interline_equation", + "image_path": "a15593f8fdca0cdc2e5a2bf415cdd52fd286427c617f23aeff5dbf4f23d44a11.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 136, + 187, + 474, + 207 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 211, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 210, + 505, + 225 + ], + "spans": [ + { + "bbox": [ + 105, + 210, + 134, + 225 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 212, + 175, + 222 + ], + "score": 0.91, + "content": "P ^ { r } \\sim \\mathcal { P } ^ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 175, + 210, + 302, + 225 + ], + "score": 1.0, + "content": "is phonemized unlabeled text,", + "type": "text" + }, + { + "bbox": [ + 303, + 212, + 325, + 223 + ], + "score": 0.94, + "content": "\\mathcal { G } ( S )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 210, + 505, + 225 + ], + "score": 1.0, + "content": "is the transcription output by the generator", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 222, + 505, + 235 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 245, + 235 + ], + "score": 1.0, + "content": "of input segment representations", + "type": "text" + }, + { + "bbox": [ + 246, + 223, + 254, + 233 + ], + "score": 0.81, + "content": "S", + "type": "inline_equation" + }, + { + "bbox": [ + 254, + 222, + 505, + 235 + ], + "score": 1.0, + "content": "for some unlabeled speech audio. The first term trains the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 234, + 505, + 246 + ], + "score": 1.0, + "content": "discriminator to assign high probability to real transcriptions, the second term encourages the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 506, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 360, + 258 + ], + "score": 1.0, + "content": "discriminator to assign low probability to generator outputs,", + "type": "text" + }, + { + "bbox": [ + 360, + 245, + 377, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { g p }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 244, + 469, + 258 + ], + "score": 1.0, + "content": "is a gradient penalty,", + "type": "text" + }, + { + "bbox": [ + 469, + 245, + 486, + 257 + ], + "score": 0.9, + "content": "\\mathcal { L } _ { s p }", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 244, + 506, + 258 + ], + "score": 1.0, + "content": "is a", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 205, + 269 + ], + "score": 1.0, + "content": "smoothness penalty and", + "type": "text" + }, + { + "bbox": [ + 206, + 256, + 223, + 267 + ], + "score": 0.91, + "content": "\\mathcal { L } _ { p d }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "is a phoneme diversity loss which we detail next. During training we", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 266, + 506, + 279 + ], + "score": 1.0, + "content": "alternate updates for the discriminator and the generator. We also alternate batches of predicted", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 369, + 289 + ], + "score": 1.0, + "content": "transcriptions from the generator and phonemized unlabeled text.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 210, + 506, + 289 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 299, + 505, + 335 + ], + "lines": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 506, + 313 + ], + "score": 1.0, + "content": "Gradient penalty. To stabilize training, we penalize the gradient norm of the discriminator with", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 311, + 504, + 324 + ], + "spans": [ + { + "bbox": [ + 105, + 312, + 473, + 324 + ], + "score": 1.0, + "content": "respect to the input [Gulrajani et al., 2017]. The penalty is computed for random samples", + "type": "text" + }, + { + "bbox": [ + 473, + 311, + 504, + 322 + ], + "score": 0.93, + "content": "\\tilde { P } \\sim \\tilde { \\mathcal { P } }", + "type": "inline_equation" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 322, + 441, + 336 + ], + "score": 1.0, + "content": "which are a linear combination of the activations of pairs of real and fake samples.2", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 298, + 506, + 336 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 237, + 339, + 373, + 367 + ], + "lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { g p } = \\underset { \\tilde { P } \\sim \\tilde { \\mathcal { P } } } { \\mathbb { E } } \\left[ \\left( \\| \\nabla \\mathcal { C } ( \\tilde { P } ) \\| - 1 \\right) ^ { 2 } \\right]", + "type": "interline_equation", + "image_path": "0a8f6af23cdb1998f6ef4e92c36819c82eb870dd5b721a5d19fc4f0384f93881.jpg" + } + ] + } + ], + "index": 19, + "virtual_lines": [ + { + "bbox": [ + 237, + 339, + 373, + 367 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 376, + 505, + 422 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 506, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 258, + 389 + ], + "score": 1.0, + "content": "Segment smoothness penalty. The", + "type": "text" + }, + { + "bbox": [ + 259, + 378, + 266, + 387 + ], + "score": 0.65, + "content": "\\mathbf { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 376, + 506, + 389 + ], + "score": 1.0, + "content": "-means segmentation of the speech audio is more granular", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 387, + 504, + 399 + ], + "score": 1.0, + "content": "than a typical phonemized transcription and neighboring representations are highly correlated. We", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 505, + 412 + ], + "score": 1.0, + "content": "therefore found it useful to add a penalty which encourages the generator to produce similar outputs", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 409, + 268, + 423 + ], + "spans": [ + { + "bbox": [ + 104, + 410, + 222, + 423 + ], + "score": 1.0, + "content": "for adjacent segments where", + "type": "text" + }, + { + "bbox": [ + 223, + 409, + 263, + 422 + ], + "score": 0.93, + "content": "p _ { t } \\in \\mathbb { R } ^ { | O | }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 410, + 268, + 423 + ], + "score": 1.0, + "content": ":", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 21.5, + "bbox_fs": [ + 104, + 376, + 506, + 423 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 235, + 426, + 376, + 455 + ], + "lines": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "spans": [ + { + "bbox": [ + 235, + 426, + 376, + 455 + ], + "score": 0.94, + "content": "\\mathcal { L } _ { s p } = \\sum _ { ( p _ { t } , p _ { t + 1 } ) \\in \\mathcal { G } ( S ) } \\| p _ { t } - p _ { t + 1 } \\| ^ { 2 }", + "type": "interline_equation", + "image_path": "dc18aa35eab1ad24f09a6f15551a94289a62fd2eada4a74b608bff6bec0d7c91.jpg" + } + ] + } + ], + "index": 24.5, + "virtual_lines": [ + { + "bbox": [ + 235, + 426, + 376, + 440.5 + ], + "spans": [], + "index": 24 + }, + { + "bbox": [ + 235, + 440.5, + 376, + 455.0 + ], + "spans": [], + "index": 25 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 464, + 505, + 498 + ], + "lines": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "Phoneme diversity loss. We also found it helpful to penalize low usage of the phoneme vocabulary", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "by the generator on the batch level. In particular, we maximize the entropy of the averaged softmax", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 486, + 503, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 154, + 500 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 154, + 487, + 197, + 499 + ], + "score": 0.93, + "content": "H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 486, + 436, + 500 + ], + "score": 1.0, + "content": "of the generator over the phoneme vocabulary across a batch", + "type": "text" + }, + { + "bbox": [ + 437, + 487, + 446, + 497 + ], + "score": 0.83, + "content": "B", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 486, + 503, + 500 + ], + "score": 1.0, + "content": "of utterances:", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 464, + 505, + 500 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 246, + 502, + 365, + 533 + ], + "lines": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "spans": [ + { + "bbox": [ + 246, + 502, + 365, + 533 + ], + "score": 0.95, + "content": "\\mathcal { L } _ { p d } = \\frac { 1 } { | B | } \\sum _ { S \\in B } - H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )", + "type": "interline_equation", + "image_path": "feda12d555ac0640b6d3cc1aeab0e1b03d7947fe4dd41ee8f6f0dbf28f5d27c6.jpg" + } + ] + } + ], + "index": 29.5, + "virtual_lines": [ + { + "bbox": [ + 246, + 502, + 365, + 517.5 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 246, + 517.5, + 365, + 533.0 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "title", + "bbox": [ + 107, + 544, + 294, + 556 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 295, + 557 + ], + "score": 1.0, + "content": "3.3 Unsupervised Cross-Validation Metric", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 564, + 506, + 598 + ], + "lines": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 106, + 564, + 506, + 576 + ], + "score": 1.0, + "content": "Our goal is to build speech recognition models without any supervision. To this end, we developed a", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "spans": [ + { + "bbox": [ + 105, + 573, + 507, + 590 + ], + "score": 1.0, + "content": "cross-validation metric which does not require labeled data. We use the metric for early stopping,", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 585, + 369, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 330, + 600 + ], + "score": 1.0, + "content": "selecting a random seed, and hyper-parameter selection", + "type": "text" + }, + { + "bbox": [ + 330, + 587, + 364, + 598 + ], + "score": 0.89, + "content": "( \\lambda , \\gamma , \\eta )", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 585, + 369, + 600 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 564, + 507, + 600 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 602, + 505, + 658 + ], + "lines": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 506, + 615 + ], + "score": 1.0, + "content": "We consider two quantities in our metric: LM negative log-likelihood (NLL) and vocabulary usage.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "spans": [ + { + "bbox": [ + 105, + 613, + 505, + 627 + ], + "score": 1.0, + "content": "LM-NLL serves as an indicator of fluency for a given transcription and it is measured with a language", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 624, + 505, + 637 + ], + "spans": [ + { + "bbox": [ + 105, + 624, + 135, + 637 + ], + "score": 1.0, + "content": "model", + "type": "text" + }, + { + "bbox": [ + 135, + 626, + 156, + 636 + ], + "score": 0.87, + "content": "p _ { L M }", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 624, + 295, + 637 + ], + "score": 1.0, + "content": "trained on phonemized text data", + "type": "text" + }, + { + "bbox": [ + 295, + 625, + 324, + 636 + ], + "score": 0.81, + "content": "( \\ S ~ 2 . 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 624, + 505, + 637 + ], + "score": 1.0, + "content": ". Vocabulary usage is the proportion of the", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 650 + ], + "score": 1.0, + "content": "phoneme vocabulary being output by the model via Viterbi decoding. Measuring vocabulary usage", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 646, + 402, + 660 + ], + "score": 1.0, + "content": "identifies degenerate models which output fluent but trivial transcriptions.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 601, + 506, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 662, + 506, + 704 + ], + "lines": [ + { + "bbox": [ + 101, + 662, + 511, + 694 + ], + "spans": [ + { + "bbox": [ + 101, + 662, + 130, + 694 + ], + "score": 1.0, + "content": "We deaudio", + "type": "text" + }, + { + "bbox": [ + 168, + 662, + 179, + 694 + ], + "score": 1.0, + "content": "rbias", + "type": "text" + }, + { + "bbox": [ + 236, + 662, + 410, + 694 + ], + "score": 1.0, + "content": "criptions for a given generator configuration . LM-NLL is measured in the standard way o", + "type": "text" + }, + { + "bbox": [ + 410, + 663, + 418, + 673 + ], + "score": 0.84, + "content": "\\mathcal { G }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 662, + 511, + 694 + ], + "score": 1.0, + "content": "and unlabeled speechr the phonemized tran-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 131, + 673, + 236, + 689 + ], + "spans": [ + { + "bbox": [ + 131, + 673, + 168, + 689 + ], + "score": 0.93, + "content": "\\{ X _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + }, + { + "bbox": [ + 179, + 673, + 236, + 689 + ], + "score": 0.93, + "content": "\\mathbf { \\bar { \\mathcal { P } } } = \\{ P _ { j } \\} _ { j = 1 } ^ { N _ { s } }", + "type": "inline_equation" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 686, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 151, + 703 + ], + "score": 1.0, + "content": "scriptions:", + "type": "text" + }, + { + "bbox": [ + 151, + 688, + 314, + 704 + ], + "score": 0.89, + "content": "\\begin{array} { r } { N L L _ { L M } ( \\mathcal { P } ) = \\frac { 1 } { N _ { s } } \\sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 314, + 686, + 342, + 706 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 342, + 687, + 505, + 703 + ], + "score": 0.92, + "content": "\\begin{array} { r } { N L L _ { L M } ( P ) = - \\frac { 1 } { M } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 70, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 132, + 86 + ], + "score": 1.0, + "content": "using", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 132, + 73, + 169, + 84 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 169, + 70, + 245, + 86 + ], + "score": 1.0, + "content": "as shorthand for", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 245, + 73, + 334, + 85 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 335, + 70, + 465, + 86 + ], + "score": 1.0, + "content": ".3 On the other hand, we use", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 466, + 73, + 505, + 85 + ], + "score": 0.9, + "content": "U ( { \\mathcal { P } } ) =", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 362, + 99 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 210, + 99 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\frac { 1 } { | O | } \\bar { \\sum _ { o \\in O } } [ o \\in \\mathcal { P } ] \\in [ 0 , 1 ] } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 210, + 83, + 348, + 96 + ], + "score": 1.0, + "content": "to denote the vocabulary usage of", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 348, + 84, + 356, + 93 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 357, + 83, + 362, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text", + "cross_page": true + } + ], + "index": 1 + } + ], + "index": 41, + "bbox_fs": [ + 101, + 662, + 511, + 706 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 72, + 504, + 96 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 132, + 86 + ], + "score": 1.0, + "content": "using", + "type": "text" + }, + { + "bbox": [ + 132, + 73, + 169, + 84 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 70, + 245, + 86 + ], + "score": 1.0, + "content": "as shorthand for", + "type": "text" + }, + { + "bbox": [ + 245, + 73, + 334, + 85 + ], + "score": 0.9, + "content": "p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 70, + 465, + 86 + ], + "score": 1.0, + "content": ".3 On the other hand, we use", + "type": "text" + }, + { + "bbox": [ + 466, + 73, + 505, + 85 + ], + "score": 0.9, + "content": "U ( { \\mathcal { P } } ) =", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 83, + 362, + 99 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 210, + 99 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\frac { 1 } { | O | } \\bar { \\sum _ { o \\in O } } [ o \\in \\mathcal { P } ] \\in [ 0 , 1 ] } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 83, + 348, + 96 + ], + "score": 1.0, + "content": "to denote the vocabulary usage of", + "type": "text" + }, + { + "bbox": [ + 348, + 84, + 356, + 93 + ], + "score": 0.84, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 83, + 362, + 96 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "score": 1.0, + "content": "In a first step, we generate phoneme transcriptions for different training checkpoints or hyper-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "score": 1.0, + "content": "parameter settings and denote the transcriptions of the configuration with the lowest vocabulary-usage", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 174, + 138 + ], + "score": 1.0, + "content": "adjusted NLL as", + "type": "text" + }, + { + "bbox": [ + 175, + 124, + 340, + 137 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\hat { \\mathcal { P } } = \\arg \\operatorname* { m i n } _ { \\mathcal { P } } N L L _ { L M } ( \\mathcal { P } ) - \\log U ( \\mathcal { P } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 124, + 505, + 138 + ], + "score": 1.0, + "content": ".4 Next, we discard model configurations", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 336, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 269, + 151 + ], + "score": 1.0, + "content": "which do not satisfy the following using", + "type": "text" + }, + { + "bbox": [ + 269, + 137, + 278, + 149 + ], + "score": 0.85, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 138, + 336, + 151 + ], + "score": 1.0, + "content": "as the anchor:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5 + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 156, + 420, + 190 + ], + "lines": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "spans": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\mathcal { P } ) < N L L _ { L M } ( \\hat { \\mathcal { P } } ) + \\log \\left( \\frac { U ( \\mathcal { P } ) } { U ( \\hat { \\mathcal { P } } ) } \\right) + \\log 1 . 2", + "type": "interline_equation", + "image_path": "b07976d07eb1264788bb86ec8c503a84fc3c183351681c48efba85a9d06b98dd.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 156, + 420, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 191, + 173.0, + 420, + 190.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "score": 1.0, + "content": "The second term on the right hand side introduces a margin over the NLL of the anchor transcription", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 208, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 208, + 160, + 222 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 208, + 294, + 223 + ], + "score": 1.0, + "content": "based on the vocabulary usage of", + "type": "text" + }, + { + "bbox": [ + 294, + 210, + 303, + 220 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 208, + 320, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 208, + 329, + 220 + ], + "score": 0.84, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 208, + 343, + 223 + ], + "score": 1.0, + "content": ": If", + "type": "text" + }, + { + "bbox": [ + 344, + 208, + 368, + 222 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 208, + 478, + 223 + ], + "score": 1.0, + "content": "is much lower compared to", + "type": "text" + }, + { + "bbox": [ + 479, + 212, + 502, + 222 + ], + "score": 0.93, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 208, + 506, + 223 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "score": 1.0, + "content": "then we allow model configurations which produce transcriptions with higher NLL compared to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 231, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 115, + 244 + ], + "score": 0.82, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 232, + 171, + 246 + ], + "score": 1.0, + "content": ". However, if", + "type": "text" + }, + { + "bbox": [ + 172, + 232, + 196, + 245 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 232, + 277, + 246 + ], + "score": 1.0, + "content": "is a lot higher than", + "type": "text" + }, + { + "bbox": [ + 277, + 233, + 301, + 245 + ], + "score": 0.92, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 232, + 505, + 246 + ], + "score": 1.0, + "content": ", then the model configuration will not satisfy the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 172, + 258 + ], + "score": 1.0, + "content": "constraint. The", + "type": "text" + }, + { + "bbox": [ + 172, + 245, + 201, + 256 + ], + "score": 0.62, + "content": "\\log 1 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "factor serves as another margin allowing checkpoints with slightly worse", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "score": 1.0, + "content": "vocabulary-usage adjusted NLL to be included.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 107, + 271, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 492, + 284 + ], + "score": 1.0, + "content": "In a final step, we take into account the length of the transcriptions: out of the configurations", + "type": "text" + }, + { + "bbox": [ + 492, + 273, + 504, + 282 + ], + "score": 0.83, + "content": "{ \\mathcal { P } } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "which satisfy the above constraint, we select the one which has the highest sum of log probability", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "score": 1.0, + "content": "without normalizing the length:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15 + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 312, + 443, + 348 + ], + "lines": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "spans": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "score": 0.94, + "content": "\\mathcal { P } ^ { * } = \\arg \\operatorname* { m a x } _ { \\mathcal { P } ^ { \\prime } } \\sum _ { j = 1 } ^ { N _ { s } } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \\dots , p _ { M } ^ { j } ]", + "type": "interline_equation", + "image_path": "2930d2c4de9bfc9242467fc6ca39076d0e51025e6eb1f2125a1e21c64638989c.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 169, + 312, + 443, + 324.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 169, + 324.0, + 443, + 336.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 169, + 336.0, + 443, + 348.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 506, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "This selects model configurations which produce phoneme sequences that score high under the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "language model but are not too long. Appendix D compares accuracy when developing with this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "score": 1.0, + "content": "metric compared to a labeled development set.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 107, + 406, + 163, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "score": 1.0, + "content": "4 Results", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 431, + 395, + 444 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "score": 1.0, + "content": "4.1 Comparison to Supervised Speech Recognition on Librispeech", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "We first test our approach on Librispeech to get a sense of how unsupervised speech recognition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "compares to the best supervised systems trained on a large amount of labeled data. Librispeech", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "is a standard benchmark in the speech recognition community which provides about 960 hours of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "transcribed read audiobooks. We use the language modeling data of Librispeech as unlabeled text", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "data for unsupervised training. In Appendix G we show that far less unlabeled text and speech audio", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "are sufficient to reach a similar level of performance. We experiment with the frozen representations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 291, + 531 + ], + "score": 1.0, + "content": "of a wav2vec 2.0 LARGE model trained on the", + "type": "text" + }, + { + "bbox": [ + 291, + 519, + 315, + 529 + ], + "score": 0.81, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "hours of Libri-Light (LL-60k) which we denote", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "as wav2vec-U LARGE. We also consider self-training over three iterations by first training an HMM", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "on the labels generated by the GANm then fine-tuning the original wav2vec 2.0 model on the labels", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "score": 1.0, + "content": "of the HMM for Librispeech followed by then fine-tuning on Libri-Light; Appendix F investigates", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "score": 1.0, + "content": "alternatives.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 506, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 295, + 591 + ], + "score": 1.0, + "content": "wav2vec-U LARGE with self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 296, + 579, + 327, + 590 + ], + "score": 0.7, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 578, + 505, + 591 + ], + "score": 1.0, + "content": ") and a Transformer language model achieves", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "WER 5.9 on test-other, the noisy test set. This shows that unsupervised speech recognition can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "perform remarkably well compared to the best supervised systems of the recent past on this much", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "studied benchmark. Also, self-training is effective even when the teacher model is unsupervised as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "per the improvement over GAN training (wav2vec-U). Interestingly, self-training on just Librispeech,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "or 960 hours of unlabeled speech audio, achieves already very good performance of WER 6.4 on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "score": 1.0, + "content": "dev-other compared to self-training on all of Libri-Light (53.2k hours) which compares at 6.0 WER.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "We note that the number of parameters trained during adversarial training is very small: the generator", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "contains only about 90k parameters for a single temporal convolution mapping to the phoneme set", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "score": 1.0, + "content": "from frozen wav2vec 2.0 representations.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 700, + 469, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 697, + 448, + 714 + ], + "spans": [ + { + "bbox": [ + 118, + 697, + 224, + 714 + ], + "score": 1.0, + "content": "3We remove SIL labels from", + "type": "text" + }, + { + "bbox": [ + 225, + 703, + 232, + 710 + ], + "score": 0.8, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 697, + 424, + 714 + ], + "score": 1.0, + "content": "when computing the NLL because SIL is not used in", + "type": "text" + }, + { + "bbox": [ + 425, + 702, + 443, + 711 + ], + "score": 0.79, + "content": "p _ { L M }", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 697, + 448, + 714 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 709, + 468, + 725 + ], + "spans": [ + { + "bbox": [ + 119, + 709, + 468, + 725 + ], + "score": 1.0, + "content": "4In practice, we used language model perplexity which is equivalent to NLL after taking the log.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 742, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 310, + 752 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 72, + 504, + 96 + ], + "lines": [], + "index": 0.5, + "bbox_fs": [ + 105, + 70, + 505, + 99 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 114 + ], + "score": 1.0, + "content": "In a first step, we generate phoneme transcriptions for different training checkpoints or hyper-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 105, + 112, + 505, + 126 + ], + "score": 1.0, + "content": "parameter settings and denote the transcriptions of the configuration with the lowest vocabulary-usage", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 124, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 105, + 124, + 174, + 138 + ], + "score": 1.0, + "content": "adjusted NLL as", + "type": "text" + }, + { + "bbox": [ + 175, + 124, + 340, + 137 + ], + "score": 0.84, + "content": "\\begin{array} { r } { \\hat { \\mathcal { P } } = \\arg \\operatorname* { m i n } _ { \\mathcal { P } } N L L _ { L M } ( \\mathcal { P } ) - \\log U ( \\mathcal { P } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 340, + 124, + 505, + 138 + ], + "score": 1.0, + "content": ".4 Next, we discard model configurations", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 336, + 151 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 269, + 151 + ], + "score": 1.0, + "content": "which do not satisfy the following using", + "type": "text" + }, + { + "bbox": [ + 269, + 137, + 278, + 149 + ], + "score": 0.85, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 138, + 336, + 151 + ], + "score": 1.0, + "content": "as the anchor:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 3.5, + "bbox_fs": [ + 105, + 102, + 506, + 151 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 191, + 156, + 420, + 190 + ], + "lines": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "spans": [ + { + "bbox": [ + 191, + 156, + 420, + 190 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\mathcal { P } ) < N L L _ { L M } ( \\hat { \\mathcal { P } } ) + \\log \\left( \\frac { U ( \\mathcal { P } ) } { U ( \\hat { \\mathcal { P } } ) } \\right) + \\log 1 . 2", + "type": "interline_equation", + "image_path": "b07976d07eb1264788bb86ec8c503a84fc3c183351681c48efba85a9d06b98dd.jpg" + } + ] + } + ], + "index": 6.5, + "virtual_lines": [ + { + "bbox": [ + 191, + 156, + 420, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 191, + 173.0, + 420, + 190.0 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 267 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 506, + 210 + ], + "score": 1.0, + "content": "The second term on the right hand side introduces a margin over the NLL of the anchor transcription", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 107, + 208, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 208, + 160, + 222 + ], + "score": 0.93, + "content": "N L L _ { L M } ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 160, + 208, + 294, + 223 + ], + "score": 1.0, + "content": "based on the vocabulary usage of", + "type": "text" + }, + { + "bbox": [ + 294, + 210, + 303, + 220 + ], + "score": 0.82, + "content": "\\mathcal { P }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 208, + 320, + 223 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 321, + 208, + 329, + 220 + ], + "score": 0.84, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 208, + 343, + 223 + ], + "score": 1.0, + "content": ": If", + "type": "text" + }, + { + "bbox": [ + 344, + 208, + 368, + 222 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 368, + 208, + 478, + 223 + ], + "score": 1.0, + "content": "is much lower compared to", + "type": "text" + }, + { + "bbox": [ + 479, + 212, + 502, + 222 + ], + "score": 0.93, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 208, + 506, + 223 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 220, + 506, + 234 + ], + "score": 1.0, + "content": "then we allow model configurations which produce transcriptions with higher NLL compared to", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 231, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 115, + 244 + ], + "score": 0.82, + "content": "\\hat { \\mathcal { P } }", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 232, + 171, + 246 + ], + "score": 1.0, + "content": ". However, if", + "type": "text" + }, + { + "bbox": [ + 172, + 232, + 196, + 245 + ], + "score": 0.93, + "content": "U ( \\hat { \\mathcal { P } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 196, + 232, + 277, + 246 + ], + "score": 1.0, + "content": "is a lot higher than", + "type": "text" + }, + { + "bbox": [ + 277, + 233, + 301, + 245 + ], + "score": 0.92, + "content": "U ( \\mathcal { P } )", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 232, + 505, + 246 + ], + "score": 1.0, + "content": ", then the model configuration will not satisfy the", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 244, + 505, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 172, + 258 + ], + "score": 1.0, + "content": "constraint. The", + "type": "text" + }, + { + "bbox": [ + 172, + 245, + 201, + 256 + ], + "score": 0.62, + "content": "\\log 1 . 2", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 244, + 505, + 258 + ], + "score": 1.0, + "content": "factor serves as another margin allowing checkpoints with slightly worse", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 297, + 268 + ], + "score": 1.0, + "content": "vocabulary-usage adjusted NLL to be included.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 196, + 506, + 268 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 271, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 105, + 270, + 504, + 284 + ], + "spans": [ + { + "bbox": [ + 105, + 270, + 492, + 284 + ], + "score": 1.0, + "content": "In a final step, we take into account the length of the transcriptions: out of the configurations", + "type": "text" + }, + { + "bbox": [ + 492, + 273, + 504, + 282 + ], + "score": 0.83, + "content": "{ \\mathcal { P } } ^ { \\prime }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "which satisfy the above constraint, we select the one which has the highest sum of log probability", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 234, + 306 + ], + "score": 1.0, + "content": "without normalizing the length:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 270, + 505, + 306 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 169, + 312, + 443, + 348 + ], + "lines": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "spans": [ + { + "bbox": [ + 169, + 312, + 443, + 348 + ], + "score": 0.94, + "content": "\\mathcal { P } ^ { * } = \\arg \\operatorname* { m a x } _ { \\mathcal { P } ^ { \\prime } } \\sum _ { j = 1 } ^ { N _ { s } } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \\dots , p _ { M } ^ { j } ]", + "type": "interline_equation", + "image_path": "2930d2c4de9bfc9242467fc6ca39076d0e51025e6eb1f2125a1e21c64638989c.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 169, + 312, + 443, + 324.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 169, + 324.0, + 443, + 336.0 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 169, + 336.0, + 443, + 348.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 354, + 506, + 388 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 505, + 367 + ], + "score": 1.0, + "content": "This selects model configurations which produce phoneme sequences that score high under the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "language model but are not too long. Appendix D compares accuracy when developing with this", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 295, + 389 + ], + "score": 1.0, + "content": "metric compared to a labeled development set.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 354, + 505, + 389 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 406, + 163, + 419 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 165, + 421 + ], + "score": 1.0, + "content": "4 Results", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "title", + "bbox": [ + 107, + 431, + 395, + 444 + ], + "lines": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "spans": [ + { + "bbox": [ + 104, + 430, + 397, + 447 + ], + "score": 1.0, + "content": "4.1 Comparison to Supervised Speech Recognition on Librispeech", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 452, + 505, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "spans": [ + { + "bbox": [ + 106, + 453, + 504, + 466 + ], + "score": 1.0, + "content": "We first test our approach on Librispeech to get a sense of how unsupervised speech recognition", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 476 + ], + "score": 1.0, + "content": "compares to the best supervised systems trained on a large amount of labeled data. Librispeech", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "is a standard benchmark in the speech recognition community which provides about 960 hours of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 506, + 498 + ], + "score": 1.0, + "content": "transcribed read audiobooks. We use the language modeling data of Librispeech as unlabeled text", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 509 + ], + "score": 1.0, + "content": "data for unsupervised training. In Appendix G we show that far less unlabeled text and speech audio", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "are sufficient to reach a similar level of performance. We experiment with the frozen representations", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 518, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 291, + 531 + ], + "score": 1.0, + "content": "of a wav2vec 2.0 LARGE model trained on the", + "type": "text" + }, + { + "bbox": [ + 291, + 519, + 315, + 529 + ], + "score": 0.81, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 518, + 505, + 531 + ], + "score": 1.0, + "content": "hours of Libri-Light (LL-60k) which we denote", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 541 + ], + "score": 1.0, + "content": "as wav2vec-U LARGE. We also consider self-training over three iterations by first training an HMM", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "spans": [ + { + "bbox": [ + 106, + 541, + 505, + 553 + ], + "score": 1.0, + "content": "on the labels generated by the GANm then fine-tuning the original wav2vec 2.0 model on the labels", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 565 + ], + "score": 1.0, + "content": "of the HMM for Librispeech followed by then fine-tuning on Libri-Light; Appendix F investigates", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 157, + 574 + ], + "score": 1.0, + "content": "alternatives.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 105, + 453, + 506, + 574 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 578, + 506, + 687 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 295, + 591 + ], + "score": 1.0, + "content": "wav2vec-U LARGE with self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 296, + 579, + 327, + 590 + ], + "score": 0.7, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 578, + 505, + 591 + ], + "score": 1.0, + "content": ") and a Transformer language model achieves", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "WER 5.9 on test-other, the noisy test set. This shows that unsupervised speech recognition can", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 104, + 600, + 506, + 613 + ], + "score": 1.0, + "content": "perform remarkably well compared to the best supervised systems of the recent past on this much", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "studied benchmark. Also, self-training is effective even when the teacher model is unsupervised as", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "per the improvement over GAN training (wav2vec-U). Interestingly, self-training on just Librispeech,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 632, + 506, + 646 + ], + "score": 1.0, + "content": "or 960 hours of unlabeled speech audio, achieves already very good performance of WER 6.4 on", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 106, + 645, + 506, + 656 + ], + "score": 1.0, + "content": "dev-other compared to self-training on all of Libri-Light (53.2k hours) which compares at 6.0 WER.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "We note that the number of parameters trained during adversarial training is very small: the generator", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "contains only about 90k parameters for a single temporal convolution mapping to the phoneme set", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 274, + 689 + ], + "score": 1.0, + "content": "from frozen wav2vec 2.0 representations.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 40.5, + "bbox_fs": [ + 104, + 578, + 506, + 689 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 505, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Table 1: WER on Librispeech dev/test sets when using 960 hours of unlabeled audio from Librispeech", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 506, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 155, + 93 + ], + "score": 1.0, + "content": "(LS-960) or", + "type": "text" + }, + { + "bbox": [ + 155, + 82, + 180, + 92 + ], + "score": 0.31, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 82, + 506, + 93 + ], + "score": 1.0, + "content": "hours from Libri-Light (LL-60k) using representations from wav2vec 2.0 LARGE.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "Librispeech provides clean dev/test sets which are less challenging than the other sets. We report", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 102, + 495, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 458, + 115 + ], + "score": 1.0, + "content": "results for GAN training only (wav2vec-U) and with subsequent self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 458, + 103, + 490, + 114 + ], + "score": 0.78, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 102, + 495, + 115 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "score": 0.983, + "html": "
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
", + "type": "table", + "image_path": "7d5f6a9816458b5d9d90dd620c97a44d70ecc200d9e2e077d3e3e3e887499080.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 122, + 504, + 191.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 191.66666666666669, + 504, + 261.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 261.33333333333337, + 504, + 331.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 106, + 345, + 306, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "score": 1.0, + "content": "4.2 Comparison to Prior Unsupervised Work", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Prior work on unsupervised speech recognition focused on the TIMIT benchmark. In order to perform", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "a direct comparison to these approaches, we report results on this benchmark as well. We consider", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "two setups to compare to previous work: in the matched setting, the unlabeled text data is simply the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "transcriptions of the unlabeled audio data but unpaired. In the unmatched setup, the unlabeled text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "score": 1.0, + "content": "data does not contain the transcriptions for the audio data which is a more realistic setting.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "We measure performance on the standard Kaldi dev and test sets (core-dev/core-test) as well as a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "score": 1.0, + "content": "slightly larger version of the test set (all-test) to be able to compare to Liu et al. [2018] and Chen et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 371, + 461 + ], + "score": 1.0, + "content": "[2019]. Further details of the two setups can be found in Appendix", + "type": "text" + }, + { + "bbox": [ + 372, + 448, + 393, + 459 + ], + "score": 0.66, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 447, + 506, + 461 + ], + "score": 1.0, + "content": ". We report performance for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "wav2vec-U with a 4-gram language model trained on the language modeling data of TIMIT and we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 469, + 291, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 254, + 482 + ], + "score": 1.0, + "content": "also consider self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 254, + 470, + 285, + 481 + ], + "score": 0.47, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 469, + 291, + 482 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "Table 2 shows that wav2vec-U outperforms prior unsupervised work in both the matched and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 379, + 509 + ], + "score": 1.0, + "content": "unmatched settings, reducing PER on all-test in the matched setup by", + "type": "text" + }, + { + "bbox": [ + 379, + 497, + 399, + 508 + ], + "score": 0.87, + "content": "57 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 498, + 505, + 509 + ], + "score": 1.0, + "content": "relative compared to Chen", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "et al. [2019]. Our method has lower performance than the best supervised methods but it performs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "still very well at PER 12 on core-test in the matched setup compared to PER 8.3 for the state of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "score": 1.0, + "content": "art [Baevski et al., 2020c].", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 555, + 297, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "score": 1.0, + "content": "4.3 Performance on non-English languages", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "To get a sense of how well the method works on non-English data, we experiment on six languages", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "of the multilingual Librispeech corpus (MLS; Pratap et al. 2020). As baseline we consider the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 364, + 610 + ], + "score": 1.0, + "content": "supervised systems of Pratap et al. [2020] trained on between", + "type": "text" + }, + { + "bbox": [ + 365, + 598, + 377, + 608 + ], + "score": 0.31, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "and 161 hours of labeled data,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "depending on the language. For adversarial learning we use 100 hours of unlabeled audio data from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "MLS for every language as well as the MLS language modeling data. As input to wav2vec-U we use", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "the representations from XLSR-53 [Conneau et al., 2020], a wav2vec 2.0 model pre-trained on 53", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "score": 1.0, + "content": "languages. Table 3 shows that wav2vec-U generalizes across a range of languages. Performance is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "score": 1.0, + "content": "lower than supervised systems but it shows the viability for other languages.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 506, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "Next, we turn to three low-resource languages, Swahili, Kyrgyz, and Tatar. Swahili is an African lan-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 351, + 691 + ], + "score": 1.0, + "content": "guage, Kyrgyz and Tatar are Turkic languages with only about", + "type": "text" + }, + { + "bbox": [ + 351, + 680, + 372, + 690 + ], + "score": 0.58, + "content": "4 . 3 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 679, + 389, + 691 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 389, + 680, + 411, + 690 + ], + "score": 0.68, + "content": "5 . 2 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "speakers, respectively.5", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 480, + 703 + ], + "score": 1.0, + "content": "We use between 1.8 hours (Kyrgyz) and 9.2 hours of unlabeled audio (Swahili), see Appendix", + "type": "text" + }, + { + "bbox": [ + 481, + 690, + 503, + 702 + ], + "score": 0.83, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 689, + 506, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 118, + 712, + 374, + 722 + ], + "lines": [ + { + "bbox": [ + 118, + 709, + 374, + 725 + ], + "spans": [ + { + "bbox": [ + 118, + 709, + 374, + 725 + ], + "score": 1.0, + "content": "5https://en.wikipedia.org/wiki/{Kyrgyz,Tatar}_language", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 741, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 505, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 505, + 82 + ], + "score": 1.0, + "content": "Table 1: WER on Librispeech dev/test sets when using 960 hours of unlabeled audio from Librispeech", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 506, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 155, + 93 + ], + "score": 1.0, + "content": "(LS-960) or", + "type": "text" + }, + { + "bbox": [ + 155, + 82, + 180, + 92 + ], + "score": 0.31, + "content": "5 3 . 2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 82, + 506, + 93 + ], + "score": 1.0, + "content": "hours from Libri-Light (LL-60k) using representations from wav2vec 2.0 LARGE.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "spans": [ + { + "bbox": [ + 104, + 92, + 506, + 105 + ], + "score": 1.0, + "content": "Librispeech provides clean dev/test sets which are less challenging than the other sets. We report", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 102, + 495, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 458, + 115 + ], + "score": 1.0, + "content": "results for GAN training only (wav2vec-U) and with subsequent self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 458, + 103, + 490, + 114 + ], + "score": 0.78, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ,", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 102, + 495, + 115 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 107, + 122, + 504, + 331 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "spans": [ + { + "bbox": [ + 107, + 122, + 504, + 331 + ], + "score": 0.983, + "html": "
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
", + "type": "table", + "image_path": "7d5f6a9816458b5d9d90dd620c97a44d70ecc200d9e2e077d3e3e3e887499080.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 122, + 504, + 191.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 191.66666666666669, + 504, + 261.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 261.33333333333337, + 504, + 331.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "title", + "bbox": [ + 106, + 345, + 306, + 358 + ], + "lines": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 345, + 308, + 360 + ], + "score": 1.0, + "content": "4.2 Comparison to Prior Unsupervised Work", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 366, + 505, + 421 + ], + "lines": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 366, + 505, + 378 + ], + "score": 1.0, + "content": "Prior work on unsupervised speech recognition focused on the TIMIT benchmark. In order to perform", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 389 + ], + "score": 1.0, + "content": "a direct comparison to these approaches, we report results on this benchmark as well. We consider", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 505, + 401 + ], + "score": 1.0, + "content": "two setups to compare to previous work: in the matched setting, the unlabeled text data is simply the", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 505, + 411 + ], + "score": 1.0, + "content": "transcriptions of the unlabeled audio data but unpaired. In the unmatched setup, the unlabeled text", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 469, + 423 + ], + "score": 1.0, + "content": "data does not contain the transcriptions for the audio data which is a more realistic setting.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 366, + 505, + 423 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 481 + ], + "lines": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 506, + 438 + ], + "score": 1.0, + "content": "We measure performance on the standard Kaldi dev and test sets (core-dev/core-test) as well as a", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 506, + 449 + ], + "score": 1.0, + "content": "slightly larger version of the test set (all-test) to be able to compare to Liu et al. [2018] and Chen et al.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 447, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 371, + 461 + ], + "score": 1.0, + "content": "[2019]. Further details of the two setups can be found in Appendix", + "type": "text" + }, + { + "bbox": [ + 372, + 448, + 393, + 459 + ], + "score": 0.66, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 447, + 506, + 461 + ], + "score": 1.0, + "content": ". We report performance for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 506, + 472 + ], + "score": 1.0, + "content": "wav2vec-U with a 4-gram language model trained on the language modeling data of TIMIT and we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 469, + 291, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 254, + 482 + ], + "score": 1.0, + "content": "also consider self-training (wav2vec-", + "type": "text" + }, + { + "bbox": [ + 254, + 470, + 285, + 481 + ], + "score": 0.47, + "content": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 469, + 291, + 482 + ], + "score": 1.0, + "content": ").", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 426, + 506, + 482 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 541 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "Table 2 shows that wav2vec-U outperforms prior unsupervised work in both the matched and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 497, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 379, + 509 + ], + "score": 1.0, + "content": "unmatched settings, reducing PER on all-test in the matched setup by", + "type": "text" + }, + { + "bbox": [ + 379, + 497, + 399, + 508 + ], + "score": 0.87, + "content": "57 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 498, + 505, + 509 + ], + "score": 1.0, + "content": "relative compared to Chen", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "et al. [2019]. Our method has lower performance than the best supervised methods but it performs", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 532 + ], + "score": 1.0, + "content": "still very well at PER 12 on core-test in the matched setup compared to PER 8.3 for the state of the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 214, + 543 + ], + "score": 1.0, + "content": "art [Baevski et al., 2020c].", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 485, + 505, + 543 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 555, + 297, + 567 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 298, + 570 + ], + "score": 1.0, + "content": "4.3 Performance on non-English languages", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 505, + 663 + ], + "lines": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 506, + 589 + ], + "score": 1.0, + "content": "To get a sense of how well the method works on non-English data, we experiment on six languages", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 506, + 599 + ], + "score": 1.0, + "content": "of the multilingual Librispeech corpus (MLS; Pratap et al. 2020). As baseline we consider the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 597, + 506, + 610 + ], + "spans": [ + { + "bbox": [ + 105, + 597, + 364, + 610 + ], + "score": 1.0, + "content": "supervised systems of Pratap et al. [2020] trained on between", + "type": "text" + }, + { + "bbox": [ + 365, + 598, + 377, + 608 + ], + "score": 0.31, + "content": "2 \\mathrm { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 597, + 506, + 610 + ], + "score": 1.0, + "content": "and 161 hours of labeled data,", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 506, + 621 + ], + "score": 1.0, + "content": "depending on the language. For adversarial learning we use 100 hours of unlabeled audio data from", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "MLS for every language as well as the MLS language modeling data. As input to wav2vec-U we use", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "spans": [ + { + "bbox": [ + 105, + 630, + 505, + 643 + ], + "score": 1.0, + "content": "the representations from XLSR-53 [Conneau et al., 2020], a wav2vec 2.0 model pre-trained on 53", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 506, + 653 + ], + "score": 1.0, + "content": "languages. Table 3 shows that wav2vec-U generalizes across a range of languages. Performance is", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 651, + 412, + 666 + ], + "score": 1.0, + "content": "lower than supervised systems but it shows the viability for other languages.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 574, + 506, + 666 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 668, + 506, + 702 + ], + "lines": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "Next, we turn to three low-resource languages, Swahili, Kyrgyz, and Tatar. Swahili is an African lan-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 679, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 351, + 691 + ], + "score": 1.0, + "content": "guage, Kyrgyz and Tatar are Turkic languages with only about", + "type": "text" + }, + { + "bbox": [ + 351, + 680, + 372, + 690 + ], + "score": 0.58, + "content": "4 . 3 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 679, + 389, + 691 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 389, + 680, + 411, + 690 + ], + "score": 0.68, + "content": "5 . 2 \\mathrm { m }", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 679, + 505, + 691 + ], + "score": 1.0, + "content": "speakers, respectively.5", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 689, + 506, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 480, + 703 + ], + "score": 1.0, + "content": "We use between 1.8 hours (Kyrgyz) and 9.2 hours of unlabeled audio (Swahili), see Appendix", + "type": "text" + }, + { + "bbox": [ + 481, + 690, + 503, + 702 + ], + "score": 0.83, + "content": "\\ S \\operatorname { E } . 1", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 689, + 506, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33, + "bbox_fs": [ + 105, + 667, + 506, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 506, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "score": 1.0, + "content": "Table 2: TIMIT Phoneme Error Rate (PER) in comparison to previous work for the matched and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 281, + 93 + ], + "score": 1.0, + "content": "unmatched training data setups (Appendix", + "type": "text" + }, + { + "bbox": [ + 281, + 82, + 304, + 93 + ], + "score": 0.77, + "content": "\\ S \\operatorname { E . 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 82, + 505, + 93 + ], + "score": 1.0, + "content": "). PER is measured on the Kaldi dev and test sets", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "(core-dev/core-test) as well as a slightly larger version of the test set (all-test) as used by some of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 103, + 491, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 154, + 116 + ], + "score": 1.0, + "content": "prior work.", + "type": "text" + }, + { + "bbox": [ + 155, + 103, + 168, + 115 + ], + "score": 0.76, + "content": "( ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 103, + 491, + 116 + ], + "score": 1.0, + "content": "indicates experiments that do not use the standard split excluding SA utterances.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "spans": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "score": 0.981, + "html": "
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
", + "type": "table", + "image_path": "77fd4860fa4ec19b6701be89b6868bc4fe53c80cb80e10678edaf6050869d0ac.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 147, + 121, + 464, + 201.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 147, + 201.66666666666669, + 464, + 282.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 147, + 282.33333333333337, + 464, + 363.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 369, + 505, + 403 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "Table 3: WER on the Multilingual Librispeech (MLS) dataset using representations from the wav2vec", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "2.0 XLSR-53 model. We consider German (de), Dutch (nl), French (fr), Spanish (es), Italian (it),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "score": 1.0, + "content": "Portuguese (pt).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "spans": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "score": 0.983, + "html": "
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
", + "type": "table", + "image_path": "b406f6c833d0575a5f43e3163dc4e9bbbbefc546de30a4f3e6945916da6d56ac.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 409, + 503, + 446.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 446.3333333333333, + 503, + 483.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 483.66666666666663, + 503, + 521.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "To compare to prior work, we measure WER for Swahili and PER for Kyrgyz and Tatar. For Tatar", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "and Kyrgyz we opted to use a reduced self-training regime for faster experimental turn-around where", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "we only perform HMM self-training and we expect better performance with the full self-training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "setup (Appendix F). Table 4 and Table 5 show that wav2vec-U achieves good performance on these", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "low-resource languages compared to previous work that utilized labeled data. We note that for Tatar", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "and Kyrgyz we use a much smaller amount of speech audio than prior work: compared to XLSR-53", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 136, + 614 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 136, + 602, + 155, + 612 + ], + "score": 0.26, + "content": "1 . 8 \\mathrm { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 601, + 460, + 614 + ], + "score": 1.0, + "content": "unlabeled data vs 17h of labeled data for Kyrgyz and 4.6h vs. 17h for Tatar.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 107, + 630, + 197, + 643 + ], + "lines": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "This paper builds on a large body of prior work which includes semi-supervised speech recognition", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "such as self-training [Kahn et al., 2020a, Xu et al., 2020b, Park et al., 2020]. Some of the earliest", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "score": 1.0, + "content": "work in self-supervised learning of speech representations was was done by van den Oord et al.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "[2018] for phoneme recognition which was simplified in Schneider et al. [2019] who applied it to full", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "speech recognition. Other work includes language model-style pre-training [Chung et al., 2019a] and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "learning fixed size representations of audio segments [Chung and Glass, 2018]. There is also work", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 23.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 742, + 308, + 750 + ], + "lines": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "spans": [ + { + "bbox": [ + 301, + 740, + 310, + 752 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 70, + 506, + 115 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 506, + 83 + ], + "score": 1.0, + "content": "Table 2: TIMIT Phoneme Error Rate (PER) in comparison to previous work for the matched and", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 82, + 505, + 93 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 281, + 93 + ], + "score": 1.0, + "content": "unmatched training data setups (Appendix", + "type": "text" + }, + { + "bbox": [ + 281, + 82, + 304, + 93 + ], + "score": 0.77, + "content": "\\ S \\operatorname { E . 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 304, + 82, + 505, + 93 + ], + "score": 1.0, + "content": "). PER is measured on the Kaldi dev and test sets", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "spans": [ + { + "bbox": [ + 106, + 93, + 505, + 104 + ], + "score": 1.0, + "content": "(core-dev/core-test) as well as a slightly larger version of the test set (all-test) as used by some of the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 103, + 491, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 154, + 116 + ], + "score": 1.0, + "content": "prior work.", + "type": "text" + }, + { + "bbox": [ + 155, + 103, + 168, + 115 + ], + "score": 0.76, + "content": "( ^ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 103, + 491, + 116 + ], + "score": 1.0, + "content": "indicates experiments that do not use the standard split excluding SA utterances.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "table_body", + "bbox": [ + 147, + 121, + 464, + 363 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "spans": [ + { + "bbox": [ + 147, + 121, + 464, + 363 + ], + "score": 0.981, + "html": "
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
", + "type": "table", + "image_path": "77fd4860fa4ec19b6701be89b6868bc4fe53c80cb80e10678edaf6050869d0ac.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 147, + 121, + 464, + 201.66666666666669 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 147, + 201.66666666666669, + 464, + 282.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 147, + 282.33333333333337, + 464, + 363.00000000000006 + ], + "spans": [], + "index": 6 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 107, + 369, + 505, + 403 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "Table 3: WER on the Multilingual Librispeech (MLS) dataset using representations from the wav2vec", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 506, + 392 + ], + "score": 1.0, + "content": "2.0 XLSR-53 model. We consider German (de), Dutch (nl), French (fr), Spanish (es), Italian (it),", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 389, + 172, + 405 + ], + "score": 1.0, + "content": "Portuguese (pt).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8 + }, + { + "type": "table_body", + "bbox": [ + 108, + 409, + 503, + 521 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "spans": [ + { + "bbox": [ + 108, + 409, + 503, + 521 + ], + "score": 0.983, + "html": "
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
", + "type": "table", + "image_path": "b406f6c833d0575a5f43e3163dc4e9bbbbefc546de30a4f3e6945916da6d56ac.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 409, + 503, + 446.3333333333333 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 446.3333333333333, + 503, + 483.66666666666663 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 483.66666666666663, + 503, + 521.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 106, + 535, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 549 + ], + "score": 1.0, + "content": "To compare to prior work, we measure WER for Swahili and PER for Kyrgyz and Tatar. For Tatar", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "and Kyrgyz we opted to use a reduced self-training regime for faster experimental turn-around where", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 505, + 572 + ], + "score": 1.0, + "content": "we only perform HMM self-training and we expect better performance with the full self-training", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 581 + ], + "score": 1.0, + "content": "setup (Appendix F). Table 4 and Table 5 show that wav2vec-U achieves good performance on these", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "low-resource languages compared to previous work that utilized labeled data. We note that for Tatar", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "and Kyrgyz we use a much smaller amount of speech audio than prior work: compared to XLSR-53", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 601, + 460, + 614 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 136, + 614 + ], + "score": 1.0, + "content": "we use", + "type": "text" + }, + { + "bbox": [ + 136, + 602, + 155, + 612 + ], + "score": 0.26, + "content": "1 . 8 \\mathrm { h }", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 601, + 460, + 614 + ], + "score": 1.0, + "content": "unlabeled data vs 17h of labeled data for Kyrgyz and 4.6h vs. 17h for Tatar.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 16, + "bbox_fs": [ + 104, + 536, + 506, + 614 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 630, + 197, + 643 + ], + "lines": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "spans": [ + { + "bbox": [ + 104, + 629, + 198, + 644 + ], + "score": 1.0, + "content": "5 Related Work", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "This paper builds on a large body of prior work which includes semi-supervised speech recognition", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 679 + ], + "score": 1.0, + "content": "such as self-training [Kahn et al., 2020a, Xu et al., 2020b, Park et al., 2020]. Some of the earliest", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 507, + 690 + ], + "score": 1.0, + "content": "work in self-supervised learning of speech representations was was done by van den Oord et al.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 689, + 506, + 701 + ], + "score": 1.0, + "content": "[2018] for phoneme recognition which was simplified in Schneider et al. [2019] who applied it to full", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "speech recognition. Other work includes language model-style pre-training [Chung et al., 2019a] and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 506, + 723 + ], + "score": 1.0, + "content": "learning fixed size representations of audio segments [Chung and Glass, 2018]. There is also work", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "on quantization of the continuous speech data [Baevski et al., 2020b,a, Liu et al., 2019, van Niekerk", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "et al., 2020, Baevski et al., 2020c, Hsu et al., 2021b] and on robustness to domain shift [Hsu et al.,", + "type": "text", + "cross_page": true + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "2021a], multilingual pre-training [Kawakami et al., 2020, Conneau et al., 2020] as well as combining", + "type": "text", + "cross_page": true + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "score": 1.0, + "content": "speech and vision [Harwath et al., 2020].", + "type": "text", + "cross_page": true + } + ], + "index": 19 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 656, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 86, + 291, + 108 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "score": 1.0, + "content": "Table 4: PER for low-resource languages,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "score": 1.0, + "content": "Tatar (tt) and Kyrgyz (ky).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "spans": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "score": 0.979, + "html": "
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
", + "type": "table", + "image_path": "2681bd270ee947c4735ec6105ffa578b603c0a8f14a89fb155bf16d2679be651.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 109, + 307, + 122.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 122.0, + 307, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 135.0, + 307, + 148.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 148.0, + 307, + 161.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 161.0, + 307, + 174.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 174.0, + 307, + 187.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 187.0, + 307, + 200.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 200.0, + 307, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 213.0, + 307, + 226.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 319, + 87, + 504, + 119 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "score": 1.0, + "content": "Table 5: WER for Swahili from the ALFFA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "score": 1.0, + "content": "corpus. We compare to the supervised baseline", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "score": 1.0, + "content": "of the ALFFA project.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "score": 0.974, + "html": "
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
", + "type": "table", + "image_path": "75c82e8aef7073d98bca3ee508ceb7960d3a4e7ae6963489d1130fb03b1695b7.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 346, + 120, + 477, + 168.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 346, + 168.5, + 477, + 217.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 305 + ], + "lines": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "spans": [ + { + "bbox": [ + 106, + 261, + 505, + 273 + ], + "score": 1.0, + "content": "on quantization of the continuous speech data [Baevski et al., 2020b,a, Liu et al., 2019, van Niekerk", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 272, + 506, + 284 + ], + "score": 1.0, + "content": "et al., 2020, Baevski et al., 2020c, Hsu et al., 2021b] and on robustness to domain shift [Hsu et al.,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 505, + 296 + ], + "score": 1.0, + "content": "2021a], multilingual pre-training [Kawakami et al., 2020, Conneau et al., 2020] as well as combining", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 272, + 306 + ], + "score": 1.0, + "content": "speech and vision [Harwath et al., 2020].", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "Learning to map speech to phonemes without supervision using adversarial learning has been explored", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "by Liu et al. [2018] who learn a mapping matrix between segment identifiers and phonemes. However,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "their work still relied on data segmented into phonemes by human annotators. This has been later", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "extended to use an automatic segmentation [Chen et al., 2019] which is iteratively refined with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "HMMs. However, cross validation is still performed using labeled data (personal communication with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "authors). We also explored HMMs to refine segmentation boundaries (Table A2) but did not find it as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "effective as self-training. Our work is in part inspired by aligning word embedding spaces of different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "languages [Mikolov et al., 2013, Artetxe et al., 2017, Conneau et al., 2018] and full unsupervised", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "score": 1.0, + "content": "machine translation [Lample et al., 2018, Artetxe et al., 2018, Conneau and Lample, 2019].", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 107, + 441, + 275, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "score": 1.0, + "content": "6 Conclusion and Future Work", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "score": 1.0, + "content": "wav2vec-U is a framework which enables building speech recognition models without labeled data.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "score": 1.0, + "content": "It embeds and segments the speech audio with self-supervised representations from wav2vec 2.0,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "learns a mapping to phonemes with adversarial learning, and cross-validates hyper-parameter choices", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "as well as early stopping with an unsupervised metric. Experiments on the standard Librispeech", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "score": 1.0, + "content": "benchmark show performance close to the state of the art models from only a few years ago, even", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "score": 1.0, + "content": "though these models relied on nearly 1,000 hours of labeled data.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "Compared to the previous best unsupervised speech recognition approach, wav2vec-U reduces TIMIT", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "phone error rate from 26.1 to 11.3. We also demonstrate the viability of our approach on several", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "languages other than English, some of which are low-resource. The ability to build speech recognition", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "models solely from unlabeled speech audio and unlabeled text drastically lowers the effort to build", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "score": 1.0, + "content": "speech technology for many more languages of the world.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "Our approach requires phonemization of the text for the language of interest. Moreover, phonemizers", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "are not available for all languages and this presents a bottleneck. To address this, future work may", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "develop phonemizers for more languages, explore phonemization approaches that generalize across", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "score": 1.0, + "content": "languages, or unsupervised training with graphemic text units such as letters.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "We explored a simple segmentation technique based on self-supervised representations, however, there", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "is a large body of research on segmentation and some of these techniques may lead to improvements", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "over our simple approach [Varadarajan et al., 2008, Zhang and Glass, 2009, Gish et al., 2009, Lee and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "Glass, 2012, Lee et al., 2015, Ondel et al., 2016, Kamper et al., 2017a,b, Kreuk et al., 2020]. Also,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "wav2vec 2.0 learns representations for fixed size units with a fixed stride, however, phonemic units", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "score": 1.0, + "content": "are of variable size. Another direction is to learn variable sized representations during pre-training.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 741, + 309, + 750 + ], + "lines": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "spans": [ + { + "bbox": [ + 302, + 741, + 309, + 752 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 86, + 291, + 108 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 85, + 292, + 98 + ], + "score": 1.0, + "content": "Table 4: PER for low-resource languages,", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "spans": [ + { + "bbox": [ + 105, + 96, + 214, + 110 + ], + "score": 1.0, + "content": "Tatar (tt) and Kyrgyz (ky).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 108, + 109, + 307, + 226 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "spans": [ + { + "bbox": [ + 108, + 109, + 307, + 226 + ], + "score": 0.979, + "html": "
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
", + "type": "table", + "image_path": "2681bd270ee947c4735ec6105ffa578b603c0a8f14a89fb155bf16d2679be651.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 108, + 109, + 307, + 122.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 108, + 122.0, + 307, + 135.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 108, + 135.0, + 307, + 148.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 108, + 148.0, + 307, + 161.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 108, + 161.0, + 307, + 174.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 108, + 174.0, + 307, + 187.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 108, + 187.0, + 307, + 200.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 108, + 200.0, + 307, + 213.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 108, + 213.0, + 307, + 226.0 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 3.25 + }, + { + "type": "table", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 319, + 87, + 504, + 119 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 320, + 86, + 505, + 98 + ], + "score": 1.0, + "content": "Table 5: WER for Swahili from the ALFFA", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "spans": [ + { + "bbox": [ + 320, + 98, + 505, + 110 + ], + "score": 1.0, + "content": "corpus. We compare to the supervised baseline", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 412, + 121 + ], + "score": 1.0, + "content": "of the ALFFA project.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + }, + { + "type": "table_body", + "bbox": [ + 346, + 120, + 477, + 217 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "spans": [ + { + "bbox": [ + 346, + 120, + 477, + 217 + ], + "score": 0.974, + "html": "
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
", + "type": "table", + "image_path": "75c82e8aef7073d98bca3ee508ceb7960d3a4e7ae6963489d1130fb03b1695b7.jpg" + } + ] + } + ], + "index": 14.5, + "virtual_lines": [ + { + "bbox": [ + 346, + 120, + 477, + 168.5 + ], + "spans": [], + "index": 14 + }, + { + "bbox": [ + 346, + 168.5, + 477, + 217.0 + ], + "spans": [], + "index": 15 + } + ] + } + ], + "index": 13.25 + }, + { + "type": "text", + "bbox": [ + 107, + 261, + 505, + 305 + ], + "lines": [], + "index": 17.5, + "bbox_fs": [ + 105, + 261, + 506, + 306 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 310, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 322 + ], + "score": 1.0, + "content": "Learning to map speech to phonemes without supervision using adversarial learning has been explored", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 506, + 334 + ], + "score": 1.0, + "content": "by Liu et al. [2018] who learn a mapping matrix between segment identifiers and phonemes. However,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "spans": [ + { + "bbox": [ + 106, + 332, + 506, + 344 + ], + "score": 1.0, + "content": "their work still relied on data segmented into phonemes by human annotators. This has been later", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 505, + 355 + ], + "score": 1.0, + "content": "extended to use an automatic segmentation [Chen et al., 2019] which is iteratively refined with", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 353, + 505, + 367 + ], + "score": 1.0, + "content": "HMMs. However, cross validation is still performed using labeled data (personal communication with", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 364, + 506, + 378 + ], + "score": 1.0, + "content": "authors). We also explored HMMs to refine segmentation boundaries (Table A2) but did not find it as", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 505, + 389 + ], + "score": 1.0, + "content": "effective as self-training. Our work is in part inspired by aligning word embedding spaces of different", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 506, + 399 + ], + "score": 1.0, + "content": "languages [Mikolov et al., 2013, Artetxe et al., 2017, Conneau et al., 2018] and full unsupervised", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 398, + 472, + 410 + ], + "score": 1.0, + "content": "machine translation [Lample et al., 2018, Artetxe et al., 2018, Conneau and Lample, 2019].", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 311, + 506, + 410 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 441, + 275, + 455 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 277, + 456 + ], + "score": 1.0, + "content": "6 Conclusion and Future Work", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 475, + 505, + 542 + ], + "lines": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 476, + 506, + 488 + ], + "score": 1.0, + "content": "wav2vec-U is a framework which enables building speech recognition models without labeled data.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 488, + 506, + 499 + ], + "score": 1.0, + "content": "It embeds and segments the speech audio with self-supervised representations from wav2vec 2.0,", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 506, + 511 + ], + "score": 1.0, + "content": "learns a mapping to phonemes with adversarial learning, and cross-validates hyper-parameter choices", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 521 + ], + "score": 1.0, + "content": "as well as early stopping with an unsupervised metric. Experiments on the standard Librispeech", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 505, + 533 + ], + "score": 1.0, + "content": "benchmark show performance close to the state of the art models from only a few years ago, even", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "spans": [ + { + "bbox": [ + 106, + 531, + 369, + 543 + ], + "score": 1.0, + "content": "though these models relied on nearly 1,000 hours of labeled data.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 476, + 506, + 543 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 547, + 504, + 602 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 506, + 560 + ], + "score": 1.0, + "content": "Compared to the previous best unsupervised speech recognition approach, wav2vec-U reduces TIMIT", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 558, + 505, + 570 + ], + "score": 1.0, + "content": "phone error rate from 26.1 to 11.3. We also demonstrate the viability of our approach on several", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 506, + 582 + ], + "score": 1.0, + "content": "languages other than English, some of which are low-resource. The ability to build speech recognition", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "models solely from unlabeled speech audio and unlabeled text drastically lowers the effort to build", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 340, + 603 + ], + "score": 1.0, + "content": "speech technology for many more languages of the world.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 546, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 505, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 607, + 505, + 619 + ], + "score": 1.0, + "content": "Our approach requires phonemization of the text for the language of interest. Moreover, phonemizers", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "are not available for all languages and this presents a bottleneck. To address this, future work may", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "develop phonemizers for more languages, explore phonemization approaches that generalize across", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "spans": [ + { + "bbox": [ + 105, + 640, + 415, + 652 + ], + "score": 1.0, + "content": "languages, or unsupervised training with graphemic text units such as letters.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 607, + 506, + 652 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 505, + 722 + ], + "lines": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 668 + ], + "score": 1.0, + "content": "We explored a simple segmentation technique based on self-supervised representations, however, there", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 680 + ], + "score": 1.0, + "content": "is a large body of research on segmentation and some of these techniques may lead to improvements", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "over our simple approach [Varadarajan et al., 2008, Zhang and Glass, 2009, Gish et al., 2009, Lee and", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "Glass, 2012, Lee et al., 2015, Ondel et al., 2016, Kamper et al., 2017a,b, Kreuk et al., 2020]. Also,", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 713 + ], + "score": 1.0, + "content": "wav2vec 2.0 learns representations for fixed size units with a fixed stride, however, phonemic units", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 503, + 725 + ], + "score": 1.0, + "content": "are of variable size. Another direction is to learn variable sized representations during pre-training.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 47.5, + "bbox_fs": [ + 105, + 655, + 506, + 725 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 339, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 506, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "We thank Zhouhan Lin for helping with initial explorations in this project, Tatiana Likhomanenko", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "score": 1.0, + "content": "for helpful discussions about self-training, Da-Rong Liu for sharing details to reproduce the setup", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "score": 1.0, + "content": "of Chen et al. [2019], Marc’Aurelio Ranzato for general helpful discussions, and Ruth Kipng’eno,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "score": 1.0, + "content": "Ruth Ndila Ndeto as well as Mark Mutitu for error analysis of our Swahili model.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 155, + 163, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 105, + 172, + 507, + 718 + ], + "lines": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "score": 1.0, + "content": "S. T. Abate, W. Menzel, and B. Tafila. An amharic speech corpus for large vocabulary continuous", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of Interspeech, 2005.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "score": 1.0, + "content": "Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "score": 1.0, + "content": "In Proc. of ICML, 2016.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. Proc. of LREC,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "score": 1.0, + "content": "M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, and E. Agirre. Learning bilingual word embeddings with (almost) no bilingual", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "score": 1.0, + "content": "data. In Proc. of ACL, 2017.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, E. Agirre, and K. Cho. Unsupervised neural machine translation. Proc. of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "score": 1.0, + "content": "A. Baevski and M. Auli. Adaptive input representations for neural language modeling. In Proc. of", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "A. Baevski, M. Auli, and A. Mohamed. Effectiveness of self-supervised pre-training for speech", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "score": 1.0, + "content": "recognition. Proc. of ICASSP, 2020a.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "score": 1.0, + "content": "A. Baevski, S. Schneider, and M. Auli. vq-wav2vec: Self-supervised learning of discrete speech", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "spans": [ + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "score": 1.0, + "content": "representations. In Proc. of ICLR, 2020b.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "A. Baevski, Y. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "spans": [ + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "score": 1.0, + "content": "learning of speech representations. In Proc. of NeurIPS, 2020c.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "L. Besacier, E. Gauthier, M. Mangeot, P. Bretier, P. Bagshaw, O. Rosec, T. Moudenc, F. Pellegrino,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "S. Voisin, E. Marsico, and P. Nocera. Speech technologies for african languages: example of a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "score": 1.0, + "content": "multilingual calculator for education. In Proc. of Interspeech, 2015.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "score": 1.0, + "content": "D. Can, V. R. Martinez, P. Papadopoulos, and S. S. Narayanan. Pykaldi: A python wrapper for kaldi.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "score": 1.0, + "content": "In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "K.-Y. Chen, C.-P. Tsai, D.-R. Liu, H.-Y. Lee, and L. shan Lee. Completely unsupervised speech", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "recognition by a generative adversarial network harmonized with iteratively refined hidden markov", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "score": 1.0, + "content": "models. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "Y. Chung, W. Weng, S. Tong, and J. R. Glass. Unsupervised cross-modal alignment of speech and", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "score": 1.0, + "content": "text embedding spaces. Proc. of NIPS, 2018.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "Y. Chung, W. Hsu, H. Tang, and J. R. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "spans": [ + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019a.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "Y.-A. Chung and J. Glass. Speech2vec: A sequence-to-sequence framework for learning word", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "score": 1.0, + "content": "embeddings from speech. Proc. of Interspeech, 2018.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "score": 1.0, + "content": "Y.-A. Chung, W.-N. Hsu, H. Tang, and J. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019b.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "score": 1.0, + "content": "A. Conneau and G. Lample. Cross-lingual language model pretraining. Proc. of NeurIPS, 2019.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 24.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 70, + 339, + 85 + ], + "lines": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "spans": [ + { + "bbox": [ + 105, + 69, + 341, + 88 + ], + "score": 1.0, + "content": "Acknowledgments and Disclosure of Funding", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 506, + 140 + ], + "lines": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 95, + 506, + 108 + ], + "score": 1.0, + "content": "We thank Zhouhan Lin for helping with initial explorations in this project, Tatiana Likhomanenko", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "spans": [ + { + "bbox": [ + 105, + 106, + 506, + 120 + ], + "score": 1.0, + "content": "for helpful discussions about self-training, Da-Rong Liu for sharing details to reproduce the setup", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 507, + 131 + ], + "score": 1.0, + "content": "of Chen et al. [2019], Marc’Aurelio Ranzato for general helpful discussions, and Ruth Kipng’eno,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 128, + 434, + 140 + ], + "score": 1.0, + "content": "Ruth Ndila Ndeto as well as Mark Mutitu for error analysis of our Swahili model.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 2.5, + "bbox_fs": [ + 105, + 95, + 507, + 140 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 155, + 163, + 168 + ], + "lines": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 165, + 170 + ], + "score": 1.0, + "content": "References", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "list", + "bbox": [ + 105, + 172, + 507, + 718 + ], + "lines": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 189 + ], + "score": 1.0, + "content": "S. T. Abate, W. Menzel, and B. Tafila. An amharic speech corpus for large vocabulary continuous", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 185, + 320, + 199 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of Interspeech, 2005.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro,", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "spans": [ + { + "bbox": [ + 114, + 214, + 507, + 230 + ], + "score": 1.0, + "content": "Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "spans": [ + { + "bbox": [ + 115, + 226, + 216, + 239 + ], + "score": 1.0, + "content": "In Proc. of ICML, 2016.", + "type": "text" + } + ], + "index": 10, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 244, + 506, + 257 + ], + "score": 1.0, + "content": "R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M.", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 115, + 255, + 506, + 269 + ], + "score": 1.0, + "content": "Tyers, and G. Weber. Common voice: A massively-multilingual speech corpus. Proc. of LREC,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "spans": [ + { + "bbox": [ + 115, + 265, + 143, + 280 + ], + "score": 1.0, + "content": "2020.", + "type": "text" + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 430, + 297 + ], + "score": 1.0, + "content": "M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. Proc. of ICML, 2017.", + "type": "text" + } + ], + "index": 14, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 505, + 317 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, and E. Agirre. Learning bilingual word embeddings with (almost) no bilingual", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "spans": [ + { + "bbox": [ + 115, + 315, + 233, + 328 + ], + "score": 1.0, + "content": "data. In Proc. of ACL, 2017.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 506, + 347 + ], + "score": 1.0, + "content": "M. Artetxe, G. Labaka, E. Agirre, and K. Cho. Unsupervised neural machine translation. Proc. of", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "spans": [ + { + "bbox": [ + 115, + 345, + 168, + 356 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "spans": [ + { + "bbox": [ + 104, + 361, + 507, + 378 + ], + "score": 1.0, + "content": "A. Baevski and M. Auli. Adaptive input representations for neural language modeling. In Proc. of", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "spans": [ + { + "bbox": [ + 115, + 374, + 168, + 386 + ], + "score": 1.0, + "content": "ICLR, 2018.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 393, + 505, + 406 + ], + "score": 1.0, + "content": "A. Baevski, M. Auli, and A. Mohamed. Effectiveness of self-supervised pre-training for speech", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "spans": [ + { + "bbox": [ + 115, + 404, + 269, + 417 + ], + "score": 1.0, + "content": "recognition. Proc. of ICASSP, 2020a.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 506, + 437 + ], + "score": 1.0, + "content": "A. Baevski, S. Schneider, and M. Auli. vq-wav2vec: Self-supervised learning of discrete speech", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "spans": [ + { + "bbox": [ + 114, + 433, + 284, + 447 + ], + "score": 1.0, + "content": "representations. In Proc. of ICLR, 2020b.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "A. Baevski, Y. Zhou, A. Mohamed, and M. Auli. wav2vec 2.0: A framework for self-supervised", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "spans": [ + { + "bbox": [ + 115, + 463, + 372, + 476 + ], + "score": 1.0, + "content": "learning of speech representations. In Proc. of NeurIPS, 2020c.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 482, + 506, + 495 + ], + "score": 1.0, + "content": "L. Besacier, E. Gauthier, M. Mangeot, P. Bretier, P. Bagshaw, O. Rosec, T. Moudenc, F. Pellegrino,", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 115, + 493, + 506, + 506 + ], + "score": 1.0, + "content": "S. Voisin, E. Marsico, and P. Nocera. Speech technologies for african languages: example of a", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "spans": [ + { + "bbox": [ + 115, + 504, + 389, + 517 + ], + "score": 1.0, + "content": "multilingual calculator for education. In Proc. of Interspeech, 2015.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "spans": [ + { + "bbox": [ + 104, + 520, + 507, + 536 + ], + "score": 1.0, + "content": "D. Can, V. R. Martinez, P. Papadopoulos, and S. S. Narayanan. Pykaldi: A python wrapper for kaldi.", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 533, + 223, + 546 + ], + "score": 1.0, + "content": "In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 506, + 565 + ], + "score": 1.0, + "content": "K.-Y. Chen, C.-P. Tsai, D.-R. Liu, H.-Y. Lee, and L. shan Lee. Completely unsupervised speech", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "spans": [ + { + "bbox": [ + 115, + 563, + 506, + 576 + ], + "score": 1.0, + "content": "recognition by a generative adversarial network harmonized with iteratively refined hidden markov", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "spans": [ + { + "bbox": [ + 115, + 574, + 273, + 587 + ], + "score": 1.0, + "content": "models. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 506, + 606 + ], + "score": 1.0, + "content": "Y. Chung, W. Weng, S. Tong, and J. R. Glass. Unsupervised cross-modal alignment of speech and", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "spans": [ + { + "bbox": [ + 115, + 604, + 298, + 617 + ], + "score": 1.0, + "content": "text embedding spaces. Proc. of NIPS, 2018.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "Y. Chung, W. Hsu, H. Tang, and J. R. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "spans": [ + { + "bbox": [ + 114, + 633, + 329, + 646 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019a.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 506, + 666 + ], + "score": 1.0, + "content": "Y.-A. Chung and J. Glass. Speech2vec: A sequence-to-sequence framework for learning word", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "spans": [ + { + "bbox": [ + 114, + 662, + 333, + 676 + ], + "score": 1.0, + "content": "embeddings from speech. Proc. of Interspeech, 2018.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 679, + 506, + 695 + ], + "score": 1.0, + "content": "Y.-A. Chung, W.-N. Hsu, H. Tang, and J. Glass. An unsupervised autoregressive model for speech", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "spans": [ + { + "bbox": [ + 115, + 692, + 330, + 705 + ], + "score": 1.0, + "content": "representation learning. Proc. of Interspeech, 2019b.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 491, + 723 + ], + "score": 1.0, + "content": "A. Conneau and G. Lample. Cross-lingual language model pretraining. Proc. of NeurIPS, 2019.", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "A. Conneau, G. Lample, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "score": 1.0, + "content": "data. Proc. of ICLR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "score": 1.0, + "content": "A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli. Unsupervised cross-lingual", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "spans": [ + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "score": 1.0, + "content": "representation learning for speech recognition. arXiv, abs/2006.13979, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "score": 1.0, + "content": "J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "score": 1.0, + "content": "transformers for language understanding. Proc. of NAACL, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "L. Dong, S. Xu, and B. Xu. Speech-transformer: A no-recurrence sequence-to-sequence model for", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of ICASSP, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "score": 1.0, + "content": "Z. Fan, M. Li, S. Zhou, and B. Xu. Exploring wav2vec 2.0 on speaker verification and language", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "score": 1.0, + "content": "identification. arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "R. Fer, P. Matejka, F. Grézl, O. Plchot, K. Vesel ˇ y, and J. H. ` Cernock ˇ y. Multilingually trained `", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "spans": [ + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "score": 1.0, + "content": "bottleneck features in spoken language recognition. Computer Speech & Language, 46, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. The DARPA", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "score": 1.0, + "content": "TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium,", + "type": "text", + "cross_page": true + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "score": 1.0, + "content": "1993.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "H. Gelas, L. Besacier, and F. Pellegrino. Developments of Swahili resources for an automatic speech", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "score": 1.0, + "content": "recognition system. In Proc. of SLTU, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "H. Gish, M. Siu, A. Chan, and W. Belfield. Unsupervised training of an hmm-based speech recognizer", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "spans": [ + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "score": 1.0, + "content": "for topic classification. In Proc. of Interspeech, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "score": 1.0, + "content": "I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial networks. Proc. of NIPS, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "Google. Google cloud: Speech-to-text. https://cloud.google.com/speech-to-text, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "score": 1.0, + "content": "Accessed: 2021-05-13.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "A. Graves, S. Fernández, and F. Gomez. Connectionist temporal classification: Labelling unsegmented", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "spans": [ + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "score": 1.0, + "content": "sequence data with recurrent neural networks. In Proc. of ICML, 2006.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "score": 1.0, + "content": "A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wu,", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "and R. Pang. Conformer: Convolution-augmented transformer for speech recognition. Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "score": 1.0, + "content": "Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "spans": [ + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "score": 1.0, + "content": "gans. Proc. of NIPS, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "score": 1.0, + "content": "W. Han, Z. Zhang, Y. Zhang, J. Yu, C.-C. Chiu, J. Qin, A. Gulati, R. Pang, and Y. Wu. Contextnet:", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Improving convolutional neural networks for automatic speech recognition with global context.", + "type": "text", + "cross_page": true + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "score": 1.0, + "content": "Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "D. Harwath and J. Glass. Towards visually grounded sub-word speech unit discovery. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "spans": [ + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "score": 1.0, + "content": "ICASSP, pages 3017–3021. IEEE, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "score": 1.0, + "content": "D. Harwath, W.-N. Hsu, and J. Glass. Learning hierarchical discrete linguistic units from visually-", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "score": 1.0, + "content": "grounded speech. In Proc. of ICLR, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "K. Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "score": 1.0, + "content": "Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "score": 1.0, + "content": "Association for Computational Linguistics.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "score": 1.0, + "content": "K. Hirsh-Pasek, D. G. Kemler Nelson, P. W. Jusczyk, K. W. Cassidy, B. Druss, and L. Kennedy.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "spans": [ + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "score": 1.0, + "content": "Clauses are perceptual units for young infants. Cognition, 26(3):269–286, 1987.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "score": 1.0, + "content": "W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V. Pratap, J. Kahn, A. Lee, R. Collobert,", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "spans": [ + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "score": 1.0, + "content": "G. Synnaeve, et al. Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training.", + "type": "text", + "cross_page": true + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2104.01027, 2021a.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "W.-N. Hsu, Y.-H. H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mohamed. Hubert: How much can a", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "score": 1.0, + "content": "bad teacher benefit ASR pre-training? In Proc. of ICASSP, 2021b.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. Proc. of ICLR,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "spans": [ + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "score": 1.0, + "content": "2016.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "score": 1.0, + "content": "H. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Trans.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "score": 1.0, + "content": "Pattern Anal. Mach. Intell., 33(1):117–128, Jan. 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "E. K. Johnson and P. W. Jusczyk. Word segmentation by 8-month-olds: When speech cues count", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "score": 1.0, + "content": "more than statistics. Journal of Memory and Language, 44(4):548–567, 2001.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "score": 1.0, + "content": "J. Johnson, M. Douze, and H. Jégou. Billion-scale similarity search with gpus. IEEE Transactions on", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "score": 1.0, + "content": "Big Data, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "P. W. Jusczyk, D. M. Houston, and M. Newsome. The beginnings of word segmentation in english-", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "spans": [ + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "score": 1.0, + "content": "learning infants. Cognitive Psychology, 39(3):159–207, 1999.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "score": 1.0, + "content": "J. Kahn, A. Lee, and A. Hannun. Self-training for end-to-end speech recognition. In Proc. of ICASSP,", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "spans": [ + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "score": 1.0, + "content": "2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "score": 1.0, + "content": "J. Kahn et al. Libri-light: A benchmark for asr with limited or no supervision. In Proc. of ICASSP,", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "score": 1.0, + "content": "2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "score": 1.0, + "content": "H. Kamper, A. Jansen, and S. Goldwater. A segmental framework for fully-unsupervised large-", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "spans": [ + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "score": 1.0, + "content": "vocabulary speech recognition. Comput. Speech Lang., 46(C), Nov. 2017a.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "H. Kamper, K. Livescu, and S. Goldwater. An embedded segmental k-means model for unsupervised", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "spans": [ + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "score": 1.0, + "content": "segmentation and clustering of speech. Proc. of ASRU, 2017b.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "score": 1.0, + "content": "K. Kawakami, L. Wang, C. Dyer, P. Blunsom, and A. van den Oord. Learning robust and multilingual", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "score": 1.0, + "content": "speech representations. Proc. of EMNLP, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "score": 1.0, + "content": "D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of ICLR, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "F. Kreuk, J. Keshet, and Y. Adi. Self-supervised contrastive learning for unsupervised phoneme", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "spans": [ + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "score": 1.0, + "content": "segmentation. Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "score": 1.0, + "content": "G. Lample, L. Denoyer, and M. Ranzato. Unsupervised machine translation using monolingual", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "score": 1.0, + "content": "corpora only. In Proc. of ICLR, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "score": 1.0, + "content": "C. Lee and J. R. Glass. A nonparametric bayesian approach to acoustic model discovery. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "spans": [ + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "score": 1.0, + "content": "ACL, 2012.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "C. Lee, T. J. O’Donnell, and J. R. Glass. Unsupervised lexicon discovery from acoustic input. TACL,", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "spans": [ + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "score": 1.0, + "content": "2015.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "score": 1.0, + "content": "M. P. Lewis, G. F. Simon, and C. D. Fennig. Ethnologue: Languages of the world, nineteenth edition.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "spans": [ + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "score": 1.0, + "content": "Online version: http://www.ethnologue.com, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "score": 1.0, + "content": "T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert. slimipl: Language-model-free", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "spans": [ + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "score": 1.0, + "content": "iterative pseudo-labeling. arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "A. H. Liu, T. Tu, H. yi Lee, and L. shan Lee. Towards unsupervised speech recognition and synthesis", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "score": 1.0, + "content": "with quantized speech representation learning. Proc. of ICASSP, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "D.-R. Liu, K.-Y. Chen, H.-Y. Lee, and L. shan Lee. Completely unsupervised phoneme recognition", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "by adversarially learning mapping relationships from audio embeddings. Proc. of Interspeech,", + "type": "text", + "cross_page": true + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "score": 1.0, + "content": "2018.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "T. Mikolov, Q. V. Le, and I. Sutskever. Exploiting similarities among languages for machine", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "score": 1.0, + "content": "translation. arXiv preprint arXiv:1309.4168, 2013.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "M. Mohri, F. Pereira, and M. Riley. Weighted finite-state transducers in speech recognition. Computer", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "spans": [ + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "score": 1.0, + "content": "Speech & Language, 16(1):69–88, 2002.", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "L. Ondel, L. Burget, and J. Cernocký. Variational inference for acoustic unit discovery. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "score": 1.0, + "content": "SLTU, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "score": 1.0, + "content": "M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli. fairseq: A fast,", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "score": 1.0, + "content": "extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "score": 1.0, + "content": "V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "score": 1.0, + "content": "domain audio books. In Proc. of ICASSP, pages 5206–5210. IEEE, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "score": 1.0, + "content": "D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le. Specaugment: A", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "score": 1.0, + "content": "simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "D. S. Park, Y. Zhang, Y. Jia, W. Han, C.-C. Chiu, B. Li, Y. Wu, and Q. V. Le. Improved noisy student", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "score": 1.0, + "content": "training for automatic speech recognition. Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "score": 1.0, + "content": "K. Park and J. Kim. g2pe. https://github.com/Kyubyong/g2p, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "score": 1.0, + "content": "L. Pepino, P. Riera, and L. Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "spans": [ + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "score": 1.0, + "content": "arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "L. Polka and J. F. Werker. Developmental changes in perception of nonnative vowel contrasts. Journal", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "spans": [ + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "score": 1.0, + "content": "of Experimental Psychology: Human perception and performance, 20(2):421, 1994.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "score": 1.0, + "content": "D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek,", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "score": 1.0, + "content": "Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech recognition toolkit.", + "type": "text", + "cross_page": true + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "score": 1.0, + "content": "In Proc. of ASRU, 2011.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "score": 1.0, + "content": "V. Pratap, A. Hannun, Q. Xu, J. Cai, J. Kahn, G. Synnaeve, V. Liptchinsky, and R. Collobert.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "score": 1.0, + "content": "Wav2letter++: A fast open-source speech recognition system. In Proc. of ICASSP, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "score": 1.0, + "content": "V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert. Mls: A large-scale multilingual dataset", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "spans": [ + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "score": 1.0, + "content": "for speech research. In Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "O. Rasanen, G. Doyle, and M. C. Frank. Unsupervised word discovery from speech using automatic", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "spans": [ + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "score": 1.0, + "content": "segmentation into syllable-like units. In Proc. of Interspeech, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "score": 1.0, + "content": "M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio. Light gated recurrent units for speech", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "score": 1.0, + "content": "recognition. IEEE Trans. on Emerging Topics in Comp. Intel., 2, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "spans": [ + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "score": 1.0, + "content": "M. Ravanelli, T. Parcollet, and Y. Bengio. The pytorch-kaldi speech recognition toolkit. Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "score": 1.0, + "content": "ICASSP, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "M. Rivière, A. Joulin, P.-E. Mazaré, and E. Dupoux. Unsupervised pretraining transfers well across", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "spans": [ + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "score": 1.0, + "content": "languages. In Proc. of ICASSP, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "S. Schneider, A. Baevski, R. Collobert, and M. Auli. wav2vec: Unsupervised pre-training for speech", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "score": 1.0, + "content": "recognition. In Proc. of Interspeech, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "score": 1.0, + "content": "N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "spans": [ + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "score": 1.0, + "content": "to prevent neural networks from overfitting. JMLR, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "score": 1.0, + "content": "G. Synnaeve, Q. Xu, J. Kahn, T. Likhomanenko, E. Grave, V. Pratap, A. Sriram, V. Liptchinsky,", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "and R. Collobert. End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern", + "type": "text", + "cross_page": true + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "score": 1.0, + "content": "Architectures. Proc. of ICML workshop on Self-supervision in Audio and Speech (SAS), 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 36 + }, + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "score": 1.0, + "content": "M. Tachbelie, S. T. Abate, and L. Besacier. Using different acoustic, lexical and language modeling", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "spans": [ + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "score": 1.0, + "content": "units for asr of an under-resourced language - amharic. Speech Communication, 56, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "score": 1.0, + "content": "Z. Tan, A. K. Sarkar, and N. Dehak. rvad: An unsupervised segment-based robust voice activity", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "spans": [ + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "score": 1.0, + "content": "detection method. Computer speech & language, 59:1–21, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "A. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "spans": [ + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "score": 1.0, + "content": "Proc. of NIPS, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "B. van Niekerk, L. Nortje, and H. Kamper. Vector-quantized neural networks for acoustic unit", + "type": "text", + "cross_page": true + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "score": 1.0, + "content": "discovery in the zerospeech 2020 challenge. Proc. of Interspeech, 2020.", + "type": "text", + "cross_page": true + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "score": 1.0, + "content": "B. Varadarajan, S. Khudanpur, and E. Dupoux. Unsupervised learning of acoustic sub-word units. In", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "score": 1.0, + "content": "Proc. of ACL, 2008.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "score": 1.0, + "content": "A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "spans": [ + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "score": 1.0, + "content": "Attention is all you need. In Proc. of NIPS, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "C. Wang, A. Wu, J. Pino, A. Baevski, M. Auli, and A. Conneau. Large-scale self- and semi-supervised", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "score": 1.0, + "content": "learning for speech translation. arXiv, 2021.", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "score": 1.0, + "content": "J. F. Werker and R. C. Tees. Cross-language speech perception: Evidence for perceptual reorganization", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "score": 1.0, + "content": "during the first year of life. Infant behavior and development, 7(1):49–63, 1984.", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "H. Xu, K. Li, Y. Wang, J. Wang, S. Kang, X. Chen, D. Povey, and S. Khudanpur. Neural network", + "type": "text", + "cross_page": true + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "score": 1.0, + "content": "language modeling with letter-based features and importance sampling. In Proc. of ICASSP, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "Q. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "score": 1.0, + "content": "M. Auli. Self-training and pre-training are complementary for speech recognition. In Proc. of", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "spans": [ + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "score": 1.0, + "content": "ICASSP, 2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert. Iterative pseudo-", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "score": 1.0, + "content": "labeling for speech recognition. Proc. of Interspeech, 2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "score": 1.0, + "content": "C.-K. Yeh, J. Chen, C. Yu, and D. Yu. Unsupervised speech recognition via segmental empirical", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "score": 1.0, + "content": "output distribution matching. In Proc. of ICLR, 2019.", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "N. Zeghidour, Q. Xu, V. Liptchinsky, N. Usunier, G. Synnaeve, and R. Collobert. Fully convolutional", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "score": 1.0, + "content": "speech recognition. arXiv, abs/1812.06864, 2018.", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "F. Zhang, Y. Wang, X. Zhang, C. Liu, Y. Saraf, and G. Zweig. Faster, simpler and more accurate", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "score": 1.0, + "content": "hybrid asr systems using wordpieces. Proc. of Interspeech, 2020a.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "score": 1.0, + "content": "Y. Zhang and J. R. Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "score": 1.0, + "content": "posteriorgrams. IEEE Workshop on Automatic Speech Recognition & Understanding, 2009.", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "score": 1.0, + "content": "Y. Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V. Le, and Y. Wu. Pushing the limits", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "score": 1.0, + "content": "of semi-supervised learning for automatic speech recognition. Proc. of NeurIPS SAS Workshop,", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "score": 1.0, + "content": "2020b.", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_end_line": true + } + ], + "index": 24.5, + "bbox_fs": [ + 104, + 172, + 507, + 723 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 43, + 507, + 724 + ], + "lines": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 73, + 506, + 87 + ], + "score": 1.0, + "content": "A. Conneau, G. Lample, M. Ranzato, L. Denoyer, and H. Jégou. Word translation without parallel", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "spans": [ + { + "bbox": [ + 116, + 84, + 225, + 96 + ], + "score": 1.0, + "content": "data. Proc. of ICLR, 2018.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 506, + 117 + ], + "score": 1.0, + "content": "A. Conneau, A. Baevski, R. Collobert, A. Mohamed, and M. Auli. Unsupervised cross-lingual", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "spans": [ + { + "bbox": [ + 114, + 114, + 427, + 128 + ], + "score": 1.0, + "content": "representation learning for speech recognition. arXiv, abs/2006.13979, 2020.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 133, + 506, + 147 + ], + "score": 1.0, + "content": "J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "spans": [ + { + "bbox": [ + 114, + 144, + 379, + 158 + ], + "score": 1.0, + "content": "transformers for language understanding. Proc. of NAACL, 2019.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 506, + 177 + ], + "score": 1.0, + "content": "L. Dong, S. Xu, and B. Xu. Speech-transformer: A no-recurrence sequence-to-sequence model for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "spans": [ + { + "bbox": [ + 114, + 173, + 306, + 188 + ], + "score": 1.0, + "content": "speech recognition. In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "spans": [ + { + "bbox": [ + 103, + 190, + 506, + 209 + ], + "score": 1.0, + "content": "Z. Fan, M. Li, S. Zhou, and B. Xu. Exploring wav2vec 2.0 on speaker verification and language", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 205, + 227, + 217 + ], + "score": 1.0, + "content": "identification. arXiv, 2021.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "spans": [ + { + "bbox": [ + 104, + 223, + 506, + 237 + ], + "score": 1.0, + "content": "R. Fer, P. Matejka, F. Grézl, O. Plchot, K. Vesel ˇ y, and J. H. ` Cernock ˇ y. Multilingually trained `", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "spans": [ + { + "bbox": [ + 114, + 234, + 493, + 248 + ], + "score": 1.0, + "content": "bottleneck features in spoken language recognition. Computer Speech & Language, 46, 2017.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "spans": [ + { + "bbox": [ + 104, + 253, + 506, + 267 + ], + "score": 1.0, + "content": "J. S. Garofolo, L. F. Lamel, W. M. Fisher, J. G. Fiscus, D. S. Pallett, and N. L. Dahlgren. The DARPA", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "spans": [ + { + "bbox": [ + 113, + 263, + 507, + 279 + ], + "score": 1.0, + "content": "TIMIT Acoustic-Phonetic Continuous Speech Corpus CDROM. Linguistic Data Consortium,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "spans": [ + { + "bbox": [ + 115, + 276, + 142, + 289 + ], + "score": 1.0, + "content": "1993.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "H. Gelas, L. Besacier, and F. Pellegrino. Developments of Swahili resources for an automatic speech", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 297, + 319 + ], + "score": 1.0, + "content": "recognition system. In Proc. of SLTU, 2012.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 104, + 325, + 506, + 339 + ], + "score": 1.0, + "content": "H. Gish, M. Siu, A. Chan, and W. Belfield. Unsupervised training of an hmm-based speech recognizer", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "spans": [ + { + "bbox": [ + 114, + 335, + 332, + 349 + ], + "score": 1.0, + "content": "for topic classification. In Proc. of Interspeech, 2009.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 367 + ], + "score": 1.0, + "content": "I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "spans": [ + { + "bbox": [ + 115, + 367, + 381, + 379 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial networks. Proc. of NIPS, 2014.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 399 + ], + "score": 1.0, + "content": "Google. Google cloud: Speech-to-text. https://cloud.google.com/speech-to-text, 2021.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 211, + 408 + ], + "score": 1.0, + "content": "Accessed: 2021-05-13.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 429 + ], + "score": 1.0, + "content": "A. Graves, S. Fernández, and F. Gomez. Connectionist temporal classification: Labelling unsegmented", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "spans": [ + { + "bbox": [ + 114, + 426, + 403, + 440 + ], + "score": 1.0, + "content": "sequence data with recurrent neural networks. In Proc. of ICML, 2006.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 507, + 459 + ], + "score": 1.0, + "content": "A. Gulati, J. Qin, C.-C. Chiu, N. Parmar, Y. Zhang, J. Yu, W. Han, S. Wang, Z. Zhang, Y. Wu,", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 114, + 456, + 506, + 470 + ], + "score": 1.0, + "content": "and R. Pang. Conformer: Convolution-augmented transformer for speech recognition. Proc. of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 114, + 468, + 192, + 480 + ], + "score": 1.0, + "content": "Interspeech, 2020.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 104, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of wasserstein", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "spans": [ + { + "bbox": [ + 114, + 497, + 227, + 510 + ], + "score": 1.0, + "content": "gans. Proc. of NIPS, 2017.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 507, + 530 + ], + "score": 1.0, + "content": "W. Han, Z. Zhang, Y. Zhang, J. Yu, C.-C. Chiu, J. Qin, A. Gulati, R. Pang, and Y. Wu. Contextnet:", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 115, + 527, + 506, + 541 + ], + "score": 1.0, + "content": "Improving convolutional neural networks for automatic speech recognition with global context.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "spans": [ + { + "bbox": [ + 116, + 539, + 226, + 551 + ], + "score": 1.0, + "content": "Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "D. Harwath and J. Glass. Towards visually grounded sub-word speech unit discovery. In Proc. of", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "spans": [ + { + "bbox": [ + 113, + 568, + 281, + 582 + ], + "score": 1.0, + "content": "ICASSP, pages 3017–3021. IEEE, 2019.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 586, + 506, + 602 + ], + "score": 1.0, + "content": "D. Harwath, W.-N. Hsu, and J. Glass. Learning hierarchical discrete linguistic units from visually-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 113, + 598, + 288, + 612 + ], + "score": 1.0, + "content": "grounded speech. In Proc. of ICLR, 2020.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 506, + 632 + ], + "score": 1.0, + "content": "K. Heafield. KenLM: Faster and smaller language model queries. In Proceedings of the Sixth", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "spans": [ + { + "bbox": [ + 115, + 629, + 507, + 643 + ], + "score": 1.0, + "content": "Workshop on Statistical Machine Translation, pages 187–197, Edinburgh, Scotland, July 2011.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "spans": [ + { + "bbox": [ + 113, + 639, + 293, + 653 + ], + "score": 1.0, + "content": "Association for Computational Linguistics.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "spans": [ + { + "bbox": [ + 104, + 657, + 507, + 673 + ], + "score": 1.0, + "content": "K. Hirsh-Pasek, D. G. Kemler Nelson, P. W. Jusczyk, K. W. Cassidy, B. Druss, and L. Kennedy.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "spans": [ + { + "bbox": [ + 115, + 670, + 439, + 684 + ], + "score": 1.0, + "content": "Clauses are perceptual units for young infants. Cognition, 26(3):269–286, 1987.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 689, + 507, + 703 + ], + "score": 1.0, + "content": "W.-N. Hsu, A. Sriram, A. Baevski, T. Likhomanenko, Q. Xu, V. Pratap, J. Kahn, A. Lee, R. Collobert,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "spans": [ + { + "bbox": [ + 113, + 696, + 507, + 716 + ], + "score": 1.0, + "content": "G. Synnaeve, et al. Robust wav2vec 2.0: Analyzing domain shift in self-supervised pre-training.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 284, + 725 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:2104.01027, 2021a.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "11", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 43, + 507, + 724 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 103, + 73, + 507, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 69, + 508, + 729 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 506, + 86 + ], + "score": 1.0, + "content": "W.-N. Hsu, Y.-H. H. Tsai, B. Bolte, R. Salakhutdinov, and A. Mohamed. Hubert: How much can a", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 383, + 96 + ], + "score": 1.0, + "content": "bad teacher benefit ASR pre-training? In Proc. of ICASSP, 2021b.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 506, + 116 + ], + "score": 1.0, + "content": "E. Jang, S. Gu, and B. Poole. Categorical reparameterization with gumbel-softmax. Proc. of ICLR,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "spans": [ + { + "bbox": [ + 114, + 113, + 144, + 127 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "spans": [ + { + "bbox": [ + 104, + 131, + 507, + 147 + ], + "score": 1.0, + "content": "H. Jegou, M. Douze, and C. Schmid. Product quantization for nearest neighbor search. IEEE Trans.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 335, + 157 + ], + "score": 1.0, + "content": "Pattern Anal. Mach. Intell., 33(1):117–128, Jan. 2011.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 506, + 176 + ], + "score": 1.0, + "content": "E. K. Johnson and P. W. Jusczyk. Word segmentation by 8-month-olds: When speech cues count", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 173, + 429, + 187 + ], + "score": 1.0, + "content": "more than statistics. Journal of Memory and Language, 44(4):548–567, 2001.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 506, + 206 + ], + "score": 1.0, + "content": "J. Johnson, M. Douze, and H. Jégou. Billion-scale similarity search with gpus. IEEE Transactions on", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "spans": [ + { + "bbox": [ + 114, + 203, + 184, + 216 + ], + "score": 1.0, + "content": "Big Data, 2019.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "spans": [ + { + "bbox": [ + 104, + 221, + 506, + 235 + ], + "score": 1.0, + "content": "P. W. Jusczyk, D. M. Houston, and M. Newsome. The beginnings of word segmentation in english-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "spans": [ + { + "bbox": [ + 114, + 233, + 366, + 246 + ], + "score": 1.0, + "content": "learning infants. Cognitive Psychology, 39(3):159–207, 1999.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 507, + 266 + ], + "score": 1.0, + "content": "J. Kahn, A. Lee, and A. Hannun. Self-training for end-to-end speech recognition. In Proc. of ICASSP,", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "spans": [ + { + "bbox": [ + 114, + 262, + 148, + 277 + ], + "score": 1.0, + "content": "2020a.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 104, + 280, + 506, + 295 + ], + "score": 1.0, + "content": "J. Kahn et al. Libri-light: A benchmark for asr with limited or no supervision. In Proc. of ICASSP,", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 147, + 305 + ], + "score": 1.0, + "content": "2020b.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "spans": [ + { + "bbox": [ + 104, + 309, + 507, + 326 + ], + "score": 1.0, + "content": "H. Kamper, A. Jansen, and S. Goldwater. A segmental framework for fully-unsupervised large-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "spans": [ + { + "bbox": [ + 114, + 323, + 419, + 336 + ], + "score": 1.0, + "content": "vocabulary speech recognition. Comput. Speech Lang., 46(C), Nov. 2017a.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 341, + 506, + 354 + ], + "score": 1.0, + "content": "H. Kamper, K. Livescu, and S. Goldwater. An embedded segmental k-means model for unsupervised", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "spans": [ + { + "bbox": [ + 114, + 352, + 368, + 367 + ], + "score": 1.0, + "content": "segmentation and clustering of speech. Proc. of ASRU, 2017b.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "spans": [ + { + "bbox": [ + 104, + 369, + 506, + 386 + ], + "score": 1.0, + "content": "K. Kawakami, L. Wang, C. Dyer, P. Blunsom, and A. van den Oord. Learning robust and multilingual", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "spans": [ + { + "bbox": [ + 115, + 383, + 310, + 395 + ], + "score": 1.0, + "content": "speech representations. Proc. of EMNLP, 2020.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "spans": [ + { + "bbox": [ + 104, + 401, + 492, + 415 + ], + "score": 1.0, + "content": "D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In Proc. of ICLR, 2015.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 420, + 506, + 434 + ], + "score": 1.0, + "content": "F. Kreuk, J. Keshet, and Y. Adi. Self-supervised contrastive learning for unsupervised phoneme", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "spans": [ + { + "bbox": [ + 114, + 431, + 285, + 444 + ], + "score": 1.0, + "content": "segmentation. Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 104, + 448, + 506, + 465 + ], + "score": 1.0, + "content": "G. Lample, L. Denoyer, and M. Ranzato. Unsupervised machine translation using monolingual", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "spans": [ + { + "bbox": [ + 115, + 460, + 270, + 474 + ], + "score": 1.0, + "content": "corpora only. In Proc. of ICLR, 2018.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "spans": [ + { + "bbox": [ + 104, + 478, + 508, + 495 + ], + "score": 1.0, + "content": "C. Lee and J. R. Glass. A nonparametric bayesian approach to acoustic model discovery. In Proc. of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "spans": [ + { + "bbox": [ + 114, + 492, + 164, + 503 + ], + "score": 1.0, + "content": "ACL, 2012.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 523 + ], + "score": 1.0, + "content": "C. Lee, T. J. O’Donnell, and J. R. Glass. Unsupervised lexicon discovery from acoustic input. TACL,", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "spans": [ + { + "bbox": [ + 114, + 520, + 144, + 534 + ], + "score": 1.0, + "content": "2015.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 539, + 507, + 553 + ], + "score": 1.0, + "content": "M. P. Lewis, G. F. Simon, and C. D. Fennig. Ethnologue: Languages of the world, nineteenth edition.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "spans": [ + { + "bbox": [ + 114, + 550, + 343, + 563 + ], + "score": 1.0, + "content": "Online version: http://www.ethnologue.com, 2016.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 506, + 583 + ], + "score": 1.0, + "content": "T. Likhomanenko, Q. Xu, J. Kahn, G. Synnaeve, and R. Collobert. slimipl: Language-model-free", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "spans": [ + { + "bbox": [ + 115, + 581, + 273, + 593 + ], + "score": 1.0, + "content": "iterative pseudo-labeling. arXiv, 2021.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "A. H. Liu, T. Tu, H. yi Lee, and L. shan Lee. Towards unsupervised speech recognition and synthesis", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "spans": [ + { + "bbox": [ + 114, + 610, + 401, + 623 + ], + "score": 1.0, + "content": "with quantized speech representation learning. Proc. of ICASSP, 2019.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 104, + 628, + 506, + 642 + ], + "score": 1.0, + "content": "D.-R. Liu, K.-Y. Chen, H.-Y. Lee, and L. shan Lee. Completely unsupervised phoneme recognition", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 114, + 640, + 507, + 654 + ], + "score": 1.0, + "content": "by adversarially learning mapping relationships from audio embeddings. Proc. of Interspeech,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "spans": [ + { + "bbox": [ + 114, + 650, + 144, + 666 + ], + "score": 1.0, + "content": "2018.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 506, + 684 + ], + "score": 1.0, + "content": "T. Mikolov, Q. V. Le, and I. Sutskever. Exploiting similarities among languages for machine", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "spans": [ + { + "bbox": [ + 115, + 681, + 323, + 695 + ], + "score": 1.0, + "content": "translation. arXiv preprint arXiv:1309.4168, 2013.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 104, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "M. Mohri, F. Pereira, and M. Riley. Weighted finite-state transducers in speech recognition. Computer", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "spans": [ + { + "bbox": [ + 113, + 709, + 282, + 725 + ], + "score": 1.0, + "content": "Speech & Language, 16(1):69–88, 2002.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 21.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 69, + 508, + 729 + ], + "lines": [], + "index": 21.5, + "bbox_fs": [ + 104, + 71, + 508, + 725 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 58, + 507, + 730 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 506, + 86 + ], + "score": 1.0, + "content": "L. Ondel, L. Burget, and J. Cernocký. Variational inference for acoustic unit discovery. In Proc. of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 170, + 96 + ], + "score": 1.0, + "content": "SLTU, 2016.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 507, + 116 + ], + "score": 1.0, + "content": "M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli. fairseq: A fast,", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "spans": [ + { + "bbox": [ + 115, + 113, + 487, + 127 + ], + "score": 1.0, + "content": "extensible toolkit for sequence modeling. In Proc. of NAACL System Demonstrations, 2019.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "spans": [ + { + "bbox": [ + 104, + 129, + 506, + 146 + ], + "score": 1.0, + "content": "V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "spans": [ + { + "bbox": [ + 115, + 142, + 412, + 156 + ], + "score": 1.0, + "content": "domain audio books. In Proc. of ICASSP, pages 5206–5210. IEEE, 2015.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "score": 1.0, + "content": "D. S. Park, W. Chan, Y. Zhang, C.-C. Chiu, B. Zoph, E. D. Cubuk, and Q. V. Le. Specaugment: A", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 507, + 186 + ], + "score": 1.0, + "content": "simple data augmentation method for automatic speech recognition. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 105, + 190, + 505, + 204 + ], + "score": 1.0, + "content": "D. S. Park, Y. Zhang, Y. Jia, W. Han, C.-C. Chiu, B. Li, Y. Wu, and Q. V. Le. Improved noisy student", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "spans": [ + { + "bbox": [ + 114, + 201, + 397, + 214 + ], + "score": 1.0, + "content": "training for automatic speech recognition. Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 406, + 233 + ], + "score": 1.0, + "content": "K. Park and J. Kim. g2pe. https://github.com/Kyubyong/g2p, 2019.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 103, + 235, + 507, + 253 + ], + "score": 1.0, + "content": "L. Pepino, P. Riera, and L. Ferrer. Emotion recognition from speech using wav2vec 2.0 embeddings.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "spans": [ + { + "bbox": [ + 114, + 248, + 169, + 262 + ], + "score": 1.0, + "content": "arXiv, 2021.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 280 + ], + "score": 1.0, + "content": "L. Polka and J. F. Werker. Developmental changes in perception of nonnative vowel contrasts. Journal", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "spans": [ + { + "bbox": [ + 114, + 277, + 455, + 292 + ], + "score": 1.0, + "content": "of Experimental Psychology: Human perception and performance, 20(2):421, 1994.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "spans": [ + { + "bbox": [ + 105, + 296, + 506, + 310 + ], + "score": 1.0, + "content": "D. Povey, A. Ghoshal, G. Boulianne, L. Burget, O. Glembek, N. Goel, M. Hannemann, P. Motlicek,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "spans": [ + { + "bbox": [ + 113, + 306, + 507, + 322 + ], + "score": 1.0, + "content": "Y. Qian, P. Schwarz, J. Silovsky, G. Stemmer, and K. Vesely. The kaldi speech recognition toolkit.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "spans": [ + { + "bbox": [ + 115, + 318, + 217, + 332 + ], + "score": 1.0, + "content": "In Proc. of ASRU, 2011.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 506, + 351 + ], + "score": 1.0, + "content": "V. Pratap, A. Hannun, Q. Xu, J. Cai, J. Kahn, G. Synnaeve, V. Liptchinsky, and R. Collobert.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "spans": [ + { + "bbox": [ + 114, + 347, + 473, + 363 + ], + "score": 1.0, + "content": "Wav2letter++: A fast open-source speech recognition system. In Proc. of ICASSP, 2019.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "spans": [ + { + "bbox": [ + 104, + 363, + 507, + 381 + ], + "score": 1.0, + "content": "V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert. Mls: A large-scale multilingual dataset", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "spans": [ + { + "bbox": [ + 115, + 377, + 321, + 391 + ], + "score": 1.0, + "content": "for speech research. In Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "spans": [ + { + "bbox": [ + 104, + 394, + 506, + 410 + ], + "score": 1.0, + "content": "O. Rasanen, G. Doyle, and M. C. Frank. Unsupervised word discovery from speech using automatic", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "spans": [ + { + "bbox": [ + 113, + 406, + 389, + 420 + ], + "score": 1.0, + "content": "segmentation into syllable-like units. In Proc. of Interspeech, 2015.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 507, + 439 + ], + "score": 1.0, + "content": "M. Ravanelli, P. Brakel, M. Omologo, and Y. Bengio. Light gated recurrent units for speech", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "spans": [ + { + "bbox": [ + 115, + 436, + 402, + 450 + ], + "score": 1.0, + "content": "recognition. IEEE Trans. on Emerging Topics in Comp. Intel., 2, 2018.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "spans": [ + { + "bbox": [ + 103, + 452, + 507, + 469 + ], + "score": 1.0, + "content": "M. Ravanelli, T. Parcollet, and Y. Bengio. The pytorch-kaldi speech recognition toolkit. Proc. of", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "spans": [ + { + "bbox": [ + 115, + 466, + 178, + 478 + ], + "score": 1.0, + "content": "ICASSP, 2019.", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 506, + 497 + ], + "score": 1.0, + "content": "M. Rivière, A. Joulin, P.-E. Mazaré, and E. Dupoux. Unsupervised pretraining transfers well across", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "spans": [ + { + "bbox": [ + 114, + 494, + 270, + 507 + ], + "score": 1.0, + "content": "languages. In Proc. of ICASSP, 2020.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 506, + 527 + ], + "score": 1.0, + "content": "S. Schneider, A. Baevski, R. Collobert, and M. Auli. wav2vec: Unsupervised pre-training for speech", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 115, + 524, + 289, + 538 + ], + "score": 1.0, + "content": "recognition. In Proc. of Interspeech, 2019.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 103, + 540, + 506, + 558 + ], + "score": 1.0, + "content": "N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "spans": [ + { + "bbox": [ + 113, + 553, + 349, + 567 + ], + "score": 1.0, + "content": "to prevent neural networks from overfitting. JMLR, 2014.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "spans": [ + { + "bbox": [ + 104, + 569, + 507, + 586 + ], + "score": 1.0, + "content": "G. Synnaeve, Q. Xu, J. Kahn, T. Likhomanenko, E. Grave, V. Pratap, A. Sriram, V. Liptchinsky,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 115, + 583, + 505, + 595 + ], + "score": 1.0, + "content": "and R. Collobert. End-to-end ASR: from Supervised to Semi-Supervised Learning with Modern", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "spans": [ + { + "bbox": [ + 115, + 593, + 495, + 607 + ], + "score": 1.0, + "content": "Architectures. Proc. of ICML workshop on Self-supervision in Audio and Speech (SAS), 2020.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "spans": [ + { + "bbox": [ + 103, + 609, + 507, + 627 + ], + "score": 1.0, + "content": "M. Tachbelie, S. T. Abate, and L. Besacier. Using different acoustic, lexical and language modeling", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "spans": [ + { + "bbox": [ + 115, + 623, + 478, + 637 + ], + "score": 1.0, + "content": "units for asr of an under-resourced language - amharic. Speech Communication, 56, 2014.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 104, + 640, + 505, + 655 + ], + "score": 1.0, + "content": "Z. Tan, A. K. Sarkar, and N. Dehak. rvad: An unsupervised segment-based robust voice activity", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "spans": [ + { + "bbox": [ + 114, + 652, + 378, + 664 + ], + "score": 1.0, + "content": "detection method. Computer speech & language, 59:1–21, 2020.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "spans": [ + { + "bbox": [ + 104, + 669, + 506, + 685 + ], + "score": 1.0, + "content": "A. van den Oord, Y. Li, and O. Vinyals. Representation learning with contrastive predictive coding.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "spans": [ + { + "bbox": [ + 114, + 680, + 202, + 695 + ], + "score": 1.0, + "content": "Proc. of NIPS, 2018.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 506, + 713 + ], + "score": 1.0, + "content": "B. van Niekerk, L. Nortje, and H. Kamper. Vector-quantized neural networks for acoustic unit", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 711, + 406, + 724 + ], + "score": 1.0, + "content": "discovery in the zerospeech 2020 challenge. Proc. of Interspeech, 2020.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 22 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 751 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 312, + 754 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 58, + 507, + 730 + ], + "lines": [], + "index": 22, + "bbox_fs": [ + 103, + 72, + 507, + 724 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 68, + 507, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 505, + 87 + ], + "score": 1.0, + "content": "B. Varadarajan, S. Khudanpur, and E. Dupoux. Unsupervised learning of acoustic sub-word units. In", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "spans": [ + { + "bbox": [ + 115, + 84, + 198, + 96 + ], + "score": 1.0, + "content": "Proc. of ACL, 2008.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "spans": [ + { + "bbox": [ + 106, + 102, + 506, + 115 + ], + "score": 1.0, + "content": "A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "spans": [ + { + "bbox": [ + 116, + 114, + 316, + 127 + ], + "score": 1.0, + "content": "Attention is all you need. In Proc. of NIPS, 2017.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 145 + ], + "score": 1.0, + "content": "C. Wang, A. Wu, J. Pino, A. Baevski, M. Auli, and A. Conneau. Large-scale self- and semi-supervised", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 295, + 155 + ], + "score": 1.0, + "content": "learning for speech translation. arXiv, 2021.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 104, + 161, + 505, + 176 + ], + "score": 1.0, + "content": "J. F. Werker and R. C. Tees. Cross-language speech perception: Evidence for perceptual reorganization", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "spans": [ + { + "bbox": [ + 116, + 173, + 439, + 186 + ], + "score": 1.0, + "content": "during the first year of life. Infant behavior and development, 7(1):49–63, 1984.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "H. Xu, K. Li, Y. Wang, J. Wang, S. Kang, X. Chen, D. Povey, and S. Khudanpur. Neural network", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 202, + 507, + 217 + ], + "score": 1.0, + "content": "language modeling with letter-based features and importance sampling. In Proc. of ICASSP, 2018.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 506, + 234 + ], + "score": 1.0, + "content": "Q. Xu, A. Baevski, T. Likhomanenko, P. Tomasello, A. Conneau, R. Collobert, G. Synnaeve, and", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 232, + 507, + 247 + ], + "score": 1.0, + "content": "M. Auli. Self-training and pre-training are complementary for speech recognition. In Proc. of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "spans": [ + { + "bbox": [ + 116, + 244, + 182, + 256 + ], + "score": 1.0, + "content": "ICASSP, 2020a.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 261, + 506, + 276 + ], + "score": 1.0, + "content": "Q. Xu, T. Likhomanenko, J. Kahn, A. Hannun, G. Synnaeve, and R. Collobert. Iterative pseudo-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "spans": [ + { + "bbox": [ + 115, + 273, + 362, + 286 + ], + "score": 1.0, + "content": "labeling for speech recognition. Proc. of Interspeech, 2020b.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 505, + 306 + ], + "score": 1.0, + "content": "C.-K. Yeh, J. Chen, C. Yu, and D. Yu. Unsupervised speech recognition via segmental empirical", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "spans": [ + { + "bbox": [ + 115, + 303, + 333, + 316 + ], + "score": 1.0, + "content": "output distribution matching. In Proc. of ICLR, 2019.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 505, + 335 + ], + "score": 1.0, + "content": "N. Zeghidour, Q. Xu, V. Liptchinsky, N. Usunier, G. Synnaeve, and R. Collobert. Fully convolutional", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "spans": [ + { + "bbox": [ + 115, + 333, + 318, + 345 + ], + "score": 1.0, + "content": "speech recognition. arXiv, abs/1812.06864, 2018.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 351, + 505, + 366 + ], + "score": 1.0, + "content": "F. Zhang, Y. Wang, X. Zhang, C. Liu, Y. Saraf, and G. Zweig. Faster, simpler and more accurate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "spans": [ + { + "bbox": [ + 115, + 362, + 384, + 376 + ], + "score": 1.0, + "content": "hybrid asr systems using wordpieces. Proc. of Interspeech, 2020a.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 395 + ], + "score": 1.0, + "content": "Y. Zhang and J. R. Glass. Unsupervised spoken keyword spotting via segmental dtw on gaussian", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "spans": [ + { + "bbox": [ + 115, + 392, + 485, + 406 + ], + "score": 1.0, + "content": "posteriorgrams. IEEE Workshop on Automatic Speech Recognition & Understanding, 2009.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 410, + 505, + 425 + ], + "score": 1.0, + "content": "Y. Zhang, J. Qin, D. S. Park, W. Han, C.-C. Chiu, R. Pang, Q. V. Le, and Y. Wu. Pushing the limits", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "spans": [ + { + "bbox": [ + 115, + 421, + 507, + 437 + ], + "score": 1.0, + "content": "of semi-supervised learning for automatic speech recognition. Proc. of NeurIPS SAS Workshop,", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "spans": [ + { + "bbox": [ + 115, + 433, + 147, + 445 + ], + "score": 1.0, + "content": "2020b.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 12.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 741, + 311, + 750 + ], + "lines": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "spans": [ + { + "bbox": [ + 299, + 740, + 313, + 754 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 68, + 507, + 448 + ], + "lines": [], + "index": 12.5, + "bbox_fs": [ + 104, + 72, + 507, + 445 + ], + "lines_deleted": true + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/QmxFsofRvW9/QmxFsofRvW9_model.json b/parse/train/QmxFsofRvW9/QmxFsofRvW9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..77eb6cf9a1cec28b0c92b40eb9596eb41cfa2747 --- /dev/null +++ b/parse/train/QmxFsofRvW9/QmxFsofRvW9_model.json @@ -0,0 +1,14209 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 704, + 1303, + 704, + 1303, + 1158, + 398, + 1158 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1288, + 1405, + 1288, + 1405, + 1594, + 298, + 1594 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1606, + 1404, + 1606, + 1404, + 1791, + 298, + 1791 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 1804, + 1403, + 1804, + 1403, + 1928, + 299, + 1928 + ], + "score": 0.969 + }, + { + "category_id": 0, + "poly": [ + 499, + 271, + 1200, + 271, + 1200, + 327, + 499, + 327 + ], + "score": 0.913 + }, + { + "category_id": 2, + "poly": [ + 334, + 1951, + 682, + 1951, + 682, + 1978, + 334, + 1978 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 299, + 1218, + 530, + 1218, + 530, + 1257, + 299, + 1257 + ], + "score": 0.903 + }, + { + "category_id": 2, + "poly": [ + 297, + 2033, + 1070, + 2033, + 1070, + 2062, + 297, + 2062 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 788, + 633, + 912, + 633, + 912, + 669, + 788, + 669 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 683, + 518, + 1018, + 518, + 1018, + 553, + 683, + 553 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 416, + 439, + 1275, + 439, + 1275, + 477, + 416, + 477 + ], + "score": 0.771 + }, + { + "category_id": 13, + "poly": [ + 376, + 1897, + 394, + 1897, + 394, + 1924, + 376, + 1924 + ], + "score": 0.46, + "latex": "\\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 267.0, + 1204.0, + 267.0, + 1204.0, + 333.0, + 496.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1949.0, + 684.0, + 1949.0, + 684.0, + 1980.0, + 332.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1215.0, + 536.0, + 1215.0, + 536.0, + 1263.0, + 292.0, + 1263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2028.0, + 1074.0, + 2028.0, + 1074.0, + 2066.0, + 295.0, + 2066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 630.0, + 919.0, + 630.0, + 919.0, + 674.0, + 782.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 704.0, + 1305.0, + 704.0, + 1305.0, + 739.0, + 394.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 734.0, + 1305.0, + 734.0, + 1305.0, + 769.0, + 394.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 763.0, + 1307.0, + 763.0, + 1307.0, + 801.0, + 393.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 792.0, + 1305.0, + 792.0, + 1305.0, + 829.0, + 393.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 825.0, + 1304.0, + 825.0, + 1304.0, + 860.0, + 394.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 857.0, + 1304.0, + 857.0, + 1304.0, + 891.0, + 393.0, + 891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 887.0, + 1306.0, + 887.0, + 1306.0, + 919.0, + 395.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 916.0, + 1306.0, + 916.0, + 1306.0, + 950.0, + 393.0, + 950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 948.0, + 1305.0, + 948.0, + 1305.0, + 978.0, + 393.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 978.0, + 1305.0, + 978.0, + 1305.0, + 1009.0, + 395.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1004.0, + 1305.0, + 1004.0, + 1305.0, + 1041.0, + 393.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1036.0, + 1305.0, + 1036.0, + 1305.0, + 1074.0, + 392.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1066.0, + 1306.0, + 1066.0, + 1306.0, + 1104.0, + 394.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1096.0, + 1305.0, + 1096.0, + 1305.0, + 1133.0, + 393.0, + 1133.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1130.0, + 927.0, + 1130.0, + 927.0, + 1161.0, + 393.0, + 1161.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1289.0, + 1403.0, + 1289.0, + 1403.0, + 1324.0, + 295.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1320.0, + 1406.0, + 1320.0, + 1406.0, + 1352.0, + 295.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1350.0, + 1405.0, + 1350.0, + 1405.0, + 1385.0, + 295.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1380.0, + 1405.0, + 1380.0, + 1405.0, + 1416.0, + 293.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1408.0, + 1406.0, + 1408.0, + 1406.0, + 1446.0, + 293.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1440.0, + 1405.0, + 1440.0, + 1405.0, + 1478.0, + 293.0, + 1478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1469.0, + 1406.0, + 1469.0, + 1406.0, + 1506.0, + 293.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1501.0, + 1405.0, + 1501.0, + 1405.0, + 1536.0, + 295.0, + 1536.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1531.0, + 1407.0, + 1531.0, + 1407.0, + 1567.0, + 295.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1560.0, + 1410.0, + 1560.0, + 1410.0, + 1597.0, + 292.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1608.0, + 1404.0, + 1608.0, + 1404.0, + 1640.0, + 297.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1639.0, + 1408.0, + 1639.0, + 1408.0, + 1674.0, + 294.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1668.0, + 1406.0, + 1668.0, + 1406.0, + 1704.0, + 293.0, + 1704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1700.0, + 1402.0, + 1700.0, + 1402.0, + 1732.0, + 296.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 1408.0, + 1728.0, + 1408.0, + 1763.0, + 294.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1372.0, + 1760.0, + 1372.0, + 1793.0, + 292.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1804.0, + 1409.0, + 1804.0, + 1409.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1833.0, + 1407.0, + 1833.0, + 1407.0, + 1871.0, + 292.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1865.0, + 1407.0, + 1865.0, + 1407.0, + 1900.0, + 294.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1895.0, + 375.0, + 1895.0, + 375.0, + 1931.0, + 294.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1895.0, + 1405.0, + 1895.0, + 1405.0, + 1931.0, + 395.0, + 1931.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 512.0, + 1024.0, + 512.0, + 1024.0, + 561.0, + 677.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 438.0, + 627.0, + 438.0, + 627.0, + 479.0, + 416.0, + 479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 435.0, + 1282.0, + 435.0, + 1282.0, + 482.0, + 640.0, + 482.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1203, + 1406, + 1203, + 1406, + 1419, + 298, + 1419 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 976, + 1404, + 976, + 1404, + 1191, + 297, + 1191 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1716, + 1405, + 1716, + 1405, + 1840, + 298, + 1840 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1854, + 1403, + 1854, + 1403, + 2008, + 299, + 2008 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 320, + 201, + 1373, + 201, + 1373, + 704, + 320, + 704 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 296, + 722, + 1404, + 722, + 1404, + 908, + 296, + 908 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 296, + 1548, + 1403, + 1548, + 1403, + 1611, + 296, + 1611 + ], + "score": 0.951 + }, + { + "category_id": 0, + "poly": [ + 298, + 1471, + 820, + 1471, + 820, + 1510, + 298, + 1510 + ], + "score": 0.935 + }, + { + "category_id": 0, + "poly": [ + 297, + 1655, + 1053, + 1655, + 1053, + 1690, + 297, + 1690 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2084, + 841, + 2084 + ], + "score": 0.721 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 859, + 2062, + 859, + 2084, + 841, + 2084 + ], + "score": 0.12 + }, + { + "category_id": 13, + "poly": [ + 1162, + 1886, + 1284, + 1886, + 1284, + 1916, + 1162, + 1916 + ], + "score": 0.92, + "latex": "g : { \\mathcal { Z } } \\mapsto { \\mathcal { C } }" + }, + { + "category_id": 13, + "poly": [ + 968, + 1855, + 1128, + 1855, + 1128, + 1886, + 968, + 1886 + ], + "score": 0.91, + "latex": "f : \\mathcal X \\mapsto \\mathcal Z" + }, + { + "category_id": 13, + "poly": [ + 1064, + 1980, + 1183, + 1980, + 1183, + 2008, + 1064, + 2008 + ], + "score": 0.9, + "latex": "q _ { 1 } , \\ldots , q _ { T }" + }, + { + "category_id": 13, + "poly": [ + 793, + 1890, + 912, + 1890, + 912, + 1916, + 793, + 1916 + ], + "score": 0.89, + "latex": "z _ { 1 } , \\dots , z _ { T }" + }, + { + "category_id": 13, + "poly": [ + 614, + 1917, + 731, + 1917, + 731, + 1946, + 614, + 1946 + ], + "score": 0.87, + "latex": "c _ { 1 } , \\ldots , c _ { T }" + }, + { + "category_id": 13, + "poly": [ + 1086, + 1919, + 1113, + 1919, + 1113, + 1946, + 1086, + 1946 + ], + "score": 0.84, + "latex": "z _ { t }" + }, + { + "category_id": 13, + "poly": [ + 406, + 1886, + 434, + 1886, + 434, + 1912, + 406, + 1912 + ], + "score": 0.81, + "latex": "X" + }, + { + "category_id": 13, + "poly": [ + 1306, + 1915, + 1371, + 1915, + 1371, + 1944, + 1306, + 1944 + ], + "score": 0.47, + "latex": "2 5 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 1304, + 755, + 1323, + 755, + 1323, + 782, + 1304, + 782 + ], + "score": 0.4, + "latex": "\\mathbf { k }" + }, + { + "category_id": 13, + "poly": [ + 481, + 1946, + 545, + 1946, + 545, + 1975, + 481, + 1975 + ], + "score": 0.31, + "latex": "2 0 \\mathrm { m s }" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 203.0, + 814.0, + 203.0, + 814.0, + 232.0, + 650.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 198.0, + 1119.0, + 198.0, + 1119.0, + 230.0, + 896.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 204.0, + 1362.0, + 204.0, + 1362.0, + 230.0, + 1184.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 650.0, + 226.0, + 783.0, + 226.0, + 783.0, + 252.0, + 650.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 220.0, + 1045.0, + 220.0, + 1045.0, + 250.0, + 896.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 226.0, + 1315.0, + 226.0, + 1315.0, + 252.0, + 1184.0, + 252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 251.0, + 545.0, + 251.0, + 545.0, + 281.0, + 343.0, + 281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 295.0, + 792.0, + 295.0, + 792.0, + 322.0, + 683.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 344.0, + 1339.0, + 344.0, + 1339.0, + 359.0, + 1192.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 377.0, + 1203.0, + 377.0, + 1203.0, + 407.0, + 1013.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 401.0, + 1148.0, + 401.0, + 1148.0, + 427.0, + 1013.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 436.0, + 511.0, + 436.0, + 511.0, + 462.0, + 385.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 447.0, + 1318.0, + 447.0, + 1318.0, + 477.0, + 1228.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 471.0, + 498.0, + 471.0, + 498.0, + 515.0, + 399.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 485.0, + 1162.0, + 485.0, + 1162.0, + 511.0, + 1017.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 501.0, + 499.0, + 501.0, + 499.0, + 556.0, + 395.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 498.0, + 765.0, + 498.0, + 765.0, + 524.0, + 606.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 508.0, + 1183.0, + 508.0, + 1183.0, + 534.0, + 1016.0, + 534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 527.0, + 1352.0, + 527.0, + 1352.0, + 556.0, + 1188.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 545.0, + 970.0, + 545.0, + 970.0, + 580.0, + 795.0, + 580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 597.0, + 937.0, + 597.0, + 937.0, + 624.0, + 822.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 681.0, + 928.0, + 681.0, + 928.0, + 705.0, + 832.0, + 705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 722.0, + 1404.0, + 722.0, + 1404.0, + 758.0, + 295.0, + 758.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 752.0, + 1303.0, + 752.0, + 1303.0, + 788.0, + 294.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1324.0, + 752.0, + 1406.0, + 752.0, + 1406.0, + 788.0, + 1324.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 784.0, + 1404.0, + 784.0, + 1404.0, + 820.0, + 295.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 849.0, + 293.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 844.0, + 1406.0, + 844.0, + 1406.0, + 880.0, + 294.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 872.0, + 1379.0, + 872.0, + 1379.0, + 911.0, + 293.0, + 911.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1465.0, + 824.0, + 1465.0, + 824.0, + 1520.0, + 289.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1651.0, + 1058.0, + 1651.0, + 1058.0, + 1700.0, + 291.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2093.0, + 839.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2093.0, + 838.0, + 2093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1201.0, + 1409.0, + 1201.0, + 1409.0, + 1242.0, + 292.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1236.0, + 1406.0, + 1236.0, + 1406.0, + 1270.0, + 295.0, + 1270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1265.0, + 1404.0, + 1265.0, + 1404.0, + 1300.0, + 295.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1295.0, + 1406.0, + 1295.0, + 1406.0, + 1333.0, + 292.0, + 1333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1327.0, + 1404.0, + 1327.0, + 1404.0, + 1358.0, + 295.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1354.0, + 1409.0, + 1354.0, + 1409.0, + 1394.0, + 291.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1385.0, + 951.0, + 1385.0, + 951.0, + 1423.0, + 293.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 976.0, + 1407.0, + 976.0, + 1407.0, + 1012.0, + 292.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1007.0, + 1405.0, + 1007.0, + 1405.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1038.0, + 1404.0, + 1038.0, + 1404.0, + 1072.0, + 295.0, + 1072.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1067.0, + 1406.0, + 1067.0, + 1406.0, + 1103.0, + 292.0, + 1103.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1096.0, + 1404.0, + 1096.0, + 1404.0, + 1135.0, + 292.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1128.0, + 1405.0, + 1128.0, + 1405.0, + 1165.0, + 294.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1160.0, + 878.0, + 1160.0, + 878.0, + 1194.0, + 295.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1714.0, + 1409.0, + 1714.0, + 1409.0, + 1755.0, + 292.0, + 1755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1746.0, + 1405.0, + 1746.0, + 1405.0, + 1783.0, + 293.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1778.0, + 1410.0, + 1778.0, + 1410.0, + 1814.0, + 293.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1810.0, + 735.0, + 1810.0, + 735.0, + 1842.0, + 295.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1854.0, + 967.0, + 1854.0, + 967.0, + 1887.0, + 297.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 1854.0, + 1404.0, + 1854.0, + 1404.0, + 1887.0, + 1129.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1882.0, + 405.0, + 1882.0, + 405.0, + 1922.0, + 292.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1882.0, + 792.0, + 1882.0, + 792.0, + 1922.0, + 435.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1882.0, + 1161.0, + 1882.0, + 1161.0, + 1922.0, + 913.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1285.0, + 1882.0, + 1407.0, + 1882.0, + 1407.0, + 1922.0, + 1285.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1914.0, + 613.0, + 1914.0, + 613.0, + 1948.0, + 293.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1914.0, + 1085.0, + 1914.0, + 1085.0, + 1948.0, + 732.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1114.0, + 1914.0, + 1305.0, + 1914.0, + 1305.0, + 1948.0, + 1114.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 1914.0, + 1407.0, + 1914.0, + 1407.0, + 1948.0, + 1372.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1946.0, + 480.0, + 1946.0, + 480.0, + 1979.0, + 295.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1946.0, + 1402.0, + 1946.0, + 1402.0, + 1979.0, + 546.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 1063.0, + 1975.0, + 1063.0, + 2012.0, + 294.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 1975.0, + 1405.0, + 1975.0, + 1405.0, + 2012.0, + 1184.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1544.0, + 1408.0, + 1544.0, + 1408.0, + 1587.0, + 293.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1577.0, + 1384.0, + 1577.0, + 1384.0, + 1614.0, + 295.0, + 1614.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1701, + 1405, + 1701, + 1405, + 2008, + 298, + 2008 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 908, + 1404, + 908, + 1404, + 1213, + 297, + 1213 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 647, + 1404, + 647, + 1404, + 894, + 297, + 894 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 462, + 1404, + 462, + 1404, + 615, + 298, + 615 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 300, + 1247, + 1367, + 1247, + 1367, + 1479, + 300, + 1479 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 301, + 279, + 1401, + 279, + 1401, + 371, + 301, + 371 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 295, + 203, + 1402, + 203, + 1402, + 265, + 295, + 265 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 298, + 406, + 920, + 406, + 920, + 439, + 298, + 439 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 300, + 1647, + 703, + 1647, + 703, + 1680, + 300, + 1680 + ], + "score": 0.921 + }, + { + "category_id": 4, + "poly": [ + 298, + 1497, + 1401, + 1497, + 1401, + 1591, + 298, + 1591 + ], + "score": 0.77 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2085, + 841, + 2085 + ], + "score": 0.652 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.328 + }, + { + "category_id": 4, + "poly": [ + 297, + 1497, + 1403, + 1497, + 1403, + 1592, + 297, + 1592 + ], + "score": 0.279 + }, + { + "category_id": 13, + "poly": [ + 390, + 204, + 486, + 204, + 486, + 233, + 390, + 233 + ], + "score": 0.91, + "latex": "{ \\mathcal { Z } } \\mapsto { \\mathcal { Q } }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 1885, + 1402, + 1885, + 1402, + 1913, + 1295, + 1913 + ], + "score": 0.89, + "latex": "K = 1 2 8" + }, + { + "category_id": 13, + "poly": [ + 597, + 769, + 622, + 769, + 622, + 805, + 597, + 805 + ], + "score": 0.89, + "latex": "c _ { t } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 584, + 1186, + 702, + 1186, + 702, + 1213, + 584, + 1213 + ], + "score": 0.89, + "latex": "c _ { 1 } , \\ldots , c _ { T }" + }, + { + "category_id": 13, + "poly": [ + 1215, + 998, + 1240, + 998, + 1240, + 1031, + 1215, + 1031 + ], + "score": 0.88, + "latex": "c _ { t } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 854, + 737, + 879, + 737, + 879, + 772, + 854, + 772 + ], + "score": 0.87, + "latex": "c _ { t } ^ { l }" + }, + { + "category_id": 13, + "poly": [ + 846, + 1859, + 964, + 1859, + 964, + 1886, + 846, + 1886 + ], + "score": 0.87, + "latex": "c _ { 1 } , \\ldots , c _ { T }" + }, + { + "category_id": 13, + "poly": [ + 1157, + 1090, + 1220, + 1090, + 1220, + 1119, + 1157, + 1119 + ], + "score": 0.87, + "latex": "7 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 542, + 712, + 660, + 712, + 660, + 740, + 542, + 740 + ], + "score": 0.86, + "latex": "c _ { 1 } , \\ldots , c _ { T }" + }, + { + "category_id": 13, + "poly": [ + 590, + 1090, + 632, + 1090, + 632, + 1119, + 590, + 1119 + ], + "score": 0.86, + "latex": "9 \\%" + }, + { + "category_id": 13, + "poly": [ + 828, + 1945, + 1013, + 1945, + 1013, + 1979, + 828, + 1979 + ], + "score": 0.84, + "latex": "i _ { t } \\in \\{ 1 , \\ldots , K \\}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1949, + 322, + 1949, + 322, + 1976, + 297, + 1976 + ], + "score": 0.84, + "latex": "c _ { t }" + }, + { + "category_id": 13, + "poly": [ + 646, + 741, + 658, + 741, + 658, + 766, + 646, + 766 + ], + "score": 0.69, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 669, + 710, + 741, + 710, + 741, + 740, + 669, + 740 + ], + "score": 0.66, + "latex": "( \\ S 2 . 1 )" + }, + { + "category_id": 13, + "poly": [ + 804, + 741, + 819, + 741, + 819, + 766, + 804, + 766 + ], + "score": 0.63, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 908, + 833, + 972, + 833, + 972, + 862, + 908, + 862 + ], + "score": 0.55, + "latex": "2 5 \\mathrm { m s }" + }, + { + "category_id": 13, + "poly": [ + 1144, + 1152, + 1157, + 1152, + 1157, + 1178, + 1144, + 1178 + ], + "score": 0.49, + "latex": "l" + }, + { + "category_id": 13, + "poly": [ + 1187, + 280, + 1234, + 280, + 1234, + 308, + 1187, + 308 + ], + "score": 0.38, + "latex": "5 3 \\mathrm { k }" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1242.0, + 385.0, + 1242.0, + 385.0, + 1281.0, + 325.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1294.0, + 338.0, + 1294.0, + 338.0, + 1362.0, + 299.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 338.0, + 1308.0, + 384.0, + 1308.0, + 384.0, + 1353.0, + 338.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 1386.0, + 381.0, + 1386.0, + 381.0, + 1419.0, + 355.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1407.0, + 1340.0, + 1407.0, + 1340.0, + 1443.0, + 413.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 763.0, + 1445.0, + 983.0, + 1445.0, + 983.0, + 1482.0, + 763.0, + 1482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 404.0, + 923.0, + 404.0, + 923.0, + 445.0, + 292.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1643.0, + 708.0, + 1643.0, + 708.0, + 1686.0, + 292.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1534.0, + 294.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1405.0, + 1529.0, + 1405.0, + 1563.0, + 294.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1557.0, + 763.0, + 1557.0, + 763.0, + 1594.0, + 293.0, + 1594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2058.0, + 860.0, + 2058.0, + 860.0, + 2089.0, + 839.0, + 2089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1535.0, + 295.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1529.0, + 1403.0, + 1529.0, + 1403.0, + 1563.0, + 295.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1558.0, + 763.0, + 1558.0, + 763.0, + 1595.0, + 294.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1702.0, + 1405.0, + 1702.0, + 1405.0, + 1737.0, + 296.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1730.0, + 1406.0, + 1730.0, + 1406.0, + 1772.0, + 292.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1763.0, + 1403.0, + 1763.0, + 1403.0, + 1799.0, + 293.0, + 1799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1793.0, + 1407.0, + 1793.0, + 1407.0, + 1829.0, + 293.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1821.0, + 1406.0, + 1821.0, + 1406.0, + 1858.0, + 293.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1853.0, + 845.0, + 1853.0, + 845.0, + 1893.0, + 293.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 965.0, + 1853.0, + 1406.0, + 1853.0, + 1406.0, + 1893.0, + 965.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1885.0, + 1294.0, + 1885.0, + 1294.0, + 1918.0, + 291.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1885.0, + 1406.0, + 1885.0, + 1406.0, + 1918.0, + 1403.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1914.0, + 1405.0, + 1914.0, + 1405.0, + 1950.0, + 293.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1946.0, + 296.0, + 1946.0, + 296.0, + 1982.0, + 293.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1946.0, + 827.0, + 1946.0, + 827.0, + 1982.0, + 323.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1946.0, + 1406.0, + 1946.0, + 1406.0, + 1982.0, + 1014.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1973.0, + 799.0, + 1973.0, + 799.0, + 2012.0, + 293.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 910.0, + 1405.0, + 910.0, + 1405.0, + 942.0, + 295.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 938.0, + 1406.0, + 938.0, + 1406.0, + 972.0, + 292.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 967.0, + 1406.0, + 967.0, + 1406.0, + 1004.0, + 292.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 997.0, + 1214.0, + 997.0, + 1214.0, + 1035.0, + 294.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 997.0, + 1408.0, + 997.0, + 1408.0, + 1035.0, + 1241.0, + 1035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 1407.0, + 1030.0, + 1407.0, + 1066.0, + 295.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1057.0, + 1406.0, + 1057.0, + 1406.0, + 1095.0, + 292.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 589.0, + 1087.0, + 589.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1087.0, + 1156.0, + 1087.0, + 1156.0, + 1125.0, + 633.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 1087.0, + 1406.0, + 1087.0, + 1406.0, + 1125.0, + 1221.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1120.0, + 1404.0, + 1120.0, + 1404.0, + 1154.0, + 294.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1150.0, + 1143.0, + 1150.0, + 1143.0, + 1186.0, + 295.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1150.0, + 1404.0, + 1150.0, + 1404.0, + 1186.0, + 1158.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1179.0, + 583.0, + 1179.0, + 583.0, + 1219.0, + 292.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1179.0, + 714.0, + 1179.0, + 714.0, + 1219.0, + 703.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 648.0, + 1404.0, + 648.0, + 1404.0, + 681.0, + 296.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 678.0, + 1406.0, + 678.0, + 1406.0, + 713.0, + 294.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 705.0, + 541.0, + 705.0, + 541.0, + 744.0, + 292.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 705.0, + 668.0, + 705.0, + 668.0, + 744.0, + 661.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 705.0, + 1406.0, + 705.0, + 1406.0, + 744.0, + 742.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 734.0, + 645.0, + 734.0, + 645.0, + 777.0, + 292.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 734.0, + 803.0, + 734.0, + 803.0, + 777.0, + 659.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 734.0, + 853.0, + 734.0, + 853.0, + 777.0, + 820.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 734.0, + 1405.0, + 734.0, + 1405.0, + 777.0, + 880.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 770.0, + 596.0, + 770.0, + 596.0, + 807.0, + 294.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 770.0, + 1404.0, + 770.0, + 1404.0, + 807.0, + 623.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 801.0, + 1405.0, + 801.0, + 1405.0, + 838.0, + 292.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 833.0, + 907.0, + 833.0, + 907.0, + 867.0, + 295.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 833.0, + 1406.0, + 833.0, + 1406.0, + 867.0, + 973.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 862.0, + 739.0, + 862.0, + 739.0, + 898.0, + 294.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 463.0, + 1408.0, + 463.0, + 1408.0, + 496.0, + 297.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 492.0, + 1404.0, + 492.0, + 1404.0, + 525.0, + 296.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 523.0, + 1405.0, + 523.0, + 1405.0, + 558.0, + 293.0, + 558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 554.0, + 1404.0, + 554.0, + 1404.0, + 587.0, + 296.0, + 587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 582.0, + 1018.0, + 582.0, + 1018.0, + 619.0, + 294.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 278.0, + 1186.0, + 278.0, + 1186.0, + 312.0, + 293.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1235.0, + 278.0, + 1405.0, + 278.0, + 1405.0, + 312.0, + 1235.0, + 312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 309.0, + 1405.0, + 309.0, + 1405.0, + 343.0, + 295.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 338.0, + 807.0, + 338.0, + 807.0, + 374.0, + 293.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 202.0, + 389.0, + 202.0, + 389.0, + 238.0, + 295.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 487.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 238.0, + 487.0, + 238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 233.0, + 1364.0, + 233.0, + 1364.0, + 268.0, + 297.0, + 268.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 202, + 1405, + 202, + 1405, + 477, + 297, + 477 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 689, + 1404, + 689, + 1404, + 904, + 298, + 904 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 918, + 1406, + 918, + 1406, + 1132, + 298, + 1132 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1656, + 1403, + 1656, + 1403, + 1810, + 298, + 1810 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1489, + 1403, + 1489, + 1403, + 1642, + 298, + 1642 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1260, + 1406, + 1260, + 1406, + 1383, + 298, + 1383 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1822, + 1402, + 1822, + 1402, + 2008, + 299, + 2008 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 302, + 583, + 1401, + 583, + 1401, + 676, + 302, + 676 + ], + "score": 0.966 + }, + { + "category_id": 0, + "poly": [ + 300, + 523, + 702, + 523, + 702, + 556, + 300, + 556 + ], + "score": 0.919 + }, + { + "category_id": 0, + "poly": [ + 300, + 1183, + 685, + 1183, + 685, + 1223, + 300, + 1223 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 299, + 1429, + 595, + 1429, + 595, + 1462, + 299, + 1462 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 841, + 2062, + 858, + 2062, + 858, + 2084, + 841, + 2084 + ], + "score": 0.763 + }, + { + "category_id": 13, + "poly": [ + 855, + 1687, + 1100, + 1687, + 1100, + 1721, + 855, + 1721 + ], + "score": 0.93, + "latex": "\\mathcal { G } ( S ) \\bar { = } [ p _ { 1 } , \\dotsc , p _ { M } ]" + }, + { + "category_id": 13, + "poly": [ + 592, + 843, + 687, + 843, + 687, + 874, + 592, + 874 + ], + "score": 0.92, + "latex": "p _ { m } \\in O" + }, + { + "category_id": 13, + "poly": [ + 465, + 385, + 732, + 385, + 732, + 416, + 465, + 416 + ], + "score": 0.91, + "latex": "S = s _ { 1 } , \\ldots , s _ { T } , S \\sim { \\mathcal { S } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 842, + 505, + 842, + 505, + 876, + 298, + 876 + ], + "score": 0.91, + "latex": "P = [ p _ { 1 } , \\cdots , p _ { M } ]" + }, + { + "category_id": 13, + "poly": [ + 747, + 1779, + 836, + 1779, + 836, + 1809, + 747, + 1809 + ], + "score": 0.91, + "latex": "M \\leq T" + }, + { + "category_id": 13, + "poly": [ + 921, + 1100, + 975, + 1100, + 975, + 1129, + 921, + 1129 + ], + "score": 0.87, + "latex": "2 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1058, + 1657, + 1248, + 1657, + 1248, + 1691, + 1058, + 1691 + ], + "score": 0.83, + "latex": "S = [ s _ { 1 } , \\dotsc , s _ { T } ]" + }, + { + "category_id": 13, + "poly": [ + 1330, + 1492, + 1352, + 1492, + 1352, + 1521, + 1330, + 1521 + ], + "score": 0.83, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 693, + 1689, + 726, + 1689, + 726, + 1715, + 693, + 1715 + ], + "score": 0.82, + "latex": "M" + }, + { + "category_id": 13, + "poly": [ + 1003, + 1948, + 1024, + 1948, + 1024, + 1973, + 1003, + 1973 + ], + "score": 0.82, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 739, + 844, + 764, + 844, + 764, + 870, + 739, + 870 + ], + "score": 0.82, + "latex": "O" + }, + { + "category_id": 13, + "poly": [ + 762, + 1659, + 786, + 1659, + 786, + 1685, + 762, + 1685 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 687, + 1719, + 712, + 1719, + 712, + 1746, + 687, + 1746 + ], + "score": 0.8, + "latex": "O" + }, + { + "category_id": 13, + "poly": [ + 428, + 1659, + 450, + 1659, + 450, + 1688, + 428, + 1688 + ], + "score": 0.8, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 642, + 1522, + 662, + 1522, + 662, + 1549, + 642, + 1549 + ], + "score": 0.79, + "latex": "\\mathcal { C }" + }, + { + "category_id": 13, + "poly": [ + 497, + 1825, + 522, + 1825, + 522, + 1852, + 497, + 1852 + ], + "score": 0.74, + "latex": "O" + }, + { + "category_id": 13, + "poly": [ + 1029, + 812, + 1056, + 812, + 1056, + 839, + 1029, + 839 + ], + "score": 0.73, + "latex": "Y" + }, + { + "category_id": 13, + "poly": [ + 722, + 873, + 784, + 873, + 784, + 903, + 722, + 903 + ], + "score": 0.73, + "latex": "\\ S \\operatorname { E } . 2" + }, + { + "category_id": 13, + "poly": [ + 1332, + 979, + 1393, + 979, + 1393, + 1010, + 1332, + 1010 + ], + "score": 0.7, + "latex": "\\ S \\ 3 . 1 \\AA ." + }, + { + "category_id": 13, + "poly": [ + 678, + 1041, + 729, + 1041, + 729, + 1071, + 678, + 1071 + ], + "score": 0.42, + "latex": "( \\ S 3 )" + }, + { + "category_id": 13, + "poly": [ + 576, + 752, + 618, + 752, + 618, + 781, + 576, + 781 + ], + "score": 0.31, + "latex": "/ \\mathrm { T } / ." + }, + { + "category_id": 15, + "poly": [ + 294.0, + 521.0, + 708.0, + 521.0, + 708.0, + 563.0, + 294.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1177.0, + 690.0, + 1177.0, + 690.0, + 1235.0, + 288.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1425.0, + 599.0, + 1425.0, + 599.0, + 1466.0, + 293.0, + 1466.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2059.0, + 862.0, + 2059.0, + 862.0, + 2091.0, + 838.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 203.0, + 1407.0, + 203.0, + 1407.0, + 239.0, + 296.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 233.0, + 1406.0, + 233.0, + 1406.0, + 268.0, + 294.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 264.0, + 1406.0, + 264.0, + 1406.0, + 300.0, + 294.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 1406.0, + 291.0, + 1406.0, + 330.0, + 294.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 324.0, + 1407.0, + 324.0, + 1407.0, + 360.0, + 295.0, + 360.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 355.0, + 1407.0, + 355.0, + 1407.0, + 392.0, + 295.0, + 392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 384.0, + 464.0, + 384.0, + 464.0, + 420.0, + 292.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 384.0, + 1407.0, + 384.0, + 1407.0, + 420.0, + 733.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 412.0, + 1406.0, + 412.0, + 1406.0, + 455.0, + 292.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 445.0, + 709.0, + 445.0, + 709.0, + 481.0, + 295.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 692.0, + 1402.0, + 692.0, + 1402.0, + 723.0, + 296.0, + 723.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 723.0, + 1404.0, + 723.0, + 1404.0, + 753.0, + 296.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 750.0, + 575.0, + 750.0, + 575.0, + 785.0, + 293.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 619.0, + 750.0, + 1408.0, + 750.0, + 1408.0, + 785.0, + 619.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 782.0, + 1406.0, + 782.0, + 1406.0, + 815.0, + 293.0, + 815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 810.0, + 1028.0, + 810.0, + 1028.0, + 844.0, + 294.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 810.0, + 1404.0, + 810.0, + 1404.0, + 844.0, + 1057.0, + 844.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 840.0, + 297.0, + 840.0, + 297.0, + 878.0, + 293.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 840.0, + 591.0, + 840.0, + 591.0, + 878.0, + 506.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 840.0, + 738.0, + 840.0, + 738.0, + 878.0, + 688.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 765.0, + 840.0, + 1406.0, + 840.0, + 1406.0, + 878.0, + 765.0, + 878.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 873.0, + 721.0, + 873.0, + 721.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 873.0, + 796.0, + 873.0, + 796.0, + 907.0, + 785.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 916.0, + 1409.0, + 916.0, + 1409.0, + 955.0, + 292.0, + 955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 949.0, + 1408.0, + 949.0, + 1408.0, + 983.0, + 295.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 978.0, + 1331.0, + 978.0, + 1331.0, + 1013.0, + 295.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 978.0, + 1408.0, + 978.0, + 1408.0, + 1013.0, + 1394.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1007.0, + 1407.0, + 1007.0, + 1407.0, + 1045.0, + 293.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1037.0, + 677.0, + 1037.0, + 677.0, + 1075.0, + 295.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 1037.0, + 1406.0, + 1037.0, + 1406.0, + 1075.0, + 730.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1070.0, + 1406.0, + 1070.0, + 1406.0, + 1105.0, + 295.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1100.0, + 920.0, + 1100.0, + 920.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 976.0, + 1100.0, + 1402.0, + 1100.0, + 1402.0, + 1135.0, + 976.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1654.0, + 427.0, + 1654.0, + 427.0, + 1693.0, + 293.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 451.0, + 1654.0, + 761.0, + 1654.0, + 761.0, + 1693.0, + 451.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 787.0, + 1654.0, + 1057.0, + 1654.0, + 1057.0, + 1693.0, + 787.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1249.0, + 1654.0, + 1407.0, + 1654.0, + 1407.0, + 1693.0, + 1249.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1685.0, + 692.0, + 1685.0, + 692.0, + 1722.0, + 294.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1685.0, + 854.0, + 1685.0, + 854.0, + 1722.0, + 727.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1722.0, + 1101.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1714.0, + 686.0, + 1714.0, + 686.0, + 1753.0, + 293.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1714.0, + 1405.0, + 1714.0, + 1405.0, + 1753.0, + 713.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1746.0, + 1407.0, + 1746.0, + 1407.0, + 1786.0, + 292.0, + 1786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1777.0, + 746.0, + 1777.0, + 746.0, + 1815.0, + 292.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1777.0, + 848.0, + 1777.0, + 848.0, + 1815.0, + 837.0, + 1815.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1489.0, + 1329.0, + 1489.0, + 1329.0, + 1526.0, + 296.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 1489.0, + 1405.0, + 1489.0, + 1405.0, + 1526.0, + 1353.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1517.0, + 641.0, + 1517.0, + 641.0, + 1558.0, + 292.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 663.0, + 1517.0, + 1405.0, + 1517.0, + 1405.0, + 1558.0, + 663.0, + 1558.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1548.0, + 1407.0, + 1548.0, + 1407.0, + 1588.0, + 293.0, + 1588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1579.0, + 1405.0, + 1579.0, + 1405.0, + 1618.0, + 293.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1610.0, + 729.0, + 1610.0, + 729.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1258.0, + 1409.0, + 1258.0, + 1409.0, + 1296.0, + 292.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1291.0, + 1408.0, + 1291.0, + 1408.0, + 1324.0, + 296.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1322.0, + 1406.0, + 1322.0, + 1406.0, + 1358.0, + 293.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1353.0, + 915.0, + 1353.0, + 915.0, + 1385.0, + 295.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1822.0, + 496.0, + 1822.0, + 496.0, + 1859.0, + 294.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 523.0, + 1822.0, + 1406.0, + 1822.0, + 1406.0, + 1859.0, + 523.0, + 1859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1852.0, + 1407.0, + 1852.0, + 1407.0, + 1891.0, + 293.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1878.0, + 1407.0, + 1878.0, + 1407.0, + 1924.0, + 292.0, + 1924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1915.0, + 1406.0, + 1915.0, + 1406.0, + 1951.0, + 294.0, + 1951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1944.0, + 1002.0, + 1944.0, + 1002.0, + 1983.0, + 292.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1983.0, + 1025.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1976.0, + 1204.0, + 1976.0, + 1204.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 584.0, + 1404.0, + 584.0, + 1404.0, + 617.0, + 297.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 614.0, + 1405.0, + 614.0, + 1405.0, + 649.0, + 294.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 648.0, + 839.0, + 648.0, + 839.0, + 678.0, + 298.0, + 678.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 587, + 1405, + 587, + 1405, + 804, + 296, + 804 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 201, + 1407, + 201, + 1407, + 356, + 297, + 356 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 296, + 1045, + 1404, + 1045, + 1404, + 1173, + 296, + 1173 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 833, + 1405, + 833, + 1405, + 931, + 297, + 931 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 1567, + 1406, + 1567, + 1406, + 1662, + 297, + 1662 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1673, + 1405, + 1673, + 1405, + 1829, + 297, + 1829 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1291, + 1404, + 1291, + 1404, + 1386, + 297, + 1386 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1840, + 1406, + 1840, + 1406, + 1958, + 298, + 1958 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 687, + 1400, + 1012, + 1400, + 1012, + 1479, + 687, + 1479 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 654, + 1187, + 1044, + 1187, + 1044, + 1266, + 654, + 1266 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 300, + 445, + 1405, + 445, + 1405, + 510, + 300, + 510 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 663, + 944, + 1034, + 944, + 1034, + 1022, + 663, + 1022 + ], + "score": 0.947 + }, + { + "category_id": 0, + "poly": [ + 299, + 1512, + 817, + 1512, + 817, + 1546, + 299, + 1546 + ], + "score": 0.929 + }, + { + "category_id": 8, + "poly": [ + 382, + 520, + 1316, + 520, + 1316, + 574, + 382, + 574 + ], + "score": 0.924 + }, + { + "category_id": 0, + "poly": [ + 298, + 390, + 478, + 390, + 478, + 423, + 298, + 423 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 331, + 1977, + 901, + 1977, + 901, + 2008, + 331, + 2008 + ], + "score": 0.912 + }, + { + "category_id": 9, + "poly": [ + 1366, + 966, + 1400, + 966, + 1400, + 996, + 1366, + 996 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1197, + 1400, + 1197, + 1400, + 1227, + 1366, + 1227 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1417, + 1400, + 1417, + 1400, + 1447, + 1366, + 1447 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1366, + 527, + 1400, + 527, + 1400, + 557, + 1366, + 557 + ], + "score": 0.861 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.742 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 859, + 2061, + 859, + 2085, + 841, + 2085 + ], + "score": 0.131 + }, + { + "category_id": 14, + "poly": [ + 685, + 1397, + 1015, + 1397, + 1015, + 1481, + 685, + 1481 + ], + "score": 0.95, + "latex": "\\mathcal { L } _ { p d } = \\frac { 1 } { | B | } \\sum _ { S \\in B } - H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )" + }, + { + "category_id": 14, + "poly": [ + 661, + 943, + 1037, + 943, + 1037, + 1021, + 661, + 1021 + ], + "score": 0.95, + "latex": "\\mathcal { L } _ { g p } = \\underset { \\tilde { P } \\sim \\tilde { \\mathcal { P } } } { \\mathbb { E } } \\left[ \\left( \\| \\nabla \\mathcal { C } ( \\tilde { P } ) \\| - 1 \\right) ^ { 2 } \\right]" + }, + { + "category_id": 13, + "poly": [ + 842, + 589, + 904, + 589, + 904, + 622, + 842, + 622 + ], + "score": 0.94, + "latex": "\\mathcal { G } ( S )" + }, + { + "category_id": 14, + "poly": [ + 653, + 1184, + 1046, + 1184, + 1046, + 1266, + 653, + 1266 + ], + "score": 0.94, + "latex": "\\mathcal { L } _ { s p } = \\sum _ { ( p _ { t } , p _ { t + 1 } ) \\in \\mathcal { G } ( S ) } \\| p _ { t } - p _ { t + 1 } \\| ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 1316, + 864, + 1401, + 864, + 1401, + 897, + 1316, + 897 + ], + "score": 0.93, + "latex": "\\tilde { P } \\sim \\tilde { \\mathcal { P } }" + }, + { + "category_id": 13, + "poly": [ + 429, + 1353, + 549, + 1353, + 549, + 1387, + 429, + 1387 + ], + "score": 0.93, + "latex": "H _ { \\mathcal { G } } ( \\mathcal { G } ( S ) )" + }, + { + "category_id": 13, + "poly": [ + 620, + 1138, + 731, + 1138, + 731, + 1173, + 620, + 1173 + ], + "score": 0.93, + "latex": "p _ { t } \\in \\mathbb { R } ^ { | O | }" + }, + { + "category_id": 13, + "poly": [ + 364, + 1871, + 467, + 1871, + 467, + 1914, + 364, + 1914 + ], + "score": 0.93, + "latex": "\\{ X _ { j } \\} _ { j = 1 } ^ { N _ { s } }" + }, + { + "category_id": 13, + "poly": [ + 499, + 1871, + 657, + 1871, + 657, + 1914, + 499, + 1914 + ], + "score": 0.93, + "latex": "\\mathbf { \\bar { \\mathcal { P } } } = \\{ P _ { j } \\} _ { j = 1 } ^ { N _ { s } }" + }, + { + "category_id": 13, + "poly": [ + 1338, + 233, + 1397, + 233, + 1397, + 267, + 1338, + 267 + ], + "score": 0.92, + "latex": "\\mathcal { G } ( S )" + }, + { + "category_id": 13, + "poly": [ + 951, + 1910, + 1405, + 1910, + 1405, + 1954, + 951, + 1954 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { N L L _ { L M } ( P ) = - \\frac { 1 } { M } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 544, + 264, + 583, + 264, + 583, + 297, + 544, + 297 + ], + "score": 0.91, + "latex": "| O |" + }, + { + "category_id": 13, + "poly": [ + 374, + 590, + 487, + 590, + 487, + 618, + 374, + 618 + ], + "score": 0.91, + "latex": "P ^ { r } \\sim \\mathcal { P } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 846, + 204, + 953, + 204, + 953, + 232, + 846, + 232 + ], + "score": 0.91, + "latex": "P ^ { r } \\sim \\mathcal { P } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 573, + 712, + 620, + 712, + 620, + 744, + 573, + 744 + ], + "score": 0.91, + "latex": "\\mathcal { L } _ { p d }" + }, + { + "category_id": 14, + "poly": [ + 378, + 521, + 1318, + 521, + 1318, + 575, + 378, + 575 + ], + "score": 0.91, + "latex": "\\operatorname* { m i n } _ { \\mathcal { G } } \\operatorname* { m a x } _ { \\mathcal { C } } \\quad \\mathbb { E } _ { \\mathcal { P } ^ { r } \\sim \\mathcal { P } ^ { r } } \\left[ \\log \\mathcal { C } ( \\boldsymbol { P } ^ { r } ) \\right] - \\underset { S \\sim \\mathcal { S } } { \\mathbb { E } } \\left[ \\log \\left( 1 - \\mathcal { C } ( \\mathcal { G } ( S ) ) \\right) \\right] - \\lambda \\mathcal { L } _ { g p } + \\gamma \\mathcal { L } _ { s p } + \\eta \\mathcal { L } _ { p d }" + }, + { + "category_id": 13, + "poly": [ + 1001, + 681, + 1048, + 681, + 1048, + 715, + 1001, + 715 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { g p }" + }, + { + "category_id": 13, + "poly": [ + 1305, + 681, + 1350, + 681, + 1350, + 714, + 1305, + 714 + ], + "score": 0.9, + "latex": "\\mathcal { L } _ { s p }" + }, + { + "category_id": 13, + "poly": [ + 421, + 1912, + 873, + 1912, + 873, + 1956, + 421, + 1956 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { N L L _ { L M } ( \\mathcal { P } ) = \\frac { 1 } { N _ { s } } \\sum _ { j = 1 } ^ { N _ { s } } N L L _ { L M } ( P _ { j } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 918, + 1631, + 1013, + 1631, + 1013, + 1663, + 918, + 1663 + ], + "score": 0.89, + "latex": "( \\lambda , \\gamma , \\eta )" + }, + { + "category_id": 13, + "poly": [ + 376, + 1739, + 434, + 1739, + 434, + 1768, + 376, + 1768 + ], + "score": 0.87, + "latex": "p _ { L M }" + }, + { + "category_id": 13, + "poly": [ + 679, + 235, + 715, + 235, + 715, + 262, + 679, + 262 + ], + "score": 0.85, + "latex": "{ \\mathcal { P } } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 1141, + 1843, + 1163, + 1843, + 1163, + 1872, + 1141, + 1872 + ], + "score": 0.84, + "latex": "\\mathcal { G }" + }, + { + "category_id": 13, + "poly": [ + 1214, + 1355, + 1239, + 1355, + 1239, + 1381, + 1214, + 1381 + ], + "score": 0.83, + "latex": "B" + }, + { + "category_id": 13, + "poly": [ + 822, + 1737, + 900, + 1737, + 900, + 1767, + 822, + 1767 + ], + "score": 0.81, + "latex": "( \\ S ~ 2 . 4 )" + }, + { + "category_id": 13, + "poly": [ + 684, + 621, + 706, + 621, + 706, + 648, + 684, + 648 + ], + "score": 0.81, + "latex": "S" + }, + { + "category_id": 13, + "poly": [ + 720, + 1050, + 739, + 1050, + 739, + 1075, + 720, + 1075 + ], + "score": 0.65, + "latex": "\\mathbf { k }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1509.0, + 821.0, + 1509.0, + 821.0, + 1549.0, + 292.0, + 1549.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 383.0, + 482.0, + 383.0, + 482.0, + 433.0, + 290.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1971.0, + 905.0, + 1971.0, + 905.0, + 2014.0, + 330.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2092.0, + 838.0, + 2092.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 585.0, + 373.0, + 585.0, + 373.0, + 626.0, + 294.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 585.0, + 841.0, + 585.0, + 841.0, + 626.0, + 488.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 905.0, + 585.0, + 1405.0, + 585.0, + 1405.0, + 626.0, + 905.0, + 626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 619.0, + 683.0, + 619.0, + 683.0, + 653.0, + 295.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 619.0, + 1403.0, + 619.0, + 1403.0, + 653.0, + 707.0, + 653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 650.0, + 1405.0, + 650.0, + 1405.0, + 684.0, + 295.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 678.0, + 1000.0, + 678.0, + 1000.0, + 719.0, + 294.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 678.0, + 1304.0, + 678.0, + 1304.0, + 719.0, + 1049.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1351.0, + 678.0, + 1408.0, + 678.0, + 1408.0, + 719.0, + 1351.0, + 719.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 709.0, + 572.0, + 709.0, + 572.0, + 748.0, + 294.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 621.0, + 709.0, + 1406.0, + 709.0, + 1406.0, + 748.0, + 621.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 739.0, + 1407.0, + 739.0, + 1407.0, + 776.0, + 293.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 771.0, + 1027.0, + 771.0, + 1027.0, + 804.0, + 294.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 201.0, + 845.0, + 201.0, + 845.0, + 239.0, + 295.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 201.0, + 1407.0, + 201.0, + 1407.0, + 239.0, + 954.0, + 239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 232.0, + 678.0, + 232.0, + 678.0, + 269.0, + 294.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 232.0, + 1337.0, + 232.0, + 1337.0, + 269.0, + 716.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 232.0, + 1408.0, + 232.0, + 1408.0, + 269.0, + 1398.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 543.0, + 261.0, + 543.0, + 300.0, + 294.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 261.0, + 1409.0, + 261.0, + 1409.0, + 300.0, + 584.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 332.0, + 292.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 322.0, + 592.0, + 322.0, + 592.0, + 358.0, + 294.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1045.0, + 719.0, + 1045.0, + 719.0, + 1083.0, + 294.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 1045.0, + 1406.0, + 1045.0, + 1406.0, + 1083.0, + 740.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1077.0, + 1402.0, + 1077.0, + 1402.0, + 1110.0, + 296.0, + 1110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1107.0, + 1404.0, + 1107.0, + 1404.0, + 1145.0, + 295.0, + 1145.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1140.0, + 619.0, + 1140.0, + 619.0, + 1175.0, + 291.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1140.0, + 745.0, + 1140.0, + 745.0, + 1175.0, + 732.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 830.0, + 1406.0, + 830.0, + 1406.0, + 872.0, + 292.0, + 872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 867.0, + 1315.0, + 867.0, + 1315.0, + 902.0, + 292.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 867.0, + 1406.0, + 867.0, + 1406.0, + 902.0, + 1402.0, + 902.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 897.0, + 1226.0, + 897.0, + 1226.0, + 934.0, + 292.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1568.0, + 1406.0, + 1568.0, + 1406.0, + 1602.0, + 296.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1594.0, + 1409.0, + 1594.0, + 1409.0, + 1640.0, + 292.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1627.0, + 917.0, + 1627.0, + 917.0, + 1667.0, + 292.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1014.0, + 1627.0, + 1025.0, + 1627.0, + 1025.0, + 1667.0, + 1014.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1671.0, + 1408.0, + 1671.0, + 1408.0, + 1711.0, + 294.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1703.0, + 1405.0, + 1703.0, + 1405.0, + 1743.0, + 292.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1735.0, + 375.0, + 1735.0, + 375.0, + 1772.0, + 294.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1735.0, + 821.0, + 1735.0, + 821.0, + 1772.0, + 435.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 1735.0, + 1405.0, + 1735.0, + 1405.0, + 1772.0, + 901.0, + 1772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1406.0, + 1760.0, + 1406.0, + 1806.0, + 292.0, + 1806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1795.0, + 1119.0, + 1795.0, + 1119.0, + 1834.0, + 294.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1290.0, + 1405.0, + 1290.0, + 1405.0, + 1330.0, + 292.0, + 1330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1324.0, + 1404.0, + 1324.0, + 1404.0, + 1358.0, + 295.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1351.0, + 428.0, + 1351.0, + 428.0, + 1391.0, + 294.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1351.0, + 1213.0, + 1351.0, + 1213.0, + 1391.0, + 550.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1240.0, + 1351.0, + 1399.0, + 1351.0, + 1399.0, + 1391.0, + 1240.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1840.0, + 363.0, + 1840.0, + 363.0, + 1930.0, + 283.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 1840.0, + 498.0, + 1840.0, + 498.0, + 1930.0, + 468.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 1840.0, + 1140.0, + 1840.0, + 1140.0, + 1930.0, + 658.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 1840.0, + 1420.0, + 1840.0, + 1420.0, + 1930.0, + 1164.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1915.0, + 420.0, + 1915.0, + 420.0, + 1955.0, + 296.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 1907.0, + 950.0, + 1907.0, + 950.0, + 1962.0, + 874.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1406.0, + 1907.0, + 1411.0, + 1907.0, + 1411.0, + 1962.0, + 1406.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 444.0, + 1409.0, + 444.0, + 1409.0, + 483.0, + 294.0, + 483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 477.0, + 1265.0, + 477.0, + 1265.0, + 514.0, + 296.0, + 514.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1258, + 1405, + 1258, + 1405, + 1594, + 297, + 1594 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 547, + 1405, + 547, + 1405, + 743, + 297, + 743 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 986, + 1406, + 986, + 1406, + 1079, + 299, + 1079 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1607, + 1406, + 1607, + 1406, + 1911, + 297, + 1911 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 283, + 1405, + 283, + 1405, + 418, + 299, + 418 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 755, + 1404, + 755, + 1404, + 849, + 299, + 849 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 531, + 439, + 1169, + 439, + 1169, + 530, + 531, + 530 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 468, + 871, + 1231, + 871, + 1231, + 966, + 468, + 966 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 291, + 200, + 1401, + 200, + 1401, + 267, + 291, + 267 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 328, + 1947, + 1304, + 1947, + 1304, + 2008, + 328, + 2008 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 298, + 1199, + 1098, + 1199, + 1098, + 1234, + 298, + 1234 + ], + "score": 0.934 + }, + { + "category_id": 0, + "poly": [ + 298, + 1128, + 455, + 1128, + 455, + 1165, + 298, + 1165 + ], + "score": 0.921 + }, + { + "category_id": 9, + "poly": [ + 1366, + 469, + 1400, + 469, + 1400, + 499, + 1366, + 499 + ], + "score": 0.886 + }, + { + "category_id": 9, + "poly": [ + 1366, + 903, + 1400, + 903, + 1400, + 933, + 1366, + 933 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 859, + 2062, + 859, + 2085, + 840, + 2085 + ], + "score": 0.803 + }, + { + "category_id": 14, + "poly": [ + 470, + 867, + 1231, + 867, + 1231, + 969, + 470, + 969 + ], + "score": 0.94, + "latex": "\\mathcal { P } ^ { * } = \\arg \\operatorname* { m a x } _ { \\mathcal { P } ^ { \\prime } } \\sum _ { j = 1 } ^ { N _ { s } } \\sum _ { t = 1 } ^ { M } \\log p _ { L M } ( p _ { t } ^ { j } ) , M = | P ^ { j } | , P ^ { j } = [ p _ { 1 } ^ { j } , \\dots , p _ { M } ^ { j } ]" + }, + { + "category_id": 14, + "poly": [ + 531, + 436, + 1168, + 436, + 1168, + 530, + 531, + 530 + ], + "score": 0.93, + "latex": "N L L _ { L M } ( \\mathcal { P } ) < N L L _ { L M } ( \\hat { \\mathcal { P } } ) + \\log \\left( \\frac { U ( \\mathcal { P } ) } { U ( \\hat { \\mathcal { P } } ) } \\right) + \\log 1 . 2" + }, + { + "category_id": 13, + "poly": [ + 298, + 580, + 446, + 580, + 446, + 618, + 298, + 618 + ], + "score": 0.93, + "latex": "N L L _ { L M } ( \\hat { \\mathcal { P } } )" + }, + { + "category_id": 13, + "poly": [ + 956, + 580, + 1023, + 580, + 1023, + 618, + 956, + 618 + ], + "score": 0.93, + "latex": "U ( \\hat { \\mathcal { P } } )" + }, + { + "category_id": 13, + "poly": [ + 478, + 645, + 545, + 645, + 545, + 682, + 478, + 682 + ], + "score": 0.93, + "latex": "U ( \\hat { \\mathcal { P } } )" + }, + { + "category_id": 13, + "poly": [ + 1331, + 589, + 1396, + 589, + 1396, + 618, + 1331, + 618 + ], + "score": 0.93, + "latex": "U ( \\mathcal { P } )" + }, + { + "category_id": 13, + "poly": [ + 771, + 649, + 838, + 649, + 838, + 683, + 771, + 683 + ], + "score": 0.92, + "latex": "U ( \\mathcal { P } )" + }, + { + "category_id": 13, + "poly": [ + 369, + 204, + 471, + 204, + 471, + 236, + 369, + 236 + ], + "score": 0.9, + "latex": "p _ { L M } ( p _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 299, + 233, + 584, + 233, + 584, + 275, + 299, + 275 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\frac { 1 } { | O | } \\bar { \\sum _ { o \\in O } } [ o \\in \\mathcal { P } ] \\in [ 0 , 1 ] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 682, + 203, + 930, + 203, + 930, + 237, + 682, + 237 + ], + "score": 0.9, + "latex": "p _ { L M } ( p _ { t } | p _ { t - 1 } , . . . , p _ { 1 } )" + }, + { + "category_id": 13, + "poly": [ + 1295, + 203, + 1403, + 203, + 1403, + 238, + 1295, + 238 + ], + "score": 0.9, + "latex": "U ( { \\mathcal { P } } ) =" + }, + { + "category_id": 13, + "poly": [ + 749, + 382, + 773, + 382, + 773, + 414, + 749, + 414 + ], + "score": 0.85, + "latex": "\\hat { \\mathcal { P } }" + }, + { + "category_id": 13, + "poly": [ + 968, + 235, + 991, + 235, + 991, + 261, + 968, + 261 + ], + "score": 0.84, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 487, + 346, + 946, + 346, + 946, + 382, + 487, + 382 + ], + "score": 0.84, + "latex": "\\begin{array} { r } { \\hat { \\mathcal { P } } = \\arg \\operatorname* { m i n } _ { \\mathcal { P } } N L L _ { L M } ( \\mathcal { P } ) - \\log U ( \\mathcal { P } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 892, + 580, + 916, + 580, + 916, + 613, + 892, + 613 + ], + "score": 0.84, + "latex": "\\hat { \\mathcal { P } }" + }, + { + "category_id": 13, + "poly": [ + 1369, + 759, + 1400, + 759, + 1400, + 784, + 1369, + 784 + ], + "score": 0.83, + "latex": "{ \\mathcal { P } } ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 297, + 644, + 322, + 644, + 322, + 678, + 297, + 678 + ], + "score": 0.82, + "latex": "\\hat { \\mathcal { P } }" + }, + { + "category_id": 13, + "poly": [ + 818, + 585, + 843, + 585, + 843, + 613, + 818, + 613 + ], + "score": 0.82, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 810, + 1442, + 876, + 1442, + 876, + 1471, + 810, + 1471 + ], + "score": 0.81, + "latex": "5 3 . 2 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 625, + 1953, + 646, + 1953, + 646, + 1973, + 625, + 1973 + ], + "score": 0.8, + "latex": "\\mathcal { P }" + }, + { + "category_id": 13, + "poly": [ + 1181, + 1952, + 1233, + 1952, + 1233, + 1977, + 1181, + 1977 + ], + "score": 0.79, + "latex": "p _ { L M }" + }, + { + "category_id": 13, + "poly": [ + 823, + 1609, + 910, + 1609, + 910, + 1639, + 823, + 1639 + ], + "score": 0.7, + "latex": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ," + }, + { + "category_id": 13, + "poly": [ + 480, + 683, + 560, + 683, + 560, + 712, + 480, + 712 + ], + "score": 0.62, + "latex": "\\log 1 . 2" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1938.0, + 624.0, + 1938.0, + 624.0, + 1985.0, + 329.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 1938.0, + 1180.0, + 1938.0, + 1180.0, + 1985.0, + 647.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1234.0, + 1938.0, + 1245.0, + 1938.0, + 1245.0, + 1985.0, + 1234.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1972.0, + 1300.0, + 1972.0, + 1300.0, + 2014.0, + 331.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1195.0, + 1104.0, + 1195.0, + 1104.0, + 1244.0, + 290.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1124.0, + 459.0, + 1124.0, + 459.0, + 1171.0, + 292.0, + 1171.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2060.0, + 862.0, + 2060.0, + 862.0, + 2091.0, + 840.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1260.0, + 1402.0, + 1260.0, + 1402.0, + 1295.0, + 295.0, + 1295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1290.0, + 1405.0, + 1290.0, + 1405.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1321.0, + 1406.0, + 1321.0, + 1406.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1385.0, + 295.0, + 1385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1380.0, + 1405.0, + 1380.0, + 1405.0, + 1415.0, + 295.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1410.0, + 1406.0, + 1410.0, + 1406.0, + 1447.0, + 292.0, + 1447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1440.0, + 809.0, + 1440.0, + 809.0, + 1476.0, + 294.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1440.0, + 1405.0, + 1440.0, + 1405.0, + 1476.0, + 877.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1472.0, + 1405.0, + 1472.0, + 1405.0, + 1505.0, + 294.0, + 1505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1503.0, + 1405.0, + 1503.0, + 1405.0, + 1537.0, + 295.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 1405.0, + 1530.0, + 1405.0, + 1570.0, + 294.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1561.0, + 438.0, + 1561.0, + 438.0, + 1596.0, + 294.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 546.0, + 1406.0, + 546.0, + 1406.0, + 584.0, + 294.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 579.0, + 297.0, + 579.0, + 297.0, + 621.0, + 294.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 579.0, + 817.0, + 579.0, + 817.0, + 621.0, + 447.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 579.0, + 891.0, + 579.0, + 891.0, + 621.0, + 844.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 579.0, + 955.0, + 579.0, + 955.0, + 621.0, + 917.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 579.0, + 1330.0, + 579.0, + 1330.0, + 621.0, + 1024.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1397.0, + 579.0, + 1408.0, + 579.0, + 1408.0, + 621.0, + 1397.0, + 621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 613.0, + 1406.0, + 613.0, + 1406.0, + 650.0, + 295.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 645.0, + 477.0, + 645.0, + 477.0, + 686.0, + 323.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 645.0, + 770.0, + 645.0, + 770.0, + 686.0, + 546.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 645.0, + 1405.0, + 645.0, + 1405.0, + 686.0, + 839.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 678.0, + 479.0, + 678.0, + 479.0, + 718.0, + 294.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 678.0, + 1405.0, + 678.0, + 1405.0, + 718.0, + 561.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 710.0, + 827.0, + 710.0, + 827.0, + 747.0, + 295.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 985.0, + 1403.0, + 985.0, + 1403.0, + 1022.0, + 293.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1018.0, + 1404.0, + 1018.0, + 1404.0, + 1052.0, + 296.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1048.0, + 820.0, + 1048.0, + 820.0, + 1082.0, + 296.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1607.0, + 822.0, + 1607.0, + 822.0, + 1643.0, + 294.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 911.0, + 1607.0, + 1404.0, + 1607.0, + 1404.0, + 1643.0, + 911.0, + 1643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 1406.0, + 1637.0, + 1406.0, + 1674.0, + 294.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1667.0, + 1407.0, + 1667.0, + 1407.0, + 1703.0, + 291.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1701.0, + 1404.0, + 1701.0, + 1404.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1730.0, + 1408.0, + 1730.0, + 1408.0, + 1764.0, + 292.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1758.0, + 1407.0, + 1758.0, + 1407.0, + 1796.0, + 292.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1792.0, + 1407.0, + 1792.0, + 1407.0, + 1823.0, + 296.0, + 1823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1818.0, + 1406.0, + 1818.0, + 1406.0, + 1856.0, + 292.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1887.0, + 294.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1880.0, + 762.0, + 1880.0, + 762.0, + 1914.0, + 295.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 284.0, + 1407.0, + 284.0, + 1407.0, + 319.0, + 294.0, + 319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 313.0, + 1403.0, + 313.0, + 1403.0, + 351.0, + 294.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 347.0, + 486.0, + 347.0, + 486.0, + 385.0, + 294.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 347.0, + 1405.0, + 347.0, + 1405.0, + 385.0, + 947.0, + 385.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 384.0, + 748.0, + 384.0, + 748.0, + 420.0, + 294.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 384.0, + 936.0, + 384.0, + 936.0, + 420.0, + 774.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 752.0, + 1368.0, + 752.0, + 1368.0, + 790.0, + 292.0, + 790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 784.0, + 1404.0, + 784.0, + 1404.0, + 824.0, + 293.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 817.0, + 652.0, + 817.0, + 652.0, + 851.0, + 295.0, + 851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 196.0, + 368.0, + 196.0, + 368.0, + 240.0, + 293.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 196.0, + 681.0, + 196.0, + 681.0, + 240.0, + 472.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 196.0, + 1294.0, + 196.0, + 1294.0, + 240.0, + 931.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 231.0, + 298.0, + 231.0, + 298.0, + 269.0, + 295.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 585.0, + 231.0, + 967.0, + 231.0, + 967.0, + 269.0, + 585.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 231.0, + 1007.0, + 231.0, + 1007.0, + 269.0, + 992.0, + 269.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 298, + 340, + 1402, + 340, + 1402, + 922, + 298, + 922 + ], + "score": 0.983, + "html": "
ModelUnlabeled dataLMdevtest
cleanothercleanother
960h - Supervised learning
DeepSpeech 2 [Amodei et al.,2016]5-gram5.3313.25
Fully Conv [Zeghidour et al., 2018]ConvLM3.089.943.2610.47
TDNN+Kaldi [Xu et al., 2018]4-gram2.717.373.127.63
SpecAugment [Park et al., 2019]RNN1-2.55.8
ContextNet [Han et al.,2020]LSTM1.93.91.94.1
Conformer [Gulati et al.,2020]LSTM2.14.31.93.9
960h - Self and semi-supervised learning
Transf.+ PL [Synnaeve et al.,2020]LL-60kCLM+Transf.2.003.652.094.11
IPL [Xu et al., 2020b]LL-60k4-gram+Transf.1.853.262.104.01
NST [Park et al., 2020]LL-60kLSTM1.63.41.73.4
wav2vec 2.0 [Baevski et al.,2020c]LL-60kTransf.1.63.01.83.3
wav2vec 2.0 + NST [Zhang et al.,2020b]LL-60kLSTM1.32.61.42.6
Unsupervised learning
wav2vec-ULARGELL-60k4-gram13.315.113.818.0
wav2vec-ULARGE+ STLL-60k4-gram3.46.03.86.5
LL-60kTransf.3.25.53.45.9
" + }, + { + "category_id": 1, + "poly": [ + 298, + 1599, + 1405, + 1599, + 1405, + 1844, + 298, + 1844 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1018, + 1403, + 1018, + 1403, + 1170, + 298, + 1170 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1185, + 1403, + 1185, + 1403, + 1338, + 298, + 1338 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1352, + 1403, + 1352, + 1403, + 1505, + 298, + 1505 + ], + "score": 0.977 + }, + { + "category_id": 6, + "poly": [ + 296, + 196, + 1405, + 196, + 1405, + 320, + 296, + 320 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1857, + 1407, + 1857, + 1407, + 1951, + 300, + 1951 + ], + "score": 0.964 + }, + { + "category_id": 0, + "poly": [ + 297, + 960, + 852, + 960, + 852, + 995, + 297, + 995 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 329, + 1978, + 1040, + 1978, + 1040, + 2008, + 329, + 2008 + ], + "score": 0.917 + }, + { + "category_id": 0, + "poly": [ + 301, + 1542, + 825, + 1542, + 825, + 1575, + 301, + 1575 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 841, + 2061, + 858, + 2061, + 858, + 2084, + 841, + 2084 + ], + "score": 0.729 + }, + { + "category_id": 13, + "poly": [ + 1055, + 1383, + 1109, + 1383, + 1109, + 1412, + 1055, + 1412 + ], + "score": 0.87, + "latex": "57 \\%" + }, + { + "category_id": 13, + "poly": [ + 1337, + 1919, + 1398, + 1919, + 1398, + 1950, + 1337, + 1950 + ], + "score": 0.83, + "latex": "\\ S \\operatorname { E } . 1" + }, + { + "category_id": 13, + "poly": [ + 1274, + 288, + 1363, + 288, + 1363, + 318, + 1274, + 318 + ], + "score": 0.78, + "latex": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T } ," + }, + { + "category_id": 13, + "poly": [ + 1083, + 1889, + 1143, + 1889, + 1143, + 1918, + 1083, + 1918 + ], + "score": 0.68, + "latex": "5 . 2 \\mathrm { m }" + }, + { + "category_id": 13, + "poly": [ + 1034, + 1246, + 1094, + 1246, + 1094, + 1276, + 1034, + 1276 + ], + "score": 0.66, + "latex": "\\ S \\operatorname { E } . 1" + }, + { + "category_id": 13, + "poly": [ + 976, + 1890, + 1036, + 1890, + 1036, + 1918, + 976, + 1918 + ], + "score": 0.58, + "latex": "4 . 3 \\mathrm { m }" + }, + { + "category_id": 13, + "poly": [ + 707, + 1307, + 794, + 1307, + 794, + 1337, + 707, + 1337 + ], + "score": 0.47, + "latex": "\\mathbf { \\partial } . \\mathbf { U } + \\mathbf { S } \\mathbf { T }" + }, + { + "category_id": 13, + "poly": [ + 411, + 761, + 520, + 761, + 520, + 788, + 411, + 788 + ], + "score": 0.43, + "latex": "2 . 0 + \\mathrm { N S T }" + }, + { + "category_id": 13, + "poly": [ + 394, + 655, + 415, + 655, + 415, + 678, + 394, + 678 + ], + "score": 0.39, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 1014, + 1662, + 1048, + 1662, + 1048, + 1689, + 1014, + 1689 + ], + "score": 0.31, + "latex": "2 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 388, + 507, + 404, + 507, + 404, + 528, + 388, + 528 + ], + "score": 0.31, + "latex": "^ +" + }, + { + "category_id": 13, + "poly": [ + 433, + 228, + 500, + 228, + 500, + 257, + 433, + 257 + ], + "score": 0.31, + "latex": "5 3 . 2 \\mathrm { k }" + }, + { + "category_id": 13, + "poly": [ + 394, + 654, + 449, + 654, + 449, + 679, + 394, + 679 + ], + "score": 0.27, + "latex": "+ \\mathrm { P L }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 195.0, + 1403.0, + 195.0, + 1403.0, + 230.0, + 293.0, + 230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 432.0, + 228.0, + 432.0, + 260.0, + 294.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 501.0, + 228.0, + 1408.0, + 228.0, + 1408.0, + 260.0, + 501.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 256.0, + 1407.0, + 256.0, + 1407.0, + 293.0, + 291.0, + 293.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 286.0, + 1273.0, + 286.0, + 1273.0, + 322.0, + 293.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 286.0, + 1377.0, + 286.0, + 1377.0, + 322.0, + 1364.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 959.0, + 856.0, + 959.0, + 856.0, + 1000.0, + 292.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.0, + 1972.0, + 1041.0, + 1972.0, + 1041.0, + 2014.0, + 330.0, + 2014.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1537.0, + 830.0, + 1537.0, + 830.0, + 1584.0, + 293.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2059.0, + 860.0, + 2059.0, + 860.0, + 2091.0, + 839.0, + 2091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1597.0, + 1406.0, + 1597.0, + 1406.0, + 1638.0, + 292.0, + 1638.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1629.0, + 1406.0, + 1629.0, + 1406.0, + 1666.0, + 293.0, + 1666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1660.0, + 1013.0, + 1660.0, + 1013.0, + 1695.0, + 292.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1660.0, + 1408.0, + 1660.0, + 1408.0, + 1695.0, + 1049.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1690.0, + 1406.0, + 1690.0, + 1406.0, + 1727.0, + 295.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1719.0, + 1406.0, + 1719.0, + 1406.0, + 1757.0, + 292.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1750.0, + 1405.0, + 1750.0, + 1405.0, + 1787.0, + 293.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1783.0, + 1406.0, + 1783.0, + 1406.0, + 1816.0, + 295.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1810.0, + 1147.0, + 1810.0, + 1147.0, + 1851.0, + 292.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1019.0, + 1404.0, + 1019.0, + 1404.0, + 1052.0, + 296.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1048.0, + 1404.0, + 1048.0, + 1404.0, + 1081.0, + 296.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1077.0, + 1404.0, + 1077.0, + 1404.0, + 1114.0, + 294.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1110.0, + 1405.0, + 1110.0, + 1405.0, + 1143.0, + 296.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1136.0, + 1304.0, + 1136.0, + 1304.0, + 1177.0, + 293.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1185.0, + 1407.0, + 1185.0, + 1407.0, + 1218.0, + 297.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1215.0, + 1407.0, + 1215.0, + 1407.0, + 1248.0, + 296.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1244.0, + 1033.0, + 1244.0, + 1033.0, + 1281.0, + 296.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1095.0, + 1244.0, + 1407.0, + 1244.0, + 1407.0, + 1281.0, + 1095.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1275.0, + 1407.0, + 1275.0, + 1407.0, + 1312.0, + 294.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1304.0, + 706.0, + 1304.0, + 706.0, + 1341.0, + 294.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1304.0, + 810.0, + 1304.0, + 810.0, + 1341.0, + 795.0, + 1341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 1404.0, + 1349.0, + 1404.0, + 1386.0, + 294.0, + 1386.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1384.0, + 1054.0, + 1384.0, + 1054.0, + 1415.0, + 294.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1384.0, + 1404.0, + 1384.0, + 1404.0, + 1415.0, + 1110.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1413.0, + 1405.0, + 1413.0, + 1405.0, + 1446.0, + 296.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1442.0, + 1405.0, + 1442.0, + 1405.0, + 1479.0, + 294.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1472.0, + 597.0, + 1472.0, + 597.0, + 1509.0, + 294.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1855.0, + 1408.0, + 1855.0, + 1408.0, + 1893.0, + 294.0, + 1893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 975.0, + 1888.0, + 975.0, + 1922.0, + 293.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 1888.0, + 1082.0, + 1888.0, + 1082.0, + 1922.0, + 1037.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1888.0, + 1404.0, + 1888.0, + 1404.0, + 1922.0, + 1144.0, + 1922.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1915.0, + 1336.0, + 1915.0, + 1336.0, + 1954.0, + 294.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1915.0, + 1408.0, + 1915.0, + 1408.0, + 1954.0, + 1399.0, + 1954.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 300, + 1138, + 1399, + 1138, + 1399, + 1449, + 300, + 1449 + ], + "score": 0.983, + "html": "
ModelLabeled data usedLMdenlfresitptAvg
Labeled training hours (full)2k1.6k1.1k918247161
Supervised learning
Pratap et al. [2020]full 5-gram6.4912.025.586.0710.5419.4910.0
Unsupervised learning0h33.3
wav2vec-U 4-gram32.540.239.858.159.843.9
wav2vec-U + ST0h4-gram11.821.414.711.326.326.318.6
" + }, + { + "category_id": 5, + "poly": [ + 409, + 338, + 1289, + 338, + 1289, + 1011, + 409, + 1011 + ], + "score": 0.981, + "html": "
ModelLMcore-devcore-testall-test
Supervised learning
LiGRU [Ravanelli et al., 2018]14.9
LiGRU [Ravanelli et al., 2019]14.2
Self and semi-supervised learning
vq-wav2vec [Baevski et al.,2020b]9.611.6
wav2vec 2.0 [Baevski et al.,2020c]7.48.3
Unsupervised learning - matched setup
EODM[Yeh et al.,2019]5-gram36.5=
GAN*[Chen et al., 2019]9-gram=48.6
GAN + HMM* [Chen et al.,2019]9-gram■-26.1
wav2vec-U4-gram17.017.816.6
wav2vec-U + ST4-gram11.312.011.3
Unsupervised learning - unmatched setup
EODM[Yeh et al., 2019]5-gram41.6■
GAN* [Chen et al.,2019]9-gram=50.0
GAN + HMM* [Chen et al., 2019]9-gram=-33.1
wav2vec-U*4-gram21.322.324.4
wav2vec-U + ST*4-gram13.815.018.6
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1822, + 1404, + 1822, + 1404, + 2008, + 299, + 2008 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1488, + 1404, + 1488, + 1404, + 1705, + 297, + 1705 + ], + "score": 0.972 + }, + { + "category_id": 6, + "poly": [ + 295, + 195, + 1406, + 195, + 1406, + 320, + 295, + 320 + ], + "score": 0.972 + }, + { + "category_id": 6, + "poly": [ + 299, + 1025, + 1405, + 1025, + 1405, + 1120, + 299, + 1120 + ], + "score": 0.926 + }, + { + "category_id": 0, + "poly": [ + 298, + 1751, + 548, + 1751, + 548, + 1788, + 298, + 1788 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 840, + 2062, + 858, + 2062, + 858, + 2085, + 840, + 2085 + ], + "score": 0.802 + }, + { + "category_id": 13, + "poly": [ + 782, + 228, + 846, + 228, + 846, + 259, + 782, + 259 + ], + "score": 0.77, + "latex": "\\ S \\operatorname { E . 1 }" + }, + { + "category_id": 13, + "poly": [ + 431, + 287, + 467, + 287, + 467, + 321, + 431, + 321 + ], + "score": 0.76, + "latex": "( ^ { * } )" + }, + { + "category_id": 13, + "poly": [ + 431, + 692, + 596, + 692, + 596, + 721, + 431, + 721 + ], + "score": 0.64, + "latex": "\\mathrm { G A N + H M M ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 429, + 871, + 502, + 871, + 502, + 898, + 429, + 898 + ], + "score": 0.58, + "latex": "\\mathrm { G A N ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 431, + 900, + 596, + 900, + 596, + 928, + 431, + 928 + ], + "score": 0.53, + "latex": "\\mathrm { G A N + H M M ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 429, + 663, + 502, + 663, + 502, + 690, + 429, + 690 + ], + "score": 0.51, + "latex": "\\mathbf { G A N ^ { * } }" + }, + { + "category_id": 13, + "poly": [ + 688, + 819, + 705, + 819, + 705, + 835, + 688, + 835 + ], + "score": 0.29, + "latex": "^ { \\ast }" + }, + { + "category_id": 13, + "poly": [ + 380, + 1673, + 432, + 1673, + 432, + 1702, + 380, + 1702 + ], + "score": 0.26, + "latex": "1 . 8 \\mathrm { h }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 195.0, + 1407.0, + 195.0, + 1407.0, + 232.0, + 294.0, + 232.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 781.0, + 228.0, + 781.0, + 260.0, + 294.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 260.0, + 847.0, + 260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 259.0, + 1404.0, + 259.0, + 1404.0, + 291.0, + 295.0, + 291.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 288.0, + 430.0, + 288.0, + 430.0, + 324.0, + 294.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 288.0, + 1366.0, + 288.0, + 1366.0, + 324.0, + 468.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1023.0, + 1405.0, + 1023.0, + 1405.0, + 1062.0, + 292.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1054.0, + 1407.0, + 1054.0, + 1407.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1083.0, + 480.0, + 1083.0, + 480.0, + 1126.0, + 292.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1748.0, + 552.0, + 1748.0, + 552.0, + 1791.0, + 291.0, + 1791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2058.0, + 862.0, + 2058.0, + 862.0, + 2090.0, + 838.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1824.0, + 1403.0, + 1824.0, + 1403.0, + 1856.0, + 297.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1854.0, + 1406.0, + 1854.0, + 1406.0, + 1888.0, + 294.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1883.0, + 1409.0, + 1883.0, + 1409.0, + 1919.0, + 293.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1914.0, + 1406.0, + 1914.0, + 1406.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1944.0, + 1408.0, + 1944.0, + 1408.0, + 1981.0, + 293.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 1408.0, + 1975.0, + 1408.0, + 2009.0, + 294.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1490.0, + 1405.0, + 1490.0, + 1405.0, + 1525.0, + 295.0, + 1525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1522.0, + 1404.0, + 1522.0, + 1404.0, + 1557.0, + 295.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1548.0, + 1405.0, + 1548.0, + 1405.0, + 1589.0, + 291.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1582.0, + 1406.0, + 1582.0, + 1406.0, + 1616.0, + 294.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1611.0, + 1406.0, + 1611.0, + 1406.0, + 1646.0, + 292.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1642.0, + 1404.0, + 1642.0, + 1404.0, + 1676.0, + 294.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1672.0, + 379.0, + 1672.0, + 379.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 1672.0, + 1278.0, + 1672.0, + 1278.0, + 1707.0, + 433.0, + 1707.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 862, + 1405, + 862, + 1405, + 1138, + 297, + 1138 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1322, + 1404, + 1322, + 1404, + 1507, + 298, + 1507 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 725, + 1405, + 725, + 1405, + 849, + 298, + 849 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 301, + 303, + 853, + 303, + 853, + 629, + 301, + 629 + ], + "score": 0.979, + "html": "
Modelttky
Supervised learning
Fer et al. [2017]42.538.7
m-CPC [Riviere et al., 2020]42.041.2
XLSR-53 [Conneau et al.,2020]5.16.1
Unsupervised learning
wav2vec-U25.724.1
wav2vec-U + HMM13.714.9
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1520, + 1402, + 1520, + 1402, + 1674, + 299, + 1674 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1823, + 1403, + 1823, + 1403, + 2007, + 299, + 2007 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1687, + 1403, + 1687, + 1403, + 1810, + 298, + 1810 + ], + "score": 0.976 + }, + { + "category_id": 5, + "poly": [ + 963, + 334, + 1327, + 334, + 1327, + 603, + 963, + 603 + ], + "score": 0.974, + "html": "
Modelsw
Supervised learning
Besacier et al. [2015]27.36
Unsupervised learning
wav2vec-U52.6
wav2vec-U + ST32.2
" + }, + { + "category_id": 6, + "poly": [ + 888, + 242, + 1402, + 242, + 1402, + 333, + 888, + 333 + ], + "score": 0.955 + }, + { + "category_id": 6, + "poly": [ + 296, + 239, + 811, + 239, + 811, + 302, + 296, + 302 + ], + "score": 0.946 + }, + { + "category_id": 0, + "poly": [ + 299, + 1225, + 765, + 1225, + 765, + 1264, + 299, + 1264 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 840, + 2061, + 859, + 2061, + 859, + 2084, + 840, + 2084 + ], + "score": 0.798 + }, + { + "category_id": 15, + "poly": [ + 889.0, + 241.0, + 1403.0, + 241.0, + 1403.0, + 274.0, + 889.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 273.0, + 1404.0, + 273.0, + 1404.0, + 306.0, + 890.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 889.0, + 300.0, + 1146.0, + 300.0, + 1146.0, + 338.0, + 889.0, + 338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 267.0, + 595.0, + 267.0, + 595.0, + 306.0, + 294.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 237.5, + 813.0, + 237.5, + 813.0, + 273.0, + 294.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1224.0, + 770.0, + 1224.0, + 770.0, + 1268.0, + 293.0, + 1268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2060.0, + 861.0, + 2060.0, + 861.0, + 2090.0, + 839.0, + 2090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 864.0, + 1405.0, + 864.0, + 1405.0, + 897.0, + 295.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 892.0, + 1408.0, + 892.0, + 1408.0, + 930.0, + 294.0, + 930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 924.0, + 1406.0, + 924.0, + 1406.0, + 957.0, + 296.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 953.0, + 1405.0, + 953.0, + 1405.0, + 988.0, + 292.0, + 988.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 982.0, + 1405.0, + 982.0, + 1405.0, + 1020.0, + 294.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1013.0, + 1406.0, + 1013.0, + 1406.0, + 1050.0, + 295.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1043.0, + 1405.0, + 1043.0, + 1405.0, + 1081.0, + 294.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1074.0, + 1406.0, + 1074.0, + 1406.0, + 1111.0, + 295.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1107.0, + 1313.0, + 1107.0, + 1313.0, + 1140.0, + 296.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1324.0, + 1408.0, + 1324.0, + 1408.0, + 1356.0, + 296.0, + 1356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1356.0, + 1406.0, + 1356.0, + 1406.0, + 1388.0, + 294.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1384.0, + 1406.0, + 1384.0, + 1406.0, + 1420.0, + 293.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1413.0, + 1404.0, + 1413.0, + 1404.0, + 1449.0, + 293.0, + 1449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1441.0, + 1405.0, + 1441.0, + 1405.0, + 1481.0, + 293.0, + 1481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1477.0, + 1026.0, + 1477.0, + 1026.0, + 1509.0, + 297.0, + 1509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 726.0, + 1405.0, + 726.0, + 1405.0, + 759.0, + 295.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 758.0, + 1408.0, + 758.0, + 1408.0, + 791.0, + 295.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 823.0, + 293.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 817.0, + 758.0, + 817.0, + 758.0, + 852.0, + 293.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1519.0, + 1406.0, + 1519.0, + 1406.0, + 1556.0, + 295.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1552.0, + 1403.0, + 1552.0, + 1403.0, + 1585.0, + 295.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1581.0, + 1406.0, + 1581.0, + 1406.0, + 1617.0, + 293.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1610.0, + 1406.0, + 1610.0, + 1406.0, + 1646.0, + 293.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1643.0, + 946.0, + 1643.0, + 946.0, + 1676.0, + 294.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1858.0, + 294.0, + 1858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1854.0, + 1407.0, + 1854.0, + 1407.0, + 1890.0, + 293.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1883.0, + 1405.0, + 1883.0, + 1405.0, + 1919.0, + 292.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1910.0, + 1408.0, + 1910.0, + 1408.0, + 1952.0, + 292.0, + 1952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1944.0, + 1404.0, + 1944.0, + 1404.0, + 1981.0, + 294.0, + 1981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1972.0, + 1398.0, + 1972.0, + 1398.0, + 2015.0, + 293.0, + 2015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1688.0, + 1405.0, + 1688.0, + 1405.0, + 1721.0, + 296.0, + 1721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1716.0, + 1404.0, + 1716.0, + 1404.0, + 1754.0, + 292.0, + 1754.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1747.0, + 1407.0, + 1747.0, + 1407.0, + 1785.0, + 292.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1780.0, + 1155.0, + 1780.0, + 1155.0, + 1812.0, + 293.0, + 1812.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 267, + 1406, + 267, + 1406, + 390, + 298, + 390 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 293, + 478, + 1409, + 478, + 1409, + 1995, + 293, + 1995 + ], + "score": 0.959 + }, + { + "category_id": 0, + "poly": [ + 298, + 197, + 944, + 197, + 944, + 237, + 298, + 237 + ], + "score": 0.928 + }, + { + "category_id": 0, + "poly": [ + 299, + 432, + 455, + 432, + 455, + 468, + 299, + 468 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2086, + 836, + 2086 + ], + "score": 0.757 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2086, + 836, + 2086 + ], + "score": 0.207 + }, + { + "category_id": 15, + "poly": [ + 293.0, + 193.0, + 948.0, + 193.0, + 948.0, + 247.0, + 293.0, + 247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 428.0, + 460.0, + 428.0, + 460.0, + 474.0, + 295.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 870.0, + 2058.0, + 870.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 265.0, + 1406.0, + 265.0, + 1406.0, + 302.0, + 292.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 295.0, + 1406.0, + 295.0, + 1406.0, + 334.0, + 292.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 326.0, + 1409.0, + 326.0, + 1409.0, + 365.0, + 292.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 358.0, + 1208.0, + 358.0, + 1208.0, + 391.0, + 292.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 480.0, + 1408.0, + 480.0, + 1408.0, + 526.0, + 289.0, + 526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 516.0, + 889.0, + 516.0, + 889.0, + 553.0, + 319.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 566.0, + 1408.0, + 566.0, + 1408.0, + 605.0, + 293.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 595.0, + 1409.0, + 595.0, + 1409.0, + 639.0, + 319.0, + 639.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 628.0, + 602.0, + 628.0, + 602.0, + 665.0, + 321.0, + 665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 679.0, + 1406.0, + 679.0, + 1406.0, + 716.0, + 294.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 711.0, + 1408.0, + 711.0, + 1408.0, + 748.0, + 322.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 738.0, + 398.0, + 738.0, + 398.0, + 778.0, + 321.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 795.0, + 1195.0, + 795.0, + 1195.0, + 827.0, + 294.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 846.0, + 1404.0, + 846.0, + 1404.0, + 883.0, + 294.0, + 883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 876.0, + 649.0, + 876.0, + 649.0, + 913.0, + 322.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 925.0, + 1408.0, + 925.0, + 1408.0, + 965.0, + 291.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 959.0, + 469.0, + 959.0, + 469.0, + 991.0, + 321.0, + 991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1004.0, + 1409.0, + 1004.0, + 1409.0, + 1050.0, + 289.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1041.0, + 468.0, + 1041.0, + 468.0, + 1073.0, + 322.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1092.0, + 1404.0, + 1092.0, + 1404.0, + 1129.0, + 294.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1124.0, + 748.0, + 1124.0, + 748.0, + 1159.0, + 321.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1172.0, + 1408.0, + 1172.0, + 1408.0, + 1215.0, + 293.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1205.0, + 791.0, + 1205.0, + 791.0, + 1242.0, + 319.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1257.0, + 1406.0, + 1257.0, + 1406.0, + 1294.0, + 293.0, + 1294.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1287.0, + 1034.0, + 1287.0, + 1034.0, + 1324.0, + 322.0, + 1324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1339.0, + 1408.0, + 1339.0, + 1408.0, + 1376.0, + 291.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1370.0, + 1408.0, + 1370.0, + 1408.0, + 1407.0, + 321.0, + 1407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1400.0, + 1082.0, + 1400.0, + 1082.0, + 1437.0, + 322.0, + 1437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1447.0, + 1409.0, + 1447.0, + 1409.0, + 1491.0, + 289.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1481.0, + 621.0, + 1481.0, + 621.0, + 1518.0, + 321.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1531.0, + 1406.0, + 1531.0, + 1406.0, + 1572.0, + 293.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1565.0, + 1406.0, + 1565.0, + 1406.0, + 1602.0, + 321.0, + 1602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1595.0, + 760.0, + 1595.0, + 760.0, + 1632.0, + 322.0, + 1632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1642.0, + 1406.0, + 1642.0, + 1406.0, + 1686.0, + 291.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1678.0, + 830.0, + 1678.0, + 830.0, + 1715.0, + 322.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1728.0, + 1406.0, + 1728.0, + 1406.0, + 1765.0, + 294.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1759.0, + 915.0, + 1759.0, + 915.0, + 1796.0, + 319.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1806.0, + 1408.0, + 1806.0, + 1408.0, + 1850.0, + 293.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1840.0, + 927.0, + 1840.0, + 927.0, + 1880.0, + 319.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1888.0, + 1406.0, + 1888.0, + 1406.0, + 1932.0, + 293.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1924.0, + 917.0, + 1924.0, + 917.0, + 1961.0, + 321.0, + 1961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1973.0, + 1365.0, + 1973.0, + 1365.0, + 2010.0, + 294.0, + 2010.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 835, + 2061, + 864, + 2061, + 864, + 2087, + 835, + 2087 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 290, + 122, + 1411, + 122, + 1411, + 2013, + 290, + 2013 + ], + "score": 0.83 + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2097.0, + 831.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 203.0, + 1406.0, + 203.0, + 1406.0, + 242.0, + 293.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 236.0, + 626.0, + 236.0, + 626.0, + 269.0, + 323.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 286.0, + 1406.0, + 286.0, + 1406.0, + 325.0, + 293.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 317.0, + 1188.0, + 317.0, + 1188.0, + 356.0, + 318.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 371.0, + 1406.0, + 371.0, + 1406.0, + 410.0, + 291.0, + 410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 400.0, + 1054.0, + 400.0, + 1054.0, + 439.0, + 318.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 454.0, + 1408.0, + 454.0, + 1408.0, + 493.0, + 293.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 483.0, + 850.0, + 483.0, + 850.0, + 524.0, + 318.0, + 524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 530.0, + 1408.0, + 530.0, + 1408.0, + 582.0, + 287.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 570.0, + 632.0, + 570.0, + 632.0, + 603.0, + 323.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 620.0, + 1408.0, + 620.0, + 1408.0, + 659.0, + 291.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 651.0, + 1370.0, + 651.0, + 1370.0, + 690.0, + 318.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 705.0, + 1408.0, + 705.0, + 1408.0, + 744.0, + 291.0, + 744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 731.0, + 1410.0, + 731.0, + 1410.0, + 775.0, + 316.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 767.0, + 395.0, + 767.0, + 395.0, + 804.0, + 321.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 817.0, + 1406.0, + 817.0, + 1406.0, + 856.0, + 293.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 850.0, + 825.0, + 850.0, + 825.0, + 887.0, + 316.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 904.0, + 1406.0, + 904.0, + 1406.0, + 943.0, + 291.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 933.0, + 923.0, + 933.0, + 923.0, + 972.0, + 318.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 989.0, + 1406.0, + 989.0, + 1406.0, + 1022.0, + 293.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1020.0, + 1060.0, + 1020.0, + 1060.0, + 1053.0, + 321.0, + 1053.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1070.0, + 1408.0, + 1070.0, + 1408.0, + 1109.0, + 293.0, + 1109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1103.0, + 588.0, + 1103.0, + 588.0, + 1136.0, + 321.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1155.0, + 1408.0, + 1155.0, + 1408.0, + 1194.0, + 293.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1184.0, + 1120.0, + 1184.0, + 1120.0, + 1223.0, + 318.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1238.0, + 1410.0, + 1238.0, + 1410.0, + 1277.0, + 293.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1269.0, + 1408.0, + 1269.0, + 1408.0, + 1308.0, + 318.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1300.0, + 536.0, + 1300.0, + 536.0, + 1335.0, + 318.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1350.0, + 1406.0, + 1350.0, + 1406.0, + 1391.0, + 291.0, + 1391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1383.0, + 632.0, + 1383.0, + 632.0, + 1418.0, + 318.0, + 1418.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1435.0, + 1410.0, + 1435.0, + 1410.0, + 1474.0, + 293.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1464.0, + 1408.0, + 1464.0, + 1408.0, + 1503.0, + 321.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1499.0, + 628.0, + 1499.0, + 628.0, + 1532.0, + 323.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1547.0, + 1406.0, + 1547.0, + 1406.0, + 1586.0, + 293.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1580.0, + 783.0, + 1580.0, + 783.0, + 1617.0, + 316.0, + 1617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1629.0, + 1408.0, + 1629.0, + 1408.0, + 1673.0, + 291.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1663.0, + 801.0, + 1663.0, + 801.0, + 1700.0, + 316.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1717.0, + 1408.0, + 1717.0, + 1408.0, + 1756.0, + 293.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1748.0, + 1410.0, + 1748.0, + 1410.0, + 1787.0, + 321.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1775.0, + 815.0, + 1775.0, + 815.0, + 1816.0, + 316.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1827.0, + 1410.0, + 1827.0, + 1410.0, + 1870.0, + 291.0, + 1870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1862.0, + 1221.0, + 1862.0, + 1221.0, + 1901.0, + 321.0, + 1901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1914.0, + 1410.0, + 1914.0, + 1410.0, + 1953.0, + 295.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1936.0, + 1410.0, + 1936.0, + 1410.0, + 1991.0, + 316.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1976.0, + 789.0, + 1976.0, + 789.0, + 2015.0, + 321.0, + 2015.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 289, + 193, + 1412, + 193, + 1412, + 2025, + 289, + 2025 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 866, + 2061, + 866, + 2087, + 836, + 2087 + ], + "score": 0.837 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 869.0, + 2057.0, + 869.0, + 2097.0, + 832.0, + 2097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 199.0, + 1408.0, + 199.0, + 1408.0, + 240.0, + 293.0, + 240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 236.0, + 1066.0, + 236.0, + 1066.0, + 268.0, + 321.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 286.0, + 1408.0, + 286.0, + 1408.0, + 324.0, + 291.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 314.0, + 400.0, + 314.0, + 400.0, + 354.0, + 319.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 364.0, + 1410.0, + 364.0, + 1410.0, + 411.0, + 291.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 399.0, + 933.0, + 399.0, + 933.0, + 437.0, + 321.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 451.0, + 1408.0, + 451.0, + 1408.0, + 489.0, + 293.0, + 489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 483.0, + 1194.0, + 483.0, + 1194.0, + 521.0, + 321.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 531.0, + 1408.0, + 531.0, + 1408.0, + 574.0, + 289.0, + 574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 564.0, + 512.0, + 564.0, + 512.0, + 600.0, + 317.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 614.0, + 1408.0, + 614.0, + 1408.0, + 654.0, + 289.0, + 654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 648.0, + 1019.0, + 648.0, + 1019.0, + 684.0, + 319.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 700.0, + 1410.0, + 700.0, + 1410.0, + 739.0, + 291.0, + 739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 729.0, + 412.0, + 729.0, + 412.0, + 771.0, + 319.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 779.0, + 1408.0, + 779.0, + 1408.0, + 821.0, + 289.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 815.0, + 410.0, + 815.0, + 410.0, + 849.0, + 321.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 861.0, + 1410.0, + 861.0, + 1410.0, + 908.0, + 289.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 898.0, + 1164.0, + 898.0, + 1164.0, + 936.0, + 319.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 948.0, + 1406.0, + 948.0, + 1406.0, + 986.0, + 293.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 978.0, + 1023.0, + 978.0, + 1023.0, + 1020.0, + 317.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1026.0, + 1408.0, + 1026.0, + 1408.0, + 1073.0, + 289.0, + 1073.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1065.0, + 862.0, + 1065.0, + 862.0, + 1099.0, + 321.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1115.0, + 1367.0, + 1115.0, + 1367.0, + 1153.0, + 291.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1167.0, + 1408.0, + 1167.0, + 1408.0, + 1206.0, + 291.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1198.0, + 794.0, + 1198.0, + 794.0, + 1236.0, + 319.0, + 1236.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1246.0, + 1408.0, + 1246.0, + 1408.0, + 1292.0, + 291.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1280.0, + 750.0, + 1280.0, + 750.0, + 1318.0, + 321.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1328.0, + 1412.0, + 1328.0, + 1412.0, + 1375.0, + 289.0, + 1375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1367.0, + 456.0, + 1367.0, + 456.0, + 1399.0, + 317.0, + 1399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1417.0, + 1408.0, + 1417.0, + 1408.0, + 1455.0, + 293.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1445.0, + 400.0, + 1445.0, + 400.0, + 1485.0, + 319.0, + 1485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1499.0, + 1410.0, + 1499.0, + 1410.0, + 1538.0, + 291.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1530.0, + 953.0, + 1530.0, + 953.0, + 1566.0, + 319.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1580.0, + 1406.0, + 1580.0, + 1406.0, + 1622.0, + 291.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1616.0, + 760.0, + 1616.0, + 760.0, + 1648.0, + 321.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1666.0, + 1408.0, + 1666.0, + 1408.0, + 1705.0, + 293.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1697.0, + 1116.0, + 1697.0, + 1116.0, + 1733.0, + 317.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1745.0, + 1408.0, + 1745.0, + 1408.0, + 1785.0, + 291.0, + 1785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1779.0, + 1410.0, + 1779.0, + 1410.0, + 1817.0, + 319.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1807.0, + 400.0, + 1807.0, + 400.0, + 1850.0, + 319.0, + 1850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1862.0, + 1406.0, + 1862.0, + 1406.0, + 1900.0, + 293.0, + 1900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1894.0, + 899.0, + 1894.0, + 899.0, + 1932.0, + 321.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1942.0, + 1408.0, + 1942.0, + 1408.0, + 1984.0, + 291.0, + 1984.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1972.0, + 784.0, + 1972.0, + 784.0, + 2015.0, + 315.0, + 2015.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 290, + 162, + 1411, + 162, + 1411, + 2029, + 290, + 2029 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 835, + 2060, + 865, + 2060, + 865, + 2087, + 835, + 2087 + ], + "score": 0.835 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2058.0, + 869.0, + 2058.0, + 869.0, + 2096.0, + 832.0, + 2096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 202.0, + 1407.0, + 202.0, + 1407.0, + 241.0, + 292.0, + 241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 235.0, + 473.0, + 235.0, + 473.0, + 268.0, + 320.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 282.0, + 1409.0, + 282.0, + 1409.0, + 323.0, + 292.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 315.0, + 1353.0, + 315.0, + 1353.0, + 354.0, + 320.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 360.0, + 1407.0, + 360.0, + 1407.0, + 407.0, + 290.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 397.0, + 1146.0, + 397.0, + 1146.0, + 436.0, + 320.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 446.0, + 1407.0, + 446.0, + 1407.0, + 485.0, + 292.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 479.0, + 1409.0, + 479.0, + 1409.0, + 518.0, + 320.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 528.0, + 1405.0, + 528.0, + 1405.0, + 567.0, + 292.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 559.0, + 1105.0, + 559.0, + 1105.0, + 596.0, + 318.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 610.0, + 1130.0, + 610.0, + 1130.0, + 649.0, + 292.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 655.0, + 1411.0, + 655.0, + 1411.0, + 704.0, + 288.0, + 704.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 690.0, + 471.0, + 690.0, + 471.0, + 729.0, + 318.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 745.0, + 1405.0, + 745.0, + 1405.0, + 778.0, + 294.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 772.0, + 1266.0, + 772.0, + 1266.0, + 813.0, + 318.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 823.0, + 1407.0, + 823.0, + 1407.0, + 862.0, + 292.0, + 862.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 850.0, + 1411.0, + 850.0, + 1411.0, + 897.0, + 316.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 884.0, + 603.0, + 884.0, + 603.0, + 923.0, + 320.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 936.0, + 1407.0, + 936.0, + 1407.0, + 975.0, + 294.0, + 975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 964.0, + 1316.0, + 964.0, + 1316.0, + 1009.0, + 318.0, + 1009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1011.0, + 1409.0, + 1011.0, + 1409.0, + 1061.0, + 290.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1048.0, + 893.0, + 1048.0, + 893.0, + 1087.0, + 320.0, + 1087.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1095.0, + 1407.0, + 1095.0, + 1407.0, + 1139.0, + 290.0, + 1139.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1128.0, + 1081.0, + 1128.0, + 1081.0, + 1169.0, + 316.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1177.0, + 1409.0, + 1177.0, + 1409.0, + 1221.0, + 290.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1212.0, + 1118.0, + 1212.0, + 1118.0, + 1251.0, + 320.0, + 1251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1257.0, + 1411.0, + 1257.0, + 1411.0, + 1305.0, + 288.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1296.0, + 497.0, + 1296.0, + 497.0, + 1329.0, + 320.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1343.0, + 1407.0, + 1343.0, + 1407.0, + 1382.0, + 292.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1374.0, + 752.0, + 1374.0, + 752.0, + 1411.0, + 318.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1425.0, + 1407.0, + 1425.0, + 1407.0, + 1464.0, + 292.0, + 1464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1456.0, + 804.0, + 1456.0, + 804.0, + 1495.0, + 320.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1501.0, + 1407.0, + 1501.0, + 1407.0, + 1550.0, + 288.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 1538.0, + 971.0, + 1538.0, + 971.0, + 1575.0, + 316.0, + 1575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1583.0, + 1409.0, + 1583.0, + 1409.0, + 1628.0, + 290.0, + 1628.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1622.0, + 1405.0, + 1622.0, + 1405.0, + 1655.0, + 322.0, + 1655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1649.0, + 1375.0, + 1649.0, + 1375.0, + 1688.0, + 320.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1694.0, + 1409.0, + 1694.0, + 1409.0, + 1743.0, + 288.0, + 1743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1731.0, + 1330.0, + 1731.0, + 1330.0, + 1770.0, + 320.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1778.0, + 1405.0, + 1778.0, + 1405.0, + 1821.0, + 290.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1813.0, + 1051.0, + 1813.0, + 1051.0, + 1847.0, + 318.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1860.0, + 1407.0, + 1860.0, + 1407.0, + 1903.0, + 290.0, + 1903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1890.0, + 563.0, + 1890.0, + 563.0, + 1932.0, + 319.0, + 1932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1944.0, + 1407.0, + 1944.0, + 1407.0, + 1983.0, + 292.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1975.0, + 1130.0, + 1975.0, + 1130.0, + 2013.0, + 320.0, + 2013.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 291, + 189, + 1411, + 189, + 1411, + 1245, + 291, + 1245 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 836, + 2061, + 865, + 2061, + 865, + 2085, + 836, + 2085 + ], + "score": 0.833 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2057.0, + 871.0, + 2057.0, + 871.0, + 2095.0, + 832.0, + 2095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 202.0, + 1405.0, + 202.0, + 1405.0, + 242.0, + 293.0, + 242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 235.0, + 550.0, + 235.0, + 550.0, + 269.0, + 321.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 285.0, + 1406.0, + 285.0, + 1406.0, + 321.0, + 296.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 317.0, + 879.0, + 317.0, + 879.0, + 353.0, + 324.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 369.0, + 1404.0, + 369.0, + 1404.0, + 405.0, + 296.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 398.0, + 822.0, + 398.0, + 822.0, + 433.0, + 322.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 449.0, + 1404.0, + 449.0, + 1404.0, + 491.0, + 291.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 482.0, + 1220.0, + 482.0, + 1220.0, + 518.0, + 324.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 535.0, + 1406.0, + 535.0, + 1406.0, + 571.0, + 294.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 563.0, + 1409.0, + 563.0, + 1409.0, + 603.0, + 322.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 616.0, + 1406.0, + 616.0, + 1406.0, + 652.0, + 296.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 645.0, + 1409.0, + 645.0, + 1409.0, + 687.0, + 320.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 680.0, + 508.0, + 680.0, + 508.0, + 712.0, + 324.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 726.0, + 1406.0, + 726.0, + 1406.0, + 769.0, + 293.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 761.0, + 1006.0, + 761.0, + 1006.0, + 797.0, + 322.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 810.0, + 1404.0, + 810.0, + 1404.0, + 852.0, + 293.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 843.0, + 925.0, + 843.0, + 925.0, + 880.0, + 322.0, + 880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 894.0, + 1404.0, + 894.0, + 1404.0, + 932.0, + 293.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 927.0, + 885.0, + 927.0, + 885.0, + 961.0, + 322.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 975.0, + 1405.0, + 975.0, + 1405.0, + 1018.0, + 292.0, + 1018.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1007.0, + 1068.0, + 1007.0, + 1068.0, + 1047.0, + 320.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1058.0, + 1405.0, + 1058.0, + 1405.0, + 1099.0, + 293.0, + 1099.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1091.0, + 1348.0, + 1091.0, + 1348.0, + 1130.0, + 321.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1140.0, + 1405.0, + 1140.0, + 1405.0, + 1181.0, + 293.0, + 1181.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1171.0, + 1409.0, + 1171.0, + 1409.0, + 1214.0, + 320.0, + 1214.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1205.0, + 409.0, + 1205.0, + 409.0, + 1238.0, + 322.0, + 1238.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/SySaJ0xCZ/SySaJ0xCZ.md b/parse/train/SySaJ0xCZ/SySaJ0xCZ.md new file mode 100644 index 0000000000000000000000000000000000000000..a13783fa5e339779c86d0f41e8ade2fb31b9e878 --- /dev/null +++ b/parse/train/SySaJ0xCZ/SySaJ0xCZ.md @@ -0,0 +1,259 @@ +# SIMPLE AND EFFICIENT ARCHITECTURE SEARCH FOR CONVOLUTIONAL NEURAL NETWORKS + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Neural networks have recently had a lot of success for many tasks. However, neural network architectures that perform well are still typically designed manually by experts in a cumbersome trial-and-error process. We propose a new method to automatically search for well-performing CNN architectures based on a simple hill climbing procedure whose operators apply network morphisms, followed by short optimization runs by cosine annealing. Surprisingly, this simple method yields competitive results, despite only requiring resources in the same order of magnitude as training a single network. E.g., on CIFAR-10, our method designs and trains networks with an error rate below $6 \%$ in only 12 hours on a single GPU; training for one day reduces this error further, to almost $5 \%$ . + +# 1 INTRODUCTION + +Neural networks have rapidly gained popularity over the last few years due to their success in a variety of tasks, such as image recognition (Krizhevsky et al., 2012), speech recognition (Hinton et al., 2012) and machine translation (Bahdanau et al., 2015). In most cases, these neural networks are still designed by hand, which is an exhausting, time-consuming process. Additionally, the vast amount of possible configurations requires expert knowledge to restrict the search. Therefore, a natural goal is to design optimization algorithms that automate this neural architecture search. + +However, most classic optimization algorithms do not apply to this problem, since the architecture search space is discrete (e.g., number of layers, layer types) and conditional (e.g., the number of parameters defining a layer depends on the layer type). Thus, methods that rely on, e.g., differentiability or independent parameters are not applicable. This led to a growing interest in using evolutionary algorithms (Real et al., 2017; Suganuma et al., 2017) and reinforcement learning (Baker et al., 2016; Cai et al., 2017; Zoph & Le, 2017) for automatically designing CNN architectures. Unfortunately, most proposed methods are either very costly (requiring hundreds or thousands of GPU days) or yield non-competitive performance. + +In this work, we aim to dramatically reduce these computational costs while still achieving competitive performance. Specifically, our contributions are as follows: + +• We propose a baseline method that randomly constructs networks and trains them with SGDR (Loshchilov & Hutter, 2017). We demonstrate that this simple baseline achieves $6 \% - 7 \%$ test error on CIFAR-10, which already rivals several existing methods for neural archictecture search. Due to its simplicity, we hope that this baseline provides a valuable starting point for the development of more sophisticated methods in the future. +• We formalize and extend the work on network morphisms (Chen et al., 2015; Wei et al., 2016; Cai et al., 2017) in order to provide popular network building blocks, such as skip connections and batch normalization. +• We propose Neural Architecture Search by Hillclimbing (NASH), a simple iterative approach that, at each step, applies a set of alternative network morphisms to the current network, trains the resulting child networks with short optimization runs of cosine annealing (Loshchilov & Hutter, 2017), and moves to the most promising child network. NASH finds and trains competitive architectures at a computational cost of the same order of magnitude as training a single network; e.g., on CIFAR-10, NASH finds and trains CNNs with an error rate below $6 \%$ in roughly 12 hours on + +a single GPU. After one day the error is reduced to almost $5 \%$ . Models from different stages of our algorithm can be combined to achieve an error of $4 . 7 \%$ within two days on a single GPU. On CIFAR-100, we achieve an error below $24 \%$ in one day and get close to $20 \%$ after two days. + +• Our method is easy to use and easy to extend, so it hopefully can serve as a basis for future work. + +We first discuss related work in Section 2. Then, we formalize the concept of network morphisms in Section 3 and propose our architecture search methods based on them in Section 4. We evaluate our methods in Section 5 and conclude in Section 6. + +# 2 RELATED WORK + +Hyperparameter optimization. Neural networks are known to be quite sensitive to the setting of various hyperparameters, such as learning rates and regularization constants. There exists a long line of research on automated methods for setting these hyperparameters, including, e.g., random search (Bergstra & Bengio, 2012), Bayesian optimization (Bergstra et al., 2011; Snoek et al., 2012), bandit-based approaches (Li et al., 2016a), and evolutionary strategies (Loshchilov & Hutter, 2016). + +Automated architecture search. In recent years, the research focus has shifted from optimizing hyperparameters to optimizing architectures. While architectural choices can be treated as categorical hyperparameters and be optimized with standard hyperparameter optimization methods (Bergstra et al., 2011; Mendoza et al., 2016), the current focus is on the development of special techniques for architectural optimization. One very popular approach is to train a reinforcement learning agent with the objective of designing well-performing convolutional neural networks (Baker et al., 2016; Zoph & Le, 2017; Cai et al., 2017). Baker et al. (2016) train an RL agent to sequentially choose the type of layers (convolution, pooling, fully connected) and their parameters. Zoph & Le (2017) use a recurrent neural network controller to sequentially generate a string representing the network architecture. Both approaches train their generated networks from scratch and evaluate their performance on a validation set, which represents a very costly step. In a follow-up work (Zoph et al., 2017), the RL agent learned to build cells, which are then used as building blocks for a neural network with a fixed global structure. Unfortunately, training an RL agent with the objective of designing architecture is extremely expensive: both Baker et al. (2016) and Zoph & Le (2017) required over 10.000 fully trained networks, requiring hundreds to thousands of GPU days. To overcome this drawback, Cai et al. (2017) proposed to apply the concept of network transformations/morphisms within RL. As in our (independent, parallel) work, the basic idea is to use the these transformation to generate new pre-trained architectures to avoid the large cost of training all networks from scratch. Compared to this work, our approach is much simpler and 15 times faster while obtaining better performance. + +Real et al. (2017) and Suganuma et al. (2017) utilized evolutionary algorithms to iteratively generate powerful networks from a small network. Operations like inserting a layer, modifying the parameters of a layer or adding skip connections serve as ”mutations” in their framework of evolution. Whereas Real et al. (2017) also used enormous computational resources (250 GPUs, 10 days), Suganuma et al. (2017) were restricted to relatively small networks due to handling a population of networks. In contrast to the previous methods where network capacity increases over time, Saxena & Verbeek (2016) start with training a large network (a ”convolution neural fabric”) and prune this in the end. Very recently, Brock et al. (2017) used hypernetworks (Ha et al., 2017) to generate the weights for a randomly sampled network architecture with the goal of eliminating the costly process of training a vast amount of networks. + +Network morphism/ transformation. Network transformations were (to our knowledge) first introduced by Chen et al. (2015) in the context of transfer learning. The authors described a function preserving operation to make a network deeper (dubbed ”Net2Deeper”) or wider (”Net2Wider”) with the goal of speeding up training and exploring network architectures. Wei et al. (2016) proposed additional operations, e.g., for handling non-idempotent activation functions or altering the kernel size and introduced the term network morphism. As mentioned above, Cai et al. (2017) used network morphisms for architecture search, though they just employ the Net2Deeper and Net2Wider operators from Chen et al. (2015) as well as altering the kernel size, i.e., they limit their search space to simple architectures without, e.g., skip connections. + +# 3 NETWORK MORPHISM + +Let $\mathcal { N } ( \mathcal { X } )$ denote a set of neural networks defined on $\mathcal { X } \subset \mathbb { R } ^ { n }$ . A network morphism is a mapping $M : \dot { \mathcal { N } } ( \dot { \mathcal { X } } ) \times \mathbb { R } ^ { k } \mathcal { N } ( \underline { { \mathcal { X } } } ) \times \mathbb { R } ^ { j }$ from a neural network $f ^ { w } \in \mathcal { N } ( \mathcal { X } )$ with parameters $w \in \mathbb { R } ^ { k }$ to another neural network $g ^ { \tilde { w } } \in \mathcal { N } ( \mathcal { X } )$ with parameters $\tilde { w } \in \mathbb { R } ^ { j }$ so that + +$$ +f ^ { w } ( x ) = g ^ { \tilde { w } } ( x ) \mathrm { f o r e v e r y } x \in \mathcal { X } . +$$ + +In the following we give a few examples of network morphisms and how standard operations for building neural networks (e.g., adding a convolutional layer) can be expressed as a network morphism. For this, let $f _ { i } ^ { w _ { i } } ( x )$ be some part of a NN $f ^ { w } ( x )$ , e.g., a layer or a subnetwork. + +Network morphism Type I. We replace $f _ { i } ^ { w _ { i } }$ by + +$$ +\tilde { f } _ { i } ^ { \tilde { w } _ { i } } ( x ) = A f _ { i } ^ { w _ { i } } ( x ) + b , +$$ + +with $\tilde { w } _ { i } = ( w _ { i } , A , b ) ^ { 1 }$ . Equation (1) obviously holds for $A = \mathbf { 1 } , b = \mathbf { 0 }$ . This morphism can be used to add a fully-connected or convolutional layer, as these layers are simply linear mappings. Chen et al. (2015) dubbed this morphism ”Net2DeeperNet”. Alternatively to the above replacement, one could also choose + +$$ +\tilde { f } _ { i } ^ { \tilde { w } _ { i } } ( x ) = C ( A f _ { i } ^ { w _ { i } } ( x ) + b ) + d , +$$ + +with $\tilde { w } _ { i } = ( w _ { i } , C , d )$ . $A , b$ are fixed, non-learnable. In this case network morphism Equation (1) holds if $C = A ^ { - 1 } , d = - C b$ . A Batch Normalization layer (or other normalization layers) can be written in the above form: $A , b$ represent the batch statistics and $C , d$ the learnable scaling and shifting. + +Network morphism Type II. Assume $f _ { i } ^ { w _ { i } }$ has the form $f _ { i } ^ { w _ { i } } ( x ) = A h ^ { w _ { h } } ( x ) + b$ for an arbitrary function $h$ . We replace $f _ { i } ^ { w _ { i } }$ , $w _ { i } = ( w _ { h } , \dot { A } , b )$ , by + +$$ +\tilde { f } _ { i } ^ { \tilde { w } _ { i } } ( x ) = \left( A \quad \tilde { A } \right) \binom { h ^ { w _ { h } } ( x ) } { \tilde { h } ^ { w _ { h } } ( x ) } + b +$$ + +with an arbitrary function $\tilde { h } ^ { w _ { \tilde { h } } } ( x )$ . The new parameters are $\tilde { w } _ { i } = ( w _ { i } , w _ { \tilde { h } } , \tilde { A } )$ . Again, Equation (1) can trivially be satisfied by setting $\tilde { A } = 0$ . We think of two modifications of a NN which can be expressed by this morphism. Firstly, a layer can be widened (i.e., increasing the number of units in a fully connected layer or the number of channels in a CNN - the Net2WiderNet transformation from Chen et al. (2015)). Think of $h ( x )$ as the layer to be widened. For example, we can then set $\tilde { h } = h$ to simply double the width. Secondly, skip-connections by concatenation as used by Huang et al. (2016) can be formulated as a network morphism. If $h ( x )$ itself is a sequence of layers, $h ( x ) = h _ { n } ( x ) \circ \cdots \circ h _ { 0 } ( x )$ , then one could choose $\tilde { h } ( x ) = x$ to realize a skip from $h _ { 0 }$ to the layer subsequent to $h _ { n }$ . + +Network morphism Type III. By definition, every idempotent function $f _ { i } ^ { w _ { i } }$ can simply be replaced by + +$$ +f _ { i } ^ { ( w _ { i } , \tilde { w } _ { i } ) } = f _ { i } ^ { \tilde { w } _ { i } } \circ f _ { i } ^ { w _ { i } } +$$ + +with the initialization $\tilde { w } _ { i } = w _ { i }$ . This trivially also holds for idempotent function without weights, e.g., Relu. + +Network morphism Type IV. Every layer $f _ { i } ^ { w _ { i } }$ is replaceable by + +$$ +\tilde { f } _ { i } ^ { \tilde { w } _ { i } } ( x ) = \lambda f _ { i } ^ { w _ { i } } ( x ) + ( 1 - \lambda ) h ^ { w _ { h } } ( x ) , \quad \tilde { w } _ { i } = ( w _ { i } , \lambda , w _ { h } ) +$$ + +with an arbitrary function $h$ and Equation (1) holds if $\lambda$ is initialized as 1. This morphism can be used to incorporate any function, especially any non-linearities. For example, Wei et al. (2016) use a special case of this operator to deal with non-linear, non-idempotent activation functions. Another example wou(2016) to sim insertioing: If $f _ { i } ^ { w _ { i } }$ f an additiv itself is a s p connection,nce of layers, $f _ { i } ^ { w _ { i } } = f _ { i _ { n } } ^ { w _ { i _ { n } } } \circ \cdot \cdot \cdot \circ f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }$ He et al., then one $h ( x ) = x$ to realize a skip from $f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }$ to the layer subsequent to $f _ { i _ { n } } ^ { w _ { i _ { n } } }$ + +Note that every combinations of the network morphisms again yields a morphism. So one could for example insert a block ”Conv-BatchNorm-Relu” subsequent to a Relu layer by using equations (2), (3) and (5). + +![](images/f8ca3cbd2a4cdfcdcf35ac65c139d715df875d1be2779a18db0872654bca035b.jpg) +Figure 1: Visualization of our method. Based on the current best model, new models are generated and trained afterwards. The best model is than updated. + +# 4 ARCHITECTURE SEARCH BY NETWORK MORPHISMS + +Our proposed algorithm is a simple hill climbing strategy (Russell & Norvig, 2009). We start with a small, (possibly) pretrained network. Then, we apply network morphisms to this initial network to generate larger ones that may perform better when trained further. These new “child” networks can be seen as neighbors of the initial “parent” network in the space of network architectures. Due to the network morphism Equation (1), the child networks start at the same performance as their parent. In essence, network morphisms can thus be seen as a way to initialize child networks to perform well, avoiding the expensive step of training them from scratch and thereby reducing the cost of their evaluation. The various child networks can then be trained further for a brief period of time to exploit the additional capacity obtained by the network morphism, and the search can move on to the best resulting child network. This constitutes one step of our proposed algorithm, which we dub Neural Architecture Search by Hill-climbing (NASH). NASH can execute this step several times until performance on a validation set saturates; we note that this greedy process may in principle get stuck in a poorly-performing region, from which it can never escape, but we did not find evidence for this in our experiments. + +Figure 1 visualizes one step of the NASH approach, and Algorithm 1 provides full details for the algorithm. In our implementation, the function $A p p l y N e t M o r p h ( m o d e l , n )$ (line 15) applies $n$ network morphisms, each of them sampled uniformly at random from the following three: + +• Make the network deeper, i.e., add a ”Conv-BatchNorm-Relu” block as described at the end of Section 3. The position where to add the block, as well as the kernel size $( \in \{ 3 , 5 \} )$ , are uniformly sampled. The number of channels is chosen to be equal to he number of channels of the closest preceding convolution. +• Make the network wider, i.e., increase the number of channels by using the network morphism type II. The conv layer to be widened, as well as the widening factor $( \in \{ 2 , 4 \} )$ are sampled uniformly at random. +• Add a skip connection from layer i to layer j (either by concatenation or addition – uniformly sampled) by using network morphism type II or IV, respectively. Layers i and j are also sampled uniformly. + +Note that the current best model is also considered as a child, i.e. our algorithm is not forced to select a new model but can rather also keep the old one if no other one improves upon it. + +# Algorithm 1 Network architecture search by hill climbing + +$1 \backslash \mathrm { ~ f u n c t i o n ~ N A S H } ( m o d e l _ { 0 } , n _ { s t e p s } , n _ { n e i g h } , n _ { N M } , e p o c h _ { n e i g h } , e p o c h _ { f i n a l } , \lambda _ { e n d } , \lambda _ { s t a r t } )$ +2: +3: # $m o d e l _ { 0 }$ , model to start with, $n _ { s t e p s } \triangleq$ number of hill climbining steps +4: # $n _ { n e i g h } \triangleq$ number of neighbours, $n _ { N M } \triangleq$ number of net. morph. applied +5: # $e p o c h _ { n e i g h } \triangleq$ number of epochs for training every neighbour +6: # $e p o c h _ { f i n a l } \triangleq$ number of epochs for final training +7: # initial LR $\lambda _ { s t a r t }$ is annealed to $\lambda _ { e n d }$ during SGDR training +8: +9: $m o d e l _ { b e s t } m o d e l _ { 0 }$ +10: # start hill climbing +11: for $i \gets 1 , \dots , n _ { s t e p s } \ { \bf d }$ o +12: #get $n _ { n e i g h }$ neighbors of $m o d e l _ { 0 }$ by applying $n _ { N M }$ network morphisms to $m o d e l _ { b e s t }$ +13: for $j \gets 1 , \dots , n _ { n e i g h } - 1$ do +14: $m o d e l _ { j } \gets A p p l y N e t M o r p h s ( m o d e l _ { b e s t } , n _ { N M } )$ +15: # train for a few epochs on training set with SGDR +16: $m o d e l _ { j } \gets \mathrm { S G D R t r a i n } ( m o d e l _ { j } , e \bar { p } o c h _ { n e i g h } , \lambda _ { s t a r t } , \lambda _ { e n d } )$ +17: end for +18: # in fact, last neighbor is always just the current best +19: $m o d e l _ { n _ { n e i g h } } \gets \mathrm { S G D R t r a i n } ( m o d e l _ { b e s t } , e p o c h _ { n e i g h } , \lambda _ { s t a r t } , \lambda _ { e n d } )$ +20: # get best model on validation set +21: $\underbrace { \Bre { \_ } { a } } { m o d } e l _ { b e s t } \underbrace { a r g m a x } _ { j = 1 , \dots , n _ { n e i g h } } \{ p e r f o r m a n c e _ { v a l i } ( m o d e l _ { j } ) \}$ +22: end for +23: # train the final model on training and validation set +24: $m o d e l _ { b e s t } \gets \mathrm { S G D R t r a i n } ( m o d e \bar { l } _ { b e s t } , e p o c h _ { f i n a l } , \lambda _ { s t a r t } , \lambda _ { e n d } )$ +25: return modelbest +26: end function + +It is important for our method that child networks only need to be trained for a few epochs2 (line 17). Hence, an optimization algorithm with good anytime performance is required. Therefore, we chose the cosine annealing strategy from Loshchilov & Hutter (2017), whereas the learning rate is implicitly restarted: the training in line 17 always starts with a learning rate $\lambda _ { s t a r t }$ which is annealed to $\lambda _ { e n d }$ after epochneigh epochs. We use the same learning rate scheduler in the final training (aside from a different number of epochs). + +While we presented our method as a simple hill-climbing method, we note that it can also be interpreted as a very simple evolutionary algorithm with a population size of $n _ { n e i g h }$ , no cross-over, network morphisms as mutations, and a selection mechanism that only considers the best-performing population member as the parent for the next generation. This interpretation also suggests several promising possibilities for extending our simple method. + +# 5 EXPERIMENTS + +We evaluate our method on CIFAR-10 and CIFAR-100. First, we investigate whether our considerations from the previous chapter coincide with empirical results. We also check if the interplay of modifying and training networks harms their eventual performance. Finally, we compare our proposed method with other automated architecture algorithms as well as hand crafted architectures. + +We use the same standard data augmentation scheme for both CIFAR datasets used by Loshchilov & Hutter (2017) in all of the following experiments. The training set (50.000 samples) is split up in training (40.000) and validation (10.000) set for the purpose of architecture search. Eventually the performance is evaluated on the test set. All experiments where run on Nvidia Titan X (Maxwell) + +GPUs, with code implemented in Keras (Chollet et al., 2015) with a TensorFlow (Abadi et al., 2015) backend. + +# 5.1 EXPERIMENTS ON CIFAR-10 + +# 5.1.1 BASELINES + +Before comparing our method to others, we run some baseline experiments to see whether our considerations from the previous chapter coincide with empirical data. + +Random model selection. First, we investigate if the simple hill climbing strategy is able to distinguish between models with high and low performance. For this, we set $n _ { n e i g h } = 1$ , i.e., there is no model selction - we simply construct random networks and train them. We then run experiments with $n _ { n e i g h } = 8$ and compare both results. All other parameters are the same in this experiment, namely $n _ { s t e p s } = 5 , n _ { N M } = 5 , e p o c h _ { n e i g h } = 1 7 , e p o c h _ { f i n a l } = 1 0 0$ . We choose $\lambda _ { s t a r t } = 0 . 0 5$ , $\lambda _ { e n d } = 0 . { \overset { \cdot } { 0 } }$ as done in Loshchilov & Hutter (2017). model0 was a simple conv net: Conv-MaxPool-Conv-MaxPool-Conv-FC-Softmax3, which is pretrained for 20 epochs, achieving $\approx 7 5 \%$ validation accuracy (up to $9 1 \%$ when trained till convergence), see Figure 5 in the appendix. If our algorithm is able to identify better networks, one would expect to get better results with the setting $n _ { n e i g h } = 8$ . + +Retraining from scratch. In the this experiment we investigate whether the ”weight inheritance” due to the network morphisms used in our algorithm harms the final performance of the final model. This weight inheritance can be seen as a strong prior on the weights and one could suspect that the new, larger model may not be able to overcome a possibly poor prior. Additionally we were interested in measuring the overhead of the architecture search process, so we compared the times for generating and training a model with the time needed when training the final model from scratch. The retraining from scratch is done for the same number of epochs as the total number of epochs spent to train the model returned by our algorithm4 . + +No SGDR. We now turn off the cosine annealing with restarts (SGDR) during the hill climbing stage, i.e., the training in line 17 of Algorithm 1 is done with a constant learning rate. We tried $\lambda \in \{ 0 . 0 1 , 0 . 0 2 5 , 0 . 0 5 \}$ , 10 runs each and averaged the results. Note that we still use the cosine decay for the final training. + +No network morphism. Lastly, we turn off the network morphism constraint for initializing the neighbor networks. In detail, we proceeded as Real et al. (2017): All weights from layer where now changes occur are inherited, whereas the weights of new/modified layers are initialized by random. + +The results for these experiments are summarized in Table 1 . The hill climbing strategy is actually able to identify better performing models. (first and second line: $5 . 7 \%$ vs. $6 . 5 \%$ ). Notice how hill climbing prefers larger models (5.7 million parameters on average vs. 4.4 million). Performance slightly decreases when the models are retrained from scratch (line 3). This experiments indicates that our algorithm does not harm the final performance of a model. Regarding the runtime, the overhead for first having to search for the architecture is roughly a factor 3. We think this is a big advantage of our method and shows that architecture search can be done in the same order of magnitude as training a single model. In line 4 we can see that SGDR plays an important role. The resulting models chosen by our algorithm when training is done with a constant learning rate perform similarly to the models without any model selection strategy $( 6 . 4 \%$ and $6 . 5 \%$ , respectively), which indicates that the performance after a few epochs on the validation set when trained without SGDR correlates less with the final performance on the test set as it is the case for training with SGDR. Indeed, we computed the Pearson correlation coefficient and obtained $R ^ { 2 } = 0 . 6 4$ for training with SGDR and and $\mathrm { \bar { \it R } ^ { 2 } = 0 . 3 6 }$ for training with a constant learning rate. See appendix A. Also, with the constant learning rate, the few epochs spent are not sufficient to improve the performance of the model. Figure 2 shows the progress while running our algorithm with and without SGDR, averaged over all runs. When turning off the network morphism constraint, performance also decreases. + +![](images/5bffa50e24c866e7a85ea7102470da8aa121189925b47b5e0807ef32b1fb1234.jpg) +Figure 2: The best model found by Algorithm 1 tracked over time (in terms of hill climbing iterations). With (red) and without (blue) using SGDR for the training within the hill climbing (line 17). Final training (line 24) is not plotted. Dashed line denotes mean, shaded area $\pm 2 \sigma$ intervalls. + +Table 1: Baseline experiments. Runtime, # params, and error rates are averaged over 10 runs (for $n _ { n e i g h } = 8 $ ) and 30 runs $( n _ { n e i g h } = 1 )$ ) runs, respectively. $n _ { s t e p s } = 5$ in all experiments. + +
algorithm settingruntime (hrs)# params (mil.)error ± std.( %)
nneigh =812.85.75.7 ± 0.35
Random networks (nneigh = 1)4.54.46.5 ± 0.76
models from line 1 retrained from scratch5.35.76.1 ± 0.92
nneigh = 8, no SGDR10.65.86.4±0.70
nneigh = 8, no net. morph.6.62.96.1 ±0.30
+ +Interestingly the number of parameters heavily decreases. This indicates that our algorithm prefers models without new parameters. + +# 5.1.2 COMPARISON TO HAND CRAFTED AND OTHER AUTOMATICALLY GENERATED ARCHITECTURES + +We now compare our algorithm against the popular wide residual networks (Zagoruyko & Komodakis, 2016), the state of the art model from Gastaldi (2017) as well as other automated architecture search methods. Beside our results for $n _ { s t e p s } = 5$ from the previous section, we also tried $n _ { s t e p s } = 8$ to generate larger models. + +For further improving the results, we take snapshots of the best models from every iteration while running our algorithm following the idea of Huang et al. (2017) when using SGDR (Loshchilov & Hutter, 2017) for training. However different from Huang et al. (2017), we do not immediately get fully trained models for free, as our snapshots are not yet trained on the validation set but rather only on the training set. Hence we spent some additional resources and train the snapshots on both sets. Afterwards the ensemble model is build by combining the snapshot models with uniform weights. Lastly, we also build an ensemble from the models returned by our algorithm across all runs. Results are listed in Table 2. + +The proposed method is able to generate competitive network architectures in only 12 hours. By spending another 12 hours, it outperforms most automated architecture search methods although all of them require (partially far) more time and GPUs. We do not reach the performance of the two handcrafted architectures as well as the ones found by Zoph & Le (2017) and Brock et al. (2017). However note that Zoph & Le (2017) spent by far more resources than we did. + +Unsurprisingly, the ensemble models perform better. It is a simple and cheap way to improve results which everyone can consider when the number of parameters is not relevant. + +Table 2: Results for CIFAR-10. For our methods the stated resources, # parameters and errors are averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models. + +
modelresources spent# params (mil.)error(%)
Shake-Shake (Gastaldi,2017)4 GPU days,2 GPUs262.9
WRN 28-10 (Loshchilov & Hutter,2017)1 GPU day36.53.86
Baker et al. (2016)80-100 GPU days116.9
Cai et al. (2017)15 GPU days19.75.7
Zoph & Le (2017)16.000-24.000 GPU days37.53.65
Real et al. (2017)2500 GPU days5.45.4
Saxena & Verbeek (2016)？217.4
Brock et al. (2017)3 GPU days16.04.0
Ours (random networks,nsteps = 5, nneigh =1)0.2 GPU days4.46.5
Ours (nsteps = 5,nneigh =8,10 runs)0.5 GPU days5.75.7
Ours (nsteps = 8,nneigh = 8,4 runs)1 GPU day19.75.2
Ours (snapshot ensemble,4 runs)2 GPU days57.84.7
Ours (ensemble across runs)4 GPU days884.4
+ +Table 3: Results for CIFAR-100. For our methods the stated resources, # parameters and errors are averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models. + +
modelresources spent# params (mil.)error(%)
15.9
Shake-Shake (Gastaldi, 2017) WRN 28-10 (Loshchilov & Hutter,2017)14 GPU days34.4
1 GPU day36.519.6
Real et al. (2017)250 GPUs40.423.7
Brock et al. (2017)3 GPU days16.020.6
Ours (nsteps = 8,nneigh = 8,5 runs)1 GPU day22.323.4
Ours (snapshot ensemble,5 runs)2 GPU days73.320.9
Ours (ensemble across runs)5 GPU days111.519.6
+ +# 5.2 EXPERIMENTS ON CIFAR-100 + +We repeat the previous experiment on CIFAR-100; hyperparameters were not changed. The results are listed in Table 3. Unfortunately most automated architecture methods did not consider CIFAR100. Our method is on a par with Real et al. (2017) after one day with a single GPU. The snapshot ensemble performs similar to Brock et al. (2017) and an ensemble model build from the 5 runs can compete with the hand crafted WRN 28-10. The performance of the Shake-Shake network (Gastaldi, 2017) is again not reached. + +# 6 CONCLUSION + +We proposed NASH, a simple and fast method for automated architecture search based on a hill climbing strategy, network morphisms, and training via SGDR. Experiments on CIFAR-10 and CIFAR-100 showed that our method yields competitive results while requiring considerably less computational resources than most alternative approaches. Our algorithm is easily extendable, e.g., by other network morphisms, evolutionary approaches for generating new models, other methods for cheap performance evaluation (such as, e.g., learning curve prediction (Klein et al., 2017) or hypernetworks (Ha et al., 2017; Brock et al., 2017)), or better resource handling strategies (such as Hyperband (Li et al., 2016b)). In this sense, we hope that our approach can serve as a basis for the development of more sophisticated methods that yield further improvements of performance. + +# REFERENCES + +Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´ Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´ berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL https://www.tensorflow.org/. Software available from tensorflow.org. + +Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ICLR, 2015. + +Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. Designing neural network architectures using reinforcement learning. ICLR 2017, 2016. + +J. Bergstra and Y. Bengio. Random search for hyper-parameter optimization. 13(1):281–305, 2012. + +J. Bergstra, R. Bardenet, Y. Bengio, and B. Kegl. Algorithms for hyper-parameter optimization. In ´ NIPS, 2011. + +Andrew Brock, Theodore Lim, James M. Ritchie, and Nick Weston. SMASH: one-shot model architecture search through hypernetworks. arXiv preprint, 2017. + +Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang. Reinforcement learning for architecture search by network transformation. 2017. + +Tianqi Chen, Ian J. Goodfellow, and Jonathon Shlens. Net2net: Accelerating learning via knowledge transfer. arXiv preprint, 2015. + +Franc¸ois Chollet et al. Keras. https://github.com/fchollet/keras, 2015. + +Xavier Gastaldi. Shake-shake regularization. ICLR 2017 Workshop, 2017. + +David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. ICLR, 2017. + +Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. + +Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. Deep neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine, 2012. + +Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. 2016. + +Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger. Snapshot ensembles: Train 1, get M for free. ICLR 2017, 2017. + +A. Klein, S. Falkner, J. T. Springenberg, and F. Hutter. Learning curve prediction with Bayesian neural networks. In International Conference on Learning Representations (ICLR) 2017 Conference Track, April 2017. + +Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 25, pp. 1097–1105. Curran Associates, Inc., 2012. + +L. Li, K. G. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar. Efficient hyperparameter optimization and infinitely many armed bandits. CoRR, abs/1603.06560, 2016a. + +Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Efficient hyperparameter optimization and infinitely many armed bandits. arXiv preprint, 2016b. +I. Loshchilov and F. Hutter. CMA-ES for hyperparameter optimization of deep neural networks. CoRR, abs/1604.07269, 2016. +I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations (ICLR) 2017 Conference Track, April 2017. +H. Mendoza, A. Klein, M. Feurer, T. Springenberg, and F. Hutter. Towards automatically-tuned neural networks. In AutoML, 2016. +Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V. Le, and Alex Kurakin. Large-scale evolution of image classifiers. arXiv preprint, 2017. +Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach (3rd Edition). Pearson, 3 edition, December 2009. ISBN 0136042597. +Shreyas Saxena and Jakob Verbeek. Convolutional neural fabrics. arXiv preprint, 2016. +J. Snoek, H. Larochelle, and R.P. Adams. Practical Bayesian optimization of machine learning algorithms. In NIPS, 2012. +Masanori Suganuma, Shinichi Shirakawa, and Tomoharu Nagao. A genetic programming approach to designing convolutional neural network architectures. arXiv preprint, 2017. +Tao Wei, Changhu Wang, Yong Rui, and Chang Wen Chen. Network morphism. arXiv preprint, 2016. +Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint, 2016. +Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. 2017. +Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning transferable architectures for scalable image recognition. 2017. + +![](images/2edbb1405729ce5dc046dc80f49bd9380081ebed874aa93e2d296d2395a7ddf0.jpg) +Figure 3: Initial network for our algorithm. + +![](images/213925428f8eea0c7c284680e760c96a1baa86df2685b0480330288a33417347.jpg) +Figure 4: Initial network for our algorithm. + +![](images/926480ca8e0fe0b1fac91e9b0c6c21dd16e70a4e522edfe39777971a2e331b4d.jpg) +Figure 5: Initial network for our algorithm. + +![](images/091ad5bd88208c13ba32695bfc45a25363ea4c4e47109a8acba5e63354b8c3d1.jpg) +Figure 6: Network generated by our algorithm with $n _ { s t e p s } = 5$ . + +![](images/e810d4abf64e5a83ec8d5f1149515b76d979a3c8fc66b0153106f09e2b3c3c2d.jpg) +Figure 7: Network generated by our algorithm with $n _ { s t e p s } = 8$ . \ No newline at end of file diff --git a/parse/train/SySaJ0xCZ/SySaJ0xCZ_content_list.json b/parse/train/SySaJ0xCZ/SySaJ0xCZ_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..535367fa7e912c9c8c5039e8c77a3eae7fe4cef9 --- /dev/null +++ b/parse/train/SySaJ0xCZ/SySaJ0xCZ_content_list.json @@ -0,0 +1,1138 @@ +[ + { + "type": "text", + "text": "SIMPLE AND EFFICIENT ARCHITECTURE SEARCH FOR CONVOLUTIONAL NEURAL NETWORKS ", + "text_level": 1, + "bbox": [ + 176, + 98, + 821, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural networks have recently had a lot of success for many tasks. However, neural network architectures that perform well are still typically designed manually by experts in a cumbersome trial-and-error process. We propose a new method to automatically search for well-performing CNN architectures based on a simple hill climbing procedure whose operators apply network morphisms, followed by short optimization runs by cosine annealing. Surprisingly, this simple method yields competitive results, despite only requiring resources in the same order of magnitude as training a single network. E.g., on CIFAR-10, our method designs and trains networks with an error rate below $6 \\%$ in only 12 hours on a single GPU; training for one day reduces this error further, to almost $5 \\%$ . ", + "bbox": [ + 233, + 267, + 764, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 435, + 336, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural networks have rapidly gained popularity over the last few years due to their success in a variety of tasks, such as image recognition (Krizhevsky et al., 2012), speech recognition (Hinton et al., 2012) and machine translation (Bahdanau et al., 2015). In most cases, these neural networks are still designed by hand, which is an exhausting, time-consuming process. Additionally, the vast amount of possible configurations requires expert knowledge to restrict the search. Therefore, a natural goal is to design optimization algorithms that automate this neural architecture search. ", + "bbox": [ + 174, + 468, + 825, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, most classic optimization algorithms do not apply to this problem, since the architecture search space is discrete (e.g., number of layers, layer types) and conditional (e.g., the number of parameters defining a layer depends on the layer type). Thus, methods that rely on, e.g., differentiability or independent parameters are not applicable. This led to a growing interest in using evolutionary algorithms (Real et al., 2017; Suganuma et al., 2017) and reinforcement learning (Baker et al., 2016; Cai et al., 2017; Zoph & Le, 2017) for automatically designing CNN architectures. Unfortunately, most proposed methods are either very costly (requiring hundreds or thousands of GPU days) or yield non-competitive performance. ", + "bbox": [ + 174, + 558, + 823, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to dramatically reduce these computational costs while still achieving competitive performance. Specifically, our contributions are as follows: ", + "bbox": [ + 173, + 676, + 821, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "• We propose a baseline method that randomly constructs networks and trains them with SGDR (Loshchilov & Hutter, 2017). We demonstrate that this simple baseline achieves $6 \\% - 7 \\%$ test error on CIFAR-10, which already rivals several existing methods for neural archictecture search. Due to its simplicity, we hope that this baseline provides a valuable starting point for the development of more sophisticated methods in the future. \n• We formalize and extend the work on network morphisms (Chen et al., 2015; Wei et al., 2016; Cai et al., 2017) in order to provide popular network building blocks, such as skip connections and batch normalization. \n• We propose Neural Architecture Search by Hillclimbing (NASH), a simple iterative approach that, at each step, applies a set of alternative network morphisms to the current network, trains the resulting child networks with short optimization runs of cosine annealing (Loshchilov & Hutter, 2017), and moves to the most promising child network. NASH finds and trains competitive architectures at a computational cost of the same order of magnitude as training a single network; e.g., on CIFAR-10, NASH finds and trains CNNs with an error rate below $6 \\%$ in roughly 12 hours on ", + "bbox": [ + 173, + 717, + 826, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a single GPU. After one day the error is reduced to almost $5 \\%$ . Models from different stages of our algorithm can be combined to achieve an error of $4 . 7 \\%$ within two days on a single GPU. On CIFAR-100, we achieve an error below $24 \\%$ in one day and get close to $20 \\%$ after two days. ", + "bbox": [ + 186, + 103, + 825, + 146 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "• Our method is easy to use and easy to extend, so it hopefully can serve as a basis for future work. ", + "bbox": [ + 173, + 160, + 821, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first discuss related work in Section 2. Then, we formalize the concept of network morphisms in Section 3 and propose our architecture search methods based on them in Section 4. We evaluate our methods in Section 5 and conclude in Section 6. ", + "bbox": [ + 174, + 191, + 823, + 234 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 267, + 341, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hyperparameter optimization. Neural networks are known to be quite sensitive to the setting of various hyperparameters, such as learning rates and regularization constants. There exists a long line of research on automated methods for setting these hyperparameters, including, e.g., random search (Bergstra & Bengio, 2012), Bayesian optimization (Bergstra et al., 2011; Snoek et al., 2012), bandit-based approaches (Li et al., 2016a), and evolutionary strategies (Loshchilov & Hutter, 2016). ", + "bbox": [ + 174, + 305, + 825, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Automated architecture search. In recent years, the research focus has shifted from optimizing hyperparameters to optimizing architectures. While architectural choices can be treated as categorical hyperparameters and be optimized with standard hyperparameter optimization methods (Bergstra et al., 2011; Mendoza et al., 2016), the current focus is on the development of special techniques for architectural optimization. One very popular approach is to train a reinforcement learning agent with the objective of designing well-performing convolutional neural networks (Baker et al., 2016; Zoph & Le, 2017; Cai et al., 2017). Baker et al. (2016) train an RL agent to sequentially choose the type of layers (convolution, pooling, fully connected) and their parameters. Zoph & Le (2017) use a recurrent neural network controller to sequentially generate a string representing the network architecture. Both approaches train their generated networks from scratch and evaluate their performance on a validation set, which represents a very costly step. In a follow-up work (Zoph et al., 2017), the RL agent learned to build cells, which are then used as building blocks for a neural network with a fixed global structure. Unfortunately, training an RL agent with the objective of designing architecture is extremely expensive: both Baker et al. (2016) and Zoph & Le (2017) required over 10.000 fully trained networks, requiring hundreds to thousands of GPU days. To overcome this drawback, Cai et al. (2017) proposed to apply the concept of network transformations/morphisms within RL. As in our (independent, parallel) work, the basic idea is to use the these transformation to generate new pre-trained architectures to avoid the large cost of training all networks from scratch. Compared to this work, our approach is much simpler and 15 times faster while obtaining better performance. ", + "bbox": [ + 173, + 382, + 825, + 645 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real et al. (2017) and Suganuma et al. (2017) utilized evolutionary algorithms to iteratively generate powerful networks from a small network. Operations like inserting a layer, modifying the parameters of a layer or adding skip connections serve as ”mutations” in their framework of evolution. Whereas Real et al. (2017) also used enormous computational resources (250 GPUs, 10 days), Suganuma et al. (2017) were restricted to relatively small networks due to handling a population of networks. In contrast to the previous methods where network capacity increases over time, Saxena & Verbeek (2016) start with training a large network (a ”convolution neural fabric”) and prune this in the end. Very recently, Brock et al. (2017) used hypernetworks (Ha et al., 2017) to generate the weights for a randomly sampled network architecture with the goal of eliminating the costly process of training a vast amount of networks. ", + "bbox": [ + 174, + 652, + 825, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Network morphism/ transformation. Network transformations were (to our knowledge) first introduced by Chen et al. (2015) in the context of transfer learning. The authors described a function preserving operation to make a network deeper (dubbed ”Net2Deeper”) or wider (”Net2Wider”) with the goal of speeding up training and exploring network architectures. Wei et al. (2016) proposed additional operations, e.g., for handling non-idempotent activation functions or altering the kernel size and introduced the term network morphism. As mentioned above, Cai et al. (2017) used network morphisms for architecture search, though they just employ the Net2Deeper and Net2Wider operators from Chen et al. (2015) as well as altering the kernel size, i.e., they limit their search space to simple architectures without, e.g., skip connections. ", + "bbox": [ + 174, + 799, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 NETWORK MORPHISM ", + "text_level": 1, + "bbox": [ + 174, + 102, + 390, + 117 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal { N } ( \\mathcal { X } )$ denote a set of neural networks defined on $\\mathcal { X } \\subset \\mathbb { R } ^ { n }$ . A network morphism is a mapping $M : \\dot { \\mathcal { N } } ( \\dot { \\mathcal { X } } ) \\times \\mathbb { R } ^ { k } \\mathcal { N } ( \\underline { { \\mathcal { X } } } ) \\times \\mathbb { R } ^ { j }$ from a neural network $f ^ { w } \\in \\mathcal { N } ( \\mathcal { X } )$ with parameters $w \\in \\mathbb { R } ^ { k }$ to another neural network $g ^ { \\tilde { w } } \\in \\mathcal { N } ( \\mathcal { X } )$ with parameters $\\tilde { w } \\in \\mathbb { R } ^ { j }$ so that ", + "bbox": [ + 173, + 132, + 825, + 175 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/01da69f32520eccbef16ce0ecf9fbbcf2637c41fb3974f8b8230761a9c0e4b45.jpg", + "text": "$$\nf ^ { w } ( x ) = g ^ { \\tilde { w } } ( x ) \\mathrm { f o r e v e r y } x \\in \\mathcal { X } .\n$$", + "text_format": "latex", + "bbox": [ + 385, + 178, + 612, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following we give a few examples of network morphisms and how standard operations for building neural networks (e.g., adding a convolutional layer) can be expressed as a network morphism. For this, let $f _ { i } ^ { w _ { i } } ( x )$ be some part of a NN $f ^ { w } ( x )$ , e.g., a layer or a subnetwork. ", + "bbox": [ + 174, + 204, + 825, + 248 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network morphism Type I. We replace $f _ { i } ^ { w _ { i } }$ by ", + "bbox": [ + 174, + 253, + 491, + 268 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/739fb4573c092613f8d0559923d6bb3981e5fbbefbec26a727c76f98d046b2eb.jpg", + "text": "$$\n\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = A f _ { i } ^ { w _ { i } } ( x ) + b ,\n$$", + "text_format": "latex", + "bbox": [ + 416, + 271, + 580, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\tilde { w } _ { i } = ( w _ { i } , A , b ) ^ { 1 }$ . Equation (1) obviously holds for $A = \\mathbf { 1 } , b = \\mathbf { 0 }$ . This morphism can be used to add a fully-connected or convolutional layer, as these layers are simply linear mappings. Chen et al. (2015) dubbed this morphism ”Net2DeeperNet”. Alternatively to the above replacement, one could also choose ", + "bbox": [ + 173, + 292, + 825, + 347 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/7999d9061602778300a91a1b4c09054ac86a6cafaa7aa40fb422aa6a160d5ca7.jpg", + "text": "$$\n\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = C ( A f _ { i } ^ { w _ { i } } ( x ) + b ) + d ,\n$$", + "text_format": "latex", + "bbox": [ + 390, + 345, + 606, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\tilde { w } _ { i } = ( w _ { i } , C , d )$ . $A , b$ are fixed, non-learnable. In this case network morphism Equation (1) holds if $C = A ^ { - 1 } , d = - C b$ . A Batch Normalization layer (or other normalization layers) can be written in the above form: $A , b$ represent the batch statistics and $C , d$ the learnable scaling and shifting. ", + "bbox": [ + 173, + 372, + 825, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network morphism Type II. Assume $f _ { i } ^ { w _ { i } }$ has the form $f _ { i } ^ { w _ { i } } ( x ) = A h ^ { w _ { h } } ( x ) + b$ for an arbitrary function $h$ . We replace $f _ { i } ^ { w _ { i } }$ , $w _ { i } = ( w _ { h } , \\dot { A } , b )$ , by ", + "bbox": [ + 173, + 434, + 820, + 464 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/f4578e30ebce96eb214241b79c83d5935f0306cd6cc67ef9b418c714979d56e3.jpg", + "text": "$$\n\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\left( A \\quad \\tilde { A } \\right) \\binom { h ^ { w _ { h } } ( x ) } { \\tilde { h } ^ { w _ { h } } ( x ) } + b\n$$", + "text_format": "latex", + "bbox": [ + 382, + 465, + 616, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with an arbitrary function $\\tilde { h } ^ { w _ { \\tilde { h } } } ( x )$ . The new parameters are $\\tilde { w } _ { i } = ( w _ { i } , w _ { \\tilde { h } } , \\tilde { A } )$ . Again, Equation (1) can trivially be satisfied by setting $\\tilde { A } = 0$ . We think of two modifications of a NN which can be expressed by this morphism. Firstly, a layer can be widened (i.e., increasing the number of units in a fully connected layer or the number of channels in a CNN - the Net2WiderNet transformation from Chen et al. (2015)). Think of $h ( x )$ as the layer to be widened. For example, we can then set $\\tilde { h } = h$ to simply double the width. Secondly, skip-connections by concatenation as used by Huang et al. (2016) can be formulated as a network morphism. If $h ( x )$ itself is a sequence of layers, $h ( x ) = h _ { n } ( x ) \\circ \\cdots \\circ h _ { 0 } ( x )$ , then one could choose $\\tilde { h } ( x ) = x$ to realize a skip from $h _ { 0 }$ to the layer subsequent to $h _ { n }$ . ", + "bbox": [ + 173, + 503, + 825, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network morphism Type III. By definition, every idempotent function $f _ { i } ^ { w _ { i } }$ can simply be replaced by ", + "bbox": [ + 169, + 643, + 823, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/eacf17365fe1bb32a28659b1c5a881dee0fdc9ac4b0b594b9ccefe3e1423f078.jpg", + "text": "$$\nf _ { i } ^ { ( w _ { i } , \\tilde { w } _ { i } ) } = f _ { i } ^ { \\tilde { w } _ { i } } \\circ f _ { i } ^ { w _ { i } }\n$$", + "text_format": "latex", + "bbox": [ + 426, + 671, + 570, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with the initialization $\\tilde { w } _ { i } = w _ { i }$ . This trivially also holds for idempotent function without weights, e.g., Relu. ", + "bbox": [ + 176, + 691, + 825, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Network morphism Type IV. Every layer $f _ { i } ^ { w _ { i } }$ is replaceable by ", + "bbox": [ + 174, + 726, + 598, + 741 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/fd47040a01a63f7e977805f8e17a1b0019034a8e830d36d15a69e1bb6c4d3d41.jpg", + "text": "$$\n\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\lambda f _ { i } ^ { w _ { i } } ( x ) + ( 1 - \\lambda ) h ^ { w _ { h } } ( x ) , \\quad \\tilde { w } _ { i } = ( w _ { i } , \\lambda , w _ { h } )\n$$", + "text_format": "latex", + "bbox": [ + 305, + 744, + 691, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with an arbitrary function $h$ and Equation (1) holds if $\\lambda$ is initialized as 1. This morphism can be used to incorporate any function, especially any non-linearities. For example, Wei et al. (2016) use a special case of this operator to deal with non-linear, non-idempotent activation functions. Another example wou(2016) to sim insertioing: If $f _ { i } ^ { w _ { i } }$ f an additiv itself is a s p connection,nce of layers, $f _ { i } ^ { w _ { i } } = f _ { i _ { n } } ^ { w _ { i _ { n } } } \\circ \\cdot \\cdot \\cdot \\circ f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }$ He et al., then one $h ( x ) = x$ to realize a skip from $f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }$ to the layer subsequent to $f _ { i _ { n } } ^ { w _ { i _ { n } } }$ ", + "bbox": [ + 173, + 765, + 825, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note that every combinations of the network morphisms again yields a morphism. So one could for example insert a block ”Conv-BatchNorm-Relu” subsequent to a Relu layer by using equations (2), (3) and (5). ", + "bbox": [ + 174, + 859, + 825, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8ca3cbd2a4cdfcdcf35ac65c139d715df875d1be2779a18db0872654bca035b.jpg", + "image_caption": [ + "Figure 1: Visualization of our method. Based on the current best model, new models are generated and trained afterwards. The best model is than updated. " + ], + "image_footnote": [], + "bbox": [ + 173, + 98, + 826, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4 ARCHITECTURE SEARCH BY NETWORK MORPHISMS ", + "text_level": 1, + "bbox": [ + 176, + 405, + 640, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed algorithm is a simple hill climbing strategy (Russell & Norvig, 2009). We start with a small, (possibly) pretrained network. Then, we apply network morphisms to this initial network to generate larger ones that may perform better when trained further. These new “child” networks can be seen as neighbors of the initial “parent” network in the space of network architectures. Due to the network morphism Equation (1), the child networks start at the same performance as their parent. In essence, network morphisms can thus be seen as a way to initialize child networks to perform well, avoiding the expensive step of training them from scratch and thereby reducing the cost of their evaluation. The various child networks can then be trained further for a brief period of time to exploit the additional capacity obtained by the network morphism, and the search can move on to the best resulting child network. This constitutes one step of our proposed algorithm, which we dub Neural Architecture Search by Hill-climbing (NASH). NASH can execute this step several times until performance on a validation set saturates; we note that this greedy process may in principle get stuck in a poorly-performing region, from which it can never escape, but we did not find evidence for this in our experiments. ", + "bbox": [ + 174, + 444, + 825, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1 visualizes one step of the NASH approach, and Algorithm 1 provides full details for the algorithm. In our implementation, the function $A p p l y N e t M o r p h ( m o d e l , n )$ (line 15) applies $n$ network morphisms, each of them sampled uniformly at random from the following three: ", + "bbox": [ + 174, + 646, + 825, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "• Make the network deeper, i.e., add a ”Conv-BatchNorm-Relu” block as described at the end of Section 3. The position where to add the block, as well as the kernel size $( \\in \\{ 3 , 5 \\} )$ , are uniformly sampled. The number of channels is chosen to be equal to he number of channels of the closest preceding convolution. \n• Make the network wider, i.e., increase the number of channels by using the network morphism type II. The conv layer to be widened, as well as the widening factor $( \\in \\{ 2 , 4 \\} )$ are sampled uniformly at random. \n• Add a skip connection from layer i to layer j (either by concatenation or addition – uniformly sampled) by using network morphism type II or IV, respectively. Layers i and j are also sampled uniformly. ", + "bbox": [ + 215, + 705, + 825, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the current best model is also considered as a child, i.e. our algorithm is not forced to select a new model but can rather also keep the old one if no other one improves upon it. ", + "bbox": [ + 176, + 895, + 823, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Network architecture search by hill climbing ", + "text_level": 1, + "bbox": [ + 173, + 103, + 557, + 117 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$1 \\backslash \\mathrm { ~ f u n c t i o n ~ N A S H } ( m o d e l _ { 0 } , n _ { s t e p s } , n _ { n e i g h } , n _ { N M } , e p o c h _ { n e i g h } , e p o c h _ { f i n a l } , \\lambda _ { e n d } , \\lambda _ { s t a r t } )$ \n2: \n3: # $m o d e l _ { 0 }$ , model to start with, $n _ { s t e p s } \\triangleq$ number of hill climbining steps \n4: # $n _ { n e i g h } \\triangleq$ number of neighbours, $n _ { N M } \\triangleq$ number of net. morph. applied \n5: # $e p o c h _ { n e i g h } \\triangleq$ number of epochs for training every neighbour \n6: # $e p o c h _ { f i n a l } \\triangleq$ number of epochs for final training \n7: # initial LR $\\lambda _ { s t a r t }$ is annealed to $\\lambda _ { e n d }$ during SGDR training \n8: \n9: $m o d e l _ { b e s t } m o d e l _ { 0 }$ \n10: # start hill climbing \n11: for $i \\gets 1 , \\dots , n _ { s t e p s } \\ { \\bf d }$ o \n12: #get $n _ { n e i g h }$ neighbors of $m o d e l _ { 0 }$ by applying $n _ { N M }$ network morphisms to $m o d e l _ { b e s t }$ \n13: for $j \\gets 1 , \\dots , n _ { n e i g h } - 1$ do \n14: $m o d e l _ { j } \\gets A p p l y N e t M o r p h s ( m o d e l _ { b e s t } , n _ { N M } )$ \n15: # train for a few epochs on training set with SGDR \n16: $m o d e l _ { j } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { j } , e \\bar { p } o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )$ \n17: end for \n18: # in fact, last neighbor is always just the current best \n19: $m o d e l _ { n _ { n e i g h } } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { b e s t } , e p o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )$ \n20: # get best model on validation set \n21: $\\underbrace { \\Bre { \\_ } { a } } { m o d } e l _ { b e s t } \\underbrace { a r g m a x } _ { j = 1 , \\dots , n _ { n e i g h } } \\{ p e r f o r m a n c e _ { v a l i } ( m o d e l _ { j } ) \\}$ \n22: end for \n23: # train the final model on training and validation set \n24: $m o d e l _ { b e s t } \\gets \\mathrm { S G D R t r a i n } ( m o d e \\bar { l } _ { b e s t } , e p o c h _ { f i n a l } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )$ \n25: return modelbest \n26: end function ", + "bbox": [ + 174, + 121, + 815, + 498 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important for our method that child networks only need to be trained for a few epochs2 (line 17). Hence, an optimization algorithm with good anytime performance is required. Therefore, we chose the cosine annealing strategy from Loshchilov & Hutter (2017), whereas the learning rate is implicitly restarted: the training in line 17 always starts with a learning rate $\\lambda _ { s t a r t }$ which is annealed to $\\lambda _ { e n d }$ after epochneigh epochs. We use the same learning rate scheduler in the final training (aside from a different number of epochs). ", + "bbox": [ + 174, + 531, + 825, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While we presented our method as a simple hill-climbing method, we note that it can also be interpreted as a very simple evolutionary algorithm with a population size of $n _ { n e i g h }$ , no cross-over, network morphisms as mutations, and a selection mechanism that only considers the best-performing population member as the parent for the next generation. This interpretation also suggests several promising possibilities for extending our simple method. ", + "bbox": [ + 174, + 621, + 825, + 690 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 715, + 326, + 731 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on CIFAR-10 and CIFAR-100. First, we investigate whether our considerations from the previous chapter coincide with empirical results. We also check if the interplay of modifying and training networks harms their eventual performance. Finally, we compare our proposed method with other automated architecture algorithms as well as hand crafted architectures. ", + "bbox": [ + 174, + 748, + 825, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the same standard data augmentation scheme for both CIFAR datasets used by Loshchilov & Hutter (2017) in all of the following experiments. The training set (50.000 samples) is split up in training (40.000) and validation (10.000) set for the purpose of architecture search. Eventually the performance is evaluated on the test set. All experiments where run on Nvidia Titan X (Maxwell) ", + "bbox": [ + 174, + 811, + 825, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GPUs, with code implemented in Keras (Chollet et al., 2015) with a TensorFlow (Abadi et al., 2015) backend. ", + "bbox": [ + 171, + 103, + 825, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 EXPERIMENTS ON CIFAR-10 ", + "text_level": 1, + "bbox": [ + 176, + 152, + 419, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1.1 BASELINES", + "text_level": 1, + "bbox": [ + 174, + 179, + 308, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before comparing our method to others, we run some baseline experiments to see whether our considerations from the previous chapter coincide with empirical data. ", + "bbox": [ + 174, + 207, + 821, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Random model selection. First, we investigate if the simple hill climbing strategy is able to distinguish between models with high and low performance. For this, we set $n _ { n e i g h } = 1$ , i.e., there is no model selction - we simply construct random networks and train them. We then run experiments with $n _ { n e i g h } = 8$ and compare both results. All other parameters are the same in this experiment, namely $n _ { s t e p s } = 5 , n _ { N M } = 5 , e p o c h _ { n e i g h } = 1 7 , e p o c h _ { f i n a l } = 1 0 0$ . We choose $\\lambda _ { s t a r t } = 0 . 0 5$ , $\\lambda _ { e n d } = 0 . { \\overset { \\cdot } { 0 } }$ as done in Loshchilov & Hutter (2017). model0 was a simple conv net: Conv-MaxPool-Conv-MaxPool-Conv-FC-Softmax3, which is pretrained for 20 epochs, achieving $\\approx 7 5 \\%$ validation accuracy (up to $9 1 \\%$ when trained till convergence), see Figure 5 in the appendix. If our algorithm is able to identify better networks, one would expect to get better results with the setting $n _ { n e i g h } = 8$ . ", + "bbox": [ + 173, + 241, + 825, + 381 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retraining from scratch. In the this experiment we investigate whether the ”weight inheritance” due to the network morphisms used in our algorithm harms the final performance of the final model. This weight inheritance can be seen as a strong prior on the weights and one could suspect that the new, larger model may not be able to overcome a possibly poor prior. Additionally we were interested in measuring the overhead of the architecture search process, so we compared the times for generating and training a model with the time needed when training the final model from scratch. The retraining from scratch is done for the same number of epochs as the total number of epochs spent to train the model returned by our algorithm4 . ", + "bbox": [ + 173, + 387, + 825, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "No SGDR. We now turn off the cosine annealing with restarts (SGDR) during the hill climbing stage, i.e., the training in line 17 of Algorithm 1 is done with a constant learning rate. We tried $\\lambda \\in \\{ 0 . 0 1 , 0 . 0 2 5 , 0 . 0 5 \\}$ , 10 runs each and averaged the results. Note that we still use the cosine decay for the final training. ", + "bbox": [ + 174, + 506, + 825, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "No network morphism. Lastly, we turn off the network morphism constraint for initializing the neighbor networks. In detail, we proceeded as Real et al. (2017): All weights from layer where now changes occur are inherited, whereas the weights of new/modified layers are initialized by random. ", + "bbox": [ + 176, + 569, + 825, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results for these experiments are summarized in Table 1 . The hill climbing strategy is actually able to identify better performing models. (first and second line: $5 . 7 \\%$ vs. $6 . 5 \\%$ ). Notice how hill climbing prefers larger models (5.7 million parameters on average vs. 4.4 million). Performance slightly decreases when the models are retrained from scratch (line 3). This experiments indicates that our algorithm does not harm the final performance of a model. Regarding the runtime, the overhead for first having to search for the architecture is roughly a factor 3. We think this is a big advantage of our method and shows that architecture search can be done in the same order of magnitude as training a single model. In line 4 we can see that SGDR plays an important role. The resulting models chosen by our algorithm when training is done with a constant learning rate perform similarly to the models without any model selection strategy $( 6 . 4 \\%$ and $6 . 5 \\%$ , respectively), which indicates that the performance after a few epochs on the validation set when trained without SGDR correlates less with the final performance on the test set as it is the case for training with SGDR. Indeed, we computed the Pearson correlation coefficient and obtained $R ^ { 2 } = 0 . 6 4$ for training with SGDR and and $\\mathrm { \\bar { \\it R } ^ { 2 } = 0 . 3 6 }$ for training with a constant learning rate. See appendix A. Also, with the constant learning rate, the few epochs spent are not sufficient to improve the performance of the model. Figure 2 shows the progress while running our algorithm with and without SGDR, averaged over all runs. When turning off the network morphism constraint, performance also decreases. ", + "bbox": [ + 174, + 617, + 825, + 853 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5bffa50e24c866e7a85ea7102470da8aa121189925b47b5e0807ef32b1fb1234.jpg", + "image_caption": [ + "Figure 2: The best model found by Algorithm 1 tracked over time (in terms of hill climbing iterations). With (red) and without (blue) using SGDR for the training within the hill climbing (line 17). Final training (line 24) is not plotted. Dashed line denotes mean, shaded area $\\pm 2 \\sigma$ intervalls. " + ], + "image_footnote": [], + "bbox": [ + 233, + 98, + 758, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/399c4ad8f3e36610cb71bf824ac3a3e315a7c2b733f1dcf09b58c5f4d7dfd8a5.jpg", + "table_caption": [ + "Table 1: Baseline experiments. Runtime, # params, and error rates are averaged over 10 runs (for $n _ { n e i g h } = 8 $ ) and 30 runs $( n _ { n e i g h } = 1 )$ ) runs, respectively. $n _ { s t e p s } = 5$ in all experiments. " + ], + "table_footnote": [], + "table_body": "
algorithm settingruntime (hrs)# params (mil.)error ± std.( %)
nneigh =812.85.75.7 ± 0.35
Random networks (nneigh = 1)4.54.46.5 ± 0.76
models from line 1 retrained from scratch5.35.76.1 ± 0.92
nneigh = 8, no SGDR10.65.86.4±0.70
nneigh = 8, no net. morph.6.62.96.1 ±0.30
", + "bbox": [ + 173, + 435, + 839, + 525 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interestingly the number of parameters heavily decreases. This indicates that our algorithm prefers models without new parameters. ", + "bbox": [ + 173, + 561, + 826, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1.2 COMPARISON TO HAND CRAFTED AND OTHER AUTOMATICALLY GENERATED ARCHITECTURES ", + "text_level": 1, + "bbox": [ + 176, + 613, + 756, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now compare our algorithm against the popular wide residual networks (Zagoruyko & Komodakis, 2016), the state of the art model from Gastaldi (2017) as well as other automated architecture search methods. Beside our results for $n _ { s t e p s } = 5$ from the previous section, we also tried $n _ { s t e p s } = 8$ to generate larger models. ", + "bbox": [ + 174, + 652, + 823, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For further improving the results, we take snapshots of the best models from every iteration while running our algorithm following the idea of Huang et al. (2017) when using SGDR (Loshchilov & Hutter, 2017) for training. However different from Huang et al. (2017), we do not immediately get fully trained models for free, as our snapshots are not yet trained on the validation set but rather only on the training set. Hence we spent some additional resources and train the snapshots on both sets. Afterwards the ensemble model is build by combining the snapshot models with uniform weights. Lastly, we also build an ensemble from the models returned by our algorithm across all runs. Results are listed in Table 2. ", + "bbox": [ + 174, + 709, + 825, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The proposed method is able to generate competitive network architectures in only 12 hours. By spending another 12 hours, it outperforms most automated architecture search methods although all of them require (partially far) more time and GPUs. We do not reach the performance of the two handcrafted architectures as well as the ones found by Zoph & Le (2017) and Brock et al. (2017). However note that Zoph & Le (2017) spent by far more resources than we did. ", + "bbox": [ + 174, + 827, + 825, + 896 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unsurprisingly, the ensemble models perform better. It is a simple and cheap way to improve results which everyone can consider when the number of parameters is not relevant. ", + "bbox": [ + 174, + 896, + 821, + 924 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/40ffbbc485084b4e3fbd3f84be67a619447ee95d3fffcf9f34d4d00c8ce93c36.jpg", + "table_caption": [ + "Table 2: Results for CIFAR-10. For our methods the stated resources, # parameters and errors are averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models. " + ], + "table_footnote": [], + "table_body": "
modelresources spent# params (mil.)error(%)
Shake-Shake (Gastaldi,2017)4 GPU days,2 GPUs262.9
WRN 28-10 (Loshchilov & Hutter,2017)1 GPU day36.53.86
Baker et al. (2016)80-100 GPU days116.9
Cai et al. (2017)15 GPU days19.75.7
Zoph & Le (2017)16.000-24.000 GPU days37.53.65
Real et al. (2017)2500 GPU days5.45.4
Saxena & Verbeek (2016)？217.4
Brock et al. (2017)3 GPU days16.04.0
Ours (random networks,nsteps = 5, nneigh =1)0.2 GPU days4.46.5
Ours (nsteps = 5,nneigh =8,10 runs)0.5 GPU days5.75.7
Ours (nsteps = 8,nneigh = 8,4 runs)1 GPU day19.75.2
Ours (snapshot ensemble,4 runs)2 GPU days57.84.7
Ours (ensemble across runs)4 GPU days884.4
", + "bbox": [ + 173, + 155, + 825, + 343 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a47d8eceb003bb23f38bfb4ad16df746a77c1f78312b59d11e2951c2db318cd.jpg", + "table_caption": [ + "Table 3: Results for CIFAR-100. For our methods the stated resources, # parameters and errors are averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models. " + ], + "table_footnote": [], + "table_body": "
modelresources spent# params (mil.)error(%)
15.9
Shake-Shake (Gastaldi, 2017) WRN 28-10 (Loshchilov & Hutter,2017)14 GPU days34.4
1 GPU day36.519.6
Real et al. (2017)250 GPUs40.423.7
Brock et al. (2017)3 GPU days16.020.6
Ours (nsteps = 8,nneigh = 8,5 runs)1 GPU day22.323.4
Ours (snapshot ensemble,5 runs)2 GPU days73.320.9
Ours (ensemble across runs)5 GPU days111.519.6
", + "bbox": [ + 179, + 405, + 818, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2 EXPERIMENTS ON CIFAR-100 ", + "text_level": 1, + "bbox": [ + 176, + 566, + 428, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We repeat the previous experiment on CIFAR-100; hyperparameters were not changed. The results are listed in Table 3. Unfortunately most automated architecture methods did not consider CIFAR100. Our method is on a par with Real et al. (2017) after one day with a single GPU. The snapshot ensemble performs similar to Brock et al. (2017) and an ensemble model build from the 5 runs can compete with the hand crafted WRN 28-10. The performance of the Shake-Shake network (Gastaldi, 2017) is again not reached. ", + "bbox": [ + 173, + 593, + 825, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 696, + 318, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed NASH, a simple and fast method for automated architecture search based on a hill climbing strategy, network morphisms, and training via SGDR. Experiments on CIFAR-10 and CIFAR-100 showed that our method yields competitive results while requiring considerably less computational resources than most alternative approaches. Our algorithm is easily extendable, e.g., by other network morphisms, evolutionary approaches for generating new models, other methods for cheap performance evaluation (such as, e.g., learning curve prediction (Klein et al., 2017) or hypernetworks (Ha et al., 2017; Brock et al., 2017)), or better resource handling strategies (such as Hyperband (Li et al., 2016b)). In this sense, we hope that our approach can serve as a basis for the development of more sophisticated methods that yield further improvements of performance. ", + "bbox": [ + 173, + 728, + 825, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 103, + 285, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S. Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´ Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vincent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´ berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015. URL https://www.tensorflow.org/. Software available from tensorflow.org. ", + "bbox": [ + 176, + 126, + 825, + 251 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. ICLR, 2015. ", + "bbox": [ + 176, + 262, + 820, + 290 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. Designing neural network architectures using reinforcement learning. ICLR 2017, 2016. ", + "bbox": [ + 173, + 300, + 820, + 329 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "J. Bergstra and Y. Bengio. Random search for hyper-parameter optimization. 13(1):281–305, 2012. ", + "bbox": [ + 173, + 338, + 821, + 354 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "J. Bergstra, R. Bardenet, Y. Bengio, and B. Kegl. Algorithms for hyper-parameter optimization. In ´ NIPS, 2011. ", + "bbox": [ + 171, + 363, + 823, + 393 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Andrew Brock, Theodore Lim, James M. Ritchie, and Nick Weston. SMASH: one-shot model architecture search through hypernetworks. arXiv preprint, 2017. ", + "bbox": [ + 169, + 402, + 823, + 431 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang. Reinforcement learning for architecture search by network transformation. 2017. ", + "bbox": [ + 171, + 441, + 823, + 470 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tianqi Chen, Ian J. Goodfellow, and Jonathon Shlens. Net2net: Accelerating learning via knowledge transfer. arXiv preprint, 2015. ", + "bbox": [ + 173, + 479, + 823, + 508 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Franc¸ois Chollet et al. Keras. https://github.com/fchollet/keras, 2015. ", + "bbox": [ + 173, + 518, + 740, + 534 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Xavier Gastaldi. Shake-shake regularization. ICLR 2017 Workshop, 2017. ", + "bbox": [ + 176, + 544, + 661, + 559 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. ICLR, 2017. ", + "bbox": [ + 176, + 569, + 658, + 583 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CVPR, 2016. ", + "bbox": [ + 173, + 593, + 821, + 622 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel rahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. Deep neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine, 2012. ", + "bbox": [ + 174, + 632, + 825, + 688 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks. 2016. ", + "bbox": [ + 174, + 699, + 821, + 727 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger. Snapshot ensembles: Train 1, get M for free. ICLR 2017, 2017. ", + "bbox": [ + 174, + 737, + 821, + 766 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "A. Klein, S. Falkner, J. T. Springenberg, and F. Hutter. Learning curve prediction with Bayesian neural networks. In International Conference on Learning Representations (ICLR) 2017 Conference Track, April 2017. ", + "bbox": [ + 176, + 776, + 821, + 819 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger (eds.), Advances in Neural Information Processing Systems 25, pp. 1097–1105. Curran Associates, Inc., 2012. ", + "bbox": [ + 174, + 829, + 825, + 885 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "L. Li, K. G. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar. Efficient hyperparameter optimization and infinitely many armed bandits. CoRR, abs/1603.06560, 2016a. ", + "bbox": [ + 174, + 895, + 820, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Efficient hyperparameter optimization and infinitely many armed bandits. arXiv preprint, 2016b. \nI. Loshchilov and F. Hutter. CMA-ES for hyperparameter optimization of deep neural networks. CoRR, abs/1604.07269, 2016. \nI. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International Conference on Learning Representations (ICLR) 2017 Conference Track, April 2017. \nH. Mendoza, A. Klein, M. Feurer, T. Springenberg, and F. Hutter. Towards automatically-tuned neural networks. In AutoML, 2016. \nEsteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V. Le, and Alex Kurakin. Large-scale evolution of image classifiers. arXiv preprint, 2017. \nStuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach (3rd Edition). Pearson, 3 edition, December 2009. ISBN 0136042597. \nShreyas Saxena and Jakob Verbeek. Convolutional neural fabrics. arXiv preprint, 2016. \nJ. Snoek, H. Larochelle, and R.P. Adams. Practical Bayesian optimization of machine learning algorithms. In NIPS, 2012. \nMasanori Suganuma, Shinichi Shirakawa, and Tomoharu Nagao. A genetic programming approach to designing convolutional neural network architectures. arXiv preprint, 2017. \nTao Wei, Changhu Wang, Yong Rui, and Chang Wen Chen. Network morphism. arXiv preprint, 2016. \nSergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint, 2016. \nBarret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. 2017. \nBarret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning transferable architectures for scalable image recognition. 2017. ", + "bbox": [ + 169, + 95, + 828, + 545 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/2edbb1405729ce5dc046dc80f49bd9380081ebed874aa93e2d296d2395a7ddf0.jpg", + "image_caption": [ + "Figure 3: Initial network for our algorithm. " + ], + "image_footnote": [], + "bbox": [ + 191, + 155, + 761, + 391 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/213925428f8eea0c7c284680e760c96a1baa86df2685b0480330288a33417347.jpg", + "image_caption": [ + "Figure 4: Initial network for our algorithm. " + ], + "image_footnote": [], + "bbox": [ + 197, + 469, + 761, + 707 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/926480ca8e0fe0b1fac91e9b0c6c21dd16e70a4e522edfe39777971a2e331b4d.jpg", + "image_caption": [ + "Figure 5: Initial network for our algorithm. " + ], + "image_footnote": [], + "bbox": [ + 410, + 169, + 581, + 825 + ], + "page_idx": 11 + }, + { + "type": "image", + "img_path": "images/091ad5bd88208c13ba32695bfc45a25363ea4c4e47109a8acba5e63354b8c3d1.jpg", + "image_caption": [ + "Figure 6: Network generated by our algorithm with $n _ { s t e p s } = 5$ . " + ], + "image_footnote": [], + "bbox": [ + 406, + 116, + 586, + 867 + ], + "page_idx": 12 + }, + { + "type": "image", + "img_path": "images/e810d4abf64e5a83ec8d5f1149515b76d979a3c8fc66b0153106f09e2b3c3c2d.jpg", + "image_caption": [ + "Figure 7: Network generated by our algorithm with $n _ { s t e p s } = 8$ . " + ], + "image_footnote": [], + "bbox": [ + 410, + 113, + 571, + 861 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/parse/train/SySaJ0xCZ/SySaJ0xCZ_middle.json b/parse/train/SySaJ0xCZ/SySaJ0xCZ_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..8d04624228111715d98a75f58ede54c2c050df81 --- /dev/null +++ b/parse/train/SySaJ0xCZ/SySaJ0xCZ_middle.json @@ -0,0 +1,27544 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "SIMPLE AND EFFICIENT ARCHITECTURE SEARCH FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 398, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 398, + 118 + ], + "score": 1.0, + "content": "CONVOLUTIONAL NEURAL NETWORKS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 322 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "score": 1.0, + "content": "Neural networks have recently had a lot of success for many tasks. However, neu-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "score": 1.0, + "content": "ral network architectures that perform well are still typically designed manually", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "score": 1.0, + "content": "by experts in a cumbersome trial-and-error process. We propose a new method", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "score": 1.0, + "content": "to automatically search for well-performing CNN architectures based on a sim-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "score": 1.0, + "content": "ple hill climbing procedure whose operators apply network morphisms, followed", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "score": 1.0, + "content": "by short optimization runs by cosine annealing. Surprisingly, this simple method", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "yields competitive results, despite only requiring resources in the same order of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "score": 1.0, + "content": "magnitude as training a single network. E.g., on CIFAR-10, our method designs", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 317, + 313 + ], + "score": 1.0, + "content": "and trains networks with an error rate below", + "type": "text" + }, + { + "bbox": [ + 318, + 300, + 333, + 311 + ], + "score": 0.86, + "content": "6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 299, + 470, + 313 + ], + "score": 1.0, + "content": "in only 12 hours on a single GPU;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 385, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 366, + 322 + ], + "score": 1.0, + "content": "training for one day reduces this error further, to almost", + "type": "text" + }, + { + "bbox": [ + 366, + 311, + 380, + 321 + ], + "score": 0.86, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 311, + 385, + 322 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9.5 + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 206, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 208, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 208, + 361 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "Neural networks have rapidly gained popularity over the last few years due to their success in a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "variety of tasks, such as image recognition (Krizhevsky et al., 2012), speech recognition (Hinton", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "et al., 2012) and machine translation (Bahdanau et al., 2015). In most cases, these neural networks", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "score": 1.0, + "content": "are still designed by hand, which is an exhausting, time-consuming process. Additionally, the vast", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "amount of possible configurations requires expert knowledge to restrict the search. Therefore, a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 426, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 482, + 438 + ], + "score": 1.0, + "content": "natural goal is to design optimization algorithms that automate this neural architecture search.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 504, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "However, most classic optimization algorithms do not apply to this problem, since the architecture", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "search space is discrete (e.g., number of layers, layer types) and conditional (e.g., the number of pa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "rameters defining a layer depends on the layer type). Thus, methods that rely on, e.g., differentiabil-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "score": 1.0, + "content": "ity or independent parameters are not applicable. This led to a growing interest in using evolutionary", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "algorithms (Real et al., 2017; Suganuma et al., 2017) and reinforcement learning (Baker et al., 2016;", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "score": 1.0, + "content": "Cai et al., 2017; Zoph & Le, 2017) for automatically designing CNN architectures. Unfortunately,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 104, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "most proposed methods are either very costly (requiring hundreds or thousands of GPU days) or", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 520, + 252, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 252, + 532 + ], + "score": 1.0, + "content": "yield non-competitive performance.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 503, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "score": 1.0, + "content": "In this work, we aim to dramatically reduce these computational costs while still achieving compet-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 547, + 364, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 364, + 559 + ], + "score": 1.0, + "content": "itive performance. Specifically, our contributions are as follows:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 106, + 568, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 107, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "• We propose a baseline method that randomly constructs networks and trains them with SGDR", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 116, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 116, + 579, + 434, + 592 + ], + "score": 1.0, + "content": "(Loshchilov & Hutter, 2017). We demonstrate that this simple baseline achieves", + "type": "text" + }, + { + "bbox": [ + 435, + 579, + 466, + 590 + ], + "score": 0.71, + "content": "6 \\% - 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "test error", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "on CIFAR-10, which already rivals several existing methods for neural archictecture search. Due", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 116, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "to its simplicity, we hope that this baseline provides a valuable starting point for the development", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 612, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 115, + 612, + 295, + 625 + ], + "score": 1.0, + "content": "of more sophisticated methods in the future.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 107, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 107, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "• We formalize and extend the work on network morphisms (Chen et al., 2015; Wei et al., 2016;", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "Cai et al., 2017) in order to provide popular network building blocks, such as skip connections", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 650, + 218, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 218, + 662 + ], + "score": 1.0, + "content": "and batch normalization.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "• We propose Neural Architecture Search by Hillclimbing (NASH), a simple iterative approach", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 115, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "that, at each step, applies a set of alternative network morphisms to the current network, trains the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 115, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "resulting child networks with short optimization runs of cosine annealing (Loshchilov & Hutter,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "2017), and moves to the most promising child network. NASH finds and trains competitive archi-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "tectures at a computational cost of the same order of magnitude as training a single network; e.g.,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 393, + 733 + ], + "score": 1.0, + "content": "on CIFAR-10, NASH finds and trains CNNs with an error rate below", + "type": "text" + }, + { + "bbox": [ + 393, + 721, + 411, + 731 + ], + "score": 0.85, + "content": "6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "in roughly 12 hours on", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38.5 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 106, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "SIMPLE AND EFFICIENT ARCHITECTURE SEARCH FOR", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 398, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 398, + 118 + ], + "score": 1.0, + "content": "CONVOLUTIONAL NEURAL NETWORKS", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 276, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 322 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 225 + ], + "score": 1.0, + "content": "Neural networks have recently had a lot of success for many tasks. However, neu-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 469, + 236 + ], + "score": 1.0, + "content": "ral network architectures that perform well are still typically designed manually", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 247 + ], + "score": 1.0, + "content": "by experts in a cumbersome trial-and-error process. We propose a new method", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "spans": [ + { + "bbox": [ + 141, + 246, + 469, + 258 + ], + "score": 1.0, + "content": "to automatically search for well-performing CNN architectures based on a sim-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 269 + ], + "score": 1.0, + "content": "ple hill climbing procedure whose operators apply network morphisms, followed", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 470, + 280 + ], + "score": 1.0, + "content": "by short optimization runs by cosine annealing. Surprisingly, this simple method", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 470, + 290 + ], + "score": 1.0, + "content": "yields competitive results, despite only requiring resources in the same order of", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 469, + 301 + ], + "score": 1.0, + "content": "magnitude as training a single network. E.g., on CIFAR-10, our method designs", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 299, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 299, + 317, + 313 + ], + "score": 1.0, + "content": "and trains networks with an error rate below", + "type": "text" + }, + { + "bbox": [ + 318, + 300, + 333, + 311 + ], + "score": 0.86, + "content": "6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 333, + 299, + 470, + 313 + ], + "score": 1.0, + "content": "in only 12 hours on a single GPU;", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 142, + 311, + 385, + 322 + ], + "spans": [ + { + "bbox": [ + 142, + 311, + 366, + 322 + ], + "score": 1.0, + "content": "training for one day reduces this error further, to almost", + "type": "text" + }, + { + "bbox": [ + 366, + 311, + 380, + 321 + ], + "score": 0.86, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 311, + 385, + 322 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 9.5, + "bbox_fs": [ + 141, + 212, + 470, + 322 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 345, + 206, + 357 + ], + "lines": [ + { + "bbox": [ + 105, + 344, + 208, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 344, + 208, + 361 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 107, + 371, + 505, + 437 + ], + "lines": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 370, + 506, + 383 + ], + "score": 1.0, + "content": "Neural networks have rapidly gained popularity over the last few years due to their success in a", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "variety of tasks, such as image recognition (Krizhevsky et al., 2012), speech recognition (Hinton", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "et al., 2012) and machine translation (Bahdanau et al., 2015). In most cases, these neural networks", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "score": 1.0, + "content": "are still designed by hand, which is an exhausting, time-consuming process. Additionally, the vast", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "amount of possible configurations requires expert knowledge to restrict the search. Therefore, a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 426, + 482, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 426, + 482, + 438 + ], + "score": 1.0, + "content": "natural goal is to design optimization algorithms that automate this neural architecture search.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 370, + 506, + 438 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 442, + 504, + 530 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 506, + 455 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 506, + 455 + ], + "score": 1.0, + "content": "However, most classic optimization algorithms do not apply to this problem, since the architecture", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 505, + 466 + ], + "score": 1.0, + "content": "search space is discrete (e.g., number of layers, layer types) and conditional (e.g., the number of pa-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 505, + 478 + ], + "score": 1.0, + "content": "rameters defining a layer depends on the layer type). Thus, methods that rely on, e.g., differentiabil-", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 489 + ], + "score": 1.0, + "content": "ity or independent parameters are not applicable. This led to a growing interest in using evolutionary", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 487, + 506, + 499 + ], + "score": 1.0, + "content": "algorithms (Real et al., 2017; Suganuma et al., 2017) and reinforcement learning (Baker et al., 2016;", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 506, + 510 + ], + "score": 1.0, + "content": "Cai et al., 2017; Zoph & Le, 2017) for automatically designing CNN architectures. Unfortunately,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 508, + 506, + 521 + ], + "spans": [ + { + "bbox": [ + 104, + 508, + 506, + 521 + ], + "score": 1.0, + "content": "most proposed methods are either very costly (requiring hundreds or thousands of GPU days) or", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 520, + 252, + 532 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 252, + 532 + ], + "score": 1.0, + "content": "yield non-competitive performance.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 25.5, + "bbox_fs": [ + 104, + 443, + 506, + 532 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 536, + 503, + 558 + ], + "lines": [ + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 549 + ], + "score": 1.0, + "content": "In this work, we aim to dramatically reduce these computational costs while still achieving compet-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 547, + 364, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 547, + 364, + 559 + ], + "score": 1.0, + "content": "itive performance. Specifically, our contributions are as follows:", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 534, + 505, + 559 + ] + }, + { + "type": "list", + "bbox": [ + 106, + 568, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 107, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 107, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "• We propose a baseline method that randomly constructs networks and trains them with SGDR", + "type": "text" + } + ], + "index": 32, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 579, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 116, + 579, + 434, + 592 + ], + "score": 1.0, + "content": "(Loshchilov & Hutter, 2017). We demonstrate that this simple baseline achieves", + "type": "text" + }, + { + "bbox": [ + 435, + 579, + 466, + 590 + ], + "score": 0.71, + "content": "6 \\% - 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 579, + 506, + 592 + ], + "score": 1.0, + "content": "test error", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 116, + 590, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 590, + 505, + 603 + ], + "score": 1.0, + "content": "on CIFAR-10, which already rivals several existing methods for neural archictecture search. Due", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 601, + 505, + 614 + ], + "spans": [ + { + "bbox": [ + 116, + 601, + 505, + 614 + ], + "score": 1.0, + "content": "to its simplicity, we hope that this baseline provides a valuable starting point for the development", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 115, + 612, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 115, + 612, + 295, + 625 + ], + "score": 1.0, + "content": "of more sophisticated methods in the future.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 107, + 628, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 107, + 628, + 505, + 640 + ], + "score": 1.0, + "content": "• We formalize and extend the work on network morphisms (Chen et al., 2015; Wei et al., 2016;", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "spans": [ + { + "bbox": [ + 115, + 639, + 505, + 652 + ], + "score": 1.0, + "content": "Cai et al., 2017) in order to provide popular network building blocks, such as skip connections", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 116, + 650, + 218, + 662 + ], + "spans": [ + { + "bbox": [ + 116, + 650, + 218, + 662 + ], + "score": 1.0, + "content": "and batch normalization.", + "type": "text" + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 678 + ], + "score": 1.0, + "content": "• We propose Neural Architecture Search by Hillclimbing (NASH), a simple iterative approach", + "type": "text" + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 676, + 506, + 690 + ], + "spans": [ + { + "bbox": [ + 115, + 676, + 506, + 690 + ], + "score": 1.0, + "content": "that, at each step, applies a set of alternative network morphisms to the current network, trains the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 688, + 506, + 700 + ], + "spans": [ + { + "bbox": [ + 115, + 688, + 506, + 700 + ], + "score": 1.0, + "content": "resulting child networks with short optimization runs of cosine annealing (Loshchilov & Hutter,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 116, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "2017), and moves to the most promising child network. NASH finds and trains competitive archi-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 708, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 115, + 708, + 506, + 724 + ], + "score": 1.0, + "content": "tectures at a computational cost of the same order of magnitude as training a single network; e.g.,", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 720, + 393, + 733 + ], + "score": 1.0, + "content": "on CIFAR-10, NASH finds and trains CNNs with an error rate below", + "type": "text" + }, + { + "bbox": [ + 393, + 721, + 411, + 731 + ], + "score": 0.85, + "content": "6 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 411, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "in roughly 12 hours on", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 38.5, + "bbox_fs": [ + 106, + 569, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 114, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 115, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 115, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "a single GPU. After one day the error is reduced to almost", + "type": "text" + }, + { + "bbox": [ + 355, + 83, + 370, + 93 + ], + "score": 0.86, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ". Models from different stages of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 329, + 106 + ], + "score": 1.0, + "content": "our algorithm can be combined to achieve an error of", + "type": "text" + }, + { + "bbox": [ + 329, + 94, + 354, + 104 + ], + "score": 0.86, + "content": "4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "within two days on a single GPU. On", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 487, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 275, + 118 + ], + "score": 1.0, + "content": "CIFAR-100, we achieve an error below", + "type": "text" + }, + { + "bbox": [ + 275, + 105, + 295, + 115 + ], + "score": 0.87, + "content": "24 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 103, + 405, + 118 + ], + "score": 1.0, + "content": "in one day and get close to", + "type": "text" + }, + { + "bbox": [ + 405, + 105, + 424, + 115 + ], + "score": 0.86, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 103, + 487, + 118 + ], + "score": 1.0, + "content": "after two days.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 503, + 139 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 504, + 140 + ], + "score": 1.0, + "content": "• Our method is easy to use and easy to extend, so it hopefully can serve as a basis for future work.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "We first discuss related work in Section 2. Then, we formalize the concept of network morphisms", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 176 + ], + "score": 1.0, + "content": "in Section 3 and propose our architecture search methods based on them in Section 4. We evaluate", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 317, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 317, + 186 + ], + "score": 1.0, + "content": "our methods in Section 5 and conclude in Section 6.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 108, + 212, + 209, + 224 + ], + "lines": [ + { + "bbox": [ + 104, + 210, + 210, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 210, + 210, + 227 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 242, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "Hyperparameter optimization. Neural networks are known to be quite sensitive to the setting of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "various hyperparameters, such as learning rates and regularization constants. There exists a long", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "line of research on automated methods for setting these hyperparameters, including, e.g., random", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "search (Bergstra & Bengio, 2012), Bayesian optimization (Bergstra et al., 2011; Snoek et al., 2012),", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "bandit-based approaches (Li et al., 2016a), and evolutionary strategies (Loshchilov & Hutter, 2016).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "score": 1.0, + "content": "Automated architecture search. In recent years, the research focus has shifted from optimizing hy-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 104, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "perparameters to optimizing architectures. While architectural choices can be treated as categorical", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "score": 1.0, + "content": "hyperparameters and be optimized with standard hyperparameter optimization methods (Bergstra", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "et al., 2011; Mendoza et al., 2016), the current focus is on the development of special techniques", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 346, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 346, + 506, + 361 + ], + "score": 1.0, + "content": "for architectural optimization. One very popular approach is to train a reinforcement learning agent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "score": 1.0, + "content": "with the objective of designing well-performing convolutional neural networks (Baker et al., 2016;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "Zoph & Le, 2017; Cai et al., 2017). Baker et al. (2016) train an RL agent to sequentially choose the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "type of layers (convolution, pooling, fully connected) and their parameters. Zoph & Le (2017) use a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "recurrent neural network controller to sequentially generate a string representing the network archi-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "tecture. Both approaches train their generated networks from scratch and evaluate their performance", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "on a validation set, which represents a very costly step. In a follow-up work (Zoph et al., 2017), the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "score": 1.0, + "content": "RL agent learned to build cells, which are then used as building blocks for a neural network with a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "fixed global structure. Unfortunately, training an RL agent with the objective of designing architec-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "ture is extremely expensive: both Baker et al. (2016) and Zoph & Le (2017) required over 10.000", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "fully trained networks, requiring hundreds to thousands of GPU days. To overcome this drawback,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 466, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 481 + ], + "score": 1.0, + "content": "Cai et al. (2017) proposed to apply the concept of network transformations/morphisms within RL.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "As in our (independent, parallel) work, the basic idea is to use the these transformation to generate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "new pre-trained architectures to avoid the large cost of training all networks from scratch. Compared", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 501, + 504, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 504, + 513 + ], + "score": 1.0, + "content": "to this work, our approach is much simpler and 15 times faster while obtaining better performance.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "Real et al. (2017) and Suganuma et al. (2017) utilized evolutionary algorithms to iteratively generate", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "powerful networks from a small network. Operations like inserting a layer, modifying the parameters", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 538, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 506, + 553 + ], + "score": 1.0, + "content": "of a layer or adding skip connections serve as ”mutations” in their framework of evolution. Whereas", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Real et al. (2017) also used enormous computational resources (250 GPUs, 10 days), Suganuma", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "et al. (2017) were restricted to relatively small networks due to handling a population of networks.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "In contrast to the previous methods where network capacity increases over time, Saxena & Verbeek", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "(2016) start with training a large network (a ”convolution neural fabric”) and prune this in the end.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "Very recently, Brock et al. (2017) used hypernetworks (Ha et al., 2017) to generate the weights for a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "randomly sampled network architecture with the goal of eliminating the costly process of training a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 617, + 209, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 209, + 628 + ], + "score": 1.0, + "content": "vast amount of networks.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36.5 + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "Network morphism/ transformation. Network transformations were (to our knowledge) first in-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "score": 1.0, + "content": "troduced by Chen et al. (2015) in the context of transfer learning. The authors described a function", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "preserving operation to make a network deeper (dubbed ”Net2Deeper”) or wider (”Net2Wider”)", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "with the goal of speeding up training and exploring network architectures. Wei et al. (2016) pro-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "posed additional operations, e.g., for handling non-idempotent activation functions or altering the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "kernel size and introduced the term network morphism. As mentioned above, Cai et al. (2017) used", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "network morphisms for architecture search, though they just employ the Net2Deeper and Net2Wider", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "operators from Chen et al. (2015) as well as altering the kernel size, i.e., they limit their search space", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 325, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 325, + 733 + ], + "score": 1.0, + "content": "to simple architectures without, e.g., skip connections.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 114, + 82, + 505, + 116 + ], + "lines": [ + { + "bbox": [ + 115, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 115, + 82, + 354, + 95 + ], + "score": 1.0, + "content": "a single GPU. After one day the error is reduced to almost", + "type": "text" + }, + { + "bbox": [ + 355, + 83, + 370, + 93 + ], + "score": 0.86, + "content": "5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 82, + 505, + 95 + ], + "score": 1.0, + "content": ". Models from different stages of", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 93, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 116, + 93, + 329, + 106 + ], + "score": 1.0, + "content": "our algorithm can be combined to achieve an error of", + "type": "text" + }, + { + "bbox": [ + 329, + 94, + 354, + 104 + ], + "score": 0.86, + "content": "4 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 93, + 505, + 106 + ], + "score": 1.0, + "content": "within two days on a single GPU. On", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 103, + 487, + 118 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 275, + 118 + ], + "score": 1.0, + "content": "CIFAR-100, we achieve an error below", + "type": "text" + }, + { + "bbox": [ + 275, + 105, + 295, + 115 + ], + "score": 0.87, + "content": "24 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 103, + 405, + 118 + ], + "score": 1.0, + "content": "in one day and get close to", + "type": "text" + }, + { + "bbox": [ + 405, + 105, + 424, + 115 + ], + "score": 0.86, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 103, + 487, + 118 + ], + "score": 1.0, + "content": "after two days.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 115, + 82, + 505, + 118 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 127, + 503, + 139 + ], + "lines": [ + { + "bbox": [ + 106, + 126, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 504, + 140 + ], + "score": 1.0, + "content": "• Our method is easy to use and easy to extend, so it hopefully can serve as a basis for future work.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3, + "bbox_fs": [ + 106, + 126, + 504, + 140 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 152, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 151, + 505, + 165 + ], + "spans": [ + { + "bbox": [ + 106, + 151, + 505, + 165 + ], + "score": 1.0, + "content": "We first discuss related work in Section 2. Then, we formalize the concept of network morphisms", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 164, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 505, + 176 + ], + "score": 1.0, + "content": "in Section 3 and propose our architecture search methods based on them in Section 4. We evaluate", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 317, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 317, + 186 + ], + "score": 1.0, + "content": "our methods in Section 5 and conclude in Section 6.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 151, + 505, + 186 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 212, + 209, + 224 + ], + "lines": [ + { + "bbox": [ + 104, + 210, + 210, + 227 + ], + "spans": [ + { + "bbox": [ + 104, + 210, + 210, + 227 + ], + "score": 1.0, + "content": "2 RELATED WORK", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 242, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 106, + 242, + 505, + 254 + ], + "score": 1.0, + "content": "Hyperparameter optimization. Neural networks are known to be quite sensitive to the setting of", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 266 + ], + "score": 1.0, + "content": "various hyperparameters, such as learning rates and regularization constants. There exists a long", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 506, + 277 + ], + "score": 1.0, + "content": "line of research on automated methods for setting these hyperparameters, including, e.g., random", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 288 + ], + "score": 1.0, + "content": "search (Bergstra & Bengio, 2012), Bayesian optimization (Bergstra et al., 2011; Snoek et al., 2012),", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "bandit-based approaches (Li et al., 2016a), and evolutionary strategies (Loshchilov & Hutter, 2016).", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10, + "bbox_fs": [ + 105, + 242, + 506, + 299 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 303, + 505, + 511 + ], + "lines": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "spans": [ + { + "bbox": [ + 105, + 302, + 505, + 317 + ], + "score": 1.0, + "content": "Automated architecture search. In recent years, the research focus has shifted from optimizing hy-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 314, + 506, + 327 + ], + "spans": [ + { + "bbox": [ + 104, + 314, + 506, + 327 + ], + "score": 1.0, + "content": "perparameters to optimizing architectures. While architectural choices can be treated as categorical", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 339 + ], + "score": 1.0, + "content": "hyperparameters and be optimized with standard hyperparameter optimization methods (Bergstra", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 505, + 349 + ], + "score": 1.0, + "content": "et al., 2011; Mendoza et al., 2016), the current focus is on the development of special techniques", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 346, + 506, + 361 + ], + "spans": [ + { + "bbox": [ + 104, + 346, + 506, + 361 + ], + "score": 1.0, + "content": "for architectural optimization. One very popular approach is to train a reinforcement learning agent", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "spans": [ + { + "bbox": [ + 105, + 357, + 506, + 371 + ], + "score": 1.0, + "content": "with the objective of designing well-performing convolutional neural networks (Baker et al., 2016;", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 505, + 382 + ], + "score": 1.0, + "content": "Zoph & Le, 2017; Cai et al., 2017). Baker et al. (2016) train an RL agent to sequentially choose the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 506, + 393 + ], + "score": 1.0, + "content": "type of layers (convolution, pooling, fully connected) and their parameters. Zoph & Le (2017) use a", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 403 + ], + "score": 1.0, + "content": "recurrent neural network controller to sequentially generate a string representing the network archi-", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 415 + ], + "score": 1.0, + "content": "tecture. Both approaches train their generated networks from scratch and evaluate their performance", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "on a validation set, which represents a very costly step. In a follow-up work (Zoph et al., 2017), the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 424, + 506, + 436 + ], + "score": 1.0, + "content": "RL agent learned to build cells, which are then used as building blocks for a neural network with a", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "fixed global structure. Unfortunately, training an RL agent with the objective of designing architec-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "ture is extremely expensive: both Baker et al. (2016) and Zoph & Le (2017) required over 10.000", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 505, + 469 + ], + "score": 1.0, + "content": "fully trained networks, requiring hundreds to thousands of GPU days. To overcome this drawback,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 466, + 505, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 481 + ], + "score": 1.0, + "content": "Cai et al. (2017) proposed to apply the concept of network transformations/morphisms within RL.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 479, + 505, + 491 + ], + "score": 1.0, + "content": "As in our (independent, parallel) work, the basic idea is to use the these transformation to generate", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "new pre-trained architectures to avoid the large cost of training all networks from scratch. Compared", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 501, + 504, + 513 + ], + "spans": [ + { + "bbox": [ + 106, + 501, + 504, + 513 + ], + "score": 1.0, + "content": "to this work, our approach is much simpler and 15 times faster while obtaining better performance.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 302, + 506, + 513 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 505, + 530 + ], + "score": 1.0, + "content": "Real et al. (2017) and Suganuma et al. (2017) utilized evolutionary algorithms to iteratively generate", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 506, + 541 + ], + "score": 1.0, + "content": "powerful networks from a small network. Operations like inserting a layer, modifying the parameters", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 538, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 104, + 538, + 506, + 553 + ], + "score": 1.0, + "content": "of a layer or adding skip connections serve as ”mutations” in their framework of evolution. Whereas", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 505, + 563 + ], + "score": 1.0, + "content": "Real et al. (2017) also used enormous computational resources (250 GPUs, 10 days), Suganuma", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "et al. (2017) were restricted to relatively small networks due to handling a population of networks.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 104, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "In contrast to the previous methods where network capacity increases over time, Saxena & Verbeek", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 596 + ], + "score": 1.0, + "content": "(2016) start with training a large network (a ”convolution neural fabric”) and prune this in the end.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "Very recently, Brock et al. (2017) used hypernetworks (Ha et al., 2017) to generate the weights for a", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 506, + 618 + ], + "score": 1.0, + "content": "randomly sampled network architecture with the goal of eliminating the costly process of training a", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 617, + 209, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 209, + 628 + ], + "score": 1.0, + "content": "vast amount of networks.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 516, + 506, + 628 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 633, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 505, + 645 + ], + "score": 1.0, + "content": "Network morphism/ transformation. Network transformations were (to our knowledge) first in-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 504, + 655 + ], + "score": 1.0, + "content": "troduced by Chen et al. (2015) in the context of transfer learning. The authors described a function", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "preserving operation to make a network deeper (dubbed ”Net2Deeper”) or wider (”Net2Wider”)", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 506, + 679 + ], + "score": 1.0, + "content": "with the goal of speeding up training and exploring network architectures. Wei et al. (2016) pro-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "posed additional operations, e.g., for handling non-idempotent activation functions or altering the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "kernel size and introduced the term network morphism. As mentioned above, Cai et al. (2017) used", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "network morphisms for architecture search, though they just employ the Net2Deeper and Net2Wider", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "operators from Chen et al. (2015) as well as altering the kernel size, i.e., they limit their search space", + "type": "text" + } + ], + "index": 49 + }, + { + "bbox": [ + 105, + 720, + 325, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 325, + 733 + ], + "score": 1.0, + "content": "to simple architectures without, e.g., skip connections.", + "type": "text" + } + ], + "index": 50 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 633, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 239, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 240, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 240, + 95 + ], + "score": 1.0, + "content": "3 NETWORK MORPHISM", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 139 + ], + "lines": [ + { + "bbox": [ + 104, + 102, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 122, + 120 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 105, + 149, + 117 + ], + "score": 0.89, + "content": "\\mathcal { N } ( \\mathcal { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 102, + 322, + 120 + ], + "score": 1.0, + "content": "denote a set of neural networks defined on", + "type": "text" + }, + { + "bbox": [ + 323, + 106, + 358, + 116 + ], + "score": 0.9, + "content": "\\mathcal { X } \\subset \\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 102, + 505, + 120 + ], + "score": 1.0, + "content": ". A network morphism is a mapping", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 244, + 128 + ], + "score": 0.91, + "content": "M : \\dot { \\mathcal { N } } ( \\dot { \\mathcal { X } } ) \\times \\mathbb { R } ^ { k } \\mathcal { N } ( \\underline { { \\mathcal { X } } } ) \\times \\mathbb { R } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 114, + 338, + 129 + ], + "score": 1.0, + "content": "from a neural network", + "type": "text" + }, + { + "bbox": [ + 339, + 116, + 391, + 128 + ], + "score": 0.93, + "content": "f ^ { w } \\in \\mathcal { N } ( \\mathcal { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 114, + 459, + 129 + ], + "score": 1.0, + "content": "with parameters", + "type": "text" + }, + { + "bbox": [ + 460, + 116, + 493, + 127 + ], + "score": 0.92, + "content": "w \\in \\mathbb { R } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 383, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 201, + 140 + ], + "score": 1.0, + "content": "another neural network", + "type": "text" + }, + { + "bbox": [ + 201, + 128, + 252, + 140 + ], + "score": 0.92, + "content": "g ^ { \\tilde { w } } \\in \\mathcal { N } ( \\mathcal { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 127, + 320, + 140 + ], + "score": 1.0, + "content": "with parameters", + "type": "text" + }, + { + "bbox": [ + 320, + 128, + 352, + 138 + ], + "score": 0.91, + "content": "\\tilde { w } \\in \\mathbb { R } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 127, + 383, + 140 + ], + "score": 1.0, + "content": "so that", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 141, + 375, + 155 + ], + "lines": [ + { + "bbox": [ + 236, + 141, + 375, + 155 + ], + "spans": [ + { + "bbox": [ + 236, + 141, + 375, + 155 + ], + "score": 0.88, + "content": "f ^ { w } ( x ) = g ^ { \\tilde { w } } ( x ) \\mathrm { f o r e v e r y } x \\in \\mathcal { X } .", + "type": "interline_equation", + "image_path": "01da69f32520eccbef16ce0ecf9fbbcf2637c41fb3974f8b8230761a9c0e4b45.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 236, + 141, + 375, + 155 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 175 + ], + "score": 1.0, + "content": "In the following we give a few examples of network morphisms and how standard operations for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 187 + ], + "score": 1.0, + "content": "building neural networks (e.g., adding a convolutional layer) can be expressed as a network mor-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 453, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 184, + 199 + ], + "score": 1.0, + "content": "phism. For this, let", + "type": "text" + }, + { + "bbox": [ + 185, + 185, + 214, + 197 + ], + "score": 0.92, + "content": "f _ { i } ^ { w _ { i } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 183, + 304, + 199 + ], + "score": 1.0, + "content": "be some part of a NN", + "type": "text" + }, + { + "bbox": [ + 304, + 185, + 330, + 197 + ], + "score": 0.91, + "content": "f ^ { w } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 183, + 453, + 199 + ], + "score": 1.0, + "content": ", e.g., a layer or a subnetwork.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 107, + 201, + 301, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 199, + 301, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 271, + 217 + ], + "score": 1.0, + "content": "Network morphism Type I. We replace", + "type": "text" + }, + { + "bbox": [ + 271, + 201, + 287, + 214 + ], + "score": 0.9, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 199, + 301, + 217 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 215, + 355, + 231 + ], + "lines": [ + { + "bbox": [ + 255, + 215, + 355, + 231 + ], + "spans": [ + { + "bbox": [ + 255, + 215, + 355, + 231 + ], + "score": 0.92, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = A f _ { i } ^ { w _ { i } } ( x ) + b ,", + "type": "interline_equation", + "image_path": "739fb4573c092613f8d0559923d6bb3981e5fbbefbec26a727c76f98d046b2eb.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 255, + 215, + 355, + 231 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 127, + 246 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 232, + 195, + 245 + ], + "score": 0.91, + "content": "\\tilde { w } _ { i } = ( w _ { i } , A , b ) ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 232, + 334, + 246 + ], + "score": 1.0, + "content": ". Equation (1) obviously holds for", + "type": "text" + }, + { + "bbox": [ + 335, + 233, + 388, + 244 + ], + "score": 0.91, + "content": "A = \\mathbf { 1 } , b = \\mathbf { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 232, + 505, + 246 + ], + "score": 1.0, + "content": ". This morphism can be used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 504, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 504, + 257 + ], + "score": 1.0, + "content": "to add a fully-connected or convolutional layer, as these layers are simply linear mappings. Chen", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 254, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 269 + ], + "score": 1.0, + "content": "et al. (2015) dubbed this morphism ”Net2DeeperNet”. Alternatively to the above replacement, one", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 264, + 180, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 180, + 279 + ], + "score": 1.0, + "content": "could also choose", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 274, + 371, + 290 + ], + "lines": [ + { + "bbox": [ + 239, + 274, + 371, + 290 + ], + "spans": [ + { + "bbox": [ + 239, + 274, + 371, + 290 + ], + "score": 0.92, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = C ( A f _ { i } ^ { w _ { i } } ( x ) + b ) + d ,", + "type": "interline_equation", + "image_path": "7999d9061602778300a91a1b4c09054ac86a6cafaa7aa40fb422aa6a160d5ca7.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 239, + 274, + 371, + 290 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 127, + 308 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 128, + 295, + 194, + 307 + ], + "score": 0.87, + "content": "\\tilde { w } _ { i } = ( w _ { i } , C , d )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 294, + 200, + 308 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 200, + 295, + 218, + 307 + ], + "score": 0.7, + "content": "A , b", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "are fixed, non-learnable. In this case network morphism Equation (1)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 305, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 104, + 305, + 141, + 319 + ], + "score": 1.0, + "content": "holds if", + "type": "text" + }, + { + "bbox": [ + 141, + 307, + 232, + 318 + ], + "score": 0.87, + "content": "C = A ^ { - 1 } , d = - C b", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 305, + 505, + 319 + ], + "score": 1.0, + "content": ". A Batch Normalization layer (or other normalization layers) can", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 229, + 330 + ], + "score": 1.0, + "content": "be written in the above form:", + "type": "text" + }, + { + "bbox": [ + 229, + 318, + 246, + 329 + ], + "score": 0.9, + "content": "A , b", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 316, + 382, + 330 + ], + "score": 1.0, + "content": "represent the batch statistics and", + "type": "text" + }, + { + "bbox": [ + 383, + 317, + 401, + 329 + ], + "score": 0.9, + "content": "C , d", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "the learnable scaling and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 326, + 143, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 143, + 343 + ], + "score": 1.0, + "content": "shifting.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 344, + 502, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 266, + 360 + ], + "score": 1.0, + "content": "Network morphism Type II. Assume", + "type": "text" + }, + { + "bbox": [ + 266, + 345, + 282, + 357 + ], + "score": 0.89, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 343, + 339, + 360 + ], + "score": 1.0, + "content": "has the form", + "type": "text" + }, + { + "bbox": [ + 339, + 345, + 439, + 357 + ], + "score": 0.94, + "content": "f _ { i } ^ { w _ { i } } ( x ) = A h ^ { w _ { h } } ( x ) + b", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 343, + 505, + 360 + ], + "score": 1.0, + "content": "for an arbitrary", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 354, + 303, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 142, + 370 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 357, + 149, + 366 + ], + "score": 0.77, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 354, + 200, + 370 + ], + "score": 1.0, + "content": ". We replace", + "type": "text" + }, + { + "bbox": [ + 200, + 357, + 216, + 369 + ], + "score": 0.85, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 354, + 221, + 370 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 221, + 356, + 285, + 368 + ], + "score": 0.88, + "content": "w _ { i } = ( w _ { h } , \\dot { A } , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 354, + 303, + 370 + ], + "score": 1.0, + "content": ", by", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 369, + 377, + 397 + ], + "lines": [ + { + "bbox": [ + 234, + 369, + 377, + 397 + ], + "spans": [ + { + "bbox": [ + 234, + 369, + 377, + 397 + ], + "score": 0.94, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\left( A \\quad \\tilde { A } \\right) \\binom { h ^ { w _ { h } } ( x ) } { \\tilde { h } ^ { w _ { h } } ( x ) } + b", + "type": "interline_equation", + "image_path": "f4578e30ebce96eb214241b79c83d5935f0306cd6cc67ef9b418c714979d56e3.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 234, + 369, + 377, + 397 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 212, + 414 + ], + "score": 1.0, + "content": "with an arbitrary function", + "type": "text" + }, + { + "bbox": [ + 212, + 398, + 243, + 412 + ], + "score": 0.94, + "content": "\\tilde { h } ^ { w _ { \\tilde { h } } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 399, + 346, + 414 + ], + "score": 1.0, + "content": ". The new parameters are", + "type": "text" + }, + { + "bbox": [ + 346, + 399, + 418, + 413 + ], + "score": 0.94, + "content": "\\tilde { w } _ { i } = ( w _ { i } , w _ { \\tilde { h } } , \\tilde { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 399, + 506, + 414 + ], + "score": 1.0, + "content": ". Again, Equation (1)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 249, + 426 + ], + "score": 1.0, + "content": "can trivially be satisfied by setting", + "type": "text" + }, + { + "bbox": [ + 249, + 412, + 279, + 424 + ], + "score": 0.91, + "content": "\\tilde { A } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 414, + 505, + 426 + ], + "score": 1.0, + "content": ". We think of two modifications of a NN which can be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "expressed by this morphism. Firstly, a layer can be widened (i.e., increasing the number of units", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "in a fully connected layer or the number of channels in a CNN - the Net2WiderNet transformation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 255, + 459 + ], + "score": 1.0, + "content": "from Chen et al. (2015)). Think of", + "type": "text" + }, + { + "bbox": [ + 255, + 446, + 276, + 459 + ], + "score": 0.92, + "content": "h ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "as the layer to be widened. For example, we can then", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 120, + 471 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 121, + 457, + 151, + 469 + ], + "score": 0.91, + "content": "\\tilde { h } = h", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "to simply double the width. Secondly, skip-connections by concatenation as used by", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 367, + 483 + ], + "score": 1.0, + "content": "Huang et al. (2016) can be formulated as a network morphism. If", + "type": "text" + }, + { + "bbox": [ + 367, + 470, + 387, + 482 + ], + "score": 0.92, + "content": "h ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "itself is a sequence of layers,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 481, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 220, + 495 + ], + "score": 0.91, + "content": "h ( x ) = h _ { n } ( x ) \\circ \\cdots \\circ h _ { 0 } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 482, + 315, + 496 + ], + "score": 1.0, + "content": ", then one could choose", + "type": "text" + }, + { + "bbox": [ + 316, + 481, + 356, + 495 + ], + "score": 0.94, + "content": "\\tilde { h } ( x ) = x", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 482, + 444, + 496 + ], + "score": 1.0, + "content": "to realize a skip from", + "type": "text" + }, + { + "bbox": [ + 444, + 483, + 456, + 494 + ], + "score": 0.87, + "content": "h _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "to the layer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 493, + 180, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 163, + 507 + ], + "score": 1.0, + "content": "subsequent to", + "type": "text" + }, + { + "bbox": [ + 164, + 495, + 176, + 505 + ], + "score": 0.87, + "content": "h _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 493, + 180, + 507 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 104, + 510, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 394, + 524 + ], + "score": 1.0, + "content": "Network morphism Type III. By definition, every idempotent function", + "type": "text" + }, + { + "bbox": [ + 395, + 511, + 410, + 523 + ], + "score": 0.91, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 509, + 506, + 524 + ], + "score": 1.0, + "content": "can simply be replaced", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 521, + 120, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 120, + 536 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5 + }, + { + "type": "interline_equation", + "bbox": [ + 261, + 532, + 349, + 548 + ], + "lines": [ + { + "bbox": [ + 261, + 532, + 349, + 548 + ], + "spans": [ + { + "bbox": [ + 261, + 532, + 349, + 548 + ], + "score": 0.92, + "content": "f _ { i } ^ { ( w _ { i } , \\tilde { w } _ { i } ) } = f _ { i } ^ { \\tilde { w } _ { i } } \\circ f _ { i } ^ { w _ { i } }", + "type": "interline_equation", + "image_path": "eacf17365fe1bb32a28659b1c5a881dee0fdc9ac4b0b594b9ccefe3e1423f078.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 261, + 532, + 349, + 548 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 548, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 195, + 561 + ], + "score": 1.0, + "content": "with the initialization", + "type": "text" + }, + { + "bbox": [ + 196, + 549, + 232, + 559 + ], + "score": 0.92, + "content": "\\tilde { w } _ { i } = w _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 546, + 506, + 561 + ], + "score": 1.0, + "content": ". This trivially also holds for idempotent function without weights,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 558, + 150, + 572 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 150, + 572 + ], + "score": 1.0, + "content": "e.g., Relu.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 366, + 587 + ], + "lines": [ + { + "bbox": [ + 104, + 572, + 366, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 279, + 591 + ], + "score": 1.0, + "content": "Network morphism Type IV. Every layer", + "type": "text" + }, + { + "bbox": [ + 279, + 576, + 295, + 588 + ], + "score": 0.91, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 572, + 366, + 591 + ], + "score": 1.0, + "content": "is replaceable by", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 590, + 423, + 604 + ], + "lines": [ + { + "bbox": [ + 187, + 590, + 423, + 604 + ], + "spans": [ + { + "bbox": [ + 187, + 590, + 423, + 604 + ], + "score": 0.87, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\lambda f _ { i } ^ { w _ { i } } ( x ) + ( 1 - \\lambda ) h ^ { w _ { h } } ( x ) , \\quad \\tilde { w } _ { i } = ( w _ { i } , \\lambda , w _ { h } )", + "type": "interline_equation", + "image_path": "fd47040a01a63f7e977805f8e17a1b0019034a8e830d36d15a69e1bb6c4d3d41.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 187, + 590, + 423, + 604 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 213, + 618 + ], + "score": 1.0, + "content": "with an arbitrary function", + "type": "text" + }, + { + "bbox": [ + 213, + 606, + 221, + 616 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 606, + 327, + 618 + ], + "score": 1.0, + "content": "and Equation (1) holds if", + "type": "text" + }, + { + "bbox": [ + 327, + 606, + 334, + 616 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "is initialized as 1. This morphism can be", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "used to incorporate any function, especially any non-linearities. For example, Wei et al. (2016) use", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 640 + ], + "score": 1.0, + "content": "a special case of this operator to deal with non-linear, non-idempotent activation functions. Another", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 639, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 161, + 676 + ], + "score": 1.0, + "content": "example wou(2016) to sim", + "type": "text" + }, + { + "bbox": [ + 201, + 639, + 228, + 676 + ], + "score": 1.0, + "content": "insertioing: If", + "type": "text" + }, + { + "bbox": [ + 228, + 651, + 244, + 663 + ], + "score": 0.91, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 639, + 289, + 676 + ], + "score": 1.0, + "content": "f an additiv itself is a s", + "type": "text" + }, + { + "bbox": [ + 309, + 639, + 364, + 676 + ], + "score": 1.0, + "content": "p connection,nce of layers,", + "type": "text" + }, + { + "bbox": [ + 365, + 650, + 464, + 664 + ], + "score": 0.92, + "content": "f _ { i } ^ { w _ { i } } = f _ { i _ { n } } ^ { w _ { i _ { n } } } \\circ \\cdot \\cdot \\cdot \\circ f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 639, + 505, + 676 + ], + "score": 1.0, + "content": "He et al., then one", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 162, + 663, + 436, + 678 + ], + "spans": [ + { + "bbox": [ + 162, + 664, + 201, + 677 + ], + "score": 0.93, + "content": "h ( x ) = x", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 664, + 289, + 678 + ], + "score": 1.0, + "content": "to realize a skip from", + "type": "text" + }, + { + "bbox": [ + 289, + 663, + 309, + 678 + ], + "score": 0.92, + "content": "f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 664, + 415, + 678 + ], + "score": 1.0, + "content": "to the layer subsequent to", + "type": "text" + }, + { + "bbox": [ + 415, + 664, + 436, + 678 + ], + "score": 0.91, + "content": "f _ { i _ { n } } ^ { w _ { i _ { n } } }", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 715 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "Note that every combinations of the network morphisms again yields a morphism. So one could for", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 691, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 706 + ], + "score": 1.0, + "content": "example insert a block ”Conv-BatchNorm-Relu” subsequent to a Relu layer by using equations (2),", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 703, + 155, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 155, + 717 + ], + "score": 1.0, + "content": "(3) and (5).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44 + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 721, + 196, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 198, + 732 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 198, + 732 + ], + "score": 1.0, + "content": "1In abuse of notation.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 239, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 240, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 240, + 95 + ], + "score": 1.0, + "content": "3 NETWORK MORPHISM", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 105, + 505, + 139 + ], + "lines": [ + { + "bbox": [ + 104, + 102, + 505, + 120 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 122, + 120 + ], + "score": 1.0, + "content": "Let", + "type": "text" + }, + { + "bbox": [ + 122, + 105, + 149, + 117 + ], + "score": 0.89, + "content": "\\mathcal { N } ( \\mathcal { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 150, + 102, + 322, + 120 + ], + "score": 1.0, + "content": "denote a set of neural networks defined on", + "type": "text" + }, + { + "bbox": [ + 323, + 106, + 358, + 116 + ], + "score": 0.9, + "content": "\\mathcal { X } \\subset \\mathbb { R } ^ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 359, + 102, + 505, + 120 + ], + "score": 1.0, + "content": ". A network morphism is a mapping", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 106, + 116, + 244, + 128 + ], + "score": 0.91, + "content": "M : \\dot { \\mathcal { N } } ( \\dot { \\mathcal { X } } ) \\times \\mathbb { R } ^ { k } \\mathcal { N } ( \\underline { { \\mathcal { X } } } ) \\times \\mathbb { R } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 114, + 338, + 129 + ], + "score": 1.0, + "content": "from a neural network", + "type": "text" + }, + { + "bbox": [ + 339, + 116, + 391, + 128 + ], + "score": 0.93, + "content": "f ^ { w } \\in \\mathcal { N } ( \\mathcal { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 391, + 114, + 459, + 129 + ], + "score": 1.0, + "content": "with parameters", + "type": "text" + }, + { + "bbox": [ + 460, + 116, + 493, + 127 + ], + "score": 0.92, + "content": "w \\in \\mathbb { R } ^ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 114, + 505, + 129 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 127, + 383, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 127, + 201, + 140 + ], + "score": 1.0, + "content": "another neural network", + "type": "text" + }, + { + "bbox": [ + 201, + 128, + 252, + 140 + ], + "score": 0.92, + "content": "g ^ { \\tilde { w } } \\in \\mathcal { N } ( \\mathcal { X } )", + "type": "inline_equation" + }, + { + "bbox": [ + 252, + 127, + 320, + 140 + ], + "score": 1.0, + "content": "with parameters", + "type": "text" + }, + { + "bbox": [ + 320, + 128, + 352, + 138 + ], + "score": 0.91, + "content": "\\tilde { w } \\in \\mathbb { R } ^ { j }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 127, + 383, + 140 + ], + "score": 1.0, + "content": "so that", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 104, + 102, + 505, + 140 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 236, + 141, + 375, + 155 + ], + "lines": [ + { + "bbox": [ + 236, + 141, + 375, + 155 + ], + "spans": [ + { + "bbox": [ + 236, + 141, + 375, + 155 + ], + "score": 0.88, + "content": "f ^ { w } ( x ) = g ^ { \\tilde { w } } ( x ) \\mathrm { f o r e v e r y } x \\in \\mathcal { X } .", + "type": "interline_equation", + "image_path": "01da69f32520eccbef16ce0ecf9fbbcf2637c41fb3974f8b8230761a9c0e4b45.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 236, + 141, + 375, + 155 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 162, + 505, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 505, + 175 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 505, + 175 + ], + "score": 1.0, + "content": "In the following we give a few examples of network morphisms and how standard operations for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 172, + 505, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 187 + ], + "score": 1.0, + "content": "building neural networks (e.g., adding a convolutional layer) can be expressed as a network mor-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 183, + 453, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 184, + 199 + ], + "score": 1.0, + "content": "phism. For this, let", + "type": "text" + }, + { + "bbox": [ + 185, + 185, + 214, + 197 + ], + "score": 0.92, + "content": "f _ { i } ^ { w _ { i } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 183, + 304, + 199 + ], + "score": 1.0, + "content": "be some part of a NN", + "type": "text" + }, + { + "bbox": [ + 304, + 185, + 330, + 197 + ], + "score": 0.91, + "content": "f ^ { w } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 183, + 453, + 199 + ], + "score": 1.0, + "content": ", e.g., a layer or a subnetwork.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 163, + 505, + 199 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 201, + 301, + 213 + ], + "lines": [ + { + "bbox": [ + 105, + 199, + 301, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 271, + 217 + ], + "score": 1.0, + "content": "Network morphism Type I. We replace", + "type": "text" + }, + { + "bbox": [ + 271, + 201, + 287, + 214 + ], + "score": 0.9, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 199, + 301, + 217 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8, + "bbox_fs": [ + 105, + 199, + 301, + 217 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 255, + 215, + 355, + 231 + ], + "lines": [ + { + "bbox": [ + 255, + 215, + 355, + 231 + ], + "spans": [ + { + "bbox": [ + 255, + 215, + 355, + 231 + ], + "score": 0.92, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = A f _ { i } ^ { w _ { i } } ( x ) + b ,", + "type": "interline_equation", + "image_path": "739fb4573c092613f8d0559923d6bb3981e5fbbefbec26a727c76f98d046b2eb.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 255, + 215, + 355, + 231 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 232, + 505, + 275 + ], + "lines": [ + { + "bbox": [ + 106, + 232, + 505, + 246 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 127, + 246 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 127, + 232, + 195, + 245 + ], + "score": 0.91, + "content": "\\tilde { w } _ { i } = ( w _ { i } , A , b ) ^ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 232, + 334, + 246 + ], + "score": 1.0, + "content": ". Equation (1) obviously holds for", + "type": "text" + }, + { + "bbox": [ + 335, + 233, + 388, + 244 + ], + "score": 0.91, + "content": "A = \\mathbf { 1 } , b = \\mathbf { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 389, + 232, + 505, + 246 + ], + "score": 1.0, + "content": ". This morphism can be used", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 243, + 504, + 257 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 504, + 257 + ], + "score": 1.0, + "content": "to add a fully-connected or convolutional layer, as these layers are simply linear mappings. Chen", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 254, + 505, + 269 + ], + "spans": [ + { + "bbox": [ + 105, + 254, + 505, + 269 + ], + "score": 1.0, + "content": "et al. (2015) dubbed this morphism ”Net2DeeperNet”. Alternatively to the above replacement, one", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 264, + 180, + 279 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 180, + 279 + ], + "score": 1.0, + "content": "could also choose", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 232, + 505, + 279 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 239, + 274, + 371, + 290 + ], + "lines": [ + { + "bbox": [ + 239, + 274, + 371, + 290 + ], + "spans": [ + { + "bbox": [ + 239, + 274, + 371, + 290 + ], + "score": 0.92, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = C ( A f _ { i } ^ { w _ { i } } ( x ) + b ) + d ,", + "type": "interline_equation", + "image_path": "7999d9061602778300a91a1b4c09054ac86a6cafaa7aa40fb422aa6a160d5ca7.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 239, + 274, + 371, + 290 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 105, + 294, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 127, + 308 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 128, + 295, + 194, + 307 + ], + "score": 0.87, + "content": "\\tilde { w } _ { i } = ( w _ { i } , C , d )", + "type": "inline_equation" + }, + { + "bbox": [ + 194, + 294, + 200, + 308 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 200, + 295, + 218, + 307 + ], + "score": 0.7, + "content": "A , b", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 294, + 506, + 308 + ], + "score": 1.0, + "content": "are fixed, non-learnable. In this case network morphism Equation (1)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 104, + 305, + 505, + 319 + ], + "spans": [ + { + "bbox": [ + 104, + 305, + 141, + 319 + ], + "score": 1.0, + "content": "holds if", + "type": "text" + }, + { + "bbox": [ + 141, + 307, + 232, + 318 + ], + "score": 0.87, + "content": "C = A ^ { - 1 } , d = - C b", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 305, + 505, + 319 + ], + "score": 1.0, + "content": ". A Batch Normalization layer (or other normalization layers) can", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 316, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 104, + 316, + 229, + 330 + ], + "score": 1.0, + "content": "be written in the above form:", + "type": "text" + }, + { + "bbox": [ + 229, + 318, + 246, + 329 + ], + "score": 0.9, + "content": "A , b", + "type": "inline_equation" + }, + { + "bbox": [ + 247, + 316, + 382, + 330 + ], + "score": 1.0, + "content": "represent the batch statistics and", + "type": "text" + }, + { + "bbox": [ + 383, + 317, + 401, + 329 + ], + "score": 0.9, + "content": "C , d", + "type": "inline_equation" + }, + { + "bbox": [ + 401, + 316, + 505, + 330 + ], + "score": 1.0, + "content": "the learnable scaling and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 326, + 143, + 343 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 143, + 343 + ], + "score": 1.0, + "content": "shifting.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 104, + 294, + 506, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 344, + 502, + 368 + ], + "lines": [ + { + "bbox": [ + 105, + 343, + 505, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 343, + 266, + 360 + ], + "score": 1.0, + "content": "Network morphism Type II. Assume", + "type": "text" + }, + { + "bbox": [ + 266, + 345, + 282, + 357 + ], + "score": 0.89, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 282, + 343, + 339, + 360 + ], + "score": 1.0, + "content": "has the form", + "type": "text" + }, + { + "bbox": [ + 339, + 345, + 439, + 357 + ], + "score": 0.94, + "content": "f _ { i } ^ { w _ { i } } ( x ) = A h ^ { w _ { h } } ( x ) + b", + "type": "inline_equation" + }, + { + "bbox": [ + 439, + 343, + 505, + 360 + ], + "score": 1.0, + "content": "for an arbitrary", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 354, + 303, + 370 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 142, + 370 + ], + "score": 1.0, + "content": "function", + "type": "text" + }, + { + "bbox": [ + 142, + 357, + 149, + 366 + ], + "score": 0.77, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 354, + 200, + 370 + ], + "score": 1.0, + "content": ". We replace", + "type": "text" + }, + { + "bbox": [ + 200, + 357, + 216, + 369 + ], + "score": 0.85, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 217, + 354, + 221, + 370 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 221, + 356, + 285, + 368 + ], + "score": 0.88, + "content": "w _ { i } = ( w _ { h } , \\dot { A } , b )", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 354, + 303, + 370 + ], + "score": 1.0, + "content": ", by", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 343, + 505, + 370 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 234, + 369, + 377, + 397 + ], + "lines": [ + { + "bbox": [ + 234, + 369, + 377, + 397 + ], + "spans": [ + { + "bbox": [ + 234, + 369, + 377, + 397 + ], + "score": 0.94, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\left( A \\quad \\tilde { A } \\right) \\binom { h ^ { w _ { h } } ( x ) } { \\tilde { h } ^ { w _ { h } } ( x ) } + b", + "type": "interline_equation", + "image_path": "f4578e30ebce96eb214241b79c83d5935f0306cd6cc67ef9b418c714979d56e3.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 234, + 369, + 377, + 397 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 399, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 106, + 398, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 212, + 414 + ], + "score": 1.0, + "content": "with an arbitrary function", + "type": "text" + }, + { + "bbox": [ + 212, + 398, + 243, + 412 + ], + "score": 0.94, + "content": "\\tilde { h } ^ { w _ { \\tilde { h } } } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 399, + 346, + 414 + ], + "score": 1.0, + "content": ". The new parameters are", + "type": "text" + }, + { + "bbox": [ + 346, + 399, + 418, + 413 + ], + "score": 0.94, + "content": "\\tilde { w } _ { i } = ( w _ { i } , w _ { \\tilde { h } } , \\tilde { A } )", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 399, + 506, + 414 + ], + "score": 1.0, + "content": ". Again, Equation (1)", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 412, + 505, + 426 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 249, + 426 + ], + "score": 1.0, + "content": "can trivially be satisfied by setting", + "type": "text" + }, + { + "bbox": [ + 249, + 412, + 279, + 424 + ], + "score": 0.91, + "content": "\\tilde { A } = 0", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 414, + 505, + 426 + ], + "score": 1.0, + "content": ". We think of two modifications of a NN which can be", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 506, + 437 + ], + "score": 1.0, + "content": "expressed by this morphism. Firstly, a layer can be widened (i.e., increasing the number of units", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 447 + ], + "score": 1.0, + "content": "in a fully connected layer or the number of channels in a CNN - the Net2WiderNet transformation", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 255, + 459 + ], + "score": 1.0, + "content": "from Chen et al. (2015)). Think of", + "type": "text" + }, + { + "bbox": [ + 255, + 446, + 276, + 459 + ], + "score": 0.92, + "content": "h ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 446, + 505, + 459 + ], + "score": 1.0, + "content": "as the layer to be widened. For example, we can then", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 505, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 120, + 471 + ], + "score": 1.0, + "content": "set", + "type": "text" + }, + { + "bbox": [ + 121, + 457, + 151, + 469 + ], + "score": 0.91, + "content": "\\tilde { h } = h", + "type": "inline_equation" + }, + { + "bbox": [ + 151, + 458, + 505, + 471 + ], + "score": 1.0, + "content": "to simply double the width. Secondly, skip-connections by concatenation as used by", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 469, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 469, + 367, + 483 + ], + "score": 1.0, + "content": "Huang et al. (2016) can be formulated as a network morphism. If", + "type": "text" + }, + { + "bbox": [ + 367, + 470, + 387, + 482 + ], + "score": 0.92, + "content": "h ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 388, + 469, + 505, + 483 + ], + "score": 1.0, + "content": "itself is a sequence of layers,", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 481, + 506, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 482, + 220, + 495 + ], + "score": 0.91, + "content": "h ( x ) = h _ { n } ( x ) \\circ \\cdots \\circ h _ { 0 } ( x )", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 482, + 315, + 496 + ], + "score": 1.0, + "content": ", then one could choose", + "type": "text" + }, + { + "bbox": [ + 316, + 481, + 356, + 495 + ], + "score": 0.94, + "content": "\\tilde { h } ( x ) = x", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 482, + 444, + 496 + ], + "score": 1.0, + "content": "to realize a skip from", + "type": "text" + }, + { + "bbox": [ + 444, + 483, + 456, + 494 + ], + "score": 0.87, + "content": "h _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 482, + 506, + 496 + ], + "score": 1.0, + "content": "to the layer", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 493, + 180, + 507 + ], + "spans": [ + { + "bbox": [ + 105, + 493, + 163, + 507 + ], + "score": 1.0, + "content": "subsequent to", + "type": "text" + }, + { + "bbox": [ + 164, + 495, + 176, + 505 + ], + "score": 0.87, + "content": "h _ { n }", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 493, + 180, + 507 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 398, + 506, + 507 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 510, + 504, + 533 + ], + "lines": [ + { + "bbox": [ + 105, + 509, + 506, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 394, + 524 + ], + "score": 1.0, + "content": "Network morphism Type III. By definition, every idempotent function", + "type": "text" + }, + { + "bbox": [ + 395, + 511, + 410, + 523 + ], + "score": 0.91, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 509, + 506, + 524 + ], + "score": 1.0, + "content": "can simply be replaced", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 521, + 120, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 120, + 536 + ], + "score": 1.0, + "content": "by", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 509, + 506, + 536 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 261, + 532, + 349, + 548 + ], + "lines": [ + { + "bbox": [ + 261, + 532, + 349, + 548 + ], + "spans": [ + { + "bbox": [ + 261, + 532, + 349, + 548 + ], + "score": 0.92, + "content": "f _ { i } ^ { ( w _ { i } , \\tilde { w } _ { i } ) } = f _ { i } ^ { \\tilde { w } _ { i } } \\circ f _ { i } ^ { w _ { i } }", + "type": "interline_equation", + "image_path": "eacf17365fe1bb32a28659b1c5a881dee0fdc9ac4b0b594b9ccefe3e1423f078.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 261, + 532, + 349, + 548 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 548, + 505, + 570 + ], + "lines": [ + { + "bbox": [ + 105, + 546, + 506, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 195, + 561 + ], + "score": 1.0, + "content": "with the initialization", + "type": "text" + }, + { + "bbox": [ + 196, + 549, + 232, + 559 + ], + "score": 0.92, + "content": "\\tilde { w } _ { i } = w _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 546, + 506, + 561 + ], + "score": 1.0, + "content": ". This trivially also holds for idempotent function without weights,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 558, + 150, + 572 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 150, + 572 + ], + "score": 1.0, + "content": "e.g., Relu.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 104, + 546, + 506, + 572 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 575, + 366, + 587 + ], + "lines": [ + { + "bbox": [ + 104, + 572, + 366, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 572, + 279, + 591 + ], + "score": 1.0, + "content": "Network morphism Type IV. Every layer", + "type": "text" + }, + { + "bbox": [ + 279, + 576, + 295, + 588 + ], + "score": 0.91, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 572, + 366, + 591 + ], + "score": 1.0, + "content": "is replaceable by", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 36, + "bbox_fs": [ + 104, + 572, + 366, + 591 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 187, + 590, + 423, + 604 + ], + "lines": [ + { + "bbox": [ + 187, + 590, + 423, + 604 + ], + "spans": [ + { + "bbox": [ + 187, + 590, + 423, + 604 + ], + "score": 0.87, + "content": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\lambda f _ { i } ^ { w _ { i } } ( x ) + ( 1 - \\lambda ) h ^ { w _ { h } } ( x ) , \\quad \\tilde { w } _ { i } = ( w _ { i } , \\lambda , w _ { h } )", + "type": "interline_equation", + "image_path": "fd47040a01a63f7e977805f8e17a1b0019034a8e830d36d15a69e1bb6c4d3d41.jpg" + } + ] + } + ], + "index": 37, + "virtual_lines": [ + { + "bbox": [ + 187, + 590, + 423, + 604 + ], + "spans": [], + "index": 37 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 606, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 606, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 213, + 618 + ], + "score": 1.0, + "content": "with an arbitrary function", + "type": "text" + }, + { + "bbox": [ + 213, + 606, + 221, + 616 + ], + "score": 0.8, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 606, + 327, + 618 + ], + "score": 1.0, + "content": "and Equation (1) holds if", + "type": "text" + }, + { + "bbox": [ + 327, + 606, + 334, + 616 + ], + "score": 0.81, + "content": "\\lambda", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 606, + 505, + 618 + ], + "score": 1.0, + "content": "is initialized as 1. This morphism can be", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 106, + 617, + 505, + 629 + ], + "score": 1.0, + "content": "used to incorporate any function, especially any non-linearities. For example, Wei et al. (2016) use", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 628, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 506, + 640 + ], + "score": 1.0, + "content": "a special case of this operator to deal with non-linear, non-idempotent activation functions. Another", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 639, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 161, + 676 + ], + "score": 1.0, + "content": "example wou(2016) to sim", + "type": "text" + }, + { + "bbox": [ + 201, + 639, + 228, + 676 + ], + "score": 1.0, + "content": "insertioing: If", + "type": "text" + }, + { + "bbox": [ + 228, + 651, + 244, + 663 + ], + "score": 0.91, + "content": "f _ { i } ^ { w _ { i } }", + "type": "inline_equation" + }, + { + "bbox": [ + 244, + 639, + 289, + 676 + ], + "score": 1.0, + "content": "f an additiv itself is a s", + "type": "text" + }, + { + "bbox": [ + 309, + 639, + 364, + 676 + ], + "score": 1.0, + "content": "p connection,nce of layers,", + "type": "text" + }, + { + "bbox": [ + 365, + 650, + 464, + 664 + ], + "score": 0.92, + "content": "f _ { i } ^ { w _ { i } } = f _ { i _ { n } } ^ { w _ { i _ { n } } } \\circ \\cdot \\cdot \\cdot \\circ f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 639, + 505, + 676 + ], + "score": 1.0, + "content": "He et al., then one", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 162, + 663, + 436, + 678 + ], + "spans": [ + { + "bbox": [ + 162, + 664, + 201, + 677 + ], + "score": 0.93, + "content": "h ( x ) = x", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 664, + 289, + 678 + ], + "score": 1.0, + "content": "to realize a skip from", + "type": "text" + }, + { + "bbox": [ + 289, + 663, + 309, + 678 + ], + "score": 0.92, + "content": "f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 664, + 415, + 678 + ], + "score": 1.0, + "content": "to the layer subsequent to", + "type": "text" + }, + { + "bbox": [ + 415, + 664, + 436, + 678 + ], + "score": 0.91, + "content": "f _ { i _ { n } } ^ { w _ { i _ { n } } }", + "type": "inline_equation" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 606, + 506, + 678 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 681, + 505, + 715 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 505, + 694 + ], + "score": 1.0, + "content": "Note that every combinations of the network morphisms again yields a morphism. So one could for", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 691, + 505, + 706 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 706 + ], + "score": 1.0, + "content": "example insert a block ”Conv-BatchNorm-Relu” subsequent to a Relu layer by using equations (2),", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 703, + 155, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 703, + 155, + 717 + ], + "score": 1.0, + "content": "(3) and (5).", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44, + "bbox_fs": [ + 105, + 681, + 505, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 78, + 506, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 78, + 506, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 506, + 250 + ], + "score": 0.973, + "type": "image", + "image_path": "f8ca3cbd2a4cdfcdcf35ac65c139d715df875d1be2779a18db0872654bca035b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 78, + 506, + 135.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 135.33333333333334, + 506, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 192.66666666666669, + 506, + 250.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 271, + 504, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 1: Visualization of our method. Based on the current best model, new models are generated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 282, + 329, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 329, + 293 + ], + "score": 1.0, + "content": "and trained afterwards. The best model is than updated.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 108, + 321, + 392, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 393, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 393, + 335 + ], + "score": 1.0, + "content": "4 ARCHITECTURE SEARCH BY NETWORK MORPHISMS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "score": 1.0, + "content": "Our proposed algorithm is a simple hill climbing strategy (Russell & Norvig, 2009). We start with a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "small, (possibly) pretrained network. Then, we apply network morphisms to this initial network to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "generate larger ones that may perform better when trained further. These new “child” networks can", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "be seen as neighbors of the initial “parent” network in the space of network architectures. Due to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "network morphism Equation (1), the child networks start at the same performance as their parent.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "In essence, network morphisms can thus be seen as a way to initialize child networks to perform", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "score": 1.0, + "content": "well, avoiding the expensive step of training them from scratch and thereby reducing the cost of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "their evaluation. The various child networks can then be trained further for a brief period of time to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "exploit the additional capacity obtained by the network morphism, and the search can move on to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "the best resulting child network. This constitutes one step of our proposed algorithm, which we dub", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "Neural Architecture Search by Hill-climbing (NASH). NASH can execute this step several times", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "until performance on a validation set saturates; we note that this greedy process may in principle get", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "stuck in a poorly-performing region, from which it can never escape, but we did not find evidence", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 496, + 217, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 217, + 508 + ], + "score": 1.0, + "content": "for this in our experiments.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Figure 1 visualizes one step of the NASH approach, and Algorithm 1 provides full details for the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 523, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 304, + 536 + ], + "score": 1.0, + "content": "algorithm. In our implementation, the function", + "type": "text" + }, + { + "bbox": [ + 304, + 523, + 424, + 535 + ], + "score": 0.31, + "content": "A p p l y N e t M o r p h ( m o d e l , n )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 524, + 496, + 536 + ], + "score": 1.0, + "content": "(line 15) applies", + "type": "text" + }, + { + "bbox": [ + 497, + 525, + 504, + 533 + ], + "score": 0.7, + "content": "n", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 534, + 468, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 468, + 546 + ], + "score": 1.0, + "content": "network morphisms, each of them sampled uniformly at random from the following three:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 132, + 559, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 133, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 133, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "• Make the network deeper, i.e., add a ”Conv-BatchNorm-Relu” block as described at the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 141, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 141, + 570, + 460, + 583 + ], + "score": 1.0, + "content": "end of Section 3. The position where to add the block, as well as the kernel size", + "type": "text" + }, + { + "bbox": [ + 460, + 570, + 500, + 583 + ], + "score": 0.85, + "content": "( \\in \\{ 3 , 5 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 570, + 505, + 583 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 142, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "are uniformly sampled. The number of channels is chosen to be equal to he number of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 592, + 328, + 604 + ], + "spans": [ + { + "bbox": [ + 142, + 592, + 328, + 604 + ], + "score": 1.0, + "content": "channels of the closest preceding convolution.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 133, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 133, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "• Make the network wider, i.e., increase the number of channels by using the network mor-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 141, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 141, + 627, + 449, + 640 + ], + "score": 1.0, + "content": "phism type II. The conv layer to be widened, as well as the widening factor", + "type": "text" + }, + { + "bbox": [ + 449, + 627, + 488, + 639 + ], + "score": 0.89, + "content": "( \\in \\{ 2 , 4 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 639, + 265, + 650 + ], + "spans": [ + { + "bbox": [ + 141, + 639, + 265, + 650 + ], + "score": 1.0, + "content": "sampled uniformly at random.", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 133, + 661, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 133, + 661, + 505, + 675 + ], + "score": 1.0, + "content": "• Add a skip connection from layer i to layer j (either by concatenation or addition – uni-", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 141, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 141, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "formly sampled) by using network morphism type II or IV, respectively. Layers i and j are", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 683, + 241, + 697 + ], + "spans": [ + { + "bbox": [ + 141, + 683, + 241, + 697 + ], + "score": 1.0, + "content": "also sampled uniformly.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 27.5 + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Note that the current best model is also considered as a child, i.e. our algorithm is not forced to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "score": 1.0, + "content": "select a new model but can rather also keep the old one if no other one improves upon it.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 78, + 506, + 250 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 78, + 506, + 250 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 78, + 506, + 250 + ], + "spans": [ + { + "bbox": [ + 106, + 78, + 506, + 250 + ], + "score": 0.973, + "type": "image", + "image_path": "f8ca3cbd2a4cdfcdcf35ac65c139d715df875d1be2779a18db0872654bca035b.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 78, + 506, + 135.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 135.33333333333334, + 506, + 192.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 192.66666666666669, + 506, + 250.00000000000003 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 271, + 504, + 294 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "Figure 1: Visualization of our method. Based on the current best model, new models are generated", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 282, + 329, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 329, + 293 + ], + "score": 1.0, + "content": "and trained afterwards. The best model is than updated.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "title", + "bbox": [ + 108, + 321, + 392, + 334 + ], + "lines": [ + { + "bbox": [ + 105, + 321, + 393, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 321, + 393, + 335 + ], + "score": 1.0, + "content": "4 ARCHITECTURE SEARCH BY NETWORK MORPHISMS", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 107, + 352, + 505, + 506 + ], + "lines": [ + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 106, + 353, + 506, + 365 + ], + "score": 1.0, + "content": "Our proposed algorithm is a simple hill climbing strategy (Russell & Norvig, 2009). We start with a", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 506, + 376 + ], + "score": 1.0, + "content": "small, (possibly) pretrained network. Then, we apply network morphisms to this initial network to", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "generate larger ones that may perform better when trained further. These new “child” networks can", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "be seen as neighbors of the initial “parent” network in the space of network architectures. Due to the", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 505, + 409 + ], + "score": 1.0, + "content": "network morphism Equation (1), the child networks start at the same performance as their parent.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 505, + 420 + ], + "score": 1.0, + "content": "In essence, network morphisms can thus be seen as a way to initialize child networks to perform", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "spans": [ + { + "bbox": [ + 105, + 418, + 506, + 431 + ], + "score": 1.0, + "content": "well, avoiding the expensive step of training them from scratch and thereby reducing the cost of", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 429, + 506, + 442 + ], + "spans": [ + { + "bbox": [ + 105, + 429, + 506, + 442 + ], + "score": 1.0, + "content": "their evaluation. The various child networks can then be trained further for a brief period of time to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 440, + 506, + 453 + ], + "spans": [ + { + "bbox": [ + 106, + 440, + 506, + 453 + ], + "score": 1.0, + "content": "exploit the additional capacity obtained by the network morphism, and the search can move on to", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 452, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 106, + 452, + 506, + 465 + ], + "score": 1.0, + "content": "the best resulting child network. This constitutes one step of our proposed algorithm, which we dub", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 461, + 505, + 475 + ], + "score": 1.0, + "content": "Neural Architecture Search by Hill-climbing (NASH). NASH can execute this step several times", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "score": 1.0, + "content": "until performance on a validation set saturates; we note that this greedy process may in principle get", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "stuck in a poorly-performing region, from which it can never escape, but we did not find evidence", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 496, + 217, + 508 + ], + "spans": [ + { + "bbox": [ + 105, + 496, + 217, + 508 + ], + "score": 1.0, + "content": "for this in our experiments.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 12.5, + "bbox_fs": [ + 105, + 353, + 506, + 508 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 512, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 512, + 505, + 525 + ], + "score": 1.0, + "content": "Figure 1 visualizes one step of the NASH approach, and Algorithm 1 provides full details for the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 523, + 504, + 536 + ], + "spans": [ + { + "bbox": [ + 106, + 524, + 304, + 536 + ], + "score": 1.0, + "content": "algorithm. In our implementation, the function", + "type": "text" + }, + { + "bbox": [ + 304, + 523, + 424, + 535 + ], + "score": 0.31, + "content": "A p p l y N e t M o r p h ( m o d e l , n )", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 524, + 496, + 536 + ], + "score": 1.0, + "content": "(line 15) applies", + "type": "text" + }, + { + "bbox": [ + 497, + 525, + 504, + 533 + ], + "score": 0.7, + "content": "n", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 534, + 468, + 546 + ], + "spans": [ + { + "bbox": [ + 106, + 534, + 468, + 546 + ], + "score": 1.0, + "content": "network morphisms, each of them sampled uniformly at random from the following three:", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 512, + 505, + 546 + ] + }, + { + "type": "list", + "bbox": [ + 132, + 559, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 133, + 559, + 505, + 571 + ], + "spans": [ + { + "bbox": [ + 133, + 559, + 505, + 571 + ], + "score": 1.0, + "content": "• Make the network deeper, i.e., add a ”Conv-BatchNorm-Relu” block as described at the", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 570, + 505, + 583 + ], + "spans": [ + { + "bbox": [ + 141, + 570, + 460, + 583 + ], + "score": 1.0, + "content": "end of Section 3. The position where to add the block, as well as the kernel size", + "type": "text" + }, + { + "bbox": [ + 460, + 570, + 500, + 583 + ], + "score": 0.85, + "content": "( \\in \\{ 3 , 5 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 500, + 570, + 505, + 583 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 142, + 582, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 142, + 582, + 505, + 594 + ], + "score": 1.0, + "content": "are uniformly sampled. The number of channels is chosen to be equal to he number of", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 142, + 592, + 328, + 604 + ], + "spans": [ + { + "bbox": [ + 142, + 592, + 328, + 604 + ], + "score": 1.0, + "content": "channels of the closest preceding convolution.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 615, + 505, + 629 + ], + "spans": [ + { + "bbox": [ + 133, + 615, + 505, + 629 + ], + "score": 1.0, + "content": "• Make the network wider, i.e., increase the number of channels by using the network mor-", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 627, + 505, + 640 + ], + "spans": [ + { + "bbox": [ + 141, + 627, + 449, + 640 + ], + "score": 1.0, + "content": "phism type II. The conv layer to be widened, as well as the widening factor", + "type": "text" + }, + { + "bbox": [ + 449, + 627, + 488, + 639 + ], + "score": 0.89, + "content": "( \\in \\{ 2 , 4 \\} )", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 627, + 505, + 640 + ], + "score": 1.0, + "content": "are", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 141, + 639, + 265, + 650 + ], + "spans": [ + { + "bbox": [ + 141, + 639, + 265, + 650 + ], + "score": 1.0, + "content": "sampled uniformly at random.", + "type": "text" + } + ], + "index": 29, + "is_list_end_line": true + }, + { + "bbox": [ + 133, + 661, + 505, + 675 + ], + "spans": [ + { + "bbox": [ + 133, + 661, + 505, + 675 + ], + "score": 1.0, + "content": "• Add a skip connection from layer i to layer j (either by concatenation or addition – uni-", + "type": "text" + } + ], + "index": 30, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 673, + 505, + 686 + ], + "spans": [ + { + "bbox": [ + 141, + 673, + 505, + 686 + ], + "score": 1.0, + "content": "formly sampled) by using network morphism type II or IV, respectively. Layers i and j are", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 141, + 683, + 241, + 697 + ], + "spans": [ + { + "bbox": [ + 141, + 683, + 241, + 697 + ], + "score": 1.0, + "content": "also sampled uniformly.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + } + ], + "index": 27.5, + "bbox_fs": [ + 133, + 559, + 505, + 697 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 709, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "Note that the current best model is also considered as a child, i.e. our algorithm is not forced to", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 461, + 733 + ], + "score": 1.0, + "content": "select a new model but can rather also keep the old one if no other one improves upon it.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 709, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 341, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 343, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 343, + 98 + ], + "score": 1.0, + "content": "Algorithm 1 Network architecture search by hill climbing", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 96, + 499, + 395 + ], + "lines": [ + { + "bbox": [ + 107, + 96, + 477, + 109 + ], + "spans": [ + { + "bbox": [ + 107, + 96, + 477, + 109 + ], + "score": 0.41, + "content": "1 \\backslash \\mathrm { ~ f u n c t i o n ~ N A S H } ( m o d e l _ { 0 } , n _ { s t e p s } , n _ { n e i g h } , n _ { N M } , e p o c h _ { n e i g h } , e p o c h _ { f i n a l } , \\lambda _ { e n d } , \\lambda _ { s t a r t } )", + "type": "inline_equation", + "image_path": "d185bf1602e246403649abcdd850b6972171aafd8e2a3db2936cb811f79dd875.jpg" + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 106, + 122, + 118 + ], + "spans": [ + { + "bbox": [ + 110, + 106, + 122, + 118 + ], + "score": 1.0, + "content": "2:", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 114, + 446, + 133 + ], + "spans": [ + { + "bbox": [ + 110, + 117, + 122, + 130 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 137, + 114, + 145, + 133 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 146, + 117, + 178, + 129 + ], + "score": 0.35, + "content": "m o d e l _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 114, + 278, + 133 + ], + "score": 1.0, + "content": ", model to start with,", + "type": "text" + }, + { + "bbox": [ + 279, + 116, + 317, + 130 + ], + "score": 0.85, + "content": "n _ { s t e p s } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 114, + 446, + 133 + ], + "score": 1.0, + "content": "number of hill climbining steps", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 130, + 449, + 144 + ], + "spans": [ + { + "bbox": [ + 110, + 130, + 121, + 143 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 136, + 130, + 145, + 144 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 146, + 130, + 185, + 142 + ], + "score": 0.77, + "content": "n _ { n e i g h } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 130, + 288, + 144 + ], + "score": 1.0, + "content": "number of neighbours,", + "type": "text" + }, + { + "bbox": [ + 288, + 130, + 323, + 142 + ], + "score": 0.82, + "content": "n _ { N M } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 130, + 449, + 144 + ], + "score": 1.0, + "content": "number of net. morph. applied", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 142, + 394, + 157 + ], + "spans": [ + { + "bbox": [ + 110, + 144, + 121, + 156 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 136, + 142, + 144, + 157 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 144, + 143, + 204, + 156 + ], + "score": 0.64, + "content": "e p o c h _ { n e i g h } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 142, + 394, + 157 + ], + "score": 1.0, + "content": "number of epochs for training every neighbour", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 154, + 347, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 157, + 121, + 169 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 136, + 154, + 144, + 171 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 144, + 156, + 202, + 169 + ], + "score": 0.72, + "content": "e p o c h _ { f i n a l } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 154, + 347, + 171 + ], + "score": 1.0, + "content": "number of epochs for final training", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 166, + 383, + 181 + ], + "spans": [ + { + "bbox": [ + 110, + 168, + 121, + 180 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 136, + 166, + 187, + 181 + ], + "score": 1.0, + "content": "# initial LR", + "type": "text" + }, + { + "bbox": [ + 187, + 168, + 212, + 179 + ], + "score": 0.91, + "content": "\\lambda _ { s t a r t }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 166, + 271, + 181 + ], + "score": 1.0, + "content": "is annealed to", + "type": "text" + }, + { + "bbox": [ + 271, + 168, + 291, + 179 + ], + "score": 0.91, + "content": "\\lambda _ { e n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 166, + 383, + 181 + ], + "score": 1.0, + "content": "during SGDR training", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 179, + 122, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 179, + 122, + 190 + ], + "score": 1.0, + "content": "8:", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 110, + 189, + 227, + 201 + ], + "spans": [ + { + "bbox": [ + 110, + 190, + 121, + 201 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 138, + 189, + 227, + 201 + ], + "score": 0.64, + "content": "m o d e l _ { b e s t } m o d e l _ { 0 }", + "type": "inline_equation", + "image_path": "989cc56ce8eb8612b27be1151844c84e03ed0ca4cbb370445a8cb2bcf6cd0d00.jpg" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 198, + 219, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 122, + 212 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 136, + 198, + 219, + 213 + ], + "score": 1.0, + "content": "# start hill climbing", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 209, + 244, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 122, + 223 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 137, + 209, + 155, + 225 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 156, + 212, + 234, + 223 + ], + "score": 0.83, + "content": "i \\gets 1 , \\dots , n _ { s t e p s } \\ { \\bf d }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 209, + 244, + 225 + ], + "score": 1.0, + "content": "o", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 221, + 496, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 122, + 234 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 151, + 221, + 173, + 236 + ], + "score": 1.0, + "content": "#get", + "type": "text" + }, + { + "bbox": [ + 173, + 224, + 201, + 234 + ], + "score": 0.7, + "content": "n _ { n e i g h }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 221, + 255, + 236 + ], + "score": 1.0, + "content": "neighbors of", + "type": "text" + }, + { + "bbox": [ + 255, + 223, + 286, + 234 + ], + "score": 0.68, + "content": "m o d e l _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 221, + 338, + 236 + ], + "score": 1.0, + "content": "by applying", + "type": "text" + }, + { + "bbox": [ + 338, + 225, + 361, + 233 + ], + "score": 0.85, + "content": "n _ { N M }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 221, + 454, + 236 + ], + "score": 1.0, + "content": "network morphisms to", + "type": "text" + }, + { + "bbox": [ + 455, + 222, + 496, + 234 + ], + "score": 0.5, + "content": "m o d e l _ { b e s t }", + "type": "inline_equation" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 232, + 281, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 122, + 245 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 152, + 232, + 170, + 247 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 171, + 234, + 263, + 245 + ], + "score": 0.84, + "content": "j \\gets 1 , \\dots , n _ { n e i g h } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 232, + 281, + 247 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 245, + 370, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 121, + 255 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 167, + 245, + 370, + 257 + ], + "score": 0.65, + "content": "m o d e l _ { j } \\gets A p p l y N e t M o r p h s ( m o d e l _ { b e s t } , n _ { N M } )", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 255, + 374, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 122, + 267 + ], + "score": 1.0, + "content": "15:", + "type": "text" + }, + { + "bbox": [ + 166, + 255, + 374, + 267 + ], + "score": 1.0, + "content": "# train for a few epochs on training set with SGDR", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 266, + 402, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 122, + 278 + ], + "score": 1.0, + "content": "16:", + "type": "text" + }, + { + "bbox": [ + 168, + 266, + 402, + 279 + ], + "score": 0.7, + "content": "m o d e l _ { j } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { j } , e \\bar { p } o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )", + "type": "inline_equation", + "image_path": "328ab971c6bec51b546f24a106ddfe7a221f18a42811b9589b2c22b48b03bf8c.jpg" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 277, + 187, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 122, + 289 + ], + "score": 1.0, + "content": "17:", + "type": "text" + }, + { + "bbox": [ + 153, + 277, + 187, + 288 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 288, + 366, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 122, + 300 + ], + "score": 1.0, + "content": "18:", + "type": "text" + }, + { + "bbox": [ + 151, + 288, + 366, + 300 + ], + "score": 1.0, + "content": "# in fact, last neighbor is always just the current best", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 299, + 417, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 122, + 311 + ], + "score": 1.0, + "content": "19:", + "type": "text" + }, + { + "bbox": [ + 153, + 299, + 417, + 312 + ], + "score": 0.62, + "content": "m o d e l _ { n _ { n e i g h } } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { b e s t } , e p o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )", + "type": "inline_equation" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 309, + 290, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 122, + 322 + ], + "score": 1.0, + "content": "20:", + "type": "text" + }, + { + "bbox": [ + 152, + 309, + 290, + 322 + ], + "score": 1.0, + "content": "# get best model on validation set", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 320, + 383, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 122, + 334 + ], + "score": 1.0, + "content": "21:", + "type": "text" + }, + { + "bbox": [ + 153, + 320, + 383, + 341 + ], + "score": 0.6, + "content": "\\underbrace { \\Bre { \\_ } { a } } { m o d } e l _ { b e s t } \\underbrace { a r g m a x } _ { j = 1 , \\dots , n _ { n e i g h } } \\{ p e r f o r m a n c e _ { v a l i } ( m o d e l _ { j } ) \\}", + "type": "inline_equation", + "image_path": "305597548bb02bd55928f77cef70b478c0f4c25ccd76325015892c958fc657d0.jpg" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 339, + 172, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 121, + 352 + ], + "score": 1.0, + "content": "22:", + "type": "text" + }, + { + "bbox": [ + 137, + 339, + 172, + 351 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 350, + 348, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 122, + 363 + ], + "score": 1.0, + "content": "23:", + "type": "text" + }, + { + "bbox": [ + 137, + 350, + 348, + 363 + ], + "score": 1.0, + "content": "# train the final model on training and validation set", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 362, + 391, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 121, + 374 + ], + "score": 1.0, + "content": "24:", + "type": "text" + }, + { + "bbox": [ + 138, + 362, + 391, + 374 + ], + "score": 0.41, + "content": "m o d e l _ { b e s t } \\gets \\mathrm { S G D R t r a i n } ( m o d e \\bar { l } _ { b e s t } , e p o c h _ { f i n a l } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 371, + 214, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 121, + 385 + ], + "score": 1.0, + "content": "25:", + "type": "text" + }, + { + "bbox": [ + 137, + 371, + 214, + 387 + ], + "score": 1.0, + "content": "return modelbest", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 383, + 180, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 121, + 394 + ], + "score": 1.0, + "content": "26:", + "type": "text" + }, + { + "bbox": [ + 122, + 383, + 180, + 396 + ], + "score": 1.0, + "content": "end function", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 421, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 432 + ], + "score": 1.0, + "content": "It is important for our method that child networks only need to be trained for a few epochs2 (line", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "17). Hence, an optimization algorithm with good anytime performance is required. Therefore, we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "chose the cosine annealing strategy from Loshchilov & Hutter (2017), whereas the learning rate is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 405, + 466 + ], + "score": 1.0, + "content": "implicitly restarted: the training in line 17 always starts with a learning rate", + "type": "text" + }, + { + "bbox": [ + 406, + 454, + 431, + 465 + ], + "score": 0.92, + "content": "\\lambda _ { s t a r t }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "which is annealed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 116, + 478 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 465, + 137, + 476 + ], + "score": 0.91, + "content": "\\lambda _ { e n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "after epochneigh epochs. We use the same learning rate scheduler in the final training (aside", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 475, + 252, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 252, + 488 + ], + "score": 1.0, + "content": "from a different number of epochs).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "While we presented our method as a simple hill-climbing method, we note that it can also be inter-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 396, + 518 + ], + "score": 1.0, + "content": "preted as a very simple evolutionary algorithm with a population size of", + "type": "text" + }, + { + "bbox": [ + 397, + 505, + 424, + 516 + ], + "score": 0.88, + "content": "n _ { n e i g h }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 502, + 505, + 518 + ], + "score": 1.0, + "content": ", no cross-over, net-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 513, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 528 + ], + "score": 1.0, + "content": "work morphisms as mutations, and a selection mechanism that only considers the best-performing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "population member as the parent for the next generation. This interpretation also suggests several", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 335, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 335, + 549 + ], + "score": 1.0, + "content": "promising possibilities for extending our simple method.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35 + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 200, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 201, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 201, + 581 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "score": 1.0, + "content": "We evaluate our method on CIFAR-10 and CIFAR-100. First, we investigate whether our consider-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "ations from the previous chapter coincide with empirical results. We also check if the interplay of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "modifying and training networks harms their eventual performance. Finally, we compare our pro-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 627, + 496, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 496, + 638 + ], + "score": 1.0, + "content": "posed method with other automated architecture algorithms as well as hand crafted architectures.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "We use the same standard data augmentation scheme for both CIFAR datasets used by Loshchilov", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "& Hutter (2017) in all of the following experiments. The training set (50.000 samples) is split up in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "training (40.000) and validation (10.000) set for the purpose of architecture search. Eventually the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "score": 1.0, + "content": "performance is evaluated on the test set. All experiments where run on Nvidia Titan X (Maxwell)", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 701, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 699, + 507, + 714 + ], + "spans": [ + { + "bbox": [ + 117, + 699, + 139, + 714 + ], + "score": 1.0, + "content": "2I.e.,", + "type": "text" + }, + { + "bbox": [ + 140, + 702, + 182, + 712 + ], + "score": 0.77, + "content": "e p o c h _ { n e i g h }", + "type": "inline_equation" + }, + { + "bbox": [ + 183, + 699, + 507, + 714 + ], + "score": 1.0, + "content": "should be small since a lot of networks need to be trained. In fact, the total number of", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 711, + 502, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 711, + 306, + 724 + ], + "score": 1.0, + "content": "epochs for training in our algorithm can be computed as", + "type": "text" + }, + { + "bbox": [ + 306, + 712, + 502, + 723 + ], + "score": 0.8, + "content": "e p o c h _ { t o t a l } = e p o c h _ { n e i g h } n _ { n e i g h } n _ { s t e p s } + e p o c h _ { f i n a l } .", + "type": "inline_equation" + } + ] + }, + { + "bbox": [ + 105, + 721, + 298, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 230, + 733 + ], + "score": 1.0, + "content": "In our later experiments we chose", + "type": "text" + }, + { + "bbox": [ + 231, + 722, + 294, + 732 + ], + "score": 0.89, + "content": "e p o c h _ { n e i g h } = 1 7", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 721, + 298, + 733 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 341, + 93 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 343, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 343, + 98 + ], + "score": 1.0, + "content": "Algorithm 1 Network architecture search by hill climbing", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "index", + "bbox": [ + 107, + 96, + 499, + 395 + ], + "lines": [ + { + "bbox": [ + 107, + 96, + 477, + 109 + ], + "spans": [ + { + "bbox": [ + 107, + 96, + 477, + 109 + ], + "score": 0.41, + "content": "1 \\backslash \\mathrm { ~ f u n c t i o n ~ N A S H } ( m o d e l _ { 0 } , n _ { s t e p s } , n _ { n e i g h } , n _ { N M } , e p o c h _ { n e i g h } , e p o c h _ { f i n a l } , \\lambda _ { e n d } , \\lambda _ { s t a r t } )", + "type": "inline_equation", + "image_path": "d185bf1602e246403649abcdd850b6972171aafd8e2a3db2936cb811f79dd875.jpg" + } + ], + "index": 1, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 106, + 122, + 118 + ], + "spans": [ + { + "bbox": [ + 110, + 106, + 122, + 118 + ], + "score": 1.0, + "content": "2:", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 114, + 446, + 133 + ], + "spans": [ + { + "bbox": [ + 110, + 117, + 122, + 130 + ], + "score": 1.0, + "content": "3:", + "type": "text" + }, + { + "bbox": [ + 137, + 114, + 145, + 133 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 146, + 117, + 178, + 129 + ], + "score": 0.35, + "content": "m o d e l _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 114, + 278, + 133 + ], + "score": 1.0, + "content": ", model to start with,", + "type": "text" + }, + { + "bbox": [ + 279, + 116, + 317, + 130 + ], + "score": 0.85, + "content": "n _ { s t e p s } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 114, + 446, + 133 + ], + "score": 1.0, + "content": "number of hill climbining steps", + "type": "text" + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 130, + 449, + 144 + ], + "spans": [ + { + "bbox": [ + 110, + 130, + 121, + 143 + ], + "score": 1.0, + "content": "4:", + "type": "text" + }, + { + "bbox": [ + 136, + 130, + 145, + 144 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 146, + 130, + 185, + 142 + ], + "score": 0.77, + "content": "n _ { n e i g h } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 130, + 288, + 144 + ], + "score": 1.0, + "content": "number of neighbours,", + "type": "text" + }, + { + "bbox": [ + 288, + 130, + 323, + 142 + ], + "score": 0.82, + "content": "n _ { N M } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 323, + 130, + 449, + 144 + ], + "score": 1.0, + "content": "number of net. morph. applied", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 142, + 394, + 157 + ], + "spans": [ + { + "bbox": [ + 110, + 144, + 121, + 156 + ], + "score": 1.0, + "content": "5:", + "type": "text" + }, + { + "bbox": [ + 136, + 142, + 144, + 157 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 144, + 143, + 204, + 156 + ], + "score": 0.64, + "content": "e p o c h _ { n e i g h } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 142, + 394, + 157 + ], + "score": 1.0, + "content": "number of epochs for training every neighbour", + "type": "text" + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 154, + 347, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 157, + 121, + 169 + ], + "score": 1.0, + "content": "6:", + "type": "text" + }, + { + "bbox": [ + 136, + 154, + 144, + 171 + ], + "score": 1.0, + "content": "#", + "type": "text" + }, + { + "bbox": [ + 144, + 156, + 202, + 169 + ], + "score": 0.72, + "content": "e p o c h _ { f i n a l } \\triangleq", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 154, + 347, + 171 + ], + "score": 1.0, + "content": "number of epochs for final training", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 166, + 383, + 181 + ], + "spans": [ + { + "bbox": [ + 110, + 168, + 121, + 180 + ], + "score": 1.0, + "content": "7:", + "type": "text" + }, + { + "bbox": [ + 136, + 166, + 187, + 181 + ], + "score": 1.0, + "content": "# initial LR", + "type": "text" + }, + { + "bbox": [ + 187, + 168, + 212, + 179 + ], + "score": 0.91, + "content": "\\lambda _ { s t a r t }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 166, + 271, + 181 + ], + "score": 1.0, + "content": "is annealed to", + "type": "text" + }, + { + "bbox": [ + 271, + 168, + 291, + 179 + ], + "score": 0.91, + "content": "\\lambda _ { e n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 166, + 383, + 181 + ], + "score": 1.0, + "content": "during SGDR training", + "type": "text" + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 179, + 122, + 190 + ], + "spans": [ + { + "bbox": [ + 110, + 179, + 122, + 190 + ], + "score": 1.0, + "content": "8:", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 110, + 189, + 227, + 201 + ], + "spans": [ + { + "bbox": [ + 110, + 190, + 121, + 201 + ], + "score": 1.0, + "content": "9:", + "type": "text" + }, + { + "bbox": [ + 138, + 189, + 227, + 201 + ], + "score": 0.64, + "content": "m o d e l _ { b e s t } m o d e l _ { 0 }", + "type": "inline_equation", + "image_path": "989cc56ce8eb8612b27be1151844c84e03ed0ca4cbb370445a8cb2bcf6cd0d00.jpg" + } + ], + "index": 9, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 198, + 219, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 200, + 122, + 212 + ], + "score": 1.0, + "content": "10:", + "type": "text" + }, + { + "bbox": [ + 136, + 198, + 219, + 213 + ], + "score": 1.0, + "content": "# start hill climbing", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 209, + 244, + 225 + ], + "spans": [ + { + "bbox": [ + 106, + 212, + 122, + 223 + ], + "score": 1.0, + "content": "11:", + "type": "text" + }, + { + "bbox": [ + 137, + 209, + 155, + 225 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 156, + 212, + 234, + 223 + ], + "score": 0.83, + "content": "i \\gets 1 , \\dots , n _ { s t e p s } \\ { \\bf d }", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 209, + 244, + 225 + ], + "score": 1.0, + "content": "o", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 221, + 496, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 222, + 122, + 234 + ], + "score": 1.0, + "content": "12:", + "type": "text" + }, + { + "bbox": [ + 151, + 221, + 173, + 236 + ], + "score": 1.0, + "content": "#get", + "type": "text" + }, + { + "bbox": [ + 173, + 224, + 201, + 234 + ], + "score": 0.7, + "content": "n _ { n e i g h }", + "type": "inline_equation" + }, + { + "bbox": [ + 201, + 221, + 255, + 236 + ], + "score": 1.0, + "content": "neighbors of", + "type": "text" + }, + { + "bbox": [ + 255, + 223, + 286, + 234 + ], + "score": 0.68, + "content": "m o d e l _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 221, + 338, + 236 + ], + "score": 1.0, + "content": "by applying", + "type": "text" + }, + { + "bbox": [ + 338, + 225, + 361, + 233 + ], + "score": 0.85, + "content": "n _ { N M }", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 221, + 454, + 236 + ], + "score": 1.0, + "content": "network morphisms to", + "type": "text" + }, + { + "bbox": [ + 455, + 222, + 496, + 234 + ], + "score": 0.5, + "content": "m o d e l _ { b e s t }", + "type": "inline_equation" + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 232, + 281, + 247 + ], + "spans": [ + { + "bbox": [ + 106, + 233, + 122, + 245 + ], + "score": 1.0, + "content": "13:", + "type": "text" + }, + { + "bbox": [ + 152, + 232, + 170, + 247 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 171, + 234, + 263, + 245 + ], + "score": 0.84, + "content": "j \\gets 1 , \\dots , n _ { n e i g h } - 1", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 232, + 281, + 247 + ], + "score": 1.0, + "content": "do", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 245, + 370, + 257 + ], + "spans": [ + { + "bbox": [ + 106, + 245, + 121, + 255 + ], + "score": 1.0, + "content": "14:", + "type": "text" + }, + { + "bbox": [ + 167, + 245, + 370, + 257 + ], + "score": 0.65, + "content": "m o d e l _ { j } \\gets A p p l y N e t M o r p h s ( m o d e l _ { b e s t } , n _ { N M } )", + "type": "inline_equation" + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 255, + 374, + 267 + ], + "spans": [ + { + "bbox": [ + 106, + 255, + 122, + 267 + ], + "score": 1.0, + "content": "15:", + "type": "text" + }, + { + "bbox": [ + 166, + 255, + 374, + 267 + ], + "score": 1.0, + "content": "# train for a few epochs on training set with SGDR", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 266, + 402, + 279 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 122, + 278 + ], + "score": 1.0, + "content": "16:", + "type": "text" + }, + { + "bbox": [ + 168, + 266, + 402, + 279 + ], + "score": 0.7, + "content": "m o d e l _ { j } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { j } , e \\bar { p } o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )", + "type": "inline_equation", + "image_path": "328ab971c6bec51b546f24a106ddfe7a221f18a42811b9589b2c22b48b03bf8c.jpg" + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 277, + 187, + 289 + ], + "spans": [ + { + "bbox": [ + 106, + 277, + 122, + 289 + ], + "score": 1.0, + "content": "17:", + "type": "text" + }, + { + "bbox": [ + 153, + 277, + 187, + 288 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 288, + 366, + 300 + ], + "spans": [ + { + "bbox": [ + 106, + 288, + 122, + 300 + ], + "score": 1.0, + "content": "18:", + "type": "text" + }, + { + "bbox": [ + 151, + 288, + 366, + 300 + ], + "score": 1.0, + "content": "# in fact, last neighbor is always just the current best", + "type": "text" + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 299, + 417, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 299, + 122, + 311 + ], + "score": 1.0, + "content": "19:", + "type": "text" + }, + { + "bbox": [ + 153, + 299, + 417, + 312 + ], + "score": 0.62, + "content": "m o d e l _ { n _ { n e i g h } } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { b e s t } , e p o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )", + "type": "inline_equation" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 309, + 290, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 122, + 322 + ], + "score": 1.0, + "content": "20:", + "type": "text" + }, + { + "bbox": [ + 152, + 309, + 290, + 322 + ], + "score": 1.0, + "content": "# get best model on validation set", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 320, + 383, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 321, + 122, + 334 + ], + "score": 1.0, + "content": "21:", + "type": "text" + }, + { + "bbox": [ + 153, + 320, + 383, + 341 + ], + "score": 0.6, + "content": "\\underbrace { \\Bre { \\_ } { a } } { m o d } e l _ { b e s t } \\underbrace { a r g m a x } _ { j = 1 , \\dots , n _ { n e i g h } } \\{ p e r f o r m a n c e _ { v a l i } ( m o d e l _ { j } ) \\}", + "type": "inline_equation", + "image_path": "305597548bb02bd55928f77cef70b478c0f4c25ccd76325015892c958fc657d0.jpg" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 339, + 172, + 352 + ], + "spans": [ + { + "bbox": [ + 106, + 339, + 121, + 352 + ], + "score": 1.0, + "content": "22:", + "type": "text" + }, + { + "bbox": [ + 137, + 339, + 172, + 351 + ], + "score": 1.0, + "content": "end for", + "type": "text" + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 350, + 348, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 122, + 363 + ], + "score": 1.0, + "content": "23:", + "type": "text" + }, + { + "bbox": [ + 137, + 350, + 348, + 363 + ], + "score": 1.0, + "content": "# train the final model on training and validation set", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 362, + 391, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 121, + 374 + ], + "score": 1.0, + "content": "24:", + "type": "text" + }, + { + "bbox": [ + 138, + 362, + 391, + 374 + ], + "score": 0.41, + "content": "m o d e l _ { b e s t } \\gets \\mathrm { S G D R t r a i n } ( m o d e \\bar { l } _ { b e s t } , e p o c h _ { f i n a l } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )", + "type": "inline_equation" + } + ], + "index": 24, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 371, + 214, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 372, + 121, + 385 + ], + "score": 1.0, + "content": "25:", + "type": "text" + }, + { + "bbox": [ + 137, + 371, + 214, + 387 + ], + "score": 1.0, + "content": "return modelbest", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 106, + 383, + 180, + 396 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 121, + 394 + ], + "score": 1.0, + "content": "26:", + "type": "text" + }, + { + "bbox": [ + 122, + 383, + 180, + 396 + ], + "score": 1.0, + "content": "end function", + "type": "text" + } + ], + "index": 26, + "is_list_start_line": true + } + ], + "index": 13.5, + "bbox_fs": [ + 106, + 96, + 496, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 421, + 505, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 421, + 505, + 432 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 505, + 432 + ], + "score": 1.0, + "content": "It is important for our method that child networks only need to be trained for a few epochs2 (line", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 432, + 505, + 444 + ], + "score": 1.0, + "content": "17). Hence, an optimization algorithm with good anytime performance is required. Therefore, we", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 442, + 505, + 455 + ], + "score": 1.0, + "content": "chose the cosine annealing strategy from Loshchilov & Hutter (2017), whereas the learning rate is", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 453, + 505, + 466 + ], + "spans": [ + { + "bbox": [ + 105, + 453, + 405, + 466 + ], + "score": 1.0, + "content": "implicitly restarted: the training in line 17 always starts with a learning rate", + "type": "text" + }, + { + "bbox": [ + 406, + 454, + 431, + 465 + ], + "score": 0.92, + "content": "\\lambda _ { s t a r t }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 453, + 505, + 466 + ], + "score": 1.0, + "content": "which is annealed", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 464, + 506, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 116, + 478 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 117, + 465, + 137, + 476 + ], + "score": 0.91, + "content": "\\lambda _ { e n d }", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 464, + 506, + 478 + ], + "score": 1.0, + "content": "after epochneigh epochs. We use the same learning rate scheduler in the final training (aside", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 475, + 252, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 252, + 488 + ], + "score": 1.0, + "content": "from a different number of epochs).", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 421, + 506, + 488 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 492, + 505, + 547 + ], + "lines": [ + { + "bbox": [ + 106, + 492, + 505, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 492, + 505, + 505 + ], + "score": 1.0, + "content": "While we presented our method as a simple hill-climbing method, we note that it can also be inter-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 502, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 396, + 518 + ], + "score": 1.0, + "content": "preted as a very simple evolutionary algorithm with a population size of", + "type": "text" + }, + { + "bbox": [ + 397, + 505, + 424, + 516 + ], + "score": 0.88, + "content": "n _ { n e i g h }", + "type": "inline_equation" + }, + { + "bbox": [ + 425, + 502, + 505, + 518 + ], + "score": 1.0, + "content": ", no cross-over, net-", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 513, + 505, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 528 + ], + "score": 1.0, + "content": "work morphisms as mutations, and a selection mechanism that only considers the best-performing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 525, + 506, + 538 + ], + "score": 1.0, + "content": "population member as the parent for the next generation. This interpretation also suggests several", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 537, + 335, + 549 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 335, + 549 + ], + "score": 1.0, + "content": "promising possibilities for extending our simple method.", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 492, + 506, + 549 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 567, + 200, + 579 + ], + "lines": [ + { + "bbox": [ + 105, + 565, + 201, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 565, + 201, + 581 + ], + "score": 1.0, + "content": "5 EXPERIMENTS", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 107, + 593, + 505, + 637 + ], + "lines": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 505, + 605 + ], + "score": 1.0, + "content": "We evaluate our method on CIFAR-10 and CIFAR-100. First, we investigate whether our consider-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 106, + 605, + 505, + 617 + ], + "score": 1.0, + "content": "ations from the previous chapter coincide with empirical results. We also check if the interplay of", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "modifying and training networks harms their eventual performance. Finally, we compare our pro-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 627, + 496, + 638 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 496, + 638 + ], + "score": 1.0, + "content": "posed method with other automated architecture algorithms as well as hand crafted architectures.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 594, + 505, + 638 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 643, + 505, + 687 + ], + "lines": [ + { + "bbox": [ + 106, + 643, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 643, + 505, + 655 + ], + "score": 1.0, + "content": "We use the same standard data augmentation scheme for both CIFAR datasets used by Loshchilov", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "& Hutter (2017) in all of the following experiments. The training set (50.000 samples) is split up in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 666, + 505, + 677 + ], + "score": 1.0, + "content": "training (40.000) and validation (10.000) set for the purpose of architecture search. Eventually the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 688 + ], + "score": 1.0, + "content": "performance is evaluated on the test set. All experiments where run on Nvidia Titan X (Maxwell)", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 643, + 505, + 688 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "GPUs, with code implemented in Keras (Chollet et al., 2015) with a TensorFlow (Abadi et al., 2015)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 145, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 145, + 106 + ], + "score": 1.0, + "content": "backend.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "title", + "bbox": [ + 108, + 121, + 257, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 258, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 258, + 134 + ], + "score": 1.0, + "content": "5.1 EXPERIMENTS ON CIFAR-10", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 142, + 189, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 190, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 190, + 156 + ], + "score": 1.0, + "content": "5.1.1 BASELINES", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 164, + 503, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "Before comparing our method to others, we run some baseline experiments to see whether our", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 175, + 389, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 389, + 187 + ], + "score": 1.0, + "content": "considerations from the previous chapter coincide with empirical data.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "Random model selection. First, we investigate if the simple hill climbing strategy is able to dis-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 411, + 217 + ], + "score": 1.0, + "content": "tinguish between models with high and low performance. For this, we set", + "type": "text" + }, + { + "bbox": [ + 411, + 203, + 460, + 215 + ], + "score": 0.9, + "content": "n _ { n e i g h } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 201, + 505, + 217 + ], + "score": 1.0, + "content": ", i.e., there", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "score": 1.0, + "content": "is no model selction - we simply construct random networks and train them. We then run ex-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 224, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 104, + 224, + 170, + 238 + ], + "score": 1.0, + "content": "periments with", + "type": "text" + }, + { + "bbox": [ + 171, + 225, + 221, + 236 + ], + "score": 0.9, + "content": "n _ { n e i g h } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 224, + 506, + 238 + ], + "score": 1.0, + "content": "and compare both results. All other parameters are the same in this", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 234, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 190, + 249 + ], + "score": 1.0, + "content": "experiment, namely", + "type": "text" + }, + { + "bbox": [ + 191, + 236, + 450, + 248 + ], + "score": 0.48, + "content": "n _ { s t e p s } = 5 , n _ { N M } = 5 , e p o c h _ { n e i g h } = 1 7 , e p o c h _ { f i n a l } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 234, + 506, + 249 + ], + "score": 1.0, + "content": ". We choose", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 164, + 258 + ], + "score": 0.79, + "content": "\\lambda _ { s t a r t } = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 246, + 167, + 259 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 167, + 247, + 214, + 258 + ], + "score": 0.82, + "content": "\\lambda _ { e n d } = 0 . { \\overset { \\cdot } { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "as done in Loshchilov & Hutter (2017). model0 was a simple conv net:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 255, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 271 + ], + "score": 1.0, + "content": "Conv-MaxPool-Conv-MaxPool-Conv-FC-Softmax3, which is pretrained for 20 epochs, achieving", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 267, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 137, + 279 + ], + "score": 0.89, + "content": "\\approx 7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 267, + 243, + 281 + ], + "score": 1.0, + "content": "validation accuracy (up to", + "type": "text" + }, + { + "bbox": [ + 243, + 268, + 263, + 279 + ], + "score": 0.89, + "content": "9 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 267, + 505, + 281 + ], + "score": 1.0, + "content": "when trained till convergence), see Figure 5 in the appendix.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "score": 1.0, + "content": "If our algorithm is able to identify better networks, one would expect to get better results with the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 290, + 187, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 135, + 304 + ], + "score": 1.0, + "content": "setting", + "type": "text" + }, + { + "bbox": [ + 136, + 291, + 182, + 303 + ], + "score": 0.9, + "content": "n _ { n e i g h } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 290, + 187, + 304 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "Retraining from scratch. In the this experiment we investigate whether the ”weight inheritance”", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "due to the network morphisms used in our algorithm harms the final performance of the final model.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "This weight inheritance can be seen as a strong prior on the weights and one could suspect that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "the new, larger model may not be able to overcome a possibly poor prior. Additionally we were", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "interested in measuring the overhead of the architecture search process, so we compared the times", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "for generating and training a model with the time needed when training the final model from scratch.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "The retraining from scratch is done for the same number of epochs as the total number of epochs", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 384, + 316, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 316, + 396 + ], + "score": 1.0, + "content": "spent to train the model returned by our algorithm4 .", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 414 + ], + "score": 1.0, + "content": "No SGDR. We now turn off the cosine annealing with restarts (SGDR) during the hill climbing", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "stage, i.e., the training in line 17 of Algorithm 1 is done with a constant learning rate. We tried", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 205, + 435 + ], + "score": 0.87, + "content": "\\lambda \\in \\{ 0 . 0 1 , 0 . 0 2 5 , 0 . 0 5 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 422, + 506, + 435 + ], + "score": 1.0, + "content": ", 10 runs each and averaged the results. Note that we still use the cosine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 433, + 217, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 217, + 447 + ], + "score": 1.0, + "content": "decay for the final training.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 108, + 451, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "No network morphism. Lastly, we turn off the network morphism constraint for initializing the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "neighbor networks. In detail, we proceeded as Real et al. (2017): All weights from layer where now", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 472, + 503, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 503, + 486 + ], + "score": 1.0, + "content": "changes occur are inherited, whereas the weights of new/modified layers are initialized by random.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "The results for these experiments are summarized in Table 1 . The hill climbing strategy is actually", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 369, + 512 + ], + "score": 1.0, + "content": "able to identify better performing models. (first and second line:", + "type": "text" + }, + { + "bbox": [ + 370, + 501, + 392, + 512 + ], + "score": 0.88, + "content": "5 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 501, + 409, + 512 + ], + "score": 1.0, + "content": "vs.", + "type": "text" + }, + { + "bbox": [ + 409, + 501, + 432, + 512 + ], + "score": 0.84, + "content": "6 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 501, + 505, + 512 + ], + "score": 1.0, + "content": "). Notice how hill", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 510, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 525 + ], + "score": 1.0, + "content": "climbing prefers larger models (5.7 million parameters on average vs. 4.4 million). Performance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "slightly decreases when the models are retrained from scratch (line 3). This experiments indicates", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "that our algorithm does not harm the final performance of a model. Regarding the runtime, the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 556 + ], + "score": 1.0, + "content": "overhead for first having to search for the architecture is roughly a factor 3. We think this is a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "big advantage of our method and shows that architecture search can be done in the same order of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "magnitude as training a single model. In line 4 we can see that SGDR plays an important role. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "resulting models chosen by our algorithm when training is done with a constant learning rate perform", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 353, + 601 + ], + "score": 1.0, + "content": "similarly to the models without any model selection strategy", + "type": "text" + }, + { + "bbox": [ + 353, + 588, + 377, + 599 + ], + "score": 0.86, + "content": "( 6 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 588, + 395, + 601 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 396, + 588, + 418, + 599 + ], + "score": 0.87, + "content": "6 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 588, + 505, + 601 + ], + "score": 1.0, + "content": ", respectively), which", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "indicates that the performance after a few epochs on the validation set when trained without SGDR", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "correlates less with the final performance on the test set as it is the case for training with SGDR.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 619, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 389, + 635 + ], + "score": 1.0, + "content": "Indeed, we computed the Pearson correlation coefficient and obtained", + "type": "text" + }, + { + "bbox": [ + 390, + 621, + 435, + 632 + ], + "score": 0.91, + "content": "R ^ { 2 } = 0 . 6 4", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 619, + 506, + 635 + ], + "score": 1.0, + "content": "for training with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 171, + 646 + ], + "score": 1.0, + "content": "SGDR and and", + "type": "text" + }, + { + "bbox": [ + 171, + 632, + 218, + 643 + ], + "score": 0.9, + "content": "\\mathrm { \\bar { \\it R } ^ { 2 } = 0 . 3 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 631, + 506, + 646 + ], + "score": 1.0, + "content": "for training with a constant learning rate. See appendix A. Also, with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "the constant learning rate, the few epochs spent are not sufficient to improve the performance of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "model. Figure 2 shows the progress while running our algorithm with and without SGDR, averaged", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "over all runs. When turning off the network morphism constraint, performance also decreases.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 39 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 690, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 688, + 310, + 703 + ], + "spans": [ + { + "bbox": [ + 118, + 688, + 310, + 703 + ], + "score": 1.0, + "content": "3By Conv we actually mean Conv-BatchNorm-Relu.", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 696, + 507, + 717 + ], + "spans": [ + { + "bbox": [ + 117, + 696, + 181, + 717 + ], + "score": 1.0, + "content": "4With the setting", + "type": "text" + }, + { + "bbox": [ + 182, + 702, + 290, + 712 + ], + "score": 0.76, + "content": "n _ { s t e p s } = 5 , e p o c h _ { n e i g h } = 1 7", + "type": "inline_equation" + }, + { + "bbox": [ + 291, + 696, + 294, + 717 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 295, + 702, + 362, + 712 + ], + "score": 0.54, + "content": "e p o c h _ { f i n a l } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 362, + 696, + 507, + 717 + ], + "score": 1.0, + "content": "and pretraining the starting network for", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 709, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 433, + 724 + ], + "score": 1.0, + "content": "20 epochs, the models that are returned by our algorithm are trained for a total number of of", + "type": "text" + }, + { + "bbox": [ + 434, + 712, + 505, + 721 + ], + "score": 0.89, + "content": "2 0 + 1 7 \\cdot 5 + 1 0 0 =", + "type": "inline_equation" + } + ] + }, + { + "bbox": [ + 105, + 719, + 153, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 153, + 734 + ], + "score": 1.0, + "content": "205 epochs.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 506, + 95 + ], + "score": 1.0, + "content": "GPUs, with code implemented in Keras (Chollet et al., 2015) with a TensorFlow (Abadi et al., 2015)", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 145, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 145, + 106 + ], + "score": 1.0, + "content": "backend.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 106 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 121, + 257, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 121, + 258, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 258, + 134 + ], + "score": 1.0, + "content": "5.1 EXPERIMENTS ON CIFAR-10", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 2 + }, + { + "type": "title", + "bbox": [ + 107, + 142, + 189, + 155 + ], + "lines": [ + { + "bbox": [ + 105, + 142, + 190, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 190, + 156 + ], + "score": 1.0, + "content": "5.1.1 BASELINES", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 164, + 503, + 186 + ], + "lines": [ + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 163, + 505, + 176 + ], + "score": 1.0, + "content": "Before comparing our method to others, we run some baseline experiments to see whether our", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 175, + 389, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 175, + 389, + 187 + ], + "score": 1.0, + "content": "considerations from the previous chapter coincide with empirical data.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 163, + 505, + 187 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 191, + 505, + 302 + ], + "lines": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 205 + ], + "score": 1.0, + "content": "Random model selection. First, we investigate if the simple hill climbing strategy is able to dis-", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 201, + 505, + 217 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 411, + 217 + ], + "score": 1.0, + "content": "tinguish between models with high and low performance. For this, we set", + "type": "text" + }, + { + "bbox": [ + 411, + 203, + 460, + 215 + ], + "score": 0.9, + "content": "n _ { n e i g h } = 1", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 201, + 505, + 217 + ], + "score": 1.0, + "content": ", i.e., there", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 506, + 227 + ], + "score": 1.0, + "content": "is no model selction - we simply construct random networks and train them. We then run ex-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 224, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 104, + 224, + 170, + 238 + ], + "score": 1.0, + "content": "periments with", + "type": "text" + }, + { + "bbox": [ + 171, + 225, + 221, + 236 + ], + "score": 0.9, + "content": "n _ { n e i g h } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 224, + 506, + 238 + ], + "score": 1.0, + "content": "and compare both results. All other parameters are the same in this", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 234, + 506, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 190, + 249 + ], + "score": 1.0, + "content": "experiment, namely", + "type": "text" + }, + { + "bbox": [ + 191, + 236, + 450, + 248 + ], + "score": 0.48, + "content": "n _ { s t e p s } = 5 , n _ { N M } = 5 , e p o c h _ { n e i g h } = 1 7 , e p o c h _ { f i n a l } = 1 0 0", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 234, + 506, + 249 + ], + "score": 1.0, + "content": ". We choose", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 246, + 505, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 164, + 258 + ], + "score": 0.79, + "content": "\\lambda _ { s t a r t } = 0 . 0 5", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 246, + 167, + 259 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 167, + 247, + 214, + 258 + ], + "score": 0.82, + "content": "\\lambda _ { e n d } = 0 . { \\overset { \\cdot } { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 246, + 505, + 259 + ], + "score": 1.0, + "content": "as done in Loshchilov & Hutter (2017). model0 was a simple conv net:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 255, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 105, + 255, + 506, + 271 + ], + "score": 1.0, + "content": "Conv-MaxPool-Conv-MaxPool-Conv-FC-Softmax3, which is pretrained for 20 epochs, achieving", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 267, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 137, + 279 + ], + "score": 0.89, + "content": "\\approx 7 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 137, + 267, + 243, + 281 + ], + "score": 1.0, + "content": "validation accuracy (up to", + "type": "text" + }, + { + "bbox": [ + 243, + 268, + 263, + 279 + ], + "score": 0.89, + "content": "9 1 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 267, + 505, + 281 + ], + "score": 1.0, + "content": "when trained till convergence), see Figure 5 in the appendix.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "spans": [ + { + "bbox": [ + 105, + 279, + 505, + 292 + ], + "score": 1.0, + "content": "If our algorithm is able to identify better networks, one would expect to get better results with the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 290, + 187, + 304 + ], + "spans": [ + { + "bbox": [ + 104, + 290, + 135, + 304 + ], + "score": 1.0, + "content": "setting", + "type": "text" + }, + { + "bbox": [ + 136, + 291, + 182, + 303 + ], + "score": 0.9, + "content": "n _ { n e i g h } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 290, + 187, + 304 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 10.5, + "bbox_fs": [ + 104, + 191, + 506, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 307, + 505, + 396 + ], + "lines": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "spans": [ + { + "bbox": [ + 105, + 306, + 505, + 320 + ], + "score": 1.0, + "content": "Retraining from scratch. In the this experiment we investigate whether the ”weight inheritance”", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 505, + 330 + ], + "score": 1.0, + "content": "due to the network morphisms used in our algorithm harms the final performance of the final model.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "spans": [ + { + "bbox": [ + 105, + 329, + 505, + 342 + ], + "score": 1.0, + "content": "This weight inheritance can be seen as a strong prior on the weights and one could suspect that", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 340, + 506, + 353 + ], + "spans": [ + { + "bbox": [ + 106, + 340, + 506, + 353 + ], + "score": 1.0, + "content": "the new, larger model may not be able to overcome a possibly poor prior. Additionally we were", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 505, + 364 + ], + "score": 1.0, + "content": "interested in measuring the overhead of the architecture search process, so we compared the times", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "for generating and training a model with the time needed when training the final model from scratch.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 505, + 385 + ], + "score": 1.0, + "content": "The retraining from scratch is done for the same number of epochs as the total number of epochs", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 384, + 316, + 396 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 316, + 396 + ], + "score": 1.0, + "content": "spent to train the model returned by our algorithm4 .", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 306, + 506, + 396 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 401, + 505, + 445 + ], + "lines": [ + { + "bbox": [ + 105, + 400, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 400, + 506, + 414 + ], + "score": 1.0, + "content": "No SGDR. We now turn off the cosine annealing with restarts (SGDR) during the hill climbing", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "stage, i.e., the training in line 17 of Algorithm 1 is done with a constant learning rate. We tried", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 422, + 506, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 205, + 435 + ], + "score": 0.87, + "content": "\\lambda \\in \\{ 0 . 0 1 , 0 . 0 2 5 , 0 . 0 5 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 422, + 506, + 435 + ], + "score": 1.0, + "content": ", 10 runs each and averaged the results. Note that we still use the cosine", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 433, + 217, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 433, + 217, + 447 + ], + "score": 1.0, + "content": "decay for the final training.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 400, + 506, + 447 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 451, + 505, + 484 + ], + "lines": [ + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "spans": [ + { + "bbox": [ + 106, + 451, + 505, + 463 + ], + "score": 1.0, + "content": "No network morphism. Lastly, we turn off the network morphism constraint for initializing the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 505, + 474 + ], + "score": 1.0, + "content": "neighbor networks. In detail, we proceeded as Real et al. (2017): All weights from layer where now", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 472, + 503, + 486 + ], + "spans": [ + { + "bbox": [ + 105, + 472, + 503, + 486 + ], + "score": 1.0, + "content": "changes occur are inherited, whereas the weights of new/modified layers are initialized by random.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 451, + 505, + 486 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 489, + 505, + 676 + ], + "lines": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 489, + 505, + 502 + ], + "score": 1.0, + "content": "The results for these experiments are summarized in Table 1 . The hill climbing strategy is actually", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 501, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 369, + 512 + ], + "score": 1.0, + "content": "able to identify better performing models. (first and second line:", + "type": "text" + }, + { + "bbox": [ + 370, + 501, + 392, + 512 + ], + "score": 0.88, + "content": "5 . 7 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 501, + 409, + 512 + ], + "score": 1.0, + "content": "vs.", + "type": "text" + }, + { + "bbox": [ + 409, + 501, + 432, + 512 + ], + "score": 0.84, + "content": "6 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 501, + 505, + 512 + ], + "score": 1.0, + "content": "). Notice how hill", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 510, + 506, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 506, + 525 + ], + "score": 1.0, + "content": "climbing prefers larger models (5.7 million parameters on average vs. 4.4 million). Performance", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 505, + 535 + ], + "score": 1.0, + "content": "slightly decreases when the models are retrained from scratch (line 3). This experiments indicates", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 533, + 506, + 546 + ], + "score": 1.0, + "content": "that our algorithm does not harm the final performance of a model. Regarding the runtime, the", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 545, + 506, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 556 + ], + "score": 1.0, + "content": "overhead for first having to search for the architecture is roughly a factor 3. We think this is a", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 556, + 506, + 568 + ], + "spans": [ + { + "bbox": [ + 106, + 556, + 506, + 568 + ], + "score": 1.0, + "content": "big advantage of our method and shows that architecture search can be done in the same order of", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 579 + ], + "score": 1.0, + "content": "magnitude as training a single model. In line 4 we can see that SGDR plays an important role. The", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "resulting models chosen by our algorithm when training is done with a constant learning rate perform", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 105, + 588, + 353, + 601 + ], + "score": 1.0, + "content": "similarly to the models without any model selection strategy", + "type": "text" + }, + { + "bbox": [ + 353, + 588, + 377, + 599 + ], + "score": 0.86, + "content": "( 6 . 4 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 588, + 395, + 601 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 396, + 588, + 418, + 599 + ], + "score": 0.87, + "content": "6 . 5 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 418, + 588, + 505, + 601 + ], + "score": 1.0, + "content": ", respectively), which", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 505, + 612 + ], + "score": 1.0, + "content": "indicates that the performance after a few epochs on the validation set when trained without SGDR", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 610, + 505, + 623 + ], + "score": 1.0, + "content": "correlates less with the final performance on the test set as it is the case for training with SGDR.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 619, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 389, + 635 + ], + "score": 1.0, + "content": "Indeed, we computed the Pearson correlation coefficient and obtained", + "type": "text" + }, + { + "bbox": [ + 390, + 621, + 435, + 632 + ], + "score": 0.91, + "content": "R ^ { 2 } = 0 . 6 4", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 619, + 506, + 635 + ], + "score": 1.0, + "content": "for training with", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 171, + 646 + ], + "score": 1.0, + "content": "SGDR and and", + "type": "text" + }, + { + "bbox": [ + 171, + 632, + 218, + 643 + ], + "score": 0.9, + "content": "\\mathrm { \\bar { \\it R } ^ { 2 } = 0 . 3 6 }", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 631, + 506, + 646 + ], + "score": 1.0, + "content": "for training with a constant learning rate. See appendix A. Also, with", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "spans": [ + { + "bbox": [ + 106, + 644, + 505, + 655 + ], + "score": 1.0, + "content": "the constant learning rate, the few epochs spent are not sufficient to improve the performance of the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 668 + ], + "score": 1.0, + "content": "model. Figure 2 shows the progress while running our algorithm with and without SGDR, averaged", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 505, + 678 + ], + "score": 1.0, + "content": "over all runs. When turning off the network morphism constraint, performance also decreases.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 39, + "bbox_fs": [ + 104, + 489, + 506, + 678 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 78, + 464, + 243 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 143, + 78, + 464, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 78, + 464, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 78, + 464, + 243 + ], + "score": 0.971, + "type": "image", + "image_path": "5bffa50e24c866e7a85ea7102470da8aa121189925b47b5e0807ef32b1fb1234.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 143, + 78, + 464, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 143, + 133.0, + 464, + 188.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 143, + 188.0, + 464, + 243.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 254, + 505, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "score": 1.0, + "content": "Figure 2: The best model found by Algorithm 1 tracked over time (in terms of hill climbing itera-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "tions). With (red) and without (blue) using SGDR for the training within the hill climbing (line 17).", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 276, + 478, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 415, + 288 + ], + "score": 1.0, + "content": "Final training (line 24) is not plotted. Dashed line denotes mean, shaded area", + "type": "text" + }, + { + "bbox": [ + 415, + 276, + 436, + 286 + ], + "score": 0.88, + "content": "\\pm 2 \\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 276, + 478, + 288 + ], + "score": 1.0, + "content": "intervalls.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 106, + 345, + 514, + 416 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 311, + 505, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Table 1: Baseline experiments. Runtime, # params, and error rates are averaged over 10 runs (for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 322, + 458, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 323, + 154, + 335 + ], + "score": 0.89, + "content": "n _ { n e i g h } = 8 ", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 322, + 208, + 335 + ], + "score": 1.0, + "content": ") and 30 runs", + "type": "text" + }, + { + "bbox": [ + 208, + 323, + 257, + 335 + ], + "score": 0.87, + "content": "( n _ { n e i g h } = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 322, + 335, + 335 + ], + "score": 1.0, + "content": ") runs, respectively.", + "type": "text" + }, + { + "bbox": [ + 335, + 323, + 380, + 335 + ], + "score": 0.91, + "content": "n _ { s t e p s } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 322, + 458, + 335 + ], + "score": 1.0, + "content": "in all experiments.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 345, + 514, + 416 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 345, + 514, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 514, + 416 + ], + "score": 0.969, + "html": "
algorithm settingruntime (hrs)# params (mil.)error ± std.( %)
nneigh =812.85.75.7 ± 0.35
Random networks (nneigh = 1)4.54.46.5 ± 0.76
models from line 1 retrained from scratch5.35.76.1 ± 0.92
nneigh = 8, no SGDR10.65.86.4±0.70
nneigh = 8, no net. morph.6.62.96.1 ±0.30
", + "type": "table", + "image_path": "399c4ad8f3e36610cb71bf824ac3a3e315a7c2b733f1dcf09b58c5f4d7dfd8a5.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 106, + 345, + 514, + 368.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 368.6666666666667, + 514, + 392.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 392.33333333333337, + 514, + 416.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.75 + }, + { + "type": "text", + "bbox": [ + 106, + 445, + 506, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "score": 1.0, + "content": "Interestingly the number of parameters heavily decreases. This indicates that our algorithm prefers", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 457, + 238, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 238, + 468 + ], + "score": 1.0, + "content": "models without new parameters.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5 + }, + { + "type": "title", + "bbox": [ + 108, + 486, + 463, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 464, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 464, + 497 + ], + "score": 1.0, + "content": "5.1.2 COMPARISON TO HAND CRAFTED AND OTHER AUTOMATICALLY GENERATED", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 498, + 212, + 508 + ], + "spans": [ + { + "bbox": [ + 139, + 498, + 212, + 508 + ], + "score": 1.0, + "content": "ARCHITECTURES", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "score": 1.0, + "content": "We now compare our algorithm against the popular wide residual networks (Zagoruyko & Ko-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "modakis, 2016), the state of the art model from Gastaldi (2017) as well as other automated archi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 295, + 553 + ], + "score": 1.0, + "content": "tecture search methods. Beside our results for", + "type": "text" + }, + { + "bbox": [ + 295, + 540, + 341, + 551 + ], + "score": 0.91, + "content": "n _ { s t e p s } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "from the previous section, we also tried", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 551, + 258, + 562 + ], + "spans": [ + { + "bbox": [ + 107, + 551, + 151, + 562 + ], + "score": 0.9, + "content": "n _ { s t e p s } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 551, + 258, + 562 + ], + "score": 1.0, + "content": "to generate larger models.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 107, + 562, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "For further improving the results, we take snapshots of the best models from every iteration while", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "running our algorithm following the idea of Huang et al. (2017) when using SGDR (Loshchilov &", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "Hutter, 2017) for training. However different from Huang et al. (2017), we do not immediately get", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "score": 1.0, + "content": "fully trained models for free, as our snapshots are not yet trained on the validation set but rather only", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "on the training set. Hence we spent some additional resources and train the snapshots on both sets.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Afterwards the ensemble model is build by combining the snapshot models with uniform weights.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "Lastly, we also build an ensemble from the models returned by our algorithm across all runs. Results", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 638, + 190, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 190, + 650 + ], + "score": 1.0, + "content": "are listed in Table 2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 710 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "The proposed method is able to generate competitive network architectures in only 12 hours. By", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "spending another 12 hours, it outperforms most automated architecture search methods although all", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "of them require (partially far) more time and GPUs. We do not reach the performance of the two", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "handcrafted architectures as well as the ones found by Zoph & Le (2017) and Brock et al. (2017).", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 421, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 421, + 711 + ], + "score": 1.0, + "content": "However note that Zoph & Le (2017) spent by far more resources than we did.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 710, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Unsurprisingly, the ensemble models perform better. It is a simple and cheap way to improve results", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 721, + 413, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 413, + 732 + ], + "score": 1.0, + "content": "which everyone can consider when the number of parameters is not relevant.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 143, + 78, + 464, + 243 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 143, + 78, + 464, + 243 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 143, + 78, + 464, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 78, + 464, + 243 + ], + "score": 0.971, + "type": "image", + "image_path": "5bffa50e24c866e7a85ea7102470da8aa121189925b47b5e0807ef32b1fb1234.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 143, + 78, + 464, + 133.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 143, + 133.0, + 464, + 188.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 143, + 188.0, + 464, + 243.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 254, + 505, + 288 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "spans": [ + { + "bbox": [ + 105, + 253, + 505, + 267 + ], + "score": 1.0, + "content": "Figure 2: The best model found by Algorithm 1 tracked over time (in terms of hill climbing itera-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "spans": [ + { + "bbox": [ + 106, + 265, + 505, + 277 + ], + "score": 1.0, + "content": "tions). With (red) and without (blue) using SGDR for the training within the hill climbing (line 17).", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 276, + 478, + 288 + ], + "spans": [ + { + "bbox": [ + 106, + 276, + 415, + 288 + ], + "score": 1.0, + "content": "Final training (line 24) is not plotted. Dashed line denotes mean, shaded area", + "type": "text" + }, + { + "bbox": [ + 415, + 276, + 436, + 286 + ], + "score": 0.88, + "content": "\\pm 2 \\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 276, + 478, + 288 + ], + "score": 1.0, + "content": "intervalls.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "table", + "bbox": [ + 106, + 345, + 514, + 416 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 105, + 311, + 505, + 334 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "spans": [ + { + "bbox": [ + 106, + 311, + 505, + 324 + ], + "score": 1.0, + "content": "Table 1: Baseline experiments. Runtime, # params, and error rates are averaged over 10 runs (for", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 322, + 458, + 335 + ], + "spans": [ + { + "bbox": [ + 107, + 323, + 154, + 335 + ], + "score": 0.89, + "content": "n _ { n e i g h } = 8 ", + "type": "inline_equation" + }, + { + "bbox": [ + 154, + 322, + 208, + 335 + ], + "score": 1.0, + "content": ") and 30 runs", + "type": "text" + }, + { + "bbox": [ + 208, + 323, + 257, + 335 + ], + "score": 0.87, + "content": "( n _ { n e i g h } = 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 322, + 335, + 335 + ], + "score": 1.0, + "content": ") runs, respectively.", + "type": "text" + }, + { + "bbox": [ + 335, + 323, + 380, + 335 + ], + "score": 0.91, + "content": "n _ { s t e p s } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 322, + 458, + 335 + ], + "score": 1.0, + "content": "in all experiments.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 345, + 514, + 416 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 345, + 514, + 416 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 514, + 416 + ], + "score": 0.969, + "html": "
algorithm settingruntime (hrs)# params (mil.)error ± std.( %)
nneigh =812.85.75.7 ± 0.35
Random networks (nneigh = 1)4.54.46.5 ± 0.76
models from line 1 retrained from scratch5.35.76.1 ± 0.92
nneigh = 8, no SGDR10.65.86.4±0.70
nneigh = 8, no net. morph.6.62.96.1 ±0.30
", + "type": "table", + "image_path": "399c4ad8f3e36610cb71bf824ac3a3e315a7c2b733f1dcf09b58c5f4d7dfd8a5.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 106, + 345, + 514, + 368.6666666666667 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 106, + 368.6666666666667, + 514, + 392.33333333333337 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 106, + 392.33333333333337, + 514, + 416.00000000000006 + ], + "spans": [], + "index": 10 + } + ] + } + ], + "index": 7.75 + }, + { + "type": "text", + "bbox": [ + 106, + 445, + 506, + 468 + ], + "lines": [ + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 505, + 458 + ], + "score": 1.0, + "content": "Interestingly the number of parameters heavily decreases. This indicates that our algorithm prefers", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 457, + 238, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 457, + 238, + 468 + ], + "score": 1.0, + "content": "models without new parameters.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 444, + 505, + 468 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 486, + 463, + 507 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 464, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 464, + 497 + ], + "score": 1.0, + "content": "5.1.2 COMPARISON TO HAND CRAFTED AND OTHER AUTOMATICALLY GENERATED", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 139, + 498, + 212, + 508 + ], + "spans": [ + { + "bbox": [ + 139, + 498, + 212, + 508 + ], + "score": 1.0, + "content": "ARCHITECTURES", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 517, + 504, + 561 + ], + "lines": [ + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 518, + 504, + 529 + ], + "score": 1.0, + "content": "We now compare our algorithm against the popular wide residual networks (Zagoruyko & Ko-", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 540 + ], + "score": 1.0, + "content": "modakis, 2016), the state of the art model from Gastaldi (2017) as well as other automated archi-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 539, + 506, + 553 + ], + "spans": [ + { + "bbox": [ + 105, + 539, + 295, + 553 + ], + "score": 1.0, + "content": "tecture search methods. Beside our results for", + "type": "text" + }, + { + "bbox": [ + 295, + 540, + 341, + 551 + ], + "score": 0.91, + "content": "n _ { s t e p s } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 539, + 506, + 553 + ], + "score": 1.0, + "content": "from the previous section, we also tried", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 551, + 258, + 562 + ], + "spans": [ + { + "bbox": [ + 107, + 551, + 151, + 562 + ], + "score": 0.9, + "content": "n _ { s t e p s } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 551, + 258, + 562 + ], + "score": 1.0, + "content": "to generate larger models.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 518, + 506, + 562 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 562, + 505, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 505, + 574 + ], + "score": 1.0, + "content": "For further improving the results, we take snapshots of the best models from every iteration while", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 584 + ], + "score": 1.0, + "content": "running our algorithm following the idea of Huang et al. (2017) when using SGDR (Loshchilov &", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "spans": [ + { + "bbox": [ + 105, + 582, + 506, + 596 + ], + "score": 1.0, + "content": "Hutter, 2017) for training. However different from Huang et al. (2017), we do not immediately get", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "spans": [ + { + "bbox": [ + 105, + 593, + 505, + 607 + ], + "score": 1.0, + "content": "fully trained models for free, as our snapshots are not yet trained on the validation set but rather only", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "on the training set. Hence we spent some additional resources and train the snapshots on both sets.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 505, + 628 + ], + "score": 1.0, + "content": "Afterwards the ensemble model is build by combining the snapshot models with uniform weights.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 627, + 505, + 639 + ], + "score": 1.0, + "content": "Lastly, we also build an ensemble from the models returned by our algorithm across all runs. Results", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 638, + 190, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 190, + 650 + ], + "score": 1.0, + "content": "are listed in Table 2.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 561, + 506, + 650 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 655, + 505, + 710 + ], + "lines": [ + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 654, + 505, + 667 + ], + "score": 1.0, + "content": "The proposed method is able to generate competitive network architectures in only 12 hours. By", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "spending another 12 hours, it outperforms most automated architecture search methods although all", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "of them require (partially far) more time and GPUs. We do not reach the performance of the two", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 505, + 700 + ], + "score": 1.0, + "content": "handcrafted architectures as well as the ones found by Zoph & Le (2017) and Brock et al. (2017).", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 699, + 421, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 421, + 711 + ], + "score": 1.0, + "content": "However note that Zoph & Le (2017) spent by far more resources than we did.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 654, + 505, + 711 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 710, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 505, + 723 + ], + "score": 1.0, + "content": "Unsurprisingly, the ensemble models perform better. It is a simple and cheap way to improve results", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 107, + 721, + 413, + 732 + ], + "spans": [ + { + "bbox": [ + 107, + 721, + 413, + 732 + ], + "score": 1.0, + "content": "which everyone can consider when the number of parameters is not relevant.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 106, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 123, + 505, + 272 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 112 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "Table 2: Results for CIFAR-10. For our methods the stated resources, # parameters and errors are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "score": 1.0, + "content": "averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 123, + 505, + 272 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 272 + ], + "score": 0.981, + "html": "
modelresources spent# params (mil.)error(%)
Shake-Shake (Gastaldi,2017)4 GPU days,2 GPUs262.9
WRN 28-10 (Loshchilov & Hutter,2017)1 GPU day36.53.86
Baker et al. (2016)80-100 GPU days116.9
Cai et al. (2017)15 GPU days19.75.7
Zoph & Le (2017)16.000-24.000 GPU days37.53.65
Real et al. (2017)2500 GPU days5.45.4
Saxena & Verbeek (2016)？217.4
Brock et al. (2017)3 GPU days16.04.0
Ours (random networks,nsteps = 5, nneigh =1)0.2 GPU days4.46.5
Ours (nsteps = 5,nneigh =8,10 runs)0.5 GPU days5.75.7
Ours (nsteps = 8,nneigh = 8,4 runs)1 GPU day19.75.2
Ours (snapshot ensemble,4 runs)2 GPU days57.84.7
Ours (ensemble across runs)4 GPU days884.4
", + "type": "table", + "image_path": "40ffbbc485084b4e3fbd3f84be67a619447ee95d3fffcf9f34d4d00c8ce93c36.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 106, + 123, + 505, + 172.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 172.66666666666666, + 505, + 222.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 222.33333333333331, + 505, + 272.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 110, + 321, + 501, + 425 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 295, + 504, + 318 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "score": 1.0, + "content": "Table 3: Results for CIFAR-100. For our methods the stated resources, # parameters and errors are", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 307, + 503, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 503, + 319 + ], + "score": 1.0, + "content": "averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 321, + 501, + 425 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 321, + 501, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 321, + 501, + 425 + ], + "score": 0.977, + "html": "
modelresources spent# params (mil.)error(%)
15.9
Shake-Shake (Gastaldi, 2017) WRN 28-10 (Loshchilov & Hutter,2017)14 GPU days34.4
1 GPU day36.519.6
Real et al. (2017)250 GPUs40.423.7
Brock et al. (2017)3 GPU days16.020.6
Ours (nsteps = 8,nneigh = 8,5 runs)1 GPU day22.323.4
Ours (snapshot ensemble,5 runs)2 GPU days73.320.9
Ours (ensemble across runs)5 GPU days111.519.6
", + "type": "table", + "image_path": "4a47d8eceb003bb23f38bfb4ad16df746a77c1f78312b59d11e2951c2db318cd.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 110, + 321, + 501, + 355.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 110, + 355.6666666666667, + 501, + 390.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 390.33333333333337, + 501, + 425.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "title", + "bbox": [ + 108, + 449, + 262, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 263, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 263, + 462 + ], + "score": 1.0, + "content": "5.2 EXPERIMENTS ON CIFAR-100", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "We repeat the previous experiment on CIFAR-100; hyperparameters were not changed. The results", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "score": 1.0, + "content": "are listed in Table 3. Unfortunately most automated architecture methods did not consider CIFAR-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 506, + 505 + ], + "score": 1.0, + "content": "100. Our method is on a par with Real et al. (2017) after one day with a single GPU. The snapshot", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "ensemble performs similar to Brock et al. (2017) and an ensemble model build from the 5 runs can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "compete with the hand crafted WRN 28-10. The performance of the Shake-Shake network (Gastaldi,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 525, + 217, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 217, + 537 + ], + "score": 1.0, + "content": "2017) is again not reached.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5 + }, + { + "type": "title", + "bbox": [ + 108, + 552, + 195, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 197, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 197, + 568 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "We proposed NASH, a simple and fast method for automated architecture search based on a hill", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 587, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 587, + 506, + 602 + ], + "score": 1.0, + "content": "climbing strategy, network morphisms, and training via SGDR. Experiments on CIFAR-10 and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "CIFAR-100 showed that our method yields competitive results while requiring considerably less", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 608, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 608, + 506, + 624 + ], + "score": 1.0, + "content": "computational resources than most alternative approaches. Our algorithm is easily extendable, e.g.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "by other network morphisms, evolutionary approaches for generating new models, other methods", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 506, + 646 + ], + "score": 1.0, + "content": "for cheap performance evaluation (such as, e.g., learning curve prediction (Klein et al., 2017) or", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "hypernetworks (Ha et al., 2017; Brock et al., 2017)), or better resource handling strategies (such as", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 653, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 667 + ], + "score": 1.0, + "content": "Hyperband (Li et al., 2016b)). In this sense, we hope that our approach can serve as a basis for the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 664, + 479, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 479, + 678 + ], + "score": 1.0, + "content": "development of more sophisticated methods that yield further improvements of performance.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 11, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 123, + 505, + 272 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 89, + 504, + 112 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "spans": [ + { + "bbox": [ + 106, + 89, + 505, + 101 + ], + "score": 1.0, + "content": "Table 2: Results for CIFAR-10. For our methods the stated resources, # parameters and errors are", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 504, + 112 + ], + "score": 1.0, + "content": "averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "table_body", + "bbox": [ + 106, + 123, + 505, + 272 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 123, + 505, + 272 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 505, + 272 + ], + "score": 0.981, + "html": "
modelresources spent# params (mil.)error(%)
Shake-Shake (Gastaldi,2017)4 GPU days,2 GPUs262.9
WRN 28-10 (Loshchilov & Hutter,2017)1 GPU day36.53.86
Baker et al. (2016)80-100 GPU days116.9
Cai et al. (2017)15 GPU days19.75.7
Zoph & Le (2017)16.000-24.000 GPU days37.53.65
Real et al. (2017)2500 GPU days5.45.4
Saxena & Verbeek (2016)？217.4
Brock et al. (2017)3 GPU days16.04.0
Ours (random networks,nsteps = 5, nneigh =1)0.2 GPU days4.46.5
Ours (nsteps = 5,nneigh =8,10 runs)0.5 GPU days5.75.7
Ours (nsteps = 8,nneigh = 8,4 runs)1 GPU day19.75.2
Ours (snapshot ensemble,4 runs)2 GPU days57.84.7
Ours (ensemble across runs)4 GPU days884.4
", + "type": "table", + "image_path": "40ffbbc485084b4e3fbd3f84be67a619447ee95d3fffcf9f34d4d00c8ce93c36.jpg" + } + ] + } + ], + "index": 3, + "virtual_lines": [ + { + "bbox": [ + 106, + 123, + 505, + 172.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 106, + 172.66666666666666, + 505, + 222.33333333333331 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 106, + 222.33333333333331, + 505, + 272.0 + ], + "spans": [], + "index": 4 + } + ] + } + ], + "index": 1.75 + }, + { + "type": "table", + "bbox": [ + 110, + 321, + 501, + 425 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 106, + 295, + 504, + 318 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "score": 1.0, + "content": "Table 3: Results for CIFAR-100. For our methods the stated resources, # parameters and errors are", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 307, + 503, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 307, + 503, + 319 + ], + "score": 1.0, + "content": "averaged over all runs. ”Resources spent” denotes training costs in case of the handcrafted models.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "table_body", + "bbox": [ + 110, + 321, + 501, + 425 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 110, + 321, + 501, + 425 + ], + "spans": [ + { + "bbox": [ + 110, + 321, + 501, + 425 + ], + "score": 0.977, + "html": "
modelresources spent# params (mil.)error(%)
15.9
Shake-Shake (Gastaldi, 2017) WRN 28-10 (Loshchilov & Hutter,2017)14 GPU days34.4
1 GPU day36.519.6
Real et al. (2017)250 GPUs40.423.7
Brock et al. (2017)3 GPU days16.020.6
Ours (nsteps = 8,nneigh = 8,5 runs)1 GPU day22.323.4
Ours (snapshot ensemble,5 runs)2 GPU days73.320.9
Ours (ensemble across runs)5 GPU days111.519.6
", + "type": "table", + "image_path": "4a47d8eceb003bb23f38bfb4ad16df746a77c1f78312b59d11e2951c2db318cd.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 110, + 321, + 501, + 355.6666666666667 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 110, + 355.6666666666667, + 501, + 390.33333333333337 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 390.33333333333337, + 501, + 425.00000000000006 + ], + "spans": [], + "index": 9 + } + ] + } + ], + "index": 6.75 + }, + { + "type": "title", + "bbox": [ + 108, + 449, + 262, + 461 + ], + "lines": [ + { + "bbox": [ + 105, + 448, + 263, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 263, + 462 + ], + "score": 1.0, + "content": "5.2 EXPERIMENTS ON CIFAR-100", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 505, + 536 + ], + "lines": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "spans": [ + { + "bbox": [ + 106, + 470, + 505, + 483 + ], + "score": 1.0, + "content": "We repeat the previous experiment on CIFAR-100; hyperparameters were not changed. The results", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "spans": [ + { + "bbox": [ + 105, + 481, + 505, + 494 + ], + "score": 1.0, + "content": "are listed in Table 3. Unfortunately most automated architecture methods did not consider CIFAR-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 491, + 506, + 505 + ], + "spans": [ + { + "bbox": [ + 106, + 491, + 506, + 505 + ], + "score": 1.0, + "content": "100. Our method is on a par with Real et al. (2017) after one day with a single GPU. The snapshot", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 503, + 505, + 515 + ], + "score": 1.0, + "content": "ensemble performs similar to Brock et al. (2017) and an ensemble model build from the 5 runs can", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 514, + 505, + 526 + ], + "score": 1.0, + "content": "compete with the hand crafted WRN 28-10. The performance of the Shake-Shake network (Gastaldi,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 525, + 217, + 537 + ], + "spans": [ + { + "bbox": [ + 106, + 525, + 217, + 537 + ], + "score": 1.0, + "content": "2017) is again not reached.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13.5, + "bbox_fs": [ + 105, + 470, + 506, + 537 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 552, + 195, + 565 + ], + "lines": [ + { + "bbox": [ + 105, + 551, + 197, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 551, + 197, + 568 + ], + "score": 1.0, + "content": "6 CONCLUSION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 577, + 505, + 677 + ], + "lines": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 505, + 591 + ], + "score": 1.0, + "content": "We proposed NASH, a simple and fast method for automated architecture search based on a hill", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 587, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 104, + 587, + 506, + 602 + ], + "score": 1.0, + "content": "climbing strategy, network morphisms, and training via SGDR. Experiments on CIFAR-10 and", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 506, + 613 + ], + "score": 1.0, + "content": "CIFAR-100 showed that our method yields competitive results while requiring considerably less", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 104, + 608, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 104, + 608, + 506, + 624 + ], + "score": 1.0, + "content": "computational resources than most alternative approaches. Our algorithm is easily extendable, e.g.,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 506, + 634 + ], + "score": 1.0, + "content": "by other network morphisms, evolutionary approaches for generating new models, other methods", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 104, + 631, + 506, + 646 + ], + "spans": [ + { + "bbox": [ + 104, + 631, + 506, + 646 + ], + "score": 1.0, + "content": "for cheap performance evaluation (such as, e.g., learning curve prediction (Klein et al., 2017) or", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 104, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "hypernetworks (Ha et al., 2017; Brock et al., 2017)), or better resource handling strategies (such as", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 653, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 505, + 667 + ], + "score": 1.0, + "content": "Hyperband (Li et al., 2016b)). In this sense, we hope that our approach can serve as a basis for the", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 664, + 479, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 664, + 479, + 678 + ], + "score": 1.0, + "content": "development of more sophisticated methods that yield further improvements of performance.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 22, + "bbox_fs": [ + 104, + 577, + 506, + 678 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vin-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "cent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 165, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 505, + 180 + ], + "score": 1.0, + "content": "berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "on heterogeneous systems, 2015. URL https://www.tensorflow.org/. Software avail-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 186, + 219, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 219, + 202 + ], + "score": 1.0, + "content": "able from tensorflow.org.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "text", + "bbox": [ + 108, + 208, + 502, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 504, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 504, + 220 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 219, + 293, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 219, + 293, + 231 + ], + "score": 1.0, + "content": "learning to align and translate. ICLR, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 502, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 504, + 251 + ], + "score": 1.0, + "content": "Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. Designing neural network architec-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 250, + 333, + 262 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 333, + 262 + ], + "score": 1.0, + "content": "tures using reinforcement learning. ICLR 2017, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 503, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "score": 1.0, + "content": "J. Bergstra and Y. Bengio. Random search for hyper-parameter optimization. 13(1):281–305, 2012.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 105, + 288, + 504, + 312 + ], + "lines": [ + { + "bbox": [ + 104, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "J. Bergstra, R. Bardenet, Y. Bengio, and B. Kegl. Algorithms for hyper-parameter optimization. In ´", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 299, + 168, + 312 + ], + "spans": [ + { + "bbox": [ + 115, + 299, + 168, + 312 + ], + "score": 1.0, + "content": "NIPS, 2011.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 104, + 319, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "Andrew Brock, Theodore Lim, James M. Ritchie, and Nick Weston. SMASH: one-shot model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 330, + 380, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 380, + 343 + ], + "score": 1.0, + "content": "architecture search through hypernetworks. arXiv preprint, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 105, + 350, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang. Reinforcement learning for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 361, + 331, + 373 + ], + "spans": [ + { + "bbox": [ + 115, + 361, + 331, + 373 + ], + "score": 1.0, + "content": "architecture search by network transformation. 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "score": 1.0, + "content": "Tianqi Chen, Ian J. Goodfellow, and Jonathon Shlens. Net2net: Accelerating learning via knowledge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 391, + 239, + 404 + ], + "spans": [ + { + "bbox": [ + 115, + 391, + 239, + 404 + ], + "score": 1.0, + "content": "transfer. arXiv preprint, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 453, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 455, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 455, + 424 + ], + "score": 1.0, + "content": "Franc¸ois Chollet et al. Keras. https://github.com/fchollet/keras, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 108, + 431, + 405, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 405, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 405, + 444 + ], + "score": 1.0, + "content": "Xavier Gastaldi. Shake-shake regularization. ICLR 2017 Workshop, 2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 108, + 451, + 403, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 405, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 405, + 464 + ], + "score": 1.0, + "content": "David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. ICLR, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 503, + 493 + ], + "lines": [ + { + "bbox": [ + 104, + 468, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 505, + 486 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 482, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 200, + 493 + ], + "score": 1.0, + "content": "nition. CVPR, 2016.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5 + }, + { + "type": "text", + "bbox": [ + 107, + 501, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel rahman Mohamed, Navdeep Jaitly, An-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 510, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 506, + 526 + ], + "score": 1.0, + "content": "drew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. Deep", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 114, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 534, + 142, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 534, + 142, + 546 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5 + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 503, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 562, + 143, + 578 + ], + "spans": [ + { + "bbox": [ + 114, + 562, + 143, + 578 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5 + }, + { + "type": "text", + "bbox": [ + 107, + 584, + 503, + 607 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 595, + 372, + 608 + ], + "spans": [ + { + "bbox": [ + 116, + 595, + 372, + 608 + ], + "score": 1.0, + "content": "Snapshot ensembles: Train 1, get M for free. ICLR 2017, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 108, + 615, + 503, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 504, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 504, + 628 + ], + "score": 1.0, + "content": "A. Klein, S. Falkner, J. T. Springenberg, and F. Hutter. Learning curve prediction with Bayesian neu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 626, + 504, + 639 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 504, + 639 + ], + "score": 1.0, + "content": "ral networks. In International Conference on Learning Representations (ICLR) 2017 Conference", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 637, + 192, + 649 + ], + "spans": [ + { + "bbox": [ + 117, + 637, + 192, + 649 + ], + "score": 1.0, + "content": "Track, April 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 657, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 669 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "lutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger (eds.),", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems 25, pp. 1097–1105. Curran Associates, Inc.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 142, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 142, + 702 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 504, + 722 + ], + "score": 1.0, + "content": "L. Li, K. G. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar. Efficient hyperparameter", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 721, + 437, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 437, + 732 + ], + "score": 1.0, + "content": "optimization and infinitely many armed bandits. CoRR, abs/1603.06560, 2016a.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 82, + 175, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 176, + 95 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 100, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 100, + 505, + 113 + ], + "score": 1.0, + "content": "Mart´ın Abadi, Ashish Agarwal, Paul Barham, Eugene Brevdo, Zhifeng Chen, Craig Citro, Greg S.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 505, + 124 + ], + "score": 1.0, + "content": "Corrado, Andy Davis, Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Ian Goodfellow, Andrew", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "spans": [ + { + "bbox": [ + 115, + 122, + 505, + 135 + ], + "score": 1.0, + "content": "Harp, Geoffrey Irving, Michael Isard, Yangqing Jia, Rafal Jozefowicz, Lukasz Kaiser, Manjunath", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "spans": [ + { + "bbox": [ + 115, + 133, + 505, + 146 + ], + "score": 1.0, + "content": "Kudlur, Josh Levenberg, Dan Mane, Rajat Monga, Sherry Moore, Derek Murray, Chris Olah, ´", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 505, + 156 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 505, + 156 + ], + "score": 1.0, + "content": "Mike Schuster, Jonathon Shlens, Benoit Steiner, Ilya Sutskever, Kunal Talwar, Paul Tucker, Vin-", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 115, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "cent Vanhoucke, Vijay Vasudevan, Fernanda Viegas, Oriol Vinyals, Pete Warden, Martin Watten- ´", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 165, + 505, + 180 + ], + "spans": [ + { + "bbox": [ + 115, + 165, + 505, + 180 + ], + "score": 1.0, + "content": "berg, Martin Wicke, Yuan Yu, and Xiaoqiang Zheng. TensorFlow: Large-scale machine learning", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 177, + 505, + 190 + ], + "spans": [ + { + "bbox": [ + 115, + 177, + 505, + 190 + ], + "score": 1.0, + "content": "on heterogeneous systems, 2015. URL https://www.tensorflow.org/. Software avail-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 186, + 219, + 202 + ], + "spans": [ + { + "bbox": [ + 115, + 186, + 219, + 202 + ], + "score": 1.0, + "content": "able from tensorflow.org.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 100, + 505, + 202 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 208, + 502, + 230 + ], + "lines": [ + { + "bbox": [ + 106, + 208, + 504, + 220 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 504, + 220 + ], + "score": 1.0, + "content": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 219, + 293, + 231 + ], + "spans": [ + { + "bbox": [ + 115, + 219, + 293, + 231 + ], + "score": 1.0, + "content": "learning to align and translate. ICLR, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 106, + 208, + 504, + 231 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 238, + 502, + 261 + ], + "lines": [ + { + "bbox": [ + 106, + 238, + 504, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 238, + 504, + 251 + ], + "score": 1.0, + "content": "Bowen Baker, Otkrist Gupta, Nikhil Naik, and Ramesh Raskar. Designing neural network architec-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 116, + 250, + 333, + 262 + ], + "spans": [ + { + "bbox": [ + 116, + 250, + 333, + 262 + ], + "score": 1.0, + "content": "tures using reinforcement learning. ICLR 2017, 2016.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 238, + 504, + 262 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 268, + 503, + 281 + ], + "lines": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 505, + 282 + ], + "score": 1.0, + "content": "J. Bergstra and Y. Bengio. Random search for hyper-parameter optimization. 13(1):281–305, 2012.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 268, + 505, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 288, + 504, + 312 + ], + "lines": [ + { + "bbox": [ + 104, + 288, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 104, + 288, + 505, + 302 + ], + "score": 1.0, + "content": "J. Bergstra, R. Bardenet, Y. Bengio, and B. Kegl. Algorithms for hyper-parameter optimization. In ´", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 299, + 168, + 312 + ], + "spans": [ + { + "bbox": [ + 115, + 299, + 168, + 312 + ], + "score": 1.0, + "content": "NIPS, 2011.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 104, + 288, + 505, + 312 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 319, + 504, + 342 + ], + "lines": [ + { + "bbox": [ + 106, + 319, + 505, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 319, + 505, + 332 + ], + "score": 1.0, + "content": "Andrew Brock, Theodore Lim, James M. Ritchie, and Nick Weston. SMASH: one-shot model", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 330, + 380, + 343 + ], + "spans": [ + { + "bbox": [ + 115, + 330, + 380, + 343 + ], + "score": 1.0, + "content": "architecture search through hypernetworks. arXiv preprint, 2017.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17.5, + "bbox_fs": [ + 106, + 319, + 505, + 343 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 350, + 504, + 373 + ], + "lines": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "Han Cai, Tianyao Chen, Weinan Zhang, Yong Yu, and Jun Wang. Reinforcement learning for", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 361, + 331, + 373 + ], + "spans": [ + { + "bbox": [ + 115, + 361, + 331, + 373 + ], + "score": 1.0, + "content": "architecture search by network transformation. 2017.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 349, + 506, + 373 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 380, + 504, + 403 + ], + "lines": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 394 + ], + "score": 1.0, + "content": "Tianqi Chen, Ian J. Goodfellow, and Jonathon Shlens. Net2net: Accelerating learning via knowledge", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 391, + 239, + 404 + ], + "spans": [ + { + "bbox": [ + 115, + 391, + 239, + 404 + ], + "score": 1.0, + "content": "transfer. arXiv preprint, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 379, + 505, + 404 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 411, + 453, + 423 + ], + "lines": [ + { + "bbox": [ + 105, + 411, + 455, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 455, + 424 + ], + "score": 1.0, + "content": "Franc¸ois Chollet et al. Keras. https://github.com/fchollet/keras, 2015.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 105, + 411, + 455, + 424 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 431, + 405, + 443 + ], + "lines": [ + { + "bbox": [ + 106, + 430, + 405, + 444 + ], + "spans": [ + { + "bbox": [ + 106, + 430, + 405, + 444 + ], + "score": 1.0, + "content": "Xavier Gastaldi. Shake-shake regularization. ICLR 2017 Workshop, 2017.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24, + "bbox_fs": [ + 106, + 430, + 405, + 444 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 451, + 403, + 462 + ], + "lines": [ + { + "bbox": [ + 106, + 450, + 405, + 464 + ], + "spans": [ + { + "bbox": [ + 106, + 450, + 405, + 464 + ], + "score": 1.0, + "content": "David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. ICLR, 2017.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25, + "bbox_fs": [ + 106, + 450, + 405, + 464 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 470, + 503, + 493 + ], + "lines": [ + { + "bbox": [ + 104, + 468, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 505, + 486 + ], + "score": 1.0, + "content": "Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 116, + 482, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 116, + 482, + 200, + 493 + ], + "score": 1.0, + "content": "nition. CVPR, 2016.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 26.5, + "bbox_fs": [ + 104, + 468, + 505, + 493 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 501, + 505, + 545 + ], + "lines": [ + { + "bbox": [ + 105, + 501, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 501, + 505, + 514 + ], + "score": 1.0, + "content": "Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdel rahman Mohamed, Navdeep Jaitly, An-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 510, + 506, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 506, + 526 + ], + "score": 1.0, + "content": "drew Senior, Vincent Vanhoucke, Patrick Nguyen, Tara Sainath, and Brian Kingsbury. Deep", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 114, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 114, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "neural networks for acoustic modeling in speech recognition. IEEE Signal Processing Magazine,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 115, + 534, + 142, + 546 + ], + "spans": [ + { + "bbox": [ + 115, + 534, + 142, + 546 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 29.5, + "bbox_fs": [ + 105, + 501, + 506, + 546 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 554, + 503, + 576 + ], + "lines": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 567 + ], + "score": 1.0, + "content": "Gao Huang, Zhuang Liu, and Kilian Q. Weinberger. Densely connected convolutional networks.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 114, + 562, + 143, + 578 + ], + "spans": [ + { + "bbox": [ + 114, + 562, + 143, + 578 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32.5, + "bbox_fs": [ + 105, + 553, + 505, + 578 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 584, + 503, + 607 + ], + "lines": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 105, + 583, + 505, + 597 + ], + "score": 1.0, + "content": "Gao Huang, Yixuan Li, Geoff Pleiss, Zhuang Liu, John E. Hopcroft, and Kilian Q. Weinberger.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 595, + 372, + 608 + ], + "spans": [ + { + "bbox": [ + 116, + 595, + 372, + 608 + ], + "score": 1.0, + "content": "Snapshot ensembles: Train 1, get M for free. ICLR 2017, 2017.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 583, + 505, + 608 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 615, + 503, + 649 + ], + "lines": [ + { + "bbox": [ + 106, + 614, + 504, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 614, + 504, + 628 + ], + "score": 1.0, + "content": "A. Klein, S. Falkner, J. T. Springenberg, and F. Hutter. Learning curve prediction with Bayesian neu-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 115, + 626, + 504, + 639 + ], + "spans": [ + { + "bbox": [ + 115, + 626, + 504, + 639 + ], + "score": 1.0, + "content": "ral networks. In International Conference on Learning Representations (ICLR) 2017 Conference", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 117, + 637, + 192, + 649 + ], + "spans": [ + { + "bbox": [ + 117, + 637, + 192, + 649 + ], + "score": 1.0, + "content": "Track, April 2017.", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 37, + "bbox_fs": [ + 106, + 614, + 504, + 649 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 657, + 505, + 701 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 669 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 669 + ], + "score": 1.0, + "content": "Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convo-", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 667, + 505, + 681 + ], + "spans": [ + { + "bbox": [ + 115, + 667, + 505, + 681 + ], + "score": 1.0, + "content": "lutional neural networks. In F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger (eds.),", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 115, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "Advances in Neural Information Processing Systems 25, pp. 1097–1105. Curran Associates, Inc.,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 142, + 702 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 142, + 702 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 656, + 506, + 702 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 502, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 504, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 504, + 722 + ], + "score": 1.0, + "content": "L. Li, K. G. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar. Efficient hyperparameter", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 116, + 721, + 437, + 732 + ], + "spans": [ + { + "bbox": [ + 116, + 721, + 437, + 732 + ], + "score": 1.0, + "content": "optimization and infinitely many armed bandits. CoRR, abs/1603.06560, 2016a.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 708, + 504, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 104, + 76, + 507, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Efficient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 94, + 469, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 469, + 107 + ], + "score": 1.0, + "content": "hyperparameter optimization and infinitely many armed bandits. arXiv preprint, 2016b.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 104, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "I. Loshchilov and F. Hutter. CMA-ES for hyperparameter optimization of deep neural networks.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 124, + 238, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 238, + 136 + ], + "score": 1.0, + "content": "CoRR, abs/1604.07269, 2016.", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 154, + 460, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 460, + 167 + ], + "score": 1.0, + "content": "Conference on Learning Representations (ICLR) 2017 Conference Track, April 2017.", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "H. Mendoza, A. Klein, M. Feurer, T. Springenberg, and F. Hutter. Towards automatically-tuned", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 183, + 259, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 183, + 259, + 196 + ], + "score": 1.0, + "content": "neural networks. In AutoML, 2016.", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 201, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 216 + ], + "score": 1.0, + "content": "Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V. Le,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 213, + 453, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 453, + 227 + ], + "score": 1.0, + "content": "and Alex Kurakin. Large-scale evolution of image classifiers. arXiv preprint, 2017.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 231, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 246 + ], + "score": 1.0, + "content": "Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach (3rd Edition). Pearson,", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 243, + 307, + 256 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 307, + 256 + ], + "score": 1.0, + "content": "3 edition, December 2009. ISBN 0136042597.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 262, + 458, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 458, + 275 + ], + "score": 1.0, + "content": "Shreyas Saxena and Jakob Verbeek. Convolutional neural fabrics. arXiv preprint, 2016.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 104, + 280, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 104, + 280, + 505, + 295 + ], + "score": 1.0, + "content": "J. Snoek, H. Larochelle, and R.P. Adams. Practical Bayesian optimization of machine learning", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 292, + 227, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 227, + 304 + ], + "score": 1.0, + "content": "algorithms. In NIPS, 2012.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "score": 1.0, + "content": "Masanori Suganuma, Shinichi Shirakawa, and Tomoharu Nagao. A genetic programming approach", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 321, + 432, + 335 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 432, + 335 + ], + "score": 1.0, + "content": "to designing convolutional neural network architectures. arXiv preprint, 2017.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 339, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 355 + ], + "score": 1.0, + "content": "Tao Wei, Changhu Wang, Yong Rui, and Chang Wen Chen. Network morphism. arXiv preprint,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 351, + 142, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 142, + 363 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 371, + 464, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 464, + 384 + ], + "score": 1.0, + "content": "Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint, 2016.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 389, + 475, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 475, + 402 + ], + "score": 1.0, + "content": "Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. 2017.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning transferable architectures", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 419, + 269, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 269, + 433 + ], + "score": 1.0, + "content": "for scalable image recognition. 2017.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 11 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 104, + 76, + 507, + 432 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "Lisha Li, Kevin G. Jamieson, Giulia DeSalvo, Afshin Rostamizadeh, and Ameet Talwalkar. Efficient", + "type": "text" + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 94, + 469, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 94, + 469, + 107 + ], + "score": 1.0, + "content": "hyperparameter optimization and infinitely many armed bandits. arXiv preprint, 2016b.", + "type": "text" + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 111, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 104, + 111, + 505, + 126 + ], + "score": 1.0, + "content": "I. Loshchilov and F. Hutter. CMA-ES for hyperparameter optimization of deep neural networks.", + "type": "text" + } + ], + "index": 2, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 124, + 238, + 136 + ], + "spans": [ + { + "bbox": [ + 116, + 124, + 238, + 136 + ], + "score": 1.0, + "content": "CoRR, abs/1604.07269, 2016.", + "type": "text" + } + ], + "index": 3, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "spans": [ + { + "bbox": [ + 105, + 142, + 506, + 155 + ], + "score": 1.0, + "content": "I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. In International", + "type": "text" + } + ], + "index": 4, + "is_list_start_line": true + }, + { + "bbox": [ + 116, + 154, + 460, + 167 + ], + "spans": [ + { + "bbox": [ + 116, + 154, + 460, + 167 + ], + "score": 1.0, + "content": "Conference on Learning Representations (ICLR) 2017 Conference Track, April 2017.", + "type": "text" + } + ], + "index": 5, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 505, + 185 + ], + "score": 1.0, + "content": "H. Mendoza, A. Klein, M. Feurer, T. Springenberg, and F. Hutter. Towards automatically-tuned", + "type": "text" + } + ], + "index": 6, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 183, + 259, + 196 + ], + "spans": [ + { + "bbox": [ + 115, + 183, + 259, + 196 + ], + "score": 1.0, + "content": "neural networks. In AutoML, 2016.", + "type": "text" + } + ], + "index": 7, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 201, + 506, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 201, + 506, + 216 + ], + "score": 1.0, + "content": "Esteban Real, Sherry Moore, Andrew Selle, Saurabh Saxena, Yutaka Leon Suematsu, Quoc V. Le,", + "type": "text" + } + ], + "index": 8, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 213, + 453, + 227 + ], + "spans": [ + { + "bbox": [ + 115, + 213, + 453, + 227 + ], + "score": 1.0, + "content": "and Alex Kurakin. Large-scale evolution of image classifiers. arXiv preprint, 2017.", + "type": "text" + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 231, + 506, + 246 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 506, + 246 + ], + "score": 1.0, + "content": "Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach (3rd Edition). Pearson,", + "type": "text" + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 243, + 307, + 256 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 307, + 256 + ], + "score": 1.0, + "content": "3 edition, December 2009. ISBN 0136042597.", + "type": "text" + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 262, + 458, + 275 + ], + "spans": [ + { + "bbox": [ + 105, + 262, + 458, + 275 + ], + "score": 1.0, + "content": "Shreyas Saxena and Jakob Verbeek. Convolutional neural fabrics. arXiv preprint, 2016.", + "type": "text" + } + ], + "index": 12, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 280, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 104, + 280, + 505, + 295 + ], + "score": 1.0, + "content": "J. Snoek, H. Larochelle, and R.P. Adams. Practical Bayesian optimization of machine learning", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 292, + 227, + 304 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 227, + 304 + ], + "score": 1.0, + "content": "algorithms. In NIPS, 2012.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 325 + ], + "score": 1.0, + "content": "Masanori Suganuma, Shinichi Shirakawa, and Tomoharu Nagao. A genetic programming approach", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 321, + 432, + 335 + ], + "spans": [ + { + "bbox": [ + 115, + 321, + 432, + 335 + ], + "score": 1.0, + "content": "to designing convolutional neural network architectures. arXiv preprint, 2017.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 339, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 105, + 339, + 505, + 355 + ], + "score": 1.0, + "content": "Tao Wei, Changhu Wang, Yong Rui, and Chang Wen Chen. Network morphism. arXiv preprint,", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 351, + 142, + 363 + ], + "spans": [ + { + "bbox": [ + 115, + 351, + 142, + 363 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 371, + 464, + 384 + ], + "spans": [ + { + "bbox": [ + 106, + 371, + 464, + 384 + ], + "score": 1.0, + "content": "Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint, 2016.", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 389, + 475, + 402 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 475, + 402 + ], + "score": 1.0, + "content": "Barret Zoph and Quoc V. Le. Neural architecture search with reinforcement learning. 2017.", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 506, + 422 + ], + "score": 1.0, + "content": "Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V. Le. Learning transferable architectures", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 419, + 269, + 433 + ], + "spans": [ + { + "bbox": [ + 115, + 419, + 269, + 433 + ], + "score": 1.0, + "content": "for scalable image recognition. 2017.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + } + ], + "index": 11, + "bbox_fs": [ + 104, + 82, + 506, + 433 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 123, + 466, + 310 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 123, + 466, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 124, + 466, + 310 + ], + "spans": [ + { + "bbox": [ + 117, + 124, + 466, + 310 + ], + "score": 0.962, + "type": "image", + "image_path": "2edbb1405729ce5dc046dc80f49bd9380081ebed874aa93e2d296d2395a7ddf0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 123, + 466, + 185.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 185.33333333333334, + 466, + 247.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 247.66666666666669, + 466, + 310.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 331, + 392, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 218, + 331, + 392, + 344 + ], + "spans": [ + { + "bbox": [ + 218, + 331, + 392, + 344 + ], + "score": 1.0, + "content": "Figure 3: Initial network for our algorithm.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 121, + 372, + 466, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 372, + 466, + 560 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 121, + 372, + 466, + 560 + ], + "spans": [ + { + "bbox": [ + 121, + 372, + 466, + 560 + ], + "score": 0.967, + "type": "image", + "image_path": "213925428f8eea0c7c284680e760c96a1baa86df2685b0480330288a33417347.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 121, + 372, + 466, + 434.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 121, + 434.6666666666667, + 466, + 497.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 121, + 497.33333333333337, + 466, + 560.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 581, + 392, + 593 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 218, + 581, + 392, + 594 + ], + "spans": [ + { + "bbox": [ + 218, + 581, + 392, + 594 + ], + "score": 1.0, + "content": "Figure 4: Initial network for our algorithm.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 82, + 435, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 437, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 437, + 96 + ], + "score": 1.0, + "content": "A CORRELATION BETWEEN VALIDATION AND TEST ACCURACY", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 617, + 205, + 630 + ], + "lines": [ + { + "bbox": [ + 105, + 614, + 206, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 206, + 632 + ], + "score": 1.0, + "content": "B SOME MODELS", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 117, + 123, + 466, + 310 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 117, + 123, + 466, + 310 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 117, + 124, + 466, + 310 + ], + "spans": [ + { + "bbox": [ + 117, + 124, + 466, + 310 + ], + "score": 0.962, + "type": "image", + "image_path": "2edbb1405729ce5dc046dc80f49bd9380081ebed874aa93e2d296d2395a7ddf0.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 117, + 123, + 466, + 185.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 117, + 185.33333333333334, + 466, + 247.66666666666669 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 117, + 247.66666666666669, + 466, + 310.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 331, + 392, + 344 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 218, + 331, + 392, + 344 + ], + "spans": [ + { + "bbox": [ + 218, + 331, + 392, + 344 + ], + "score": 1.0, + "content": "Figure 3: Initial network for our algorithm.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 121, + 372, + 466, + 560 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 121, + 372, + 466, + 560 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 121, + 372, + 466, + 560 + ], + "spans": [ + { + "bbox": [ + 121, + 372, + 466, + 560 + ], + "score": 0.967, + "type": "image", + "image_path": "213925428f8eea0c7c284680e760c96a1baa86df2685b0480330288a33417347.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 121, + 372, + 466, + 434.6666666666667 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 121, + 434.6666666666667, + 466, + 497.33333333333337 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 121, + 497.33333333333337, + 466, + 560.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 581, + 392, + 593 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 218, + 581, + 392, + 594 + ], + "spans": [ + { + "bbox": [ + 218, + 581, + 392, + 594 + ], + "score": 1.0, + "content": "Figure 4: Initial network for our algorithm.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 251, + 134, + 356, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 251, + 134, + 356, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 251, + 134, + 356, + 654 + ], + "spans": [ + { + "bbox": [ + 251, + 134, + 356, + 654 + ], + "score": 0.939, + "type": "image", + "image_path": "926480ca8e0fe0b1fac91e9b0c6c21dd16e70a4e522edfe39777971a2e331b4d.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 251, + 134, + 356, + 654 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 665, + 392, + 677 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 217, + 663, + 392, + 678 + ], + "spans": [ + { + "bbox": [ + 217, + 663, + 392, + 678 + ], + "score": 1.0, + "content": "Figure 5: Initial network for our algorithm.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + } + ], + "index": 0.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 251, + 134, + 356, + 654 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 251, + 134, + 356, + 654 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 251, + 134, + 356, + 654 + ], + "spans": [ + { + "bbox": [ + 251, + 134, + 356, + 654 + ], + "score": 0.939, + "type": "image", + "image_path": "926480ca8e0fe0b1fac91e9b0c6c21dd16e70a4e522edfe39777971a2e331b4d.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 251, + 134, + 356, + 654 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 219, + 665, + 392, + 677 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 217, + 663, + 392, + 678 + ], + "spans": [ + { + "bbox": [ + 217, + 663, + 392, + 678 + ], + "score": 1.0, + "content": "Figure 5: Initial network for our algorithm.", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + } + ], + "index": 0.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 249, + 92, + 359, + 687 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 249, + 92, + 359, + 687 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 249, + 92, + 359, + 687 + ], + "spans": [ + { + "bbox": [ + 249, + 92, + 359, + 687 + ], + "score": 0.928, + "type": "image", + "image_path": "091ad5bd88208c13ba32695bfc45a25363ea4c4e47109a8acba5e63354b8c3d1.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 249, + 92, + 359, + 687 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 178, + 697, + 433, + 709 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 177, + 695, + 434, + 712 + ], + "spans": [ + { + "bbox": [ + 177, + 695, + 385, + 712 + ], + "score": 1.0, + "content": "Figure 6: Network generated by our algorithm with", + "type": "text" + }, + { + "bbox": [ + 386, + 698, + 430, + 710 + ], + "score": 0.91, + "content": "n _ { s t e p s } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 695, + 434, + 712 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + } + ], + "index": 0.5 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 249, + 92, + 359, + 687 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 249, + 92, + 359, + 687 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 249, + 92, + 359, + 687 + ], + "spans": [ + { + "bbox": [ + 249, + 92, + 359, + 687 + ], + "score": 0.928, + "type": "image", + "image_path": "091ad5bd88208c13ba32695bfc45a25363ea4c4e47109a8acba5e63354b8c3d1.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 249, + 92, + 359, + 687 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 178, + 697, + 433, + 709 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 177, + 695, + 434, + 712 + ], + "spans": [ + { + "bbox": [ + 177, + 695, + 385, + 712 + ], + "score": 1.0, + "content": "Figure 6: Network generated by our algorithm with", + "type": "text" + }, + { + "bbox": [ + 386, + 698, + 430, + 710 + ], + "score": 0.91, + "content": "n _ { s t e p s } = 5", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 695, + 434, + 712 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + } + ], + "index": 0.5 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 251, + 90, + 350, + 682 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 251, + 90, + 350, + 682 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 251, + 90, + 350, + 682 + ], + "spans": [ + { + "bbox": [ + 251, + 90, + 350, + 682 + ], + "score": 0.917, + "type": "image", + "image_path": "e810d4abf64e5a83ec8d5f1149515b76d979a3c8fc66b0153106f09e2b3c3c2d.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 251, + 90, + 350, + 682 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 178, + 697, + 433, + 709 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 177, + 695, + 434, + 712 + ], + "spans": [ + { + "bbox": [ + 177, + 695, + 385, + 712 + ], + "score": 1.0, + "content": "Figure 7: Network generated by our algorithm with", + "type": "text" + }, + { + "bbox": [ + 386, + 698, + 430, + 710 + ], + "score": 0.9, + "content": "n _ { s t e p s } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 695, + 434, + 712 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + } + ], + "index": 0.5 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 763 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 251, + 90, + 350, + 682 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 251, + 90, + 350, + 682 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 251, + 90, + 350, + 682 + ], + "spans": [ + { + "bbox": [ + 251, + 90, + 350, + 682 + ], + "score": 0.917, + "type": "image", + "image_path": "e810d4abf64e5a83ec8d5f1149515b76d979a3c8fc66b0153106f09e2b3c3c2d.jpg" + } + ] + } + ], + "index": 0, + "virtual_lines": [ + { + "bbox": [ + 251, + 90, + 350, + 682 + ], + "spans": [], + "index": 0 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 178, + 697, + 433, + 709 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 177, + 695, + 434, + 712 + ], + "spans": [ + { + "bbox": [ + 177, + 695, + 385, + 712 + ], + "score": 1.0, + "content": "Figure 7: Network generated by our algorithm with", + "type": "text" + }, + { + "bbox": [ + 386, + 698, + 430, + 710 + ], + "score": 0.9, + "content": "n _ { s t e p s } = 8", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 695, + 434, + 712 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + } + ], + "index": 0.5 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/SySaJ0xCZ/SySaJ0xCZ_model.json b/parse/train/SySaJ0xCZ/SySaJ0xCZ_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4f48c322ebe58f8c5d198c54298b89c8ae29f01b --- /dev/null +++ b/parse/train/SySaJ0xCZ/SySaJ0xCZ_model.json @@ -0,0 +1,15136 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 591, + 1302, + 591, + 1302, + 896, + 398, + 896 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1230, + 1402, + 1230, + 1402, + 1474, + 298, + 1474 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1031, + 1404, + 1031, + 1404, + 1215, + 298, + 1215 + ], + "score": 0.976 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1399, + 219, + 1399, + 324, + 301, + 324 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 295, + 1490, + 1398, + 1490, + 1398, + 1551, + 295, + 1551 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 297, + 1580, + 1406, + 1580, + 1406, + 2034, + 297, + 2034 + ], + "score": 0.919 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.902 + }, + { + "category_id": 0, + "poly": [ + 302, + 959, + 573, + 959, + 573, + 994, + 302, + 994 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 773, + 519, + 926, + 519, + 926, + 553, + 773, + 553 + ], + "score": 0.866 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 857, + 2088, + 857, + 2112, + 842, + 2112 + ], + "score": 0.643 + }, + { + "category_id": 13, + "poly": [ + 884, + 836, + 925, + 836, + 925, + 864, + 884, + 864 + ], + "score": 0.86, + "latex": "6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1019, + 866, + 1058, + 866, + 1058, + 894, + 1019, + 894 + ], + "score": 0.86, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1094, + 2003, + 1142, + 2003, + 1142, + 2032, + 1094, + 2032 + ], + "score": 0.85, + "latex": "6 \\%" + }, + { + "category_id": 13, + "poly": [ + 1209, + 1611, + 1296, + 1611, + 1296, + 1640, + 1209, + 1640 + ], + "score": 0.71, + "latex": "6 \\% - 7 \\%" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 218.0, + 1403.0, + 218.0, + 1403.0, + 273.0, + 295.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 270.0, + 1107.0, + 270.0, + 1107.0, + 330.0, + 294.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 956.0, + 579.0, + 956.0, + 579.0, + 1003.0, + 294.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 518.0, + 932.0, + 518.0, + 932.0, + 556.0, + 769.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 2088.0, + 858.0, + 2088.0, + 858.0, + 2116.0, + 841.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 591.0, + 1305.0, + 591.0, + 1305.0, + 625.0, + 394.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 620.0, + 1305.0, + 620.0, + 1305.0, + 656.0, + 393.0, + 656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 651.0, + 1306.0, + 651.0, + 1306.0, + 687.0, + 392.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 684.0, + 1305.0, + 684.0, + 1305.0, + 717.0, + 394.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 712.0, + 1307.0, + 712.0, + 1307.0, + 748.0, + 392.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 743.0, + 1307.0, + 743.0, + 1307.0, + 779.0, + 393.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 774.0, + 1307.0, + 774.0, + 1307.0, + 808.0, + 392.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 806.0, + 1305.0, + 806.0, + 1305.0, + 838.0, + 394.0, + 838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 833.0, + 883.0, + 833.0, + 883.0, + 870.0, + 394.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 926.0, + 833.0, + 1306.0, + 833.0, + 1306.0, + 870.0, + 926.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 866.0, + 1018.0, + 866.0, + 1018.0, + 897.0, + 395.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 866.0, + 1071.0, + 866.0, + 1071.0, + 897.0, + 1059.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1231.0, + 1406.0, + 1231.0, + 1406.0, + 1265.0, + 296.0, + 1265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1262.0, + 1403.0, + 1262.0, + 1403.0, + 1296.0, + 294.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1291.0, + 1403.0, + 1291.0, + 1403.0, + 1328.0, + 293.0, + 1328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1319.0, + 1406.0, + 1319.0, + 1406.0, + 1359.0, + 292.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1353.0, + 1406.0, + 1353.0, + 1406.0, + 1387.0, + 294.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1419.0, + 292.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1413.0, + 1407.0, + 1413.0, + 1407.0, + 1448.0, + 291.0, + 1448.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1445.0, + 700.0, + 1445.0, + 700.0, + 1479.0, + 296.0, + 1479.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1029.0, + 1408.0, + 1029.0, + 1408.0, + 1066.0, + 293.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1061.0, + 1405.0, + 1061.0, + 1405.0, + 1096.0, + 294.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1091.0, + 1404.0, + 1091.0, + 1404.0, + 1126.0, + 292.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1123.0, + 1406.0, + 1123.0, + 1406.0, + 1159.0, + 294.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1152.0, + 1406.0, + 1152.0, + 1406.0, + 1188.0, + 294.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1185.0, + 1340.0, + 1185.0, + 1340.0, + 1217.0, + 296.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1485.0, + 1403.0, + 1485.0, + 1403.0, + 1527.0, + 293.0, + 1527.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1522.0, + 1013.0, + 1522.0, + 1013.0, + 1553.0, + 296.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1581.0, + 1404.0, + 1581.0, + 1404.0, + 1615.0, + 299.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1610.0, + 1208.0, + 1610.0, + 1208.0, + 1645.0, + 323.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1610.0, + 1406.0, + 1610.0, + 1406.0, + 1645.0, + 1297.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1641.0, + 1404.0, + 1641.0, + 1404.0, + 1676.0, + 323.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1672.0, + 1404.0, + 1672.0, + 1404.0, + 1707.0, + 323.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1701.0, + 820.0, + 1701.0, + 820.0, + 1737.0, + 321.0, + 1737.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1745.0, + 1404.0, + 1745.0, + 1404.0, + 1780.0, + 299.0, + 1780.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1775.0, + 1404.0, + 1775.0, + 1404.0, + 1812.0, + 321.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1806.0, + 606.0, + 1806.0, + 606.0, + 1840.0, + 323.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1850.0, + 1404.0, + 1850.0, + 1404.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1880.0, + 1407.0, + 1880.0, + 1407.0, + 1918.0, + 321.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1912.0, + 1406.0, + 1912.0, + 1406.0, + 1947.0, + 321.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1944.0, + 1403.0, + 1944.0, + 1403.0, + 1975.0, + 324.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1968.0, + 1406.0, + 1968.0, + 1406.0, + 2012.0, + 320.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 2001.0, + 1093.0, + 2001.0, + 1093.0, + 2038.0, + 321.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 2001.0, + 1404.0, + 2001.0, + 1404.0, + 2038.0, + 1143.0, + 2038.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 842, + 1405, + 842, + 1405, + 1422, + 297, + 1422 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1759, + 1404, + 1759, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1437, + 1404, + 1437, + 1404, + 1743, + 298, + 1743 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 299, + 674, + 1403, + 674, + 1403, + 827, + 299, + 827 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 317, + 229, + 1405, + 229, + 1405, + 323, + 317, + 323 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 299, + 424, + 1400, + 424, + 1400, + 517, + 299, + 517 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 296, + 355, + 1398, + 355, + 1398, + 388, + 296, + 388 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 300, + 589, + 581, + 589, + 581, + 624, + 300, + 624 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 856, + 75, + 856, + 105, + 298, + 105 + ], + "score": 0.733 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2112, + 841, + 2112 + ], + "score": 0.678 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.277 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.198 + }, + { + "category_id": 13, + "poly": [ + 766, + 292, + 820, + 292, + 820, + 320, + 766, + 320 + ], + "score": 0.87, + "latex": "24 \\%" + }, + { + "category_id": 13, + "poly": [ + 1126, + 292, + 1180, + 292, + 1180, + 320, + 1126, + 320 + ], + "score": 0.86, + "latex": "20 \\%" + }, + { + "category_id": 13, + "poly": [ + 987, + 231, + 1028, + 231, + 1028, + 259, + 987, + 259 + ], + "score": 0.86, + "latex": "5 \\%" + }, + { + "category_id": 13, + "poly": [ + 916, + 262, + 984, + 262, + 984, + 290, + 916, + 290 + ], + "score": 0.86, + "latex": "4 . 7 \\%" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 585.0, + 586.0, + 585.0, + 586.0, + 631.0, + 291.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2121.0, + 839.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 839.0, + 1403.0, + 839.0, + 1403.0, + 881.0, + 292.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 873.0, + 1406.0, + 873.0, + 1406.0, + 909.0, + 291.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 903.0, + 1405.0, + 903.0, + 1405.0, + 942.0, + 294.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 933.0, + 1405.0, + 933.0, + 1405.0, + 971.0, + 292.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 962.0, + 1407.0, + 962.0, + 1407.0, + 1003.0, + 291.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 994.0, + 1406.0, + 994.0, + 1406.0, + 1033.0, + 294.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1062.0, + 294.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1056.0, + 1407.0, + 1056.0, + 1407.0, + 1093.0, + 294.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1087.0, + 1405.0, + 1087.0, + 1405.0, + 1122.0, + 294.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1118.0, + 1406.0, + 1118.0, + 1406.0, + 1153.0, + 294.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1147.0, + 1405.0, + 1147.0, + 1405.0, + 1182.0, + 294.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1178.0, + 1407.0, + 1178.0, + 1407.0, + 1213.0, + 295.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1208.0, + 1403.0, + 1208.0, + 1403.0, + 1243.0, + 295.0, + 1243.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1239.0, + 1405.0, + 1239.0, + 1405.0, + 1275.0, + 295.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1270.0, + 1403.0, + 1270.0, + 1403.0, + 1305.0, + 295.0, + 1305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1337.0, + 292.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1366.0, + 295.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1361.0, + 1405.0, + 1361.0, + 1405.0, + 1396.0, + 294.0, + 1396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1392.0, + 1400.0, + 1392.0, + 1400.0, + 1427.0, + 295.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1761.0, + 1404.0, + 1761.0, + 1404.0, + 1794.0, + 296.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1790.0, + 1402.0, + 1790.0, + 1402.0, + 1822.0, + 296.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1855.0, + 293.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1850.0, + 1406.0, + 1850.0, + 1406.0, + 1887.0, + 293.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1919.0, + 294.0, + 1919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 1405.0, + 1911.0, + 1405.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1404.0, + 1971.0, + 1404.0, + 2011.0, + 293.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 905.0, + 2002.0, + 905.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1436.0, + 1404.0, + 1436.0, + 1404.0, + 1473.0, + 292.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1467.0, + 1406.0, + 1467.0, + 1406.0, + 1504.0, + 292.0, + 1504.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1495.0, + 1406.0, + 1495.0, + 1406.0, + 1537.0, + 291.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1529.0, + 1405.0, + 1529.0, + 1405.0, + 1565.0, + 293.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 1405.0, + 1561.0, + 1405.0, + 1596.0, + 293.0, + 1596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1590.0, + 1406.0, + 1590.0, + 1406.0, + 1624.0, + 291.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1620.0, + 1404.0, + 1620.0, + 1404.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1653.0, + 1405.0, + 1653.0, + 1405.0, + 1685.0, + 297.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1680.0, + 1406.0, + 1680.0, + 1406.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1714.0, + 581.0, + 1714.0, + 581.0, + 1745.0, + 296.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 674.0, + 1404.0, + 674.0, + 1404.0, + 707.0, + 297.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 703.0, + 1405.0, + 703.0, + 1405.0, + 741.0, + 294.0, + 741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 733.0, + 1407.0, + 733.0, + 1407.0, + 772.0, + 294.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 801.0, + 293.0, + 801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 832.0, + 295.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 230.0, + 986.0, + 230.0, + 986.0, + 264.0, + 320.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 264.0, + 1029.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 261.0, + 915.0, + 261.0, + 915.0, + 295.0, + 323.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 985.0, + 261.0, + 1404.0, + 261.0, + 1404.0, + 295.0, + 985.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 288.0, + 765.0, + 288.0, + 765.0, + 328.0, + 322.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 288.0, + 1125.0, + 288.0, + 1125.0, + 328.0, + 821.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 288.0, + 1355.0, + 288.0, + 1355.0, + 328.0, + 1181.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 422.0, + 1404.0, + 422.0, + 1404.0, + 460.0, + 295.0, + 460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 456.0, + 1404.0, + 456.0, + 1404.0, + 490.0, + 295.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 486.0, + 881.0, + 486.0, + 881.0, + 519.0, + 294.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 352.0, + 1402.0, + 352.0, + 1402.0, + 391.0, + 295.0, + 391.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1109, + 1405, + 1109, + 1405, + 1407, + 296, + 1407 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 1684, + 1405, + 1684, + 1405, + 1883, + 297, + 1883 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 647, + 1404, + 647, + 1404, + 766, + 297, + 766 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 820, + 1404, + 820, + 1404, + 946, + 296, + 946 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 452, + 1404, + 452, + 1404, + 548, + 298, + 548 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 296, + 293, + 1404, + 293, + 1404, + 387, + 296, + 387 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1893, + 1403, + 1893, + 1403, + 1987, + 299, + 1987 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 650, + 1029, + 1050, + 1029, + 1050, + 1102, + 650, + 1102 + ], + "score": 0.954 + }, + { + "category_id": 8, + "poly": [ + 712, + 601, + 989, + 601, + 989, + 642, + 712, + 642 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 297, + 957, + 1397, + 957, + 1397, + 1024, + 297, + 1024 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 666, + 765, + 1035, + 765, + 1035, + 807, + 666, + 807 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 728, + 1481, + 970, + 1481, + 970, + 1523, + 728, + 1523 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 291, + 1418, + 1401, + 1418, + 1401, + 1481, + 291, + 1481 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 302, + 1524, + 1403, + 1524, + 1403, + 1586, + 302, + 1586 + ], + "score": 0.936 + }, + { + "category_id": 8, + "poly": [ + 659, + 395, + 1040, + 395, + 1040, + 432, + 659, + 432 + ], + "score": 0.935 + }, + { + "category_id": 8, + "poly": [ + 520, + 1642, + 1177, + 1642, + 1177, + 1680, + 520, + 1680 + ], + "score": 0.93 + }, + { + "category_id": 1, + "poly": [ + 298, + 1598, + 1018, + 1598, + 1018, + 1633, + 298, + 1633 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 299, + 226, + 665, + 226, + 665, + 261, + 299, + 261 + ], + "score": 0.924 + }, + { + "category_id": 2, + "poly": [ + 297, + 74, + 857, + 74, + 857, + 106, + 297, + 106 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 335, + 2005, + 547, + 2005, + 547, + 2033, + 335, + 2033 + ], + "score": 0.91 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1647, + 1400, + 1647, + 1400, + 1676, + 1366, + 1676 + ], + "score": 0.896 + }, + { + "category_id": 9, + "poly": [ + 1366, + 607, + 1400, + 607, + 1400, + 636, + 1366, + 636 + ], + "score": 0.889 + }, + { + "category_id": 9, + "poly": [ + 1366, + 399, + 1400, + 399, + 1400, + 429, + 1366, + 429 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 772, + 1400, + 772, + 1400, + 801, + 1366, + 801 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1051, + 1400, + 1051, + 1400, + 1080, + 1366, + 1080 + ], + "score": 0.877 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1490, + 1400, + 1490, + 1400, + 1519, + 1366, + 1519 + ], + "score": 0.858 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.669 + }, + { + "category_id": 1, + "poly": [ + 298, + 560, + 837, + 560, + 837, + 594, + 298, + 594 + ], + "score": 0.616 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.461 + }, + { + "category_id": 0, + "poly": [ + 298, + 560, + 837, + 560, + 837, + 594, + 298, + 594 + ], + "score": 0.353 + }, + { + "category_id": 13, + "poly": [ + 590, + 1108, + 675, + 1108, + 675, + 1147, + 590, + 1147 + ], + "score": 0.94, + "latex": "\\tilde { h } ^ { w _ { \\tilde { h } } } ( x )" + }, + { + "category_id": 14, + "poly": [ + 652, + 1027, + 1049, + 1027, + 1049, + 1103, + 652, + 1103 + ], + "score": 0.94, + "latex": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\left( A \\quad \\tilde { A } \\right) \\binom { h ^ { w _ { h } } ( x ) } { \\tilde { h } ^ { w _ { h } } ( x ) } + b" + }, + { + "category_id": 13, + "poly": [ + 963, + 1109, + 1162, + 1109, + 1162, + 1148, + 963, + 1148 + ], + "score": 0.94, + "latex": "\\tilde { w } _ { i } = ( w _ { i } , w _ { \\tilde { h } } , \\tilde { A } )" + }, + { + "category_id": 13, + "poly": [ + 878, + 1338, + 989, + 1338, + 989, + 1376, + 878, + 1376 + ], + "score": 0.94, + "latex": "\\tilde { h } ( x ) = x" + }, + { + "category_id": 13, + "poly": [ + 943, + 959, + 1221, + 959, + 1221, + 994, + 943, + 994 + ], + "score": 0.94, + "latex": "f _ { i } ^ { w _ { i } } ( x ) = A h ^ { w _ { h } } ( x ) + b" + }, + { + "category_id": 13, + "poly": [ + 450, + 1846, + 559, + 1846, + 559, + 1881, + 450, + 1881 + ], + "score": 0.93, + "latex": "h ( x ) = x" + }, + { + "category_id": 13, + "poly": [ + 942, + 324, + 1087, + 324, + 1087, + 358, + 942, + 358 + ], + "score": 0.93, + "latex": "f ^ { w } \\in \\mathcal { N } ( \\mathcal { X } )" + }, + { + "category_id": 14, + "poly": [ + 727, + 1478, + 972, + 1478, + 972, + 1524, + 727, + 1524 + ], + "score": 0.92, + "latex": "f _ { i } ^ { ( w _ { i } , \\tilde { w } _ { i } ) } = f _ { i } ^ { \\tilde { w } _ { i } } \\circ f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 805, + 1844, + 859, + 1844, + 859, + 1885, + 805, + 1885 + ], + "score": 0.92, + "latex": "f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }" + }, + { + "category_id": 14, + "poly": [ + 665, + 762, + 1033, + 762, + 1033, + 806, + 665, + 806 + ], + "score": 0.92, + "latex": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = C ( A f _ { i } ^ { w _ { i } } ( x ) + b ) + d ," + }, + { + "category_id": 13, + "poly": [ + 561, + 356, + 701, + 356, + 701, + 389, + 561, + 389 + ], + "score": 0.92, + "latex": "g ^ { \\tilde { w } } \\in \\mathcal { N } ( \\mathcal { X } )" + }, + { + "category_id": 13, + "poly": [ + 1022, + 1307, + 1077, + 1307, + 1077, + 1341, + 1022, + 1341 + ], + "score": 0.92, + "latex": "h ( x )" + }, + { + "category_id": 13, + "poly": [ + 514, + 514, + 596, + 514, + 596, + 549, + 514, + 549 + ], + "score": 0.92, + "latex": "f _ { i } ^ { w _ { i } } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1014, + 1806, + 1290, + 1806, + 1290, + 1847, + 1014, + 1847 + ], + "score": 0.92, + "latex": "f _ { i } ^ { w _ { i } } = f _ { i _ { n } } ^ { w _ { i _ { n } } } \\circ \\cdot \\cdot \\cdot \\circ f _ { i _ { 0 } } ^ { w _ { i _ { 0 } } }" + }, + { + "category_id": 14, + "poly": [ + 710, + 598, + 987, + 598, + 987, + 642, + 710, + 642 + ], + "score": 0.92, + "latex": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = A f _ { i } ^ { w _ { i } } ( x ) + b ," + }, + { + "category_id": 13, + "poly": [ + 711, + 1241, + 768, + 1241, + 768, + 1275, + 711, + 1275 + ], + "score": 0.92, + "latex": "h ( x )" + }, + { + "category_id": 13, + "poly": [ + 545, + 1526, + 647, + 1526, + 647, + 1554, + 545, + 1554 + ], + "score": 0.92, + "latex": "\\tilde { w } _ { i } = w _ { i }" + }, + { + "category_id": 13, + "poly": [ + 1278, + 323, + 1371, + 323, + 1371, + 353, + 1278, + 353 + ], + "score": 0.92, + "latex": "w \\in \\mathbb { R } ^ { k }" + }, + { + "category_id": 13, + "poly": [ + 337, + 1272, + 421, + 1272, + 421, + 1305, + 337, + 1305 + ], + "score": 0.91, + "latex": "\\tilde { h } = h" + }, + { + "category_id": 13, + "poly": [ + 354, + 647, + 543, + 647, + 543, + 682, + 354, + 682 + ], + "score": 0.91, + "latex": "\\tilde { w } _ { i } = ( w _ { i } , A , b ) ^ { 1 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 323, + 680, + 323, + 680, + 358, + 297, + 358 + ], + "score": 0.91, + "latex": "M : \\dot { \\mathcal { N } } ( \\dot { \\mathcal { X } } ) \\times \\mathbb { R } ^ { k } \\mathcal { N } ( \\underline { { \\mathcal { X } } } ) \\times \\mathbb { R } ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 636, + 1809, + 679, + 1809, + 679, + 1844, + 636, + 1844 + ], + "score": 0.91, + "latex": "f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 931, + 649, + 1080, + 649, + 1080, + 680, + 931, + 680 + ], + "score": 0.91, + "latex": "A = \\mathbf { 1 } , b = \\mathbf { 0 }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1341, + 612, + 1341, + 612, + 1376, + 297, + 1376 + ], + "score": 0.91, + "latex": "h ( x ) = h _ { n } ( x ) \\circ \\cdots \\circ h _ { 0 } ( x )" + }, + { + "category_id": 13, + "poly": [ + 776, + 1601, + 820, + 1601, + 820, + 1635, + 776, + 1635 + ], + "score": 0.91, + "latex": "f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 891, + 356, + 978, + 356, + 978, + 384, + 891, + 384 + ], + "score": 0.91, + "latex": "\\tilde { w } \\in \\mathbb { R } ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 1098, + 1420, + 1140, + 1420, + 1140, + 1454, + 1098, + 1454 + ], + "score": 0.91, + "latex": "f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 694, + 1146, + 775, + 1146, + 775, + 1179, + 694, + 1179 + ], + "score": 0.91, + "latex": "\\tilde { A } = 0" + }, + { + "category_id": 13, + "poly": [ + 846, + 514, + 919, + 514, + 919, + 548, + 846, + 548 + ], + "score": 0.91, + "latex": "f ^ { w } ( x )" + }, + { + "category_id": 13, + "poly": [ + 1155, + 1846, + 1212, + 1846, + 1212, + 1885, + 1155, + 1885 + ], + "score": 0.91, + "latex": "f _ { i _ { n } } ^ { w _ { i _ { n } } }" + }, + { + "category_id": 13, + "poly": [ + 1064, + 883, + 1114, + 883, + 1114, + 914, + 1064, + 914 + ], + "score": 0.9, + "latex": "C , d" + }, + { + "category_id": 13, + "poly": [ + 755, + 561, + 798, + 561, + 798, + 596, + 755, + 596 + ], + "score": 0.9, + "latex": "f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 898, + 295, + 997, + 295, + 997, + 323, + 898, + 323 + ], + "score": 0.9, + "latex": "\\mathcal { X } \\subset \\mathbb { R } ^ { n }" + }, + { + "category_id": 13, + "poly": [ + 638, + 884, + 686, + 884, + 686, + 914, + 638, + 914 + ], + "score": 0.9, + "latex": "A , b" + }, + { + "category_id": 13, + "poly": [ + 341, + 294, + 416, + 294, + 416, + 326, + 341, + 326 + ], + "score": 0.89, + "latex": "\\mathcal { N } ( \\mathcal { X } )" + }, + { + "category_id": 13, + "poly": [ + 741, + 960, + 784, + 960, + 784, + 993, + 741, + 993 + ], + "score": 0.89, + "latex": "f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 616, + 991, + 794, + 991, + 794, + 1024, + 616, + 1024 + ], + "score": 0.88, + "latex": "w _ { i } = ( w _ { h } , \\dot { A } , b )" + }, + { + "category_id": 14, + "poly": [ + 656, + 392, + 1042, + 392, + 1042, + 432, + 656, + 432 + ], + "score": 0.88, + "latex": "f ^ { w } ( x ) = g ^ { \\tilde { w } } ( x ) \\mathrm { f o r e v e r y } x \\in \\mathcal { X } ." + }, + { + "category_id": 13, + "poly": [ + 394, + 853, + 645, + 853, + 645, + 884, + 394, + 884 + ], + "score": 0.87, + "latex": "C = A ^ { - 1 } , d = - C b" + }, + { + "category_id": 13, + "poly": [ + 456, + 1376, + 489, + 1376, + 489, + 1404, + 456, + 1404 + ], + "score": 0.87, + "latex": "h _ { n }" + }, + { + "category_id": 13, + "poly": [ + 1236, + 1343, + 1268, + 1343, + 1268, + 1374, + 1236, + 1374 + ], + "score": 0.87, + "latex": "h _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 356, + 821, + 540, + 821, + 540, + 853, + 356, + 853 + ], + "score": 0.87, + "latex": "\\tilde { w } _ { i } = ( w _ { i } , C , d )" + }, + { + "category_id": 14, + "poly": [ + 522, + 1639, + 1177, + 1639, + 1177, + 1680, + 522, + 1680 + ], + "score": 0.87, + "latex": "\\tilde { f } _ { i } ^ { \\tilde { w } _ { i } } ( x ) = \\lambda f _ { i } ^ { w _ { i } } ( x ) + ( 1 - \\lambda ) h ^ { w _ { h } } ( x ) , \\quad \\tilde { w } _ { i } = ( w _ { i } , \\lambda , w _ { h } )" + }, + { + "category_id": 13, + "poly": [ + 558, + 992, + 602, + 992, + 602, + 1025, + 558, + 1025 + ], + "score": 0.85, + "latex": "f _ { i } ^ { w _ { i } }" + }, + { + "category_id": 13, + "poly": [ + 910, + 1686, + 930, + 1686, + 930, + 1712, + 910, + 1712 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 594, + 1686, + 614, + 1686, + 614, + 1713, + 594, + 1713 + ], + "score": 0.8, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 397, + 992, + 415, + 992, + 415, + 1018, + 397, + 1018 + ], + "score": 0.77, + "latex": "h" + }, + { + "category_id": 13, + "poly": [ + 558, + 822, + 607, + 822, + 607, + 853, + 558, + 853 + ], + "score": 0.7, + "latex": "A , b" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 226.0, + 668.0, + 226.0, + 668.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 109.0, + 297.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 2001.0, + 552.0, + 2001.0, + 552.0, + 2035.0, + 333.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 553.0, + 754.0, + 553.0, + 754.0, + 603.0, + 292.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 553.0, + 838.0, + 553.0, + 838.0, + 603.0, + 799.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1110.0, + 589.0, + 1110.0, + 589.0, + 1152.0, + 295.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1110.0, + 962.0, + 1110.0, + 962.0, + 1152.0, + 676.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1110.0, + 1406.0, + 1110.0, + 1406.0, + 1152.0, + 1163.0, + 1152.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1150.0, + 693.0, + 1150.0, + 693.0, + 1185.0, + 295.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 1150.0, + 1405.0, + 1150.0, + 1405.0, + 1185.0, + 776.0, + 1185.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1180.0, + 1406.0, + 1180.0, + 1406.0, + 1216.0, + 294.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1209.0, + 1405.0, + 1209.0, + 1405.0, + 1244.0, + 294.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1239.0, + 710.0, + 1239.0, + 710.0, + 1275.0, + 294.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1239.0, + 1405.0, + 1239.0, + 1405.0, + 1275.0, + 769.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1274.0, + 336.0, + 1274.0, + 336.0, + 1311.0, + 293.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 1274.0, + 1403.0, + 1274.0, + 1403.0, + 1311.0, + 422.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1305.0, + 1021.0, + 1305.0, + 1021.0, + 1344.0, + 293.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1305.0, + 1403.0, + 1305.0, + 1403.0, + 1344.0, + 1078.0, + 1344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1339.0, + 296.0, + 1339.0, + 296.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1339.0, + 877.0, + 1339.0, + 877.0, + 1380.0, + 613.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1339.0, + 1235.0, + 1339.0, + 1235.0, + 1380.0, + 990.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1339.0, + 1406.0, + 1339.0, + 1406.0, + 1380.0, + 1269.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1370.0, + 455.0, + 1370.0, + 455.0, + 1411.0, + 294.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 490.0, + 1370.0, + 502.0, + 1370.0, + 502.0, + 1411.0, + 490.0, + 1411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1685.0, + 593.0, + 1685.0, + 593.0, + 1718.0, + 296.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 1685.0, + 909.0, + 1685.0, + 909.0, + 1718.0, + 615.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1718.0, + 931.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1715.0, + 1405.0, + 1715.0, + 1405.0, + 1749.0, + 296.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1746.0, + 1406.0, + 1746.0, + 1406.0, + 1778.0, + 294.0, + 1778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1775.0, + 449.0, + 1775.0, + 449.0, + 1880.0, + 294.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1775.0, + 635.0, + 1775.0, + 635.0, + 1880.0, + 560.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1775.0, + 804.0, + 1775.0, + 804.0, + 1880.0, + 680.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1775.0, + 1013.0, + 1775.0, + 1013.0, + 1880.0, + 860.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 1775.0, + 1403.0, + 1775.0, + 1403.0, + 1880.0, + 1291.0, + 1880.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1845.0, + 449.0, + 1845.0, + 449.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 560.0, + 1845.0, + 804.0, + 1845.0, + 804.0, + 1885.0, + 560.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 1846.0, + 1154.0, + 1846.0, + 1154.0, + 1884.0, + 860.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1236.0, + 1801.0, + 1406.0, + 1801.0, + 1406.0, + 1845.0, + 1236.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 1842.0, + 859.0, + 1842.0, + 859.0, + 1866.5, + 818.0, + 1866.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1159.0, + 1837.5, + 1214.0, + 1837.5, + 1214.0, + 1869.5, + 1159.0, + 1869.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 1858.5, + 838.0, + 1858.5, + 838.0, + 1883.5, + 802.0, + 1883.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1860.0, + 1195.0, + 1860.0, + 1195.0, + 1886.0, + 1153.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 645.0, + 353.0, + 645.0, + 353.0, + 686.0, + 295.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 544.0, + 645.0, + 930.0, + 645.0, + 930.0, + 686.0, + 544.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 645.0, + 1405.0, + 645.0, + 1405.0, + 686.0, + 1081.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 676.0, + 1402.0, + 676.0, + 1402.0, + 715.0, + 294.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 706.0, + 1404.0, + 706.0, + 1404.0, + 748.0, + 292.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 736.0, + 502.0, + 736.0, + 502.0, + 775.0, + 294.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 818.0, + 355.0, + 818.0, + 355.0, + 857.0, + 294.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 541.0, + 818.0, + 557.0, + 818.0, + 557.0, + 857.0, + 541.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 818.0, + 1406.0, + 818.0, + 1406.0, + 857.0, + 608.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 848.0, + 393.0, + 848.0, + 393.0, + 888.0, + 291.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 646.0, + 848.0, + 1405.0, + 848.0, + 1405.0, + 888.0, + 646.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 879.0, + 637.0, + 879.0, + 637.0, + 917.0, + 291.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 879.0, + 1063.0, + 879.0, + 1063.0, + 917.0, + 687.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1115.0, + 879.0, + 1405.0, + 879.0, + 1405.0, + 917.0, + 1115.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 908.0, + 399.0, + 908.0, + 399.0, + 953.0, + 293.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 453.0, + 1404.0, + 453.0, + 1404.0, + 488.0, + 294.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 480.0, + 1405.0, + 480.0, + 1405.0, + 520.0, + 292.0, + 520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 511.0, + 513.0, + 511.0, + 513.0, + 553.0, + 292.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 511.0, + 845.0, + 511.0, + 845.0, + 553.0, + 597.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 511.0, + 1259.0, + 511.0, + 1259.0, + 553.0, + 920.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 285.0, + 340.0, + 285.0, + 340.0, + 334.0, + 290.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 285.0, + 897.0, + 285.0, + 897.0, + 334.0, + 417.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 285.0, + 1405.0, + 285.0, + 1405.0, + 334.0, + 998.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 319.0, + 941.0, + 319.0, + 941.0, + 359.0, + 681.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 319.0, + 1277.0, + 319.0, + 1277.0, + 359.0, + 1088.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1372.0, + 319.0, + 1405.0, + 319.0, + 1405.0, + 359.0, + 1372.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 353.0, + 560.0, + 353.0, + 560.0, + 390.0, + 294.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 702.0, + 353.0, + 890.0, + 353.0, + 890.0, + 390.0, + 702.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 353.0, + 1065.0, + 353.0, + 1065.0, + 390.0, + 979.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1893.0, + 1405.0, + 1893.0, + 1405.0, + 1930.0, + 293.0, + 1930.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1922.0, + 1404.0, + 1922.0, + 1404.0, + 1963.0, + 292.0, + 1963.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1953.0, + 432.0, + 1953.0, + 432.0, + 1992.0, + 293.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 953.0, + 740.0, + 953.0, + 740.0, + 1000.0, + 292.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 785.0, + 953.0, + 942.0, + 953.0, + 942.0, + 1000.0, + 785.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 953.0, + 1403.0, + 953.0, + 1403.0, + 1000.0, + 1222.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 984.0, + 396.0, + 984.0, + 396.0, + 1030.0, + 293.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 416.0, + 984.0, + 557.0, + 984.0, + 557.0, + 1030.0, + 416.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 984.0, + 615.0, + 984.0, + 615.0, + 1030.0, + 603.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 984.0, + 843.0, + 984.0, + 843.0, + 1030.0, + 795.0, + 1030.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1415.0, + 1097.0, + 1415.0, + 1097.0, + 1458.0, + 293.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1141.0, + 1415.0, + 1406.0, + 1415.0, + 1406.0, + 1458.0, + 1141.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1448.0, + 336.0, + 1448.0, + 336.0, + 1490.0, + 293.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1518.0, + 544.0, + 1518.0, + 544.0, + 1560.0, + 293.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 648.0, + 1518.0, + 1407.0, + 1518.0, + 1407.0, + 1560.0, + 648.0, + 1560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1552.0, + 418.0, + 1552.0, + 418.0, + 1589.0, + 291.0, + 1589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1591.0, + 775.0, + 1591.0, + 775.0, + 1644.0, + 291.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1591.0, + 1019.0, + 1591.0, + 1019.0, + 1644.0, + 821.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 553.0, + 754.0, + 553.0, + 754.0, + 603.0, + 292.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 799.0, + 553.0, + 838.0, + 553.0, + 838.0, + 603.0, + 799.0, + 603.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 980, + 1405, + 980, + 1405, + 1408, + 298, + 1408 + ], + "score": 0.98 + }, + { + "category_id": 3, + "poly": [ + 296, + 219, + 1408, + 219, + 1408, + 697, + 296, + 697 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 1424, + 1405, + 1424, + 1405, + 1516, + 299, + 1516 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 367, + 1555, + 1404, + 1555, + 1404, + 1934, + 367, + 1934 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 300, + 1972, + 1401, + 1972, + 1401, + 2034, + 300, + 2034 + ], + "score": 0.943 + }, + { + "category_id": 0, + "poly": [ + 300, + 894, + 1090, + 894, + 1090, + 930, + 300, + 930 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.764 + }, + { + "category_id": 4, + "poly": [ + 295, + 754, + 1402, + 754, + 1402, + 817, + 295, + 817 + ], + "score": 0.745 + }, + { + "category_id": 1, + "poly": [ + 295, + 754, + 1402, + 754, + 1402, + 817, + 295, + 817 + ], + "score": 0.278 + }, + { + "category_id": 13, + "poly": [ + 1249, + 1743, + 1356, + 1743, + 1356, + 1777, + 1249, + 1777 + ], + "score": 0.89, + "latex": "( \\in \\{ 2 , 4 \\} )" + }, + { + "category_id": 13, + "poly": [ + 1280, + 1586, + 1390, + 1586, + 1390, + 1620, + 1280, + 1620 + ], + "score": 0.85, + "latex": "( \\in \\{ 3 , 5 \\} )" + }, + { + "category_id": 13, + "poly": [ + 1381, + 1461, + 1401, + 1461, + 1401, + 1482, + 1381, + 1482 + ], + "score": 0.7, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 1053, + 1455, + 1178, + 1455, + 1178, + 1488, + 1053, + 1488 + ], + "score": 0.55, + "latex": "( m o d e l , n )" + }, + { + "category_id": 13, + "poly": [ + 846, + 1455, + 1180, + 1455, + 1180, + 1487, + 846, + 1487 + ], + "score": 0.31, + "latex": "A p p l y N e t M o r p h ( m o d e l , n )" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 231.0, + 654.0, + 231.0, + 654.0, + 270.0, + 564.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 234.0, + 904.0, + 234.0, + 904.0, + 268.0, + 815.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 568.0, + 258.0, + 659.0, + 258.0, + 659.0, + 285.0, + 568.0, + 285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 266.0, + 795.0, + 266.0, + 795.0, + 292.0, + 709.0, + 292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 261.0, + 911.0, + 261.0, + 911.0, + 284.0, + 819.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 408.0, + 427.0, + 408.0, + 427.0, + 457.0, + 304.0, + 457.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 400.0, + 564.0, + 400.0, + 564.0, + 430.0, + 430.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.0, + 387.0, + 661.0, + 387.0, + 661.0, + 440.0, + 565.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 401.0, + 795.0, + 401.0, + 795.0, + 431.0, + 709.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 387.0, + 909.0, + 387.0, + 909.0, + 440.0, + 817.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 376.0, + 1015.0, + 376.0, + 1015.0, + 553.0, + 979.0, + 553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 403.0, + 1206.0, + 403.0, + 1206.0, + 451.0, + 1088.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 395.0, + 1351.0, + 395.0, + 1351.0, + 423.0, + 1218.0, + 423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1085.0, + 435.0, + 1181.0, + 435.0, + 1181.0, + 465.0, + 1085.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 447.0, + 403.0, + 447.0, + 403.0, + 477.0, + 306.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 459.0, + 1400.0, + 459.0, + 1400.0, + 476.0, + 1362.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 495.0, + 365.0, + 495.0, + 365.0, + 512.0, + 348.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 480.0, + 506.0, + 480.0, + 506.0, + 517.0, + 488.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1276.0, + 475.0, + 1293.0, + 475.0, + 1293.0, + 511.0, + 1276.0, + 511.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 513.0, + 626.0, + 513.0, + 626.0, + 550.0, + 608.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 512.0, + 754.0, + 512.0, + 754.0, + 551.0, + 736.0, + 551.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 511.0, + 881.0, + 511.0, + 881.0, + 550.0, + 865.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 590.0, + 538.0, + 590.0, + 538.0, + 599.0, + 528.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 597.0, + 661.0, + 597.0, + 661.0, + 627.0, + 569.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 596.0, + 795.0, + 596.0, + 795.0, + 622.0, + 709.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 818.0, + 599.0, + 911.0, + 599.0, + 911.0, + 625.0, + 818.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 610.0, + 635.0, + 634.0, + 635.0, + 634.0, + 657.0, + 610.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 334.0, + 542.0, + 334.0, + 542.0, + 369.5, + 401.0, + 369.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 329.0, + 1328.0, + 329.0, + 1328.0, + 360.5, + 1192.0, + 360.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 566.25, + 565.5, + 693.25, + 565.5, + 693.25, + 607.0, + 566.25, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.75, + 565.0, + 945.75, + 565.0, + 945.75, + 607.5, + 817.75, + 607.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 893.0, + 1094.0, + 893.0, + 1094.0, + 933.0, + 294.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 754.0, + 1404.0, + 754.0, + 1404.0, + 789.0, + 295.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 784.0, + 916.0, + 784.0, + 916.0, + 816.0, + 297.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 981.0, + 1407.0, + 981.0, + 1407.0, + 1016.0, + 296.0, + 1016.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1011.0, + 1406.0, + 1011.0, + 1406.0, + 1045.0, + 293.0, + 1045.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1042.0, + 1406.0, + 1042.0, + 1406.0, + 1078.0, + 292.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1072.0, + 1405.0, + 1072.0, + 1405.0, + 1106.0, + 293.0, + 1106.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1105.0, + 1403.0, + 1105.0, + 1403.0, + 1137.0, + 295.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1136.0, + 1403.0, + 1136.0, + 1403.0, + 1167.0, + 296.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1162.0, + 1407.0, + 1162.0, + 1407.0, + 1199.0, + 293.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1193.0, + 1407.0, + 1193.0, + 1407.0, + 1230.0, + 293.0, + 1230.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1224.0, + 1407.0, + 1224.0, + 1407.0, + 1260.0, + 295.0, + 1260.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1256.0, + 1406.0, + 1256.0, + 1406.0, + 1292.0, + 295.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1283.0, + 1405.0, + 1283.0, + 1405.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1317.0, + 1405.0, + 1317.0, + 1405.0, + 1353.0, + 296.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1346.0, + 1406.0, + 1346.0, + 1406.0, + 1382.0, + 295.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1378.0, + 603.0, + 1378.0, + 603.0, + 1412.0, + 293.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1423.0, + 1405.0, + 1423.0, + 1405.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1456.0, + 845.0, + 1456.0, + 845.0, + 1489.0, + 295.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1456.0, + 1380.0, + 1456.0, + 1380.0, + 1489.0, + 1181.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1456.0, + 1405.0, + 1456.0, + 1405.0, + 1489.0, + 1402.0, + 1489.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1486.0, + 1302.0, + 1486.0, + 1302.0, + 1519.0, + 295.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1555.0, + 1404.0, + 1555.0, + 1404.0, + 1587.0, + 370.0, + 1587.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1584.0, + 1279.0, + 1584.0, + 1279.0, + 1621.0, + 393.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1391.0, + 1584.0, + 1404.0, + 1584.0, + 1404.0, + 1621.0, + 1391.0, + 1621.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1618.0, + 1404.0, + 1618.0, + 1404.0, + 1650.0, + 395.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1645.0, + 912.0, + 1645.0, + 912.0, + 1680.0, + 396.0, + 1680.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1710.0, + 1403.0, + 1710.0, + 1403.0, + 1749.0, + 370.0, + 1749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1743.0, + 1248.0, + 1743.0, + 1248.0, + 1779.0, + 394.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1357.0, + 1743.0, + 1405.0, + 1743.0, + 1405.0, + 1779.0, + 1357.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1776.0, + 737.0, + 1776.0, + 737.0, + 1808.0, + 394.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1838.0, + 1404.0, + 1838.0, + 1404.0, + 1875.0, + 370.0, + 1875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1870.0, + 1404.0, + 1870.0, + 1404.0, + 1907.0, + 394.0, + 1907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1899.0, + 671.0, + 1899.0, + 671.0, + 1938.0, + 393.0, + 1938.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 1283.0, + 2000.0, + 1283.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 754.0, + 1404.0, + 754.0, + 1404.0, + 789.0, + 295.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 784.0, + 916.0, + 784.0, + 916.0, + 816.0, + 297.0, + 816.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1369, + 1403, + 1369, + 1403, + 1522, + 299, + 1522 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1170, + 1403, + 1170, + 1403, + 1354, + 298, + 1354 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1649, + 1403, + 1649, + 1403, + 1772, + 299, + 1772 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1788, + 1403, + 1788, + 1403, + 1911, + 299, + 1911 + ], + "score": 0.969 + }, + { + "category_id": 2, + "poly": [ + 298, + 1949, + 1403, + 1949, + 1403, + 2035, + 298, + 2035 + ], + "score": 0.954 + }, + { + "category_id": 0, + "poly": [ + 300, + 1575, + 557, + 1575, + 557, + 1610, + 300, + 1610 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 299, + 268, + 1388, + 268, + 1388, + 1099, + 299, + 1099 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.823 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.751 + }, + { + "category_id": 0, + "poly": [ + 296, + 228, + 949, + 228, + 949, + 261, + 296, + 261 + ], + "score": 0.27 + }, + { + "category_id": 6, + "poly": [ + 296, + 228, + 949, + 228, + 949, + 261, + 296, + 261 + ], + "score": 0.231 + }, + { + "category_id": 5, + "poly": [ + 299, + 268, + 1388, + 268, + 1388, + 1099, + 299, + 1099 + ], + "score": 0.129, + "html": "
1: function NASH( modelo,nsteps, nneigh, nNM, epochneigh, epochfinal, Xend, Xstart )
2:
3:# modelo = model to start with， nsteps 叁 number of hill climbining steps
4:# nneigh ≌ number of neighbours, nNM ≌ number of net. morph. applied
5:# epochneigh = number of epochs for training every neighbour
6:# epoch final ≌ number of epochs for final training
7:# initial LR Xstart is annealed to Xend during SGDR training
8:
9:modelbest ←modelo
10:# start hill climbing
11:for i ← 1,...,nsteps do
12:#get nneigh neighbors of modelo by applying nNM network morphisms to modelbest
13:for j ← 1,...,nneigh-1 do
14:modelj ← ApplyNetMorphs(modelbest,nNM)
15:# train for a few epochs on training set with SGDR
16:modelj ← SGDRtrain(modelj,epochneigh, Xstart, Xend)
17:end for
18:# in fact, last neighbor is always just the current best
19:modelnneigh ← SGDRtrain(modelbest, epochneigh, Xstart, Xend)
20:# get best model on validation set
21:modelbest ← arg max {performancevali(modelj)} j=1,..,.nneigh
22:end for
23:# train the final model on training and validation set
24:modelbest ← SGDRtrain(modelbest,epoch final, Xstart, Xend)
25:return modelbest
26: end function
" + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 104, + 298, + 104 + ], + "score": 0.117 + }, + { + "category_id": 13, + "poly": [ + 1128, + 1262, + 1198, + 1262, + 1198, + 1293, + 1128, + 1293 + ], + "score": 0.92, + "latex": "\\lambda _ { s t a r t }" + }, + { + "category_id": 13, + "poly": [ + 325, + 1293, + 382, + 1293, + 382, + 1324, + 325, + 1324 + ], + "score": 0.91, + "latex": "\\lambda _ { e n d }" + }, + { + "category_id": 13, + "poly": [ + 521, + 468, + 591, + 468, + 591, + 498, + 521, + 498 + ], + "score": 0.91, + "latex": "\\lambda _ { s t a r t }" + }, + { + "category_id": 13, + "poly": [ + 754, + 467, + 811, + 467, + 811, + 498, + 754, + 498 + ], + "score": 0.91, + "latex": "\\lambda _ { e n d }" + }, + { + "category_id": 13, + "poly": [ + 642, + 2006, + 819, + 2006, + 819, + 2036, + 642, + 2036 + ], + "score": 0.89, + "latex": "e p o c h _ { n e i g h } = 1 7" + }, + { + "category_id": 13, + "poly": [ + 1103, + 1405, + 1180, + 1405, + 1180, + 1434, + 1103, + 1434 + ], + "score": 0.88, + "latex": "n _ { n e i g h }" + }, + { + "category_id": 13, + "poly": [ + 941, + 625, + 1003, + 625, + 1003, + 649, + 941, + 649 + ], + "score": 0.85, + "latex": "n _ { N M }" + }, + { + "category_id": 13, + "poly": [ + 775, + 324, + 881, + 324, + 881, + 362, + 775, + 362 + ], + "score": 0.85, + "latex": "n _ { s t e p s } \\triangleq" + }, + { + "category_id": 13, + "poly": [ + 475, + 652, + 732, + 652, + 732, + 682, + 475, + 682 + ], + "score": 0.84, + "latex": "j \\gets 1 , \\dots , n _ { n e i g h } - 1" + }, + { + "category_id": 13, + "poly": [ + 434, + 590, + 652, + 590, + 652, + 621, + 434, + 621 + ], + "score": 0.83, + "latex": "i \\gets 1 , \\dots , n _ { s t e p s } \\ { \\bf d }" + }, + { + "category_id": 13, + "poly": [ + 801, + 362, + 898, + 362, + 898, + 396, + 801, + 396 + ], + "score": 0.82, + "latex": "n _ { N M } \\triangleq" + }, + { + "category_id": 13, + "poly": [ + 852, + 1979, + 1397, + 1979, + 1397, + 2009, + 852, + 2009 + ], + "score": 0.8, + "latex": "e p o c h _ { t o t a l } = e p o c h _ { n e i g h } n _ { n e i g h } n _ { s t e p s } + e p o c h _ { f i n a l } ." + }, + { + "category_id": 13, + "poly": [ + 406, + 362, + 516, + 362, + 516, + 397, + 406, + 397 + ], + "score": 0.77, + "latex": "n _ { n e i g h } \\triangleq" + }, + { + "category_id": 13, + "poly": [ + 389, + 1950, + 508, + 1950, + 508, + 1980, + 389, + 1980 + ], + "score": 0.77, + "latex": "e p o c h _ { n e i g h }" + }, + { + "category_id": 13, + "poly": [ + 401, + 434, + 563, + 434, + 563, + 470, + 401, + 470 + ], + "score": 0.72, + "latex": "e p o c h _ { f i n a l } \\triangleq" + }, + { + "category_id": 14, + "poly": [ + 468, + 740, + 1119, + 740, + 1119, + 775, + 468, + 775 + ], + "score": 0.7, + "latex": "m o d e l _ { j } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { j } , e \\bar { p } o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )" + }, + { + "category_id": 13, + "poly": [ + 482, + 624, + 560, + 624, + 560, + 650, + 482, + 650 + ], + "score": 0.7, + "latex": "n _ { n e i g h }" + }, + { + "category_id": 13, + "poly": [ + 710, + 620, + 797, + 620, + 797, + 650, + 710, + 650 + ], + "score": 0.68, + "latex": "m o d e l _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 466, + 681, + 1028, + 681, + 1028, + 714, + 466, + 714 + ], + "score": 0.65, + "latex": "m o d e l _ { j } \\gets A p p l y N e t M o r p h s ( m o d e l _ { b e s t } , n _ { N M } )" + }, + { + "category_id": 14, + "poly": [ + 385, + 527, + 633, + 527, + 633, + 559, + 385, + 559 + ], + "score": 0.64, + "latex": "m o d e l _ { b e s t } m o d e l _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 401, + 398, + 568, + 398, + 568, + 434, + 401, + 434 + ], + "score": 0.64, + "latex": "e p o c h _ { n e i g h } \\triangleq" + }, + { + "category_id": 13, + "poly": [ + 426, + 832, + 1160, + 832, + 1160, + 867, + 426, + 867 + ], + "score": 0.62, + "latex": "m o d e l _ { n _ { n e i g h } } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { b e s t } , e p o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )" + }, + { + "category_id": 14, + "poly": [ + 425, + 890, + 1064, + 890, + 1064, + 948, + 425, + 948 + ], + "score": 0.6, + "latex": "\\underbrace { \\Bre { \\_ } { a } } { m o d } e l _ { b e s t } \\underbrace { a r g m a x } _ { j = 1 , \\dots , n _ { n e i g h } } \\{ p e r f o r m a n c e _ { v a l i } ( m o d e l _ { j } ) \\}" + }, + { + "category_id": 13, + "poly": [ + 1264, + 619, + 1380, + 619, + 1380, + 650, + 1264, + 650 + ], + "score": 0.5, + "latex": "m o d e l _ { b e s t }" + }, + { + "category_id": 13, + "poly": [ + 385, + 1006, + 1087, + 1006, + 1087, + 1039, + 385, + 1039 + ], + "score": 0.41, + "latex": "m o d e l _ { b e s t } \\gets \\mathrm { S G D R t r a i n } ( m o d e \\bar { l } _ { b e s t } , e p o c h _ { f i n a l } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )" + }, + { + "category_id": 14, + "poly": [ + 425, + 832, + 1161, + 832, + 1161, + 867, + 425, + 867 + ], + "score": 0.41, + "latex": "m o d e l _ { n _ { n e i g h } } \\gets \\mathrm { S G D R t r a i n } ( m o d e l _ { b e s t } , e p o c h _ { n e i g h } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )" + }, + { + "category_id": 14, + "poly": [ + 298, + 267, + 1327, + 267, + 1327, + 303, + 298, + 303 + ], + "score": 0.41, + "latex": "1 \\backslash \\mathrm { ~ f u n c t i o n ~ N A S H } ( m o d e l _ { 0 } , n _ { s t e p s } , n _ { n e i g h } , n _ { N M } , e p o c h _ { n e i g h } , e p o c h _ { f i n a l } , \\lambda _ { e n d } , \\lambda _ { s t a r t } )" + }, + { + "category_id": 13, + "poly": [ + 406, + 325, + 496, + 325, + 496, + 359, + 406, + 359 + ], + "score": 0.35, + "latex": "m o d e l _ { 0 }" + }, + { + "category_id": 14, + "poly": [ + 385, + 1006, + 1087, + 1006, + 1087, + 1040, + 385, + 1040 + ], + "score": 0.31, + "latex": "m o d e l _ { b e s t } \\gets \\mathrm { S G D R t r a i n } ( m o d e \\bar { l } _ { b e s t } , e p o c h _ { f i n a l } , \\lambda _ { s t a r t } , \\lambda _ { e n d } )" + }, + { + "category_id": 13, + "poly": [ + 583, + 923, + 722, + 923, + 722, + 948, + 583, + 948 + ], + "score": 0.28, + "latex": "j { = } 1 , { \\ldots } , n _ { n e i g h }" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 1942.0, + 388.0, + 1942.0, + 388.0, + 1985.0, + 327.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1942.0, + 1409.0, + 1942.0, + 1409.0, + 1985.0, + 509.0, + 1985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1976.0, + 851.0, + 1976.0, + 851.0, + 2013.0, + 292.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 1976.0, + 1407.0, + 1976.0, + 1407.0, + 2013.0, + 1398.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2005.0, + 641.0, + 2005.0, + 641.0, + 2037.0, + 294.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 2005.0, + 829.0, + 2005.0, + 829.0, + 2037.0, + 820.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1572.0, + 561.0, + 1572.0, + 561.0, + 1616.0, + 292.0, + 1616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 218.0, + 953.0, + 218.0, + 953.0, + 274.0, + 293.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 218.0, + 953.0, + 218.0, + 953.0, + 274.0, + 293.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1368.0, + 1405.0, + 1368.0, + 1405.0, + 1403.0, + 295.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1397.0, + 1102.0, + 1397.0, + 1102.0, + 1439.0, + 292.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1397.0, + 1404.0, + 1397.0, + 1404.0, + 1439.0, + 1181.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1427.0, + 1404.0, + 1427.0, + 1404.0, + 1468.0, + 293.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1459.0, + 1407.0, + 1459.0, + 1407.0, + 1497.0, + 293.0, + 1497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1492.0, + 931.0, + 1492.0, + 931.0, + 1526.0, + 294.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1171.0, + 1404.0, + 1171.0, + 1404.0, + 1202.0, + 294.0, + 1202.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1200.0, + 1405.0, + 1200.0, + 1405.0, + 1235.0, + 296.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1229.0, + 1405.0, + 1229.0, + 1405.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1261.0, + 1127.0, + 1261.0, + 1127.0, + 1296.0, + 294.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1199.0, + 1261.0, + 1405.0, + 1261.0, + 1405.0, + 1296.0, + 1199.0, + 1296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1290.0, + 324.0, + 1290.0, + 324.0, + 1329.0, + 293.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1290.0, + 1407.0, + 1290.0, + 1407.0, + 1329.0, + 383.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1322.0, + 700.0, + 1322.0, + 700.0, + 1357.0, + 294.0, + 1357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1650.0, + 1403.0, + 1650.0, + 1403.0, + 1683.0, + 297.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1681.0, + 1405.0, + 1681.0, + 1405.0, + 1714.0, + 295.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1405.0, + 1712.0, + 1405.0, + 1745.0, + 294.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1742.0, + 1378.0, + 1742.0, + 1378.0, + 1774.0, + 294.0, + 1774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1787.0, + 1403.0, + 1787.0, + 1403.0, + 1822.0, + 295.0, + 1822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1817.0, + 1404.0, + 1817.0, + 1404.0, + 1854.0, + 294.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1851.0, + 1404.0, + 1851.0, + 1404.0, + 1883.0, + 295.0, + 1883.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1403.0, + 1881.0, + 1403.0, + 1913.0, + 294.0, + 1913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 296.0, + 339.0, + 296.0, + 339.0, + 329.0, + 307.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 326.0, + 341.0, + 326.0, + 341.0, + 362.0, + 306.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 318.0, + 405.0, + 318.0, + 405.0, + 370.0, + 381.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 318.0, + 774.0, + 318.0, + 774.0, + 370.0, + 497.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 318.0, + 1241.0, + 318.0, + 1241.0, + 370.0, + 882.0, + 370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 363.0, + 338.0, + 363.0, + 338.0, + 398.0, + 307.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 362.0, + 405.0, + 362.0, + 405.0, + 401.0, + 380.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 362.0, + 800.0, + 362.0, + 800.0, + 401.0, + 517.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 362.0, + 1249.0, + 362.0, + 1249.0, + 401.0, + 899.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 400.0, + 338.0, + 400.0, + 338.0, + 434.0, + 307.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 396.0, + 400.0, + 396.0, + 400.0, + 438.0, + 380.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 569.0, + 396.0, + 1095.0, + 396.0, + 1095.0, + 438.0, + 569.0, + 438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 437.0, + 338.0, + 437.0, + 338.0, + 470.0, + 307.0, + 470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 430.0, + 400.0, + 430.0, + 400.0, + 476.0, + 378.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 430.0, + 965.0, + 430.0, + 965.0, + 476.0, + 564.0, + 476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 307.0, + 467.0, + 338.0, + 467.0, + 338.0, + 500.0, + 307.0, + 500.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 463.0, + 520.0, + 463.0, + 520.0, + 503.0, + 380.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 592.0, + 463.0, + 753.0, + 463.0, + 753.0, + 503.0, + 592.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 463.0, + 1066.0, + 463.0, + 1066.0, + 503.0, + 812.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 498.0, + 339.0, + 498.0, + 339.0, + 529.0, + 306.0, + 529.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 528.0, + 338.0, + 528.0, + 338.0, + 560.0, + 306.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 557.0, + 339.0, + 557.0, + 339.0, + 590.0, + 297.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 552.0, + 611.0, + 552.0, + 611.0, + 594.0, + 380.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 589.0, + 339.0, + 589.0, + 339.0, + 622.0, + 296.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 583.0, + 433.0, + 583.0, + 433.0, + 627.0, + 381.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 583.0, + 678.0, + 583.0, + 678.0, + 627.0, + 653.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 618.0, + 339.0, + 618.0, + 339.0, + 652.0, + 296.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 614.0, + 481.0, + 614.0, + 481.0, + 658.0, + 422.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 614.0, + 709.0, + 614.0, + 709.0, + 658.0, + 561.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 798.0, + 614.0, + 940.0, + 614.0, + 940.0, + 658.0, + 798.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 614.0, + 1263.0, + 614.0, + 1263.0, + 658.0, + 1004.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1381.0, + 614.0, + 1385.0, + 614.0, + 1385.0, + 658.0, + 1381.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 649.0, + 339.0, + 649.0, + 339.0, + 682.0, + 296.0, + 682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 647.0, + 474.0, + 647.0, + 474.0, + 687.0, + 424.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 647.0, + 782.0, + 647.0, + 782.0, + 687.0, + 733.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 681.0, + 338.0, + 681.0, + 338.0, + 710.0, + 297.0, + 710.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 676.0, + 1032.0, + 676.0, + 1032.0, + 715.0, + 1029.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 710.0, + 339.0, + 710.0, + 339.0, + 743.0, + 297.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 709.0, + 1039.0, + 709.0, + 1039.0, + 743.0, + 463.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 740.0, + 339.0, + 740.0, + 339.0, + 773.0, + 296.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 771.0, + 339.0, + 771.0, + 339.0, + 804.0, + 296.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 770.0, + 521.0, + 770.0, + 521.0, + 802.0, + 426.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 801.0, + 339.0, + 801.0, + 339.0, + 834.0, + 296.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 422.0, + 800.0, + 1017.0, + 800.0, + 1017.0, + 835.0, + 422.0, + 835.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 831.0, + 339.0, + 831.0, + 339.0, + 864.0, + 296.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 825.0, + 424.0, + 825.0, + 424.0, + 875.0, + 419.0, + 875.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 862.0, + 339.0, + 862.0, + 339.0, + 895.0, + 295.0, + 895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 861.0, + 808.0, + 861.0, + 808.0, + 896.0, + 424.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 893.0, + 339.0, + 893.0, + 339.0, + 928.0, + 295.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 944.0, + 338.0, + 944.0, + 338.0, + 978.0, + 295.0, + 978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 943.0, + 478.0, + 943.0, + 478.0, + 977.0, + 383.0, + 977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 976.0, + 339.0, + 976.0, + 339.0, + 1010.0, + 295.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 974.0, + 969.0, + 974.0, + 969.0, + 1010.0, + 382.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1006.0, + 338.0, + 1006.0, + 338.0, + 1040.0, + 295.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1036.0, + 338.0, + 1036.0, + 338.0, + 1071.0, + 295.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 1032.0, + 596.0, + 1032.0, + 596.0, + 1075.0, + 382.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1069.0, + 338.0, + 1069.0, + 338.0, + 1097.0, + 297.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 340.0, + 1064.0, + 502.0, + 1064.0, + 502.0, + 1100.0, + 340.0, + 1100.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 533, + 1404, + 533, + 1404, + 840, + 297, + 840 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1359, + 1405, + 1359, + 1405, + 1878, + 298, + 1878 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 854, + 1404, + 854, + 1404, + 1101, + 297, + 1101 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1115, + 1403, + 1115, + 1403, + 1238, + 299, + 1238 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 300, + 1253, + 1404, + 1253, + 1404, + 1345, + 300, + 1345 + ], + "score": 0.971 + }, + { + "category_id": 2, + "poly": [ + 296, + 1918, + 1402, + 1918, + 1402, + 2033, + 296, + 2033 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 299, + 456, + 1399, + 456, + 1399, + 519, + 299, + 519 + ], + "score": 0.95 + }, + { + "category_id": 1, + "poly": [ + 292, + 229, + 1404, + 229, + 1404, + 293, + 292, + 293 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 299, + 397, + 527, + 397, + 527, + 431, + 299, + 431 + ], + "score": 0.906 + }, + { + "category_id": 0, + "poly": [ + 301, + 338, + 714, + 338, + 714, + 370, + 301, + 370 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.813 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.791 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.133 + }, + { + "category_id": 13, + "poly": [ + 1084, + 1725, + 1211, + 1725, + 1211, + 1756, + 1084, + 1756 + ], + "score": 0.91, + "latex": "R ^ { 2 } = 0 . 6 4" + }, + { + "category_id": 13, + "poly": [ + 378, + 810, + 506, + 810, + 506, + 843, + 378, + 843 + ], + "score": 0.9, + "latex": "n _ { n e i g h } = 8" + }, + { + "category_id": 13, + "poly": [ + 1144, + 566, + 1279, + 566, + 1279, + 598, + 1144, + 598 + ], + "score": 0.9, + "latex": "n _ { n e i g h } = 1" + }, + { + "category_id": 13, + "poly": [ + 477, + 1756, + 607, + 1756, + 607, + 1787, + 477, + 1787 + ], + "score": 0.9, + "latex": "\\mathrm { \\bar { \\it R } ^ { 2 } = 0 . 3 6 }" + }, + { + "category_id": 13, + "poly": [ + 475, + 627, + 614, + 627, + 614, + 658, + 475, + 658 + ], + "score": 0.9, + "latex": "n _ { n e i g h } = 8" + }, + { + "category_id": 13, + "poly": [ + 676, + 747, + 731, + 747, + 731, + 777, + 676, + 777 + ], + "score": 0.89, + "latex": "9 1 \\%" + }, + { + "category_id": 13, + "poly": [ + 295, + 747, + 381, + 747, + 381, + 777, + 295, + 777 + ], + "score": 0.89, + "latex": "\\approx 7 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 1206, + 1978, + 1405, + 1978, + 1405, + 2005, + 1206, + 2005 + ], + "score": 0.89, + "latex": "2 0 + 1 7 \\cdot 5 + 1 0 0 =" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1392, + 1091, + 1392, + 1091, + 1423, + 1028, + 1423 + ], + "score": 0.88, + "latex": "5 . 7 \\%" + }, + { + "category_id": 13, + "poly": [ + 1100, + 1635, + 1162, + 1635, + 1162, + 1666, + 1100, + 1666 + ], + "score": 0.87, + "latex": "6 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 297, + 1176, + 571, + 1176, + 571, + 1210, + 297, + 1210 + ], + "score": 0.87, + "latex": "\\lambda \\in \\{ 0 . 0 1 , 0 . 0 2 5 , 0 . 0 5 \\}" + }, + { + "category_id": 13, + "poly": [ + 982, + 1635, + 1049, + 1635, + 1049, + 1666, + 982, + 1666 + ], + "score": 0.86, + "latex": "( 6 . 4 \\%" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1392, + 1201, + 1392, + 1201, + 1423, + 1138, + 1423 + ], + "score": 0.84, + "latex": "6 . 5 \\%" + }, + { + "category_id": 13, + "poly": [ + 466, + 688, + 597, + 688, + 597, + 718, + 466, + 718 + ], + "score": 0.82, + "latex": "\\lambda _ { e n d } = 0 . { \\overset { \\cdot } { 0 } }" + }, + { + "category_id": 13, + "poly": [ + 297, + 687, + 456, + 687, + 456, + 717, + 297, + 717 + ], + "score": 0.79, + "latex": "\\lambda _ { s t a r t } = 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 506, + 1950, + 808, + 1950, + 808, + 1980, + 506, + 1980 + ], + "score": 0.76, + "latex": "n _ { s t e p s } = 5 , e p o c h _ { n e i g h } = 1 7" + }, + { + "category_id": 13, + "poly": [ + 820, + 1951, + 1006, + 1951, + 1006, + 1980, + 820, + 1980 + ], + "score": 0.54, + "latex": "e p o c h _ { f i n a l } = 1 0 0" + }, + { + "category_id": 13, + "poly": [ + 531, + 656, + 1251, + 656, + 1251, + 690, + 531, + 690 + ], + "score": 0.48, + "latex": "n _ { s t e p s } = 5 , n _ { N M } = 5 , e p o c h _ { n e i g h } = 1 7 , e p o c h _ { f i n a l } = 1 0 0" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1912.0, + 863.0, + 1912.0, + 863.0, + 1954.0, + 329.0, + 1954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1936.0, + 505.0, + 1936.0, + 505.0, + 1993.0, + 325.0, + 1993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1936.0, + 819.0, + 1936.0, + 819.0, + 1993.0, + 809.0, + 1993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1936.0, + 1409.0, + 1936.0, + 1409.0, + 1993.0, + 1007.0, + 1993.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1970.0, + 1205.0, + 1970.0, + 1205.0, + 2013.0, + 293.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1998.0, + 426.0, + 1998.0, + 426.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 395.0, + 529.0, + 395.0, + 529.0, + 434.0, + 294.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 338.0, + 717.0, + 338.0, + 717.0, + 374.0, + 294.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 531.0, + 1405.0, + 531.0, + 1405.0, + 570.0, + 294.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 561.0, + 1143.0, + 561.0, + 1143.0, + 603.0, + 292.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 561.0, + 1405.0, + 561.0, + 1405.0, + 603.0, + 1280.0, + 603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 595.0, + 1406.0, + 595.0, + 1406.0, + 631.0, + 294.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 623.0, + 474.0, + 623.0, + 474.0, + 662.0, + 291.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 615.0, + 623.0, + 1406.0, + 623.0, + 1406.0, + 662.0, + 615.0, + 662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 652.0, + 530.0, + 652.0, + 530.0, + 694.0, + 292.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1252.0, + 652.0, + 1407.0, + 652.0, + 1407.0, + 694.0, + 1252.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 684.0, + 296.0, + 684.0, + 296.0, + 722.0, + 291.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 457.0, + 684.0, + 465.0, + 684.0, + 465.0, + 722.0, + 457.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 684.0, + 1405.0, + 684.0, + 1405.0, + 722.0, + 598.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 709.0, + 1406.0, + 709.0, + 1406.0, + 755.0, + 292.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 382.0, + 744.0, + 675.0, + 744.0, + 675.0, + 783.0, + 382.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 744.0, + 1404.0, + 744.0, + 1404.0, + 783.0, + 732.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 777.0, + 1405.0, + 777.0, + 1405.0, + 812.0, + 294.0, + 812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 807.0, + 377.0, + 807.0, + 377.0, + 845.0, + 291.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 507.0, + 807.0, + 520.0, + 807.0, + 520.0, + 845.0, + 507.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 1403.0, + 1359.0, + 1403.0, + 1397.0, + 293.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1394.0, + 1027.0, + 1394.0, + 1027.0, + 1424.0, + 294.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 1394.0, + 1137.0, + 1394.0, + 1137.0, + 1424.0, + 1092.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1394.0, + 1403.0, + 1394.0, + 1403.0, + 1424.0, + 1202.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1417.0, + 1407.0, + 1417.0, + 1407.0, + 1461.0, + 292.0, + 1461.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1455.0, + 1405.0, + 1455.0, + 1405.0, + 1487.0, + 296.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1482.0, + 1406.0, + 1482.0, + 1406.0, + 1518.0, + 293.0, + 1518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1514.0, + 1407.0, + 1514.0, + 1407.0, + 1546.0, + 293.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1579.0, + 296.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1575.0, + 1405.0, + 1575.0, + 1405.0, + 1609.0, + 293.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1604.0, + 1405.0, + 1604.0, + 1405.0, + 1642.0, + 292.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1635.0, + 981.0, + 1635.0, + 981.0, + 1671.0, + 293.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1635.0, + 1099.0, + 1635.0, + 1099.0, + 1671.0, + 1050.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1163.0, + 1635.0, + 1405.0, + 1635.0, + 1405.0, + 1671.0, + 1163.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1664.0, + 1403.0, + 1664.0, + 1403.0, + 1700.0, + 293.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1696.0, + 1403.0, + 1696.0, + 1403.0, + 1732.0, + 295.0, + 1732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1722.0, + 1083.0, + 1722.0, + 1083.0, + 1764.0, + 291.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1722.0, + 1406.0, + 1722.0, + 1406.0, + 1764.0, + 1212.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1753.0, + 476.0, + 1753.0, + 476.0, + 1795.0, + 292.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 1753.0, + 1406.0, + 1753.0, + 1406.0, + 1795.0, + 608.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1789.0, + 1403.0, + 1789.0, + 1403.0, + 1821.0, + 296.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1816.0, + 1405.0, + 1816.0, + 1405.0, + 1856.0, + 293.0, + 1856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1847.0, + 1405.0, + 1847.0, + 1405.0, + 1884.0, + 292.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 852.0, + 1405.0, + 852.0, + 1405.0, + 890.0, + 292.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 885.0, + 1404.0, + 885.0, + 1404.0, + 918.0, + 294.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 914.0, + 1405.0, + 914.0, + 1405.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 947.0, + 1406.0, + 947.0, + 1406.0, + 981.0, + 295.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 979.0, + 1404.0, + 979.0, + 1404.0, + 1012.0, + 294.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1008.0, + 1404.0, + 1008.0, + 1404.0, + 1041.0, + 295.0, + 1041.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1038.0, + 1404.0, + 1038.0, + 1404.0, + 1071.0, + 295.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 879.0, + 1069.0, + 879.0, + 1102.0, + 293.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1113.0, + 1407.0, + 1113.0, + 1407.0, + 1151.0, + 293.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1179.0, + 294.0, + 1179.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1174.0, + 296.0, + 1174.0, + 296.0, + 1211.0, + 293.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 1174.0, + 1407.0, + 1174.0, + 1407.0, + 1211.0, + 572.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1203.0, + 605.0, + 1203.0, + 605.0, + 1244.0, + 293.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1253.0, + 1404.0, + 1253.0, + 1404.0, + 1287.0, + 295.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 1404.0, + 1285.0, + 1404.0, + 1318.0, + 294.0, + 1318.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1312.0, + 1399.0, + 1312.0, + 1399.0, + 1352.0, + 294.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 454.0, + 1403.0, + 454.0, + 1403.0, + 490.0, + 295.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 488.0, + 1083.0, + 488.0, + 1083.0, + 522.0, + 294.0, + 522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1406.0, + 226.0, + 1406.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 258.0, + 404.0, + 258.0, + 404.0, + 296.0, + 293.0, + 296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1562, + 1403, + 1562, + 1403, + 1804, + 298, + 1804 + ], + "score": 0.977 + }, + { + "category_id": 3, + "poly": [ + 399, + 217, + 1289, + 217, + 1289, + 676, + 399, + 676 + ], + "score": 0.971 + }, + { + "category_id": 5, + "poly": [ + 296, + 961, + 1429, + 961, + 1429, + 1156, + 296, + 1156 + ], + "score": 0.969, + "html": "
algorithm settingruntime (hrs)# params (mil.)error ± std.( %)
nneigh =812.85.75.7 ± 0.35
Random networks (nneigh = 1)4.54.46.5 ± 0.76
models from line 1 retrained from scratch5.35.76.1 ± 0.92
nneigh = 8, no SGDR10.65.86.4±0.70
nneigh = 8, no net. morph.6.62.96.1 ±0.30
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1437, + 1402, + 1437, + 1402, + 1559, + 299, + 1559 + ], + "score": 0.963 + }, + { + "category_id": 4, + "poly": [ + 296, + 706, + 1403, + 706, + 1403, + 800, + 296, + 800 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 297, + 1237, + 1407, + 1237, + 1407, + 1300, + 297, + 1300 + ], + "score": 0.941 + }, + { + "category_id": 6, + "poly": [ + 294, + 866, + 1404, + 866, + 1404, + 930, + 294, + 930 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 105, + 297, + 105 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 299, + 1820, + 1403, + 1820, + 1403, + 1974, + 299, + 1974 + ], + "score": 0.818 + }, + { + "category_id": 1, + "poly": [ + 299, + 1974, + 1399, + 1974, + 1399, + 2034, + 299, + 2034 + ], + "score": 0.804 + }, + { + "category_id": 0, + "poly": [ + 300, + 1351, + 1287, + 1351, + 1287, + 1410, + 300, + 1410 + ], + "score": 0.803 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 858, + 2087, + 858, + 2111, + 841, + 2111 + ], + "score": 0.653 + }, + { + "category_id": 2, + "poly": [ + 841, + 2087, + 859, + 2087, + 859, + 2111, + 841, + 2111 + ], + "score": 0.293 + }, + { + "category_id": 1, + "poly": [ + 300, + 1351, + 1287, + 1351, + 1287, + 1410, + 300, + 1410 + ], + "score": 0.142 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 104, + 297, + 104 + ], + "score": 0.093 + }, + { + "category_id": 13, + "poly": [ + 822, + 1501, + 949, + 1501, + 949, + 1533, + 822, + 1533 + ], + "score": 0.91, + "latex": "n _ { s t e p s } = 5" + }, + { + "category_id": 13, + "poly": [ + 932, + 898, + 1056, + 898, + 1056, + 931, + 932, + 931 + ], + "score": 0.91, + "latex": "n _ { s t e p s } = 5" + }, + { + "category_id": 13, + "poly": [ + 298, + 1531, + 422, + 1531, + 422, + 1563, + 298, + 1563 + ], + "score": 0.9, + "latex": "n _ { s t e p s } = 8" + }, + { + "category_id": 13, + "poly": [ + 298, + 898, + 428, + 898, + 428, + 931, + 298, + 931 + ], + "score": 0.89, + "latex": "n _ { n e i g h } = 8 " + }, + { + "category_id": 13, + "poly": [ + 1155, + 769, + 1212, + 769, + 1212, + 797, + 1155, + 797 + ], + "score": 0.88, + "latex": "\\pm 2 \\sigma" + }, + { + "category_id": 13, + "poly": [ + 314, + 1122, + 442, + 1122, + 442, + 1153, + 314, + 1153 + ], + "score": 0.88, + "latex": "n _ { n e i g h } = 8" + }, + { + "category_id": 13, + "poly": [ + 580, + 898, + 715, + 898, + 715, + 931, + 580, + 931 + ], + "score": 0.87, + "latex": "( n _ { n e i g h } = 1 )" + }, + { + "category_id": 13, + "poly": [ + 314, + 1090, + 443, + 1090, + 443, + 1121, + 314, + 1121 + ], + "score": 0.86, + "latex": "n _ { n e i g h } = 8" + }, + { + "category_id": 13, + "poly": [ + 313, + 999, + 443, + 999, + 443, + 1030, + 313, + 1030 + ], + "score": 0.81, + "latex": "\\overline { { n _ { n e i g h } = 8 } }" + }, + { + "category_id": 13, + "poly": [ + 527, + 1028, + 668, + 1028, + 668, + 1062, + 527, + 1062 + ], + "score": 0.55, + "latex": "( n _ { n e i g h } = 1 )" + }, + { + "category_id": 15, + "poly": [ + 430.0, + 213.0, + 502.0, + 213.0, + 502.0, + 251.0, + 430.0, + 251.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 400.0, + 259.0, + 443.0, + 259.0, + 443.0, + 589.0, + 400.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 266.0, + 501.0, + 266.0, + 501.0, + 306.0, + 428.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 318.0, + 502.0, + 318.0, + 502.0, + 358.0, + 428.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 427.0, + 370.0, + 503.0, + 370.0, + 503.0, + 414.0, + 427.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 425.0, + 502.0, + 425.0, + 502.0, + 465.0, + 428.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 456.0, + 714.0, + 456.0, + 714.0, + 464.0, + 706.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 478.0, + 500.0, + 478.0, + 500.0, + 518.0, + 428.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 532.0, + 502.0, + 532.0, + 502.0, + 572.0, + 428.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 428.0, + 585.0, + 502.0, + 585.0, + 502.0, + 624.0, + 428.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 616.0, + 554.0, + 616.0, + 554.0, + 650.0, + 509.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 616.0, + 644.0, + 616.0, + 644.0, + 650.0, + 598.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 616.0, + 733.0, + 616.0, + 733.0, + 650.0, + 685.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 615.0, + 822.0, + 615.0, + 822.0, + 649.0, + 774.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 865.0, + 615.0, + 912.0, + 615.0, + 912.0, + 649.0, + 865.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 615.0, + 1003.0, + 615.0, + 1003.0, + 649.0, + 954.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 616.0, + 1090.0, + 616.0, + 1090.0, + 650.0, + 1044.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1133.0, + 616.0, + 1179.0, + 616.0, + 1179.0, + 650.0, + 1133.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1221.0, + 615.0, + 1267.0, + 615.0, + 1267.0, + 650.0, + 1221.0, + 650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 635.0, + 979.0, + 635.0, + 979.0, + 681.0, + 796.0, + 681.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 703.0, + 1404.0, + 703.0, + 1404.0, + 742.0, + 293.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 738.0, + 1403.0, + 738.0, + 1403.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 768.0, + 1154.0, + 768.0, + 1154.0, + 802.0, + 295.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 768.0, + 1330.0, + 768.0, + 1330.0, + 802.0, + 1213.0, + 802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 865.0, + 1405.0, + 865.0, + 1405.0, + 901.0, + 296.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 895.0, + 297.0, + 895.0, + 297.0, + 933.0, + 292.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 895.0, + 579.0, + 895.0, + 579.0, + 933.0, + 429.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 895.0, + 931.0, + 895.0, + 931.0, + 933.0, + 716.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 895.0, + 1274.0, + 895.0, + 1274.0, + 933.0, + 1057.0, + 933.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1347.0, + 1291.0, + 1347.0, + 1291.0, + 1383.0, + 295.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1384.0, + 591.0, + 1384.0, + 591.0, + 1412.0, + 387.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1561.0, + 1404.0, + 1561.0, + 1404.0, + 1595.0, + 296.0, + 1595.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1590.0, + 1407.0, + 1590.0, + 1407.0, + 1624.0, + 292.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1618.0, + 1407.0, + 1618.0, + 1407.0, + 1656.0, + 292.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1648.0, + 1405.0, + 1648.0, + 1405.0, + 1687.0, + 292.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1680.0, + 1405.0, + 1680.0, + 1405.0, + 1718.0, + 294.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1404.0, + 1712.0, + 1404.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1743.0, + 1404.0, + 1743.0, + 1404.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1774.0, + 528.0, + 1774.0, + 528.0, + 1807.0, + 293.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1439.0, + 1402.0, + 1439.0, + 1402.0, + 1471.0, + 297.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1469.0, + 1404.0, + 1469.0, + 1404.0, + 1501.0, + 295.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1498.0, + 821.0, + 1498.0, + 821.0, + 1537.0, + 292.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1498.0, + 1408.0, + 1498.0, + 1408.0, + 1537.0, + 950.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1531.0, + 297.0, + 1531.0, + 297.0, + 1563.0, + 293.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1531.0, + 719.0, + 1531.0, + 719.0, + 1563.0, + 423.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1234.0, + 1405.0, + 1234.0, + 1405.0, + 1274.0, + 294.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1270.0, + 663.0, + 1270.0, + 663.0, + 1302.0, + 296.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1818.0, + 1403.0, + 1818.0, + 1403.0, + 1855.0, + 297.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1849.0, + 1404.0, + 1849.0, + 1404.0, + 1884.0, + 294.0, + 1884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1916.0, + 295.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1171.0, + 1942.0, + 1171.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1969.0, + 1403.0, + 1969.0, + 1403.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 2003.0, + 1149.0, + 2003.0, + 1149.0, + 2034.0, + 298.0, + 2034.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1347.0, + 1291.0, + 1347.0, + 1291.0, + 1383.0, + 295.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 1384.0, + 591.0, + 1384.0, + 591.0, + 1412.0, + 387.0, + 1412.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 5, + "poly": [ + 295, + 343, + 1405, + 343, + 1405, + 757, + 295, + 757 + ], + "score": 0.981, + "html": "
modelresources spent# params (mil.)error(%)
Shake-Shake (Gastaldi,2017)4 GPU days,2 GPUs262.9
WRN 28-10 (Loshchilov & Hutter,2017)1 GPU day36.53.86
Baker et al. (2016)80-100 GPU days116.9
Cai et al. (2017)15 GPU days19.75.7
Zoph & Le (2017)16.000-24.000 GPU days37.53.65
Real et al. (2017)2500 GPU days5.45.4
Saxena & Verbeek (2016)？217.4
Brock et al. (2017)3 GPU days16.04.0
Ours (random networks,nsteps = 5, nneigh =1)0.2 GPU days4.46.5
Ours (nsteps = 5,nneigh =8,10 runs)0.5 GPU days5.75.7
Ours (nsteps = 8,nneigh = 8,4 runs)1 GPU day19.75.2
Ours (snapshot ensemble,4 runs)2 GPU days57.84.7
Ours (ensemble across runs)4 GPU days884.4
" + }, + { + "category_id": 5, + "poly": [ + 306, + 892, + 1394, + 892, + 1394, + 1182, + 306, + 1182 + ], + "score": 0.977, + "html": "
modelresources spent# params (mil.)error(%)
15.9
Shake-Shake (Gastaldi, 2017) WRN 28-10 (Loshchilov & Hutter,2017)14 GPU days34.4
1 GPU day36.519.6
Real et al. (2017)250 GPUs40.423.7
Brock et al. (2017)3 GPU days16.020.6
Ours (nsteps = 8,nneigh = 8,5 runs)1 GPU day22.323.4
Ours (snapshot ensemble,5 runs)2 GPU days73.320.9
Ours (ensemble across runs)5 GPU days111.519.6
" + }, + { + "category_id": 1, + "poly": [ + 297, + 1306, + 1405, + 1306, + 1405, + 1490, + 297, + 1490 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1604, + 1404, + 1604, + 1404, + 1881, + 297, + 1881 + ], + "score": 0.966 + }, + { + "category_id": 6, + "poly": [ + 295, + 822, + 1400, + 822, + 1400, + 886, + 295, + 886 + ], + "score": 0.943 + }, + { + "category_id": 6, + "poly": [ + 296, + 248, + 1401, + 248, + 1401, + 312, + 296, + 312 + ], + "score": 0.94 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.912 + }, + { + "category_id": 0, + "poly": [ + 300, + 1249, + 730, + 1249, + 730, + 1282, + 300, + 1282 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 1534, + 544, + 1534, + 544, + 1571, + 300, + 1571 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2111, + 840, + 2111 + ], + "score": 0.789 + }, + { + "category_id": 13, + "poly": [ + 370, + 646, + 600, + 646, + 600, + 674, + 370, + 674 + ], + "score": 0.82, + "latex": "( n _ { s t e p s } = 5 , n _ { n e i g h } = 8" + }, + { + "category_id": 13, + "poly": [ + 553, + 620, + 788, + 620, + 788, + 647, + 553, + 647 + ], + "score": 0.59, + "latex": "\\underline { { n _ { s t e p s } } } = 5 , n _ { n e i g h } = 1 )" + }, + { + "category_id": 13, + "poly": [ + 394, + 1085, + 517, + 1085, + 517, + 1117, + 394, + 1117 + ], + "score": 0.48, + "latex": "\\overline { { n _ { s t e p s } = 8 } }" + }, + { + "category_id": 13, + "poly": [ + 371, + 674, + 481, + 674, + 481, + 700, + 371, + 700 + ], + "score": 0.42, + "latex": "( n _ { s t e p s } = 8" + }, + { + "category_id": 13, + "poly": [ + 487, + 675, + 600, + 675, + 600, + 700, + 487, + 700 + ], + "score": 0.35, + "latex": "n _ { n e i g h } = 8" + }, + { + "category_id": 13, + "poly": [ + 370, + 674, + 610, + 674, + 610, + 701, + 370, + 701 + ], + "score": 0.32, + "latex": "( n _ { s t e p s } = 8 , n _ { n e i g h } = 8 ," + }, + { + "category_id": 13, + "poly": [ + 527, + 1086, + 656, + 1086, + 656, + 1117, + 527, + 1117 + ], + "score": 0.29, + "latex": "\\overline { { n _ { n e i g h } = 8 } }" + }, + { + "category_id": 13, + "poly": [ + 553, + 620, + 660, + 620, + 660, + 646, + 553, + 646 + ], + "score": 0.26, + "latex": "n _ { s t e p s } = 5" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 818.0, + 1405.0, + 818.0, + 1405.0, + 858.0, + 294.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 853.0, + 1399.0, + 853.0, + 1399.0, + 887.0, + 294.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 248.0, + 1403.0, + 248.0, + 1403.0, + 283.0, + 296.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 1400.0, + 279.0, + 1400.0, + 313.0, + 294.0, + 313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1247.0, + 733.0, + 1247.0, + 733.0, + 1284.0, + 293.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1531.0, + 549.0, + 1531.0, + 549.0, + 1579.0, + 292.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1307.0, + 1405.0, + 1307.0, + 1405.0, + 1342.0, + 295.0, + 1342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1337.0, + 1405.0, + 1337.0, + 1405.0, + 1373.0, + 294.0, + 1373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1366.0, + 1406.0, + 1366.0, + 1406.0, + 1403.0, + 295.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1398.0, + 1405.0, + 1398.0, + 1405.0, + 1432.0, + 294.0, + 1432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1428.0, + 1405.0, + 1428.0, + 1405.0, + 1463.0, + 292.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1459.0, + 603.0, + 1459.0, + 603.0, + 1494.0, + 295.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1605.0, + 1405.0, + 1605.0, + 1405.0, + 1642.0, + 296.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1631.0, + 1407.0, + 1631.0, + 1407.0, + 1673.0, + 291.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1664.0, + 1406.0, + 1664.0, + 1406.0, + 1703.0, + 294.0, + 1703.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1691.0, + 1408.0, + 1691.0, + 1408.0, + 1736.0, + 291.0, + 1736.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1727.0, + 1406.0, + 1727.0, + 1406.0, + 1763.0, + 292.0, + 1763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1753.0, + 1406.0, + 1753.0, + 1406.0, + 1796.0, + 291.0, + 1796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1786.0, + 1407.0, + 1786.0, + 1407.0, + 1824.0, + 291.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1816.0, + 1405.0, + 1816.0, + 1405.0, + 1855.0, + 294.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1847.0, + 1332.0, + 1847.0, + 1332.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 300, + 280, + 1404, + 280, + 1404, + 555, + 300, + 555 + ], + "score": 0.957 + }, + { + "category_id": 1, + "poly": [ + 298, + 1392, + 1403, + 1392, + 1403, + 1516, + 298, + 1516 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 297, + 1308, + 1398, + 1308, + 1398, + 1370, + 297, + 1370 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 301, + 578, + 1397, + 578, + 1397, + 641, + 301, + 641 + ], + "score": 0.909 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 105, + 298, + 105 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 299, + 1624, + 1399, + 1624, + 1399, + 1687, + 299, + 1687 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 300, + 1198, + 1125, + 1198, + 1125, + 1231, + 300, + 1231 + ], + "score": 0.898 + }, + { + "category_id": 1, + "poly": [ + 300, + 1253, + 1122, + 1253, + 1122, + 1286, + 300, + 1286 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 297, + 662, + 1397, + 662, + 1397, + 727, + 297, + 727 + ], + "score": 0.896 + }, + { + "category_id": 1, + "poly": [ + 295, + 1058, + 1401, + 1058, + 1401, + 1121, + 295, + 1121 + ], + "score": 0.895 + }, + { + "category_id": 1, + "poly": [ + 301, + 1709, + 1398, + 1709, + 1398, + 1803, + 301, + 1803 + ], + "score": 0.89 + }, + { + "category_id": 1, + "poly": [ + 299, + 1539, + 1399, + 1539, + 1399, + 1601, + 299, + 1601 + ], + "score": 0.888 + }, + { + "category_id": 1, + "poly": [ + 297, + 1144, + 1260, + 1144, + 1260, + 1177, + 297, + 1177 + ], + "score": 0.882 + }, + { + "category_id": 1, + "poly": [ + 293, + 802, + 1402, + 802, + 1402, + 867, + 293, + 867 + ], + "score": 0.877 + }, + { + "category_id": 1, + "poly": [ + 295, + 747, + 1398, + 747, + 1398, + 783, + 295, + 783 + ], + "score": 0.873 + }, + { + "category_id": 1, + "poly": [ + 291, + 887, + 1401, + 887, + 1401, + 952, + 291, + 952 + ], + "score": 0.87 + }, + { + "category_id": 1, + "poly": [ + 293, + 973, + 1401, + 973, + 1401, + 1037, + 293, + 1037 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 298, + 1825, + 1403, + 1825, + 1403, + 1949, + 298, + 1949 + ], + "score": 0.827 + }, + { + "category_id": 0, + "poly": [ + 300, + 228, + 488, + 228, + 488, + 261, + 300, + 261 + ], + "score": 0.811 + }, + { + "category_id": 1, + "poly": [ + 299, + 1972, + 1397, + 1972, + 1397, + 2034, + 299, + 2034 + ], + "score": 0.787 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.774 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 227.0, + 491.0, + 227.0, + 491.0, + 265.0, + 295.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 1404.0, + 279.0, + 1404.0, + 316.0, + 294.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 309.0, + 1405.0, + 309.0, + 1405.0, + 345.0, + 321.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 341.0, + 1405.0, + 341.0, + 1405.0, + 377.0, + 321.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 371.0, + 1405.0, + 371.0, + 1405.0, + 407.0, + 321.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 403.0, + 1404.0, + 403.0, + 1404.0, + 436.0, + 323.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 431.0, + 1404.0, + 431.0, + 1404.0, + 468.0, + 320.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 459.0, + 1405.0, + 459.0, + 1405.0, + 502.0, + 320.0, + 502.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 493.0, + 1404.0, + 493.0, + 1404.0, + 528.0, + 320.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 519.0, + 610.0, + 519.0, + 610.0, + 562.0, + 321.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1393.0, + 1404.0, + 1393.0, + 1404.0, + 1430.0, + 294.0, + 1430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1419.0, + 1408.0, + 1419.0, + 1408.0, + 1463.0, + 320.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 318.0, + 1454.0, + 1407.0, + 1454.0, + 1407.0, + 1493.0, + 318.0, + 1493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1484.0, + 396.0, + 1484.0, + 396.0, + 1517.0, + 321.0, + 1517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1301.0, + 1403.0, + 1301.0, + 1403.0, + 1350.0, + 291.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1340.0, + 558.0, + 1340.0, + 558.0, + 1371.0, + 323.0, + 1371.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 578.0, + 1401.0, + 578.0, + 1401.0, + 613.0, + 297.0, + 613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 610.0, + 814.0, + 610.0, + 814.0, + 643.0, + 322.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1622.0, + 1403.0, + 1622.0, + 1403.0, + 1661.0, + 294.0, + 1661.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1653.0, + 1036.0, + 1653.0, + 1036.0, + 1689.0, + 324.0, + 1689.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1197.0, + 1126.0, + 1197.0, + 1126.0, + 1235.0, + 295.0, + 1235.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1252.0, + 1125.0, + 1252.0, + 1125.0, + 1290.0, + 295.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 662.0, + 1401.0, + 662.0, + 1401.0, + 698.0, + 296.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 695.0, + 927.0, + 695.0, + 927.0, + 728.0, + 323.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1097.0, + 294.0, + 1097.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1088.0, + 666.0, + 1088.0, + 666.0, + 1123.0, + 322.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1707.0, + 1402.0, + 1707.0, + 1402.0, + 1746.0, + 295.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1740.0, + 1402.0, + 1740.0, + 1402.0, + 1777.0, + 321.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1772.0, + 535.0, + 1772.0, + 535.0, + 1803.0, + 325.0, + 1803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1537.0, + 1403.0, + 1537.0, + 1403.0, + 1576.0, + 294.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1563.0, + 398.0, + 1563.0, + 398.0, + 1606.0, + 319.0, + 1606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1142.0, + 1265.0, + 1142.0, + 1265.0, + 1180.0, + 294.0, + 1180.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 801.0, + 1405.0, + 801.0, + 1405.0, + 841.0, + 291.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 833.0, + 467.0, + 833.0, + 467.0, + 867.0, + 322.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 747.0, + 1403.0, + 747.0, + 1403.0, + 786.0, + 294.0, + 786.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 887.0, + 1404.0, + 887.0, + 1404.0, + 923.0, + 295.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 918.0, + 1056.0, + 918.0, + 1056.0, + 953.0, + 322.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 970.0, + 1406.0, + 970.0, + 1406.0, + 1010.0, + 293.0, + 1010.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1004.0, + 922.0, + 1004.0, + 922.0, + 1037.0, + 322.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1824.0, + 1404.0, + 1824.0, + 1404.0, + 1861.0, + 292.0, + 1861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1853.0, + 1405.0, + 1853.0, + 1405.0, + 1892.0, + 321.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1886.0, + 1407.0, + 1886.0, + 1407.0, + 1923.0, + 320.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1917.0, + 396.0, + 1917.0, + 396.0, + 1950.0, + 321.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1969.0, + 1401.0, + 1969.0, + 1401.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 1216.0, + 2003.0, + 1216.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 857, + 75, + 857, + 106, + 298, + 106 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 866, + 2088, + 866, + 2113, + 836, + 2113 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 291, + 212, + 1409, + 212, + 1409, + 1202, + 291, + 1202 + ], + "score": 0.733 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 262.0, + 1303.0, + 262.0, + 1303.0, + 298.0, + 320.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 311.0, + 1403.0, + 311.0, + 1403.0, + 351.0, + 291.0, + 351.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 346.0, + 663.0, + 346.0, + 663.0, + 378.0, + 323.0, + 378.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 397.0, + 1406.0, + 397.0, + 1406.0, + 432.0, + 293.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 428.0, + 1279.0, + 428.0, + 1279.0, + 464.0, + 323.0, + 464.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 477.0, + 1404.0, + 477.0, + 1404.0, + 515.0, + 293.0, + 515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 511.0, + 722.0, + 511.0, + 722.0, + 545.0, + 321.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 561.0, + 1407.0, + 561.0, + 1407.0, + 601.0, + 292.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 593.0, + 1259.0, + 593.0, + 1259.0, + 631.0, + 321.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 643.0, + 1406.0, + 643.0, + 1406.0, + 684.0, + 293.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 677.0, + 853.0, + 677.0, + 853.0, + 712.0, + 322.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 728.0, + 1274.0, + 728.0, + 1274.0, + 764.0, + 294.0, + 764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 779.0, + 1404.0, + 779.0, + 1404.0, + 821.0, + 291.0, + 821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 812.0, + 633.0, + 812.0, + 633.0, + 847.0, + 322.0, + 847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 861.0, + 1404.0, + 861.0, + 1404.0, + 904.0, + 292.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 894.0, + 1201.0, + 894.0, + 1201.0, + 932.0, + 320.0, + 932.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 944.0, + 1404.0, + 944.0, + 1404.0, + 987.0, + 292.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 977.0, + 396.0, + 977.0, + 396.0, + 1011.0, + 321.0, + 1011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1031.0, + 1289.0, + 1031.0, + 1289.0, + 1067.0, + 296.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1083.0, + 1322.0, + 1083.0, + 1322.0, + 1118.0, + 296.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1135.0, + 1406.0, + 1135.0, + 1406.0, + 1174.0, + 293.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1165.0, + 748.0, + 1165.0, + 748.0, + 1204.0, + 320.0, + 1204.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 338, + 1034, + 1295, + 1034, + 1295, + 1556, + 338, + 1556 + ], + "score": 0.967 + }, + { + "category_id": 3, + "poly": [ + 325, + 346, + 1297, + 346, + 1297, + 863, + 325, + 863 + ], + "score": 0.962 + }, + { + "category_id": 4, + "poly": [ + 610, + 920, + 1091, + 920, + 1091, + 956, + 610, + 956 + ], + "score": 0.92 + }, + { + "category_id": 4, + "poly": [ + 610, + 1615, + 1090, + 1615, + 1090, + 1649, + 610, + 1649 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.849 + }, + { + "category_id": 1, + "poly": [ + 298, + 1714, + 571, + 1714, + 571, + 1751, + 298, + 1751 + ], + "score": 0.692 + }, + { + "category_id": 0, + "poly": [ + 296, + 229, + 1211, + 229, + 1211, + 262, + 296, + 262 + ], + "score": 0.686 + }, + { + "category_id": 4, + "poly": [ + 453, + 343, + 1273, + 343, + 1273, + 373, + 453, + 373 + ], + "score": 0.28 + }, + { + "category_id": 2, + "poly": [ + 296, + 229, + 1211, + 229, + 1211, + 262, + 296, + 262 + ], + "score": 0.177 + }, + { + "category_id": 2, + "poly": [ + 298, + 1714, + 571, + 1714, + 571, + 1751, + 298, + 1751 + ], + "score": 0.156 + }, + { + "category_id": 13, + "poly": [ + 1027, + 343, + 1164, + 343, + 1164, + 372, + 1027, + 372 + ], + "score": 0.66, + "latex": "\\mathsf { R } ^ { \\sim } 2 { = } 0 . 6 3 5" + }, + { + "category_id": 13, + "poly": [ + 1013, + 1038, + 1152, + 1038, + 1152, + 1066, + 1013, + 1066 + ], + "score": 0.4, + "latex": "\\mathord { \\mathsf { R } } ^ { \\mathord { \\sim } } 2 { = } 0 . 3 6 1" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 1034.0, + 1012.0, + 1034.0, + 1012.0, + 1069.0, + 466.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1153.0, + 1034.0, + 1261.0, + 1034.0, + 1261.0, + 1069.0, + 1153.0, + 1069.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1017.0, + 1079.0, + 1031.0, + 1079.0, + 1031.0, + 1093.0, + 1017.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 744.0, + 1106.0, + 755.0, + 1106.0, + 755.0, + 1116.0, + 744.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1111.0, + 440.0, + 1111.0, + 440.0, + 1144.0, + 368.0, + 1144.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1132.0, + 1119.0, + 1132.0, + 1119.0, + 1143.0, + 1109.0, + 1143.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1246.0, + 1148.0, + 1256.0, + 1148.0, + 1256.0, + 1157.0, + 1246.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1165.0, + 629.0, + 1165.0, + 629.0, + 1178.0, + 616.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 657.0, + 1160.0, + 670.0, + 1160.0, + 670.0, + 1174.0, + 657.0, + 1174.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 977.0, + 1161.0, + 992.0, + 1161.0, + 992.0, + 1175.0, + 977.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1137.0, + 1168.0, + 1148.0, + 1168.0, + 1148.0, + 1177.0, + 1137.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1177.0, + 1183.0, + 1177.0, + 1183.0, + 1190.0, + 1170.0, + 1190.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1184.0, + 441.0, + 1184.0, + 441.0, + 1217.0, + 368.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 1195.0, + 840.0, + 1195.0, + 840.0, + 1208.0, + 826.0, + 1208.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1227.0, + 657.0, + 1227.0, + 657.0, + 1240.0, + 644.0, + 1240.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1229.0, + 1167.0, + 1229.0, + 1167.0, + 1239.0, + 1157.0, + 1239.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1244.0, + 1074.0, + 1244.0, + 1074.0, + 1254.0, + 1063.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1255.0, + 1107.0, + 1255.0, + 1107.0, + 1264.0, + 1098.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 1271.0, + 1046.0, + 1271.0, + 1046.0, + 1281.0, + 1036.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 1278.0, + 379.0, + 1278.0, + 379.0, + 1358.0, + 336.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1289.0, + 597.0, + 1289.0, + 597.0, + 1298.0, + 587.0, + 1298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1322.0, + 715.0, + 1322.0, + 715.0, + 1336.0, + 703.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1332.0, + 440.0, + 1332.0, + 440.0, + 1365.0, + 368.0, + 1365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1345.0, + 613.0, + 1345.0, + 613.0, + 1358.0, + 600.0, + 1358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1359.0, + 694.0, + 1359.0, + 694.0, + 1372.0, + 681.0, + 1372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1377.0, + 607.0, + 1377.0, + 607.0, + 1388.0, + 598.0, + 1388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1399.0, + 505.0, + 1399.0, + 505.0, + 1412.0, + 492.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 1405.0, + 440.0, + 1405.0, + 440.0, + 1439.0, + 368.0, + 1439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1466.0, + 483.0, + 1466.0, + 483.0, + 1480.0, + 470.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 506.0, + 1488.0, + 564.0, + 1488.0, + 564.0, + 1523.0, + 506.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 1488.0, + 665.0, + 1488.0, + 665.0, + 1522.0, + 606.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1488.0, + 766.0, + 1488.0, + 766.0, + 1522.0, + 706.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 807.0, + 1488.0, + 865.0, + 1488.0, + 865.0, + 1522.0, + 807.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 1489.0, + 964.0, + 1489.0, + 964.0, + 1521.0, + 909.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 1488.0, + 1066.0, + 1488.0, + 1066.0, + 1522.0, + 1007.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1107.0, + 1488.0, + 1167.0, + 1488.0, + 1167.0, + 1522.0, + 1107.0, + 1522.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 1488.0, + 1266.0, + 1488.0, + 1266.0, + 1523.0, + 1208.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 1512.0, + 982.0, + 1512.0, + 982.0, + 1553.0, + 747.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 339.0, + 1026.0, + 339.0, + 1026.0, + 375.0, + 456.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 339.0, + 1272.0, + 339.0, + 1272.0, + 375.0, + 1165.0, + 375.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 386.0, + 1193.0, + 386.0, + 1193.0, + 399.0, + 1178.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 398.0, + 438.0, + 398.0, + 438.0, + 429.0, + 358.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1247.0, + 445.0, + 1258.0, + 445.0, + 1258.0, + 455.0, + 1247.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 355.0, + 449.0, + 441.0, + 449.0, + 441.0, + 487.0, + 355.0, + 487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 925.0, + 485.0, + 936.0, + 485.0, + 936.0, + 496.0, + 925.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 503.0, + 439.0, + 503.0, + 439.0, + 598.0, + 324.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 505.0, + 989.0, + 505.0, + 989.0, + 521.0, + 975.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 511.0, + 1072.0, + 511.0, + 1072.0, + 519.0, + 1063.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 525.0, + 1035.0, + 525.0, + 1035.0, + 535.0, + 1024.0, + 535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 560.0, + 438.0, + 560.0, + 438.0, + 591.0, + 354.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 584.0, + 439.0, + 584.0, + 439.0, + 659.0, + 320.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 623.0, + 1086.0, + 623.0, + 1086.0, + 633.0, + 1077.0, + 633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 650.0, + 506.0, + 650.0, + 506.0, + 658.0, + 497.0, + 658.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 659.0, + 991.0, + 659.0, + 991.0, + 669.0, + 982.0, + 669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 668.0, + 443.0, + 668.0, + 443.0, + 702.0, + 357.0, + 702.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 676.0, + 482.0, + 676.0, + 482.0, + 687.0, + 472.0, + 687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 675.0, + 813.0, + 675.0, + 813.0, + 685.0, + 804.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 745.0, + 688.0, + 755.0, + 688.0, + 755.0, + 699.0, + 745.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 703.0, + 813.0, + 703.0, + 813.0, + 713.0, + 804.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 719.0, + 440.0, + 719.0, + 440.0, + 757.0, + 356.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 841.0, + 734.0, + 851.0, + 734.0, + 851.0, + 742.0, + 841.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 572.0, + 749.0, + 582.0, + 749.0, + 582.0, + 759.0, + 572.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 745.0, + 712.0, + 745.0, + 712.0, + 760.0, + 698.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 775.0, + 445.0, + 775.0, + 445.0, + 809.0, + 357.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 771.0, + 546.0, + 771.0, + 546.0, + 785.0, + 532.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 503.0, + 793.0, + 573.0, + 793.0, + 573.0, + 828.0, + 503.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 794.0, + 753.0, + 794.0, + 753.0, + 826.0, + 686.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 794.0, + 935.0, + 794.0, + 935.0, + 826.0, + 868.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 794.0, + 1118.0, + 794.0, + 1118.0, + 828.0, + 1048.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 793.0, + 1300.0, + 793.0, + 1300.0, + 828.0, + 1231.0, + 828.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 818.0, + 981.0, + 818.0, + 981.0, + 860.0, + 748.0, + 860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 606.0, + 920.0, + 1090.0, + 920.0, + 1090.0, + 957.0, + 606.0, + 957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 1614.0, + 1090.0, + 1614.0, + 1090.0, + 1651.0, + 607.0, + 1651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1214.0, + 225.0, + 1214.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 458.0, + 340.0, + 1026.0, + 340.0, + 1026.0, + 377.0, + 458.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1165.0, + 340.0, + 1269.0, + 340.0, + 1269.0, + 377.0, + 1165.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1214.0, + 225.0, + 1214.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1708.0, + 573.0, + 1708.0, + 573.0, + 1757.0, + 292.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1708.0, + 573.0, + 1708.0, + 573.0, + 1757.0, + 292.0, + 1757.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 698, + 373, + 991, + 373, + 991, + 1819, + 698, + 1819 + ], + "score": 0.939 + }, + { + "category_id": 4, + "poly": [ + 611, + 1848, + 1091, + 1848, + 1091, + 1882, + 611, + 1882 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.898 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.839 + }, + { + "category_id": 15, + "poly": [ + 764.0, + 381.0, + 928.0, + 381.0, + 928.0, + 430.0, + 764.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 780.0, + 486.0, + 912.0, + 486.0, + 912.0, + 532.0, + 780.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 717.0, + 593.0, + 974.0, + 593.0, + 974.0, + 637.0, + 717.0, + 637.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 681.0, + 854.0, + 681.0, + 854.0, + 693.0, + 840.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 703.0, + 920.0, + 703.0, + 920.0, + 740.0, + 769.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 782.0, + 855.0, + 782.0, + 855.0, + 805.0, + 835.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 806.0, + 949.0, + 806.0, + 949.0, + 846.0, + 742.0, + 846.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 911.0, + 920.0, + 911.0, + 920.0, + 958.0, + 775.0, + 958.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 714.0, + 1020.0, + 980.0, + 1020.0, + 980.0, + 1061.0, + 714.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 766.0, + 1127.0, + 927.0, + 1127.0, + 927.0, + 1164.0, + 766.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1231.0, + 949.0, + 1231.0, + 949.0, + 1273.0, + 743.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 1334.0, + 920.0, + 1334.0, + 920.0, + 1381.0, + 774.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1439.0, + 980.0, + 1439.0, + 980.0, + 1484.0, + 713.0, + 1484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1523.0, + 855.0, + 1523.0, + 855.0, + 1545.0, + 837.0, + 1545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1547.0, + 929.0, + 1547.0, + 929.0, + 1592.0, + 762.0, + 1592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1631.0, + 855.0, + 1631.0, + 855.0, + 1650.0, + 837.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1653.0, + 915.0, + 1653.0, + 915.0, + 1695.0, + 778.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 1735.0, + 855.0, + 1735.0, + 855.0, + 1756.0, + 837.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 1758.0, + 922.0, + 1758.0, + 922.0, + 1802.0, + 771.0, + 1802.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 604.0, + 1844.0, + 1091.0, + 1844.0, + 1091.0, + 1886.0, + 604.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 694, + 257, + 999, + 257, + 999, + 1911, + 694, + 1911 + ], + "score": 0.928 + }, + { + "category_id": 4, + "poly": [ + 495, + 1937, + 1205, + 1937, + 1205, + 1972, + 495, + 1972 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 857, + 75, + 857, + 105, + 297, + 105 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.855 + }, + { + "category_id": 13, + "poly": [ + 1073, + 1939, + 1195, + 1939, + 1195, + 1973, + 1073, + 1973 + ], + "score": 0.91, + "latex": "n _ { s t e p s } = 5" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1931.0, + 1072.0, + 1931.0, + 1072.0, + 1978.0, + 492.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1931.0, + 1207.0, + 1931.0, + 1207.0, + 1978.0, + 1196.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 859.0, + 73.0, + 859.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 4, + "poly": [ + 495, + 1937, + 1205, + 1937, + 1205, + 1972, + 495, + 1972 + ], + "score": 0.918 + }, + { + "category_id": 3, + "poly": [ + 698, + 252, + 973, + 252, + 973, + 1896, + 698, + 1896 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 297, + 76, + 857, + 76, + 857, + 105, + 297, + 105 + ], + "score": 0.901 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.851 + }, + { + "category_id": 13, + "poly": [ + 1073, + 1939, + 1195, + 1939, + 1195, + 1973, + 1073, + 1973 + ], + "score": 0.9, + "latex": "n _ { s t e p s } = 8" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1931.0, + 1072.0, + 1931.0, + 1072.0, + 1978.0, + 492.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1196.0, + 1931.0, + 1207.0, + 1931.0, + 1207.0, + 1978.0, + 1196.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 306.0, + 914.0, + 306.0, + 914.0, + 331.0, + 888.0, + 331.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 309.0, + 939.0, + 309.0, + 939.0, + 482.0, + 861.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 478.0, + 875.0, + 478.0, + 875.0, + 513.0, + 843.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 531.0, + 873.0, + 531.0, + 873.0, + 554.0, + 851.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 569.0, + 869.0, + 569.0, + 869.0, + 584.0, + 855.0, + 584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 793.0, + 671.0, + 836.0, + 671.0, + 836.0, + 722.0, + 793.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 800.0, + 780.0, + 824.0, + 780.0, + 824.0, + 794.0, + 800.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 784.0, + 938.0, + 784.0, + 938.0, + 1155.0, + 829.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 809.0, + 810.0, + 809.0, + 810.0, + 829.0, + 767.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 1130.0, + 893.0, + 1130.0, + 893.0, + 1247.0, + 834.0, + 1247.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1245.0, + 919.0, + 1245.0, + 919.0, + 1557.0, + 857.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 878.0, + 1245.0, + 894.0, + 1245.0, + 894.0, + 1259.0, + 878.0, + 1259.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 900.0, + 1545.0, + 945.0, + 1545.0, + 945.0, + 1675.0, + 900.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 1655.0, + 955.0, + 1655.0, + 955.0, + 1905.0, + 890.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.25, + 693.5, + 902.25, + 693.5, + 902.25, + 831.0, + 743.25, + 831.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 72.0, + 859.0, + 72.0, + 859.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2122.0, + 832.0, + 2122.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/iaO86DUuKi/iaO86DUuKi_content_list.json b/parse/train/iaO86DUuKi/iaO86DUuKi_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd601bcfd1514088e73849b290af126647df7d3b --- /dev/null +++ b/parse/train/iaO86DUuKi/iaO86DUuKi_content_list.json @@ -0,0 +1,3610 @@ +[ + { + "type": "text", + "text": "CONSERVATIVE SAFETY CRITICS FOR EXPLORATION ", + "text_level": 1, + "bbox": [ + 174, + 98, + 805, + 121 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Homanga Bharadhwaj1∗, Aviral Kumar2, Nicholas Rhinehart2, Sergey Levine2, Florian Shkurti1, Animesh Garg1 ", + "bbox": [ + 183, + 143, + 620, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Toronto, Vector Institute 2University of California Berkeley homanga@cs.toronto.edu ", + "bbox": [ + 184, + 175, + 442, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 253, + 544, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Safe exploration presents a major challenge in reinforcement learning (RL): when active data collection requires deploying partially trained policies, we must ensure that these policies avoid catastrophically unsafe regions, while still enabling trial and error learning. In this paper, we target the problem of safe exploration in RL by learning a conservative safety estimate of environment states through a critic, and provably upper bound the likelihood of catastrophic failures at every training iteration. We theoretically characterize the tradeoff between safety and policy improvement, show that the safety constraints are likely to be satisfied with high probability during training, derive provable convergence bounds for our approach, which is no worse asymptotically than standard RL, and demonstrate the efficacy of the proposed approach on a suite of challenging navigation, manipulation, and locomotion tasks. Empirically, we show that the proposed approach can achieve competitive task performance while incurring significantly lower catastrophic failure rates during training than prior methods. Videos are at this url https: //sites.google.com/view/conservative-safety-critics/ ", + "bbox": [ + 233, + 286, + 764, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 527, + 336, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reinforcement learning (RL) is a powerful framework for learning-based control because it can enable agents to learn to make decisions automatically through trial and error. However, in the real world, the cost of those trials – and those errors – can be quite high: a quadruped learning to run as fast as possible, might fall down and crash, and then be unable to attempt further trials due to extensive physical damage. However, learning complex skills without any failures at all is likely impossible. Even humans and animals regularly experience failure, but quickly learn from their mistakes and behave cautiously in risky situations. In this paper, our goal is to develop safe exploration methods for RL that similarly exhibit conservative behavior, erring on the side of caution in particularly dangerous settings, and limiting the number of catastrophic failures. ", + "bbox": [ + 174, + 560, + 825, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A number of previous approaches have tackled this problem of safe exploration, often by formulating the problem as a constrained Markov decision process (CMDP) (Garcıa & Fernandez ´ , 2015; Altman, 1999). However, most of these approaches require additional assumptions, like assuming access to a function that can be queried to check if a state is safe (Thananjeyan et al., 2020), assuming access to a default safe controller (Koller et al., 2018; Berkenkamp et al., 2017), assuming knowledge of all the unsafe states (Fisac et al., 2019), and only obtaining safe policies after training converges, while being unsafe during the training process (Tessler et al., 2018; Dalal et al., 2018). ", + "bbox": [ + 174, + 693, + 825, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a general safe RL algorithm, with bounds on the probability of failures during training. Our method only assumes access to a sparse (e.g., binary) indicator for catastrophic failure, in the standard RL setting. We train a conservative safety critic that overestimates the probability of catastrophic failure, building on tools in the recently proposed conservative Q-learning framework (Kumar et al., 2020) for offline RL. In order to bound the likelihood of catastrophic failures at every iteration, we impose a KL-divergence constraint on successive policy updates so that the stationary distribution of states induced by the old and the new policies are not arbitrarily different. Based on the safety critic’s value, we consider a chance constraint denoting probability of failure, and optimize the policy through primal-dual gradient descent. ", + "bbox": [ + 173, + 797, + 825, + 895 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 103, + 823, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions in this paper are designing an algorithm that we refer to as Conservative Safety Critics (CSC), that learns a conservative estimate of how safe a state is, using this conservative estimate for safe-exploration and policy updates, and theoretically providing upper bounds on the probability of failures throughout training. Through empirical evaluation in five separate simulated robotic control domains spanning manipulation, navigation, and locomotion, we show that CSC is able to learn effective policies while reducing the rate of catastrophic failures by up to $50 \\%$ over prior safe exploration methods. ", + "bbox": [ + 174, + 138, + 825, + 236 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 256, + 339, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We describe the problem setting of a constrained MDP (Altman, 1999) specific to our approach and the conservative Q learning (Kumar et al., 2020) framework that we build on in our algorithm. ", + "bbox": [ + 176, + 286, + 823, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Constrained MDPs. We take a constrained RL view of safety (Garcıa & Fernandez ´ , 2015; Achiam et al., 2017), and define safe exploration as the process of ensuring the constraints of the constrained MDP (CMDP) are satisfied while exploring the environment to collect data samples. A CMDP is a tuple $( \\mathcal S , \\mathcal A , P , R , \\gamma , \\mu , \\mathcal C )$ , where $s$ is the state space, $\\mathcal { A }$ is the action space, $P : \\bar { \\mathcal { S } } \\times \\mathcal { A } \\times \\mathcal { S } [ 0 , 1 ]$ is a transition kernel, $R : S \\times \\mathcal { A } \\mathbb { R }$ is a task reward function, $\\gamma \\in ( 0 , 1 )$ is a discount factor, $\\mu$ is a starting state distribution, and ${ \\mathcal { C } } = \\{ ( c _ { i } : { \\mathcal { S } } \\{ 0 , 1 \\} , \\chi _ { i } \\in \\mathbb { R } ) | i \\in \\mathbb { Z } \\}$ is a set of (safety) constraints that the agent must satisfy, with constraint functions $c _ { i }$ taking values either 0 (alive) or 1 (failure) and limits $\\chi _ { i }$ defining the maximal allowable amount of non-satisfaction, in terms of expected probability of failure. A stochastic policy $\\pi : S { \\mathcal { P } } ( A )$ is a mapping from states to action distributions, and the set of all stationary policies is denoted by $\\Pi$ . Without loss of generality, we can consider a single constraint, where $\\mathcal { C }$ denotes the constraint satisfaction function ${ \\mathsf { \\bar { C } } } : { \\mathcal { S } } \\to \\{ 0 , 1 \\}$ , $( C \\equiv \\mathbb { 1 } \\{ f a i l u r e \\} )$ similar to the task reward function, and an upper limit $\\chi$ . Note that since we assume only a sparse binary indicator of failure from the environment $C ( s )$ , in purely online training, the agent must fail a few times during training, and hence 0 failures is impossible. However, we will discuss how we can minimize the number of failures to a small rate, for constraint satisfaction. ", + "bbox": [ + 174, + 321, + 825, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We define discounted state distribution of a policy $\\pi$ as $\\begin{array} { r } { d ^ { \\pi } ( s ) = ( 1 - \\gamma ) \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } P ( s _ { t } = s | \\pi ) . } \\end{array}$ the state value function as $V _ { R } ^ { \\pi } ( s ) ~ = ~ \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ R ( \\tau ) | s _ { 0 } = s \\right]$ , the state-action value function as $Q _ { R } ^ { \\pi } ( s , a ) = \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ R ( \\tau ) | s _ { 0 } = \\dot { s } , a _ { 0 } = a \\right]$ , and the advantage function as ${ \\cal A } _ { R } ^ { \\pi } ( s , a ) = { \\cal Q } _ { R } ^ { \\pi } ( s , a ) -$ $V _ { R } ^ { \\pi } ( s )$ . We define similar quantities for the constraint function, as $V _ { C }$ , $Q _ { C }$ , and $A _ { C }$ . So, we have $\\begin{array} { r } { V _ { R } ^ { \\pi } ( \\mu ) = \\mathbb E _ { \\tau \\sim \\pi } \\left[ \\sum _ { t = 0 } ^ { \\infty } R ( s _ { t } , a _ { t } ) \\right] } \\end{array}$ and $V _ { C } ^ { \\pi } ( \\mu )$ denoting the average episodic failures, which can also be interpreted as expected probability of failure since $\\begin{array} { r } { V _ { C } ^ { \\pi } ( \\mu ) = \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ \\sum _ { t = 0 } ^ { \\infty } C ( s _ { t } ) \\right] = } \\end{array}$ $\\mathbb { E } _ { \\tau \\sim \\pi } [ \\mathbb { 1 } \\{ f a i l u r e \\} ] = \\mathbb { P } ( f a i l u r e | \\mu )$ . For policy parameterized as $\\pi _ { \\phi }$ , we denote $d ^ { \\pi } ( s )$ as $\\rho _ { \\phi } ( s )$ . Note that although $C : \\mathcal { S } \\{ 0 , 1 \\}$ takes on binary values in our setting, the function $V _ { C } ^ { \\pi } ( \\mu )$ is a continuous function of the policy $\\pi$ . ", + "bbox": [ + 173, + 536, + 825, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conservative Q Learning. CQL (Kumar et al., 2020) is a method for offline/batch RL (Lange et al., 2012; Levine et al., 2020) that aims to learn a $Q$ -function such that the expected value of a policy under the learned $Q$ function lower bounds its true value, preventing over-estimation due to out-ofdistribution actions as a result. In addition to training Q-functions via standard Bellman error, CQL minimizes the expected $Q$ -values under a particular distribution of actions, $\\mu ( a | s )$ , and maximizes the expected $\\mathrm { Q }$ -value under the on-policy distribution, $\\pi ( a | s )$ . CQL in and of itself might lead to unsafe exploration, whereas we will show in Section 3, how the theoretical tool introduced in CQL can be used to devise a safe RL algorithm. ", + "bbox": [ + 173, + 670, + 825, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3 THE CONSERVATIVE SAFE-EXPLORATION FRAMEWORK ", + "text_level": 1, + "bbox": [ + 176, + 801, + 669, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section we describe our safe exploration framework. The safety constraint $C ( s )$ defined in Section 2 is an indicator of catastrophic failure: $C ( s ) = 1$ when a state $s$ is unsafe and $C ( s ) = 0$ when it is not, and we ideally desire $C ( s ) \\ = \\ 0 \\ \\forall s \\in \\mathcal { S }$ that the agent visits. Since we do not make any assumptions in the problem structure for RL (for example a known dynamics model), formulate the constraint as we cannot guarantee this, but can at best reduce the probability of failure in every episode. So, we $V _ { C } ^ { \\pi } ( \\mu ) = \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ \\sum _ { t = 0 } ^ { \\infty } C ( s _ { t } ) \\right] \\le \\chi$ , where $\\chi \\in [ 0 , 1 )$ denotes probability of failure. Our approach is motivated by the insight that by being “conservative” with respect to how safe a state is, and hence by over-estimating this probability of failure, we can effectively ensure constrained exploration. ", + "bbox": [ + 174, + 825, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/91609bec77fec3004e9edc944501c37f5adebfd82e5aeedd77ca62fdebbeeae1.jpg", + "image_caption": [ + "Figure 1: CSC (Algorithm 1). $e n v . s t e p ( a )$ steps the simulator to the next state $s ^ { \\prime }$ and provides $R ( s , a )$ and $C \\bar { ( } s ^ { \\prime } )$ values to the agent. If $C ( s ^ { \\prime } ) = 1$ (failure), episode terminates. $Q _ { C }$ is the learned safety critic. " + ], + "image_footnote": [], + "bbox": [ + 250, + 109, + 753, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 291, + 823, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 1 provides an overview of the approach. The key idea of our algorithm is to train a conservative safety critic denoted as $Q _ { C } ( s , a )$ , that overestimates how unsafe a particular state is and modifies the exploration strategy to appropriately account for this safety under-estimate (by overestimating the probability of failure). During policy evaluation in the environment, we use the safety critic $Q _ { C } ( s , a )$ to reduce the chance of catastrophic failures by checking whether taking action $a$ in state $s$ has $Q _ { C } ( s , a )$ less than a threshold $\\epsilon$ . If not, we re-sample $a$ from the current policy $\\pi ( a | s )$ . ", + "bbox": [ + 173, + 327, + 825, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now discuss our algorithm more formally. We start by discussing the procedure for learning the safety critic $Q _ { C }$ , then discuss how we incorporate this in the policy gradient updates, and finally discuss how we perform safe exploration (Garcıa & Fernandez ´ , 2015) during policy execution in the environment. ", + "bbox": [ + 173, + 416, + 825, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall objective. Our objective is to learn an optimal policy $\\pi ^ { * }$ that maximizes task rewards, while respecting the constraint on expected probability of failures. ", + "bbox": [ + 171, + 479, + 823, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/a448bcb6cda6d2504ba423279f0159f97289f72769ec5c89973a0520887253bb.jpg", + "text": "$$\n\\pi ^ { * } = \\underset { \\pi \\in \\Pi _ { C } } { \\operatorname { a r g m a x } } V _ { R } ^ { \\pi } ( \\mu ) \\quad \\mathrm { ~ w h e r e ~ } \\quad \\Pi _ { C } = \\{ \\pi \\in \\Pi : V _ { C } ^ { \\pi } ( \\mu ) \\leq \\chi \\}\n$$", + "text_format": "latex", + "bbox": [ + 284, + 507, + 714, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning the safety critic. The safety critic $Q _ { C }$ is used to obtain an estimate of how unsafe a particular state is, by providing an estimate of probability of failure, that will be used to guide exploration. We desire the estimates to be “conservative”, in the sense that the probability of failure should be an over-estimate of the actual probability so that the agent can err on the side of caution while exploring. To train such a critic $Q _ { C }$ , we incorporate tools from CQL to estimate $Q _ { C }$ through updates similar to those obtained by reversing the sign of $\\alpha$ in Equation 2 of $\\operatorname { C Q L } ( { \\mathcal { H } } )$ (Kumar et al., 2020). This gives us an upper bound on $Q _ { C }$ instead of a lower bound, as ensured by CQL. We denote the over-estimated advantage corresponding to this safety critic as $\\hat { A } _ { C }$ . ", + "bbox": [ + 173, + 540, + 825, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally the safety critic is trained via the following objective, where the objective inside arg min is called $C Q L ( \\zeta )$ , $\\zeta$ parameterizes $Q _ { C }$ , and $k$ denotes the $k ^ { \\mathrm { { t h } } }$ update iteration. ", + "bbox": [ + 171, + 661, + 823, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/774992334fe08d658ca93bb0c6b117443c7864f57165b6790a001a83cd08e1ab.jpg", + "text": "$$\n\\begin{array} { r l } & { \\hat { Q } _ { C } ^ { k + 1 } \\gets \\underset { Q _ { C } } { \\mathrm { a r g m i n } } \\quad \\cdot \\left( - \\mathbb { E } _ { s \\sim \\mathcal { D } _ { e n v } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q _ { C } ( s , a ) ] + \\mathbb { E } _ { ( s , a ) \\sim \\mathcal { D } _ { e n v } } [ Q _ { C } ( s , a ) ] \\right) } \\\\ & { \\quad \\quad \\quad + \\frac { 1 } { 2 } \\mathbb { E } _ { ( s , a , s ^ { \\prime } , c ) \\sim \\mathcal { D } _ { e n v } } \\left[ \\left( Q _ { C } ( s , a ) - \\hat { \\mathcal { B } } ^ { \\pi _ { \\phi } } \\hat { Q } _ { C } ^ { k } ( s , a ) \\right) ^ { 2 } \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 227, + 690, + 766, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\hat { B } ^ { \\pi _ { \\phi } }$ is the empirical Bellman operator discussed in section 3.1 and equation 2 of Kumar et al. (2020). $\\alpha$ is a weight that varies the importance of the first term in equation 2, and controls the magnitude of value over-estimation, as we now highlight in red above. For states sampled from the replay buffer $\\mathcal { D } _ { e n v }$ , the first term seeks to maximize the expectation of $Q _ { C }$ over actions sampled from the current policy, while the second term seeks to minimize the expectation of $Q _ { C }$ over actions sampled from the replay buffer. $\\mathcal { D } _ { e n v }$ can include off-policy data, and also offline-data (if available). We interleave the gradient descent updates for training of $Q _ { C }$ , with gradient ascent updates for policy $\\pi _ { \\phi }$ and gradient descent updates for Lagrange multiplier $\\lambda$ , which we describe next. ", + "bbox": [ + 173, + 756, + 825, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Policy learning. Since we want to learn policies that obey the constraint we set in terms of the safety critic, we can solve the objective in equation 1 via: ", + "bbox": [ + 173, + 875, + 821, + 902 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/fa14a3dceb04d6bf2e846b4e5ecbb618dc8bd4cf1348c7699ed84b4144da18bd.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi } , a \\sim \\pi _ { \\phi } } \\big [ \\bar { A } _ { R } ^ { \\pi _ { \\phi } } ( s , a ) \\big ] \\ \\quad \\mathrm { s . t . } \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi } , a \\sim \\pi _ { \\phi } } Q _ { C } ( s , a ) \\leq \\chi\n$$", + "text_format": "latex", + "bbox": [ + 281, + 901, + 717, + 928 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can construct a Lagrangian and solve the policy optimization problem through primal dual gradient descent ", + "bbox": [ + 171, + 103, + 823, + 131 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/b118c2098667472503db7f75d7c0e582c96adbef8fdf57670448d5d9248512a3.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\operatorname* { m i n } _ { \\lambda \\geq 0 } \\mathbb { E } _ { s \\sim \\rho _ { \\phi } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi } } ( s , a ) - \\lambda \\left( Q _ { C } ( s , a ) - \\chi \\right) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 308, + 130, + 671, + 155 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can apply vanilla policy gradients or some actor-critic style Q-function approximator for optimization. Here, $Q _ { C }$ is the safety critic trained through CQL as described in equation 2. We defer specific implementation details for policy learning to the final paragraph of this section. ", + "bbox": [ + 174, + 155, + 825, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 CSC: safe exploration with conservative safety critics ", + "text_level": 1, + "bbox": [ + 174, + 212, + 614, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1: Initialize $V _ { \\theta } ^ { r }$ (task value fn), $Q _ { \\zeta } ^ { s }$ (safety critic), policy $\\pi _ { \\phi }$ , $\\lambda$ , $\\mathcal { D } _ { e n v }$ , thresholds $\\epsilon , \\delta , \\chi$ . \n2: Set $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\chi$ . . $> \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ denotes avg. failures in the previous epoch. \n3: for epochs until convergence do . Execute actions in the environment. Collect on-policy samples. \n4: for episode $e$ in $\\{ 1 , . . . , \\mathbf { M } \\}$ do \n5: Set $\\epsilon ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) )$ \n6: Sample $a \\sim \\pi _ { \\phi _ { o l d } } ( s )$ . Execute $a$ iff $Q _ { C } ( s , a ) \\leq \\epsilon$ . Else, resample $a$ . \n7: Obtain next state $s ^ { \\prime }$ ${ \\bf \\omega } ^ { \\prime } , \\dot { r } = R ( s , a )$ , $c = { C } ( { s } ^ { \\prime } )$ . \n8: $\\mathcal { D } _ { e n v } \\gets \\mathcal { D } _ { e n v } \\cup \\{ ( s , a , s ^ { \\prime } , r , c ) \\}$ . If available, $\\mathcal { D } _ { e n v }$ can be seeded with off-policy/offline data \n9: 10: end forStore th average episodic failures $\\begin{array} { r } { \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\sum _ { e = 1 } ^ { M } \\hat { V } _ { C } ^ { e } } \\end{array}$ \n11: $t$ . Policy and Q function updates using $\\mathcal { D } _ { e n v }$ \n12: Gradient ascent on $\\phi$ and (Optionally) add Entropy regularization (Appendix A.2) \n13: Gradient updates for the Q-function $\\zeta : = \\zeta - \\eta _ { Q } \\nabla _ { \\zeta } { \\bar { C } } Q L ( \\zeta )$ \n14: Gradient descent step on Lagrange multiplier $\\lambda$ (Appendix A.2) \n15: end for \n16: $\\phi _ { o l d } \\phi$ \n17: end for ", + "bbox": [ + 176, + 231, + 825, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Executing rollouts (i.e., safe exploration). Since we are interested in minimizing the number of constraint violations while exploring the environment, we do not simply execute the learned policy iterate in the environment for active data collection. Rather, we query the safety critic $Q _ { C }$ to obtain an estimate of how unsafe an action is and choose an action that is safe via rejection sampling. Formally, we sample an action $a \\sim \\pi _ { \\phi _ { o l d } } ( s )$ , and check if $Q _ { C } ( s , a ) \\leq \\epsilon$ . ", + "bbox": [ + 173, + 468, + 825, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We keep re-sampling actions $\\pi _ { \\phi _ { o l d } } ( s )$ until this condition is met, and once met, we execute that action in the environment. In practice, we execute this loop for 100 iterations, and choose the action $a$ among all actions in state $s$ for which $Q _ { C } ( s , a ) \\leq \\epsilon$ and the value of $Q _ { C } ( s , a )$ is minimum. If no such action $a$ is found that maintains $Q _ { C } ( s , a ) \\ \\leq \\ \\epsilon$ , we just choose $a$ for which $Q _ { C } ( s , a )$ is minimum (although above the threshold). ", + "bbox": [ + 173, + 545, + 825, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\epsilon$ is a threshold that varies across iterations and is defined as $\\epsilon = ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) )$ where, $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ Cis the average episodic failures in the previous epoch, denoting a sample estimate of the true $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ . This value of $\\epsilon$ is theoretically obtained such that Lemma 1 holds. ", + "bbox": [ + 173, + 622, + 825, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the replay buffer $\\mathcal { D } _ { e n v }$ , we store tuples of the form $( s , a , s ^ { \\prime } , r , c )$ , where $s$ is the previous state, $a$ is the action executed, $s ^ { \\prime }$ is the next state, $r$ is the task reward from the environment, and $c = { C } ( s ^ { \\prime } )$ , the constraint value. In our setting, $c$ is binary, with 0 denoting a live agent and 1 denoting failure. ", + "bbox": [ + 176, + 678, + 823, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall algorithm. Our overall algorithm, shown in Algorithm 1, executes policy rollouts in the environment by respecting the constraint $Q _ { C } ( s , a ) \\leq \\epsilon$ , stores the observed data tuples in the replay buffer $\\mathcal { D } _ { e n v }$ , and uses the collected tuples to train a safety value function $Q _ { C }$ using equation 2, update the policy and the dual variable $\\lambda$ following the optimization objective in equation 6. ", + "bbox": [ + 173, + 727, + 825, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation details. Here, we discuss the specifics of the implementation for policy optimization. We consider the surrogate policy improvement problem Sutton (2020): ", + "bbox": [ + 171, + 790, + 821, + 819 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/2ded846775164de4f1d5a936c569e1dd16ef25844bc57e6b8da60c7fd2c30967.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 279, + 819, + 519, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, we have introduced a $D _ { \\mathrm { K L } }$ constraint to ensure successive policies are close in order to help obtain bounds on the expected failures of the new policy in terms of the expected failures of the old policy in Section 4. We replace the $D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) )$ term by its second order Taylor expansion (expressed in terms of the Fisher Information Matrix $\\pmb { F }$ ) and enforce the resulting constraint exactly (Schulman et al., 2015a). Following equation 22 (Appendix A.2) we have, ", + "bbox": [ + 174, + 869, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 103, + 712, + 118 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/a9e56b740939c472195883b14154f854e960b1f764e64ef2ea874afb213bd8ac.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\quad \\mathrm { s . t . } \\quad V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi\n$$", + "text_format": "latex", + "bbox": [ + 171, + 118, + 803, + 150 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/7d6c3d422fea239ba91d2420f5ecd0fe703b5e260e4f5599623ac2da30c60638.jpg", + "text": "$$\n\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta\n$$", + "text_format": "latex", + "bbox": [ + 212, + 151, + 475, + 169 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We replace the true $A _ { C }$ by the learned over-estimated $\\hat { A } _ { C }$ , and consider the Lagrangian dual of this constrained problem, which we can solve by alternating gradient descent as shown below. ", + "bbox": [ + 176, + 170, + 825, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/6e5be88ae245f5e805adf9c5c5c1fb6b3bc02674d5ee1d015a40daf63011b219.jpg", + "text": "$$\n\\begin{array} { r l r } { { \\operatorname* { m a x } \\operatorname* { m i n } _ { \\lambda \\geq 0 } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } \\big ( s , a \\big ) \\Big ] - \\lambda ( V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\big ( \\mu \\big ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ \\hat { A } _ { C } ( s , a ) \\Big ] - \\chi ) } } \\\\ { \\cdot \\ L _ { \\delta } } & { \\frac { 1 } { 2 } ( \\phi - \\phi _ { o l d } ) ^ { T } F ( \\phi - \\phi _ { o l d } ) \\leq \\delta } & { ( 6 ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 199, + 826, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that although we use FIM for the updates, we can also apply vanilla policy gradients or some actor-critic style Q-function approximator to optimize equation 6. Detailed derivations of the gradient updates are in Appendix A.2. ", + "bbox": [ + 174, + 262, + 825, + 305 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 THEORETICAL ANALYSIS ", + "text_level": 1, + "bbox": [ + 176, + 333, + 416, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we aim to theoretically analyze our approach, showing that the expected probability of failures is bounded after each policy update throughout the learning process, while ensuring that the convergence rate to the optimal solution is only mildly bottlenecked by the additional safety constraint. ", + "bbox": [ + 173, + 371, + 825, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our main result, stated in Theorem 1, bounds the expected probability of failure of the policy that results from Equation 5. To prove this, we first state a Lemma that shows that the constraints in Equation 5 are satisfied with high probability during the policy updates. Detailed proofs of all the Lemmas and Theorems are in Appendix A.1. ", + "bbox": [ + 173, + 433, + 825, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Notation. Let $\\begin{array} { r } { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ( s , a ) | } \\end{array}$ and $\\Delta$ be the amount of overestimation in the expected advantage value generated from the safety critic, $\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ \\hat { A } _ { C } ( s , a ) ]$ as per equation 2, such that $\\Delta = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ \\hat { A } _ { C } ( s , a ) - A _ { C } ( s , a ) ]$ . Let $\\zeta$ denote the sampling error in the estimation of $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)$ by its sample estimate $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ (i.e. $\\zeta = | \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) - V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) | )$ and $N$ be the number of samples used in the estimation of $V _ { C }$ . Let $\\mathrm { R e g } _ { C } ( T )$ be the total cumulative failures incurred by running Algorithm 1 until $T$ samples are collected from the environment. We first show that when using Algorithm 1, we can upper bound the expectation probability of failure for each policy iterate πφold . ", + "bbox": [ + 173, + 496, + 826, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 1. If we follow Algorithm $I$ , during policy updates via Equation 5, the following is satisfied with high probability $\\geq 1 - \\omega$ ", + "bbox": [ + 173, + 631, + 823, + 659 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/60c89c0caff0a81a410dce2ebf22cbeda8d3295e1b0f79c71efe7b59303815f6.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma }\n$$", + "text_format": "latex", + "bbox": [ + 287, + 659, + 710, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\zeta$ captures sampling error in the estimation of $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)$ and we have $\\begin{array} { r } { \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | } } \\end{array}$ , where $C ^ { \\prime }$ is a constant independent of $\\omega$ obtained from union bounds and concentration inequalities (Kumar et al., 2020) and $N$ is the number of samples used in the estimation of $V _ { C }$ . ", + "bbox": [ + 173, + 695, + 825, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This lemma intuitively implies that the constraint on the safety critic in equation 5 is satisfied with a high probability, when we note that the RHS can be made small as $N$ becomes large. ", + "bbox": [ + 174, + 760, + 823, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lemma 1 had a bound in terms of $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ for the old policy $\\pi _ { \\phi _ { o l d } }$ , but not for the updated policy $\\pi _ { \\phi _ { n e w } }$ . We now show that the expected probability of failure for the policy $\\pi _ { \\phi _ { n e w } }$ resulting from solving equation 5, $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)$ is bounded with a high probability. ", + "bbox": [ + 173, + 795, + 825, + 842 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. Consider policy updates that solve the constrained optimization problem defined in Equation 5. With high probability $\\geq 1 - \\omega$ , we have the following upper bound on expected probability of failure $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu )$ for $\\pi _ { \\phi _ { n e w } }$ during every policy update iteration:√ ", + "bbox": [ + 173, + 849, + 825, + 893 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/86f8df6986698d4d85a2774b702495c31d34edc1b5c9f40486bc5dd63db63f01.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad w h e r e \\quad \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }\n$$", + "text_format": "latex", + "bbox": [ + 248, + 892, + 750, + 929 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "So far we have shown that, with high probability, we can satisfy the constraint in the objective during policy updates (Lemma 1) and obtain an upper bound on the expected probability of failure of the updated policy $\\pi _ { \\phi _ { n e w } }$ (Theorem 1). The key insight from Theorem 1 is that if we execute policy $\\pi _ { \\phi _ { n e w } }$ in the environment, the probability of failing is upper-bounded by a small number depending on the specified safety threshold $\\chi$ . Since the probability of failure is bounded, if we execute $\\pi _ { \\phi _ { n e w } }$ for multiple episodes, the total number of failures is bounded as well. ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We now bound the task performance in terms of policy return and show that incorporating and satisfying safety constraints during learning does not severely affect the convergence rate to the optimal solution for task performance. Theorem 2 builds upon and relies on the assumptions in (Agarwal et al., 2019) and extends it to our constrained policy updates in equation 5. ", + "bbox": [ + 174, + 194, + 825, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 2 (Convergence rate for policy gradient updates with the safety constraint). If we run the policy gradient updates through equation 5, for policy $\\pi _ { \\phi }$ , with $\\mu$ as the starting state distribution, with $\\phi ^ { ( 0 ) } = 0$ , learning rate $\\eta > 0$ , and choose $\\alpha$ as mentioned in the discussion of Theorem $I$ , then for all policy update iterations $T > 0$ we have, with probability $\\geq 1 - \\omega$ , ", + "bbox": [ + 173, + 253, + 825, + 313 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/b20cf45ac0b523bb22604d6a9d9933b8de39af16bdf09352a429bb2bfaa8b388.jpg", + "text": "$$\nV _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\mathcal { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + K \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T } w h e r e K \\leq ( 1 - \\chi ) + \\frac { 4 \\sqrt { 2 \\delta } \\gamma } { ( 1 - \\gamma ) ^ { 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 181, + 313, + 816, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since the value of the dual variables $\\lambda$ strictly decreases during gradient descent updates (Algorithm 1), PT −1t=0 λ(t) is upper-bounded. So, we see that the additional term proportional to K introduced in the convergence rate (compared to (Agarwal et al., 2019)) due to the safety constraint is upper bounded, and can be made small with a high probability by choosing $\\alpha$ appropriately. In addition, we note that the safety threshold $\\chi$ helps tradeoff the convergence rate by modifying the magnitude of $K$ (a low $\\chi$ means a stricter safety threshold, and a higher value of $K$ , implying a larger RHS and slower convergence). We discuss some practical considerations of the theoretical results in Appendix A.4. ", + "bbox": [ + 173, + 357, + 825, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "So far we have demonstrated that the resulting policy iterates from our algorithm all satisfy the desired safety constraint of the CMDP which allows for a maximum safety violation of $\\chi$ for every intermediate policy. While this result ensures that the probability of failures is bounded, it does not elaborate on the total failures incurred by the algorithm. In our next result, we show that the cumulative failures until a certain number of samples $T$ of the algorithm grows sublinearly when executing Algorithm 1, provided the safety threshold $\\chi$ is set in the right way. ", + "bbox": [ + 173, + 478, + 825, + 563 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 3. [Number of cumulative safety failures grows sublinearly] Let √ $\\chi$ in Algorithm 1 be timedependent such that $\\chi _ { t } \\stackrel { \\cdot } { = } \\mathcal { O } ( 1 / \\sqrt { t } )$ . Then, the total number of cumulative safety violations until when $T$ transition samples have been collected by Algorithm 1, $R e g _ { C } ( T )$ , scales sub-linearly with $T$ , i.e., $R e g _ { C } ( T ) = \\mathcal { O } ( \\sqrt { | \\mathcal { S } | | \\mathcal { A } | T } )$ . ", + "bbox": [ + 174, + 566, + 825, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A proof is provided in Appendix A.1. Theorem 3 is in many ways similar to a typical regret bound for exploration (Russo, 2019; Jaksch et al., 2010), though it measures the total number of safety violations. This means that training with Algorithm 1 will converge to a “safe” policy that incurs no failures at a quick, $\\mathcal { O } ( \\sqrt { T } )$ rate. ", + "bbox": [ + 174, + 636, + 825, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 715, + 326, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through experiments on continuous control environments of varying complexity, we aim to empirically evaluate the agreement between empirical performance and theoretical guidance by understanding the following questions: ", + "bbox": [ + 176, + 746, + 823, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "• How safe is CSC in terms of constraint satisfaction during training? • How does learning of safe policies trade-off with task performance during training? ", + "bbox": [ + 212, + 801, + 777, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1 EXPERIMENTAL SETUP ", + "text_level": 1, + "bbox": [ + 174, + 848, + 377, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Environments. In each environment, shown in Figure 2, we define a task objective that the agent must achieve and a criteria for catastrophic failure. The goal is to solve the task without dying. In point agent/car navigation avoiding traps, the agent must navigate a maze while avoiding traps. The agent has a health counter that decreases every timestep that it spends within a trap. When the counter hits 0, the agent gets trapped and dies. In Panda push without toppling, a 7-DoF Franka Emika Panda arm must push a vertically placed block across the table to a goal location without the block toppling over. Failure is defined as when the block topples. In Panda push within boundary, the Panda arm must be controlled to push a block across the table to a goal location without the block going outside a rectangular constraint region. Failure occurs when the block center of mass $( ( x , y )$ position) move outside the constraint region. In Laikago walk without falling, an 18-DoF Laikago quadruped robot must walk without falling. The agent is rewarded for walking as fast as possible (or trotting) and failure occurs when the robot falls. Since quadruped walking is an extremely challenging task, for all the baselines, we initialize the agent’s policy with a controller that has been trained to keep the agent standing, while not in motion. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d1837df1508b762b060f01b1fe0a959a844ef439597b23c4e8921dd11c577212.jpg", + "image_caption": [ + "Figure 2: Illustrations of the five environments in our experiments: (a) 2D Point agent navigation avoiding traps. (b) Car navigation avoiding traps. (c) Panda push without toppling. (d) Panda push within boundary. (e) Laikago walk without falling. " + ], + "image_footnote": [], + "bbox": [ + 194, + 106, + 803, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 238, + 825, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines and comparisons. We compare CSC to three prior methods: constrained policy optimization $( C P O )$ (Achiam et al., 2017), a standard unconstrained RL method (Schulman et al., 2015a) which we call Base (comparison with SAC (Haarnoja et al., 2018) in Appendix Figure 7), an algorithm similar to Base, called BaseShaped that modifies the reward $R ( s , a )$ as $R ( s , a ) { \\bar { } } - P C ( s )$ where $P = 1 0$ and $C ( s )$ is 1 when a failure occurs and is 0 otherwise. We also consider a method that extends Leave No Trace (Eysenbach et al., 2017) to our setting, which we refer to as $\\varrho$ ensembles. This last comparison is the most similar to our approach, in that it also implements a safety critic (adapted from LNT’s backward critic), but instead of using our conservative updates, the safety critic uses an ensemble for epistemic uncertainty estimation, as proposed by Eysenbach et al. (2017). ", + "bbox": [ + 173, + 383, + 825, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "There are other safe RL approaches which we cannot compare against, as they make multiple additional assumptions, such as the availability of a function that can be queried to determine if a state is safe or not Thananjeyan et al. (2020), availability of a default safe policy for the task Koller et al. (2018); Berkenkamp et al. (2017), and prior knowledge of the location of unsafe states (Fisac et al., 2019). In addition to the baselines (Figure 3), we analyze variants of our algorithm with different safety thresholds through ablation studies (Figure 4). We also analyze CSC and the baselines by seeding with a small amount of offline data in the Appendix A.10. ", + "bbox": [ + 173, + 530, + 825, + 627 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2 EMPIRICAL RESULTS ", + "text_level": 1, + "bbox": [ + 176, + 651, + 361, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparable or better performance with significantly lower failures during training. In Figure 3, we observe that CSC has significantly lower average failures per episode, and hence lower cumulative failures during the entire training process. Although the failures are significantly lower for our method, task performance and convergence of average task rewards is comparable to or better than all prior methods, including the Base method, corresponding to an unconstrained RL algorithm. While the $C P O$ and $Q$ -ensembles baselines also achieve near 0 average failures eventually, we see that CSC achieves this very early on during training. ", + "bbox": [ + 173, + 681, + 825, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CSC trades off performance with safety constraint satisfaction, based on the safety-threshold $\\chi$ . In Figure 4, we plot variants of our method with different safety constraint thresholds $\\chi$ . Observe that: (a) when the threshold is set to a lower value (stricter constraint), the number of avg. failures per episode decreases in all the environments, and (b) the convergence rate of the task reward is lower when the safety threshold is stricter. These observations empirically complement our theoretical guarantees in Theorems 1 and 2. We note that there are quite a few failures even in the case where $\\chi = 0 . 0$ , which is to be expected in practice because in the initial stages of training there is high function approximation error in the learned critic $Q _ { C }$ . However, we observe that the average episodic failures quickly drop below the specified threshold after about 500 episodes of training. ", + "bbox": [ + 174, + 785, + 825, + 910 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1915f199d15cdc9b08086f3e3cd493783f1f46e394d26eb495b9dd45c10d011a.jpg", + "image_caption": [ + "Figure 3: Top row: Average task rewards (higher is better). Bottom row: Average catastrophic failures (lower is better). x-axis: Number of episodes (each episode has 500 steps). Results on four of the five environments we consider for our experiments. For each environment, we plot the average task reward, the average episodic failures, and the cumulative episodic failures. The safety threshold is $\\chi = 0 . 0 3$ for all the baselines in all the environments. Results are over four random seeds. Detailed results including plots of cumulative failures are in Fig. 6 of the Appendix. " + ], + "image_footnote": [], + "bbox": [ + 189, + 102, + 792, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 397, + 344, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We discuss prior safe RL and safe control methods under three subheadings ", + "bbox": [ + 176, + 435, + 669, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assuming prior domain knowledge of the problem structure. Prior works have attempted to solve safe exploration in the presence of structural assumptions about the environment or safety structures. For example, Koller et al. (2018); Berkenkamp et al. (2017) assume access to a safe set of environment states, and a default safe policy, while in Fisac et al. (2018); Dean et al. (2019), knowledge of system dynamics is assumed and (Fisac et al., 2019) assume access to a distance metric on the state space. SAVED (Thananjeyan et al., 2020) learns a kernel density estimate over unsafe states, and assumes access to a set of user demonstrations and a user specified function that can be queried to determine whether a state is safe or not. In contrast to these approaches, our method does not assume any prior knowledge from the user, or domain knowledge of the problem setting, except a binary signal from the environment indicating when a catastrophic failure has occurred. ", + "bbox": [ + 174, + 458, + 825, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Assuming a continuous safety cost function. CPO (Achiam et al., 2017), and (Chow et al., 2019) assume a cost function can be queried from the environment at every time-step and the objective is to keep the cumulative costs within a certain limit. This assumption limits the generality of the method in scenarios where only minimal feedback, such as binary reward feedback is provided (additional details in section A.3). ", + "bbox": [ + 174, + 603, + 823, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(Stooke et al., 2020) devise a general modification to the Lagrangian by incorporating two additional terms in the optimization of the dual variable. SAMBA (Cowen-Rivers et al., 2020) has a learned GP dynamics model and a continuous constraint cost function that encodes safety. The objective is to minimize task cost function while maintaining the $\\mathrm { C V A R } _ { \\alpha }$ of cumulative costs below a threshold. In the work of Dalal et al. (2018); Paternain et al. (2019b;a); Grbic & Risi (2020), only the optimal policy is learned to be safe, and there are no safety constraint satisfactions during training. In contrast to these approaches, we assume only a binary signal from the environment indicating when a catastrophic failure has occurred. Instead of minimizing expected costs, our constraint formulation directly seeks to constrain the expected probability of failure. ", + "bbox": [ + 174, + 680, + 825, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Safety through recoverability. Prior works have attempted to devise resetting mechanisms to recover the policy to a base configuration from (near) a potentially unsafe state. LNT (Eysenbach et al., 2017) trains both a forward policy for solving a task, and a reset goal-conditioned policy that kicks in when the agent is in an unsafe state and learns an ensemble of critics, which is substantially more complex than our approach of a learned safety critic, which can give rise to a simple but provable safe exploration algorithm. Concurrently to us, SQRL (Srinivasan et al., 2020) developed an approach also using safety critics such that during the pre-training phase, the agent explores both safe and unsafe states in the environment for training the critic. ", + "bbox": [ + 174, + 811, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72a5113a9ad7cab90ac8d5079ca9ac07eef0a9672209e5a6290e3581099b0d19.jpg", + "image_caption": [ + "Figure 4: Top row: Average task rewards (higher is better). Bottom row: Average catastrophic failures (lower is better). $\\mathbf { X }$ -axis: Number of episodes (each episode has 500 steps). Results on four of the five environments we consider for our experiments. For each environment we plot the average task reward, the average episodic failures, and the cumulative episodic failures. All the plots are for our method (CSC) with different safety thresholds $\\chi$ , specified in the legend. From the plots it is evident that our method can naturally trade-off safety for task performance depending on how strict the safety threshold is set to. Results are over four random seeds. Detailed results including plots of cumulative failures are in Fig. 5 of the Appendix. " + ], + "image_footnote": [], + "bbox": [ + 199, + 111, + 756, + 291 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "In control theory, a number of prior works have focused on Hamilton-Jacobi-Isaacs (HJI) reachability analysis (Bansal et al., 2017) for providing safety constraint satisfactions and obtaining control inputs for dynamical systems (Herbert et al., 2019; Bajcsy et al., 2019; Leung et al., 2018). Our method does not require knowledge of the system dynamics or regularity conditions on the statespace, which are crucial for computing unsafe states using HJI reachability. ", + "bbox": [ + 174, + 397, + 825, + 467 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "7 DISCUSSION, LIMITATIONS, AND CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 484, + 598, + 501 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We introduced a safe exploration algorithm to learn a conservative safety critic that estimates the probability of failure for each candidate state-action tuple, and uses this to constrain policy evaluation and policy improvement. We provably demonstrated that the probability of failures is bounded throughout training and provided convergence results showing how ensuring safety does not severely bottleneck task performance. We empirically validated our theoretical results and showed that we achieve high task performance while incurring low accidents during training. ", + "bbox": [ + 174, + 512, + 825, + 595 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "While our theoretical results demonstrated that the probability of failures is bounded with a high probability, one limitation is that we still observe non-zero failures empirically even when the threshold $\\chi$ is set to 0. This is primarily because of neural network function approximation error in the early stages of training the safety critic, which we cannot account for precisely in the theoretical results, and also due to the fact that we bound the probability of failures, which in practice means that the number of failures is also bounded, but non-zero. We also showed that if we set the constraint threshold in an appropriate time-varying manner, training with CSC incurs cumulative failures that scales at most sub-linearly with the number of transition samples in the environment. ", + "bbox": [ + 174, + 603, + 825, + 714 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Although our approach bounds the probability of failure and is general in the sense that it does not assume access any user-specified constraint function, in situations where the task is difficult to solve, for example due to stability concerns of the agent, our approach will fail without additional assumptions. In such situations, some interesting future work directions would be to develop a curriculum of tasks to start with simple tasks where safety is easier to achieve, and gradually move towards more difficult tasks, such that the learned knowledge from previous tasks is not forgotten. ", + "bbox": [ + 174, + 720, + 825, + 804 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "ACKNOWLEDGEMENT ", + "text_level": 1, + "bbox": [ + 176, + 820, + 326, + 833 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Vector Institute, Toronto and the Department of Computer Science, University of Toronto for compute support. We thank Glen Berseth and Kevin Xie for helpful initial discussions about the project, Alexandra Volokhova, Arthur Allshire, Mayank Mittal, Samarth Sinha, and Irene Zhang for feedback on the paper, and other members of the UofT CS Robotics Group for insightful discussions during internal presentations and reading group sessions. Finally, we are grateful to the anonymous ICLR 2021 reviewers for their feedback in helping improve the paper. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 176, + 102, + 287, + 118 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Joshua Achiam, David Held, Aviv Tamar, and Pieter Abbeel. Constrained policy optimization. arXiv preprint arXiv:1705.10528, 2017. ", + "bbox": [ + 171, + 126, + 823, + 155 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alekh Agarwal, Sham M Kakade, Jason D Lee, and Gaurav Mahajan. Optimality and approximation with policy gradient methods in markov decision processes. arXiv preprint arXiv:1908.00261, 2019. ", + "bbox": [ + 176, + 165, + 823, + 207 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Eitan Altman. Constrained Markov decision processes, volume 7. CRC Press, 1999. ", + "bbox": [ + 174, + 217, + 727, + 232 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Andrea Bajcsy, Somil Bansal, Eli Bronstein, Varun Tolani, and Claire J Tomlin. An efficient reachability-based framework for provably safe autonomous navigation in unknown environments. In 2019 IEEE 58th Conference on Decision and Control (CDC), pp. 1758–1765. IEEE, 2019. ", + "bbox": [ + 174, + 241, + 825, + 297 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Somil Bansal, Mo Chen, Sylvia Herbert, and Claire J Tomlin. Hamilton-jacobi reachability: A brief overview and recent advances. In 2017 IEEE 56th Annual Conference on Decision and Control (CDC), pp. 2242–2253. IEEE, 2017. ", + "bbox": [ + 174, + 308, + 825, + 351 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Felix Berkenkamp, Matteo Turchetta, Angela Schoellig, and Andreas Krause. Safe model-based reinforcement learning with stability guarantees. In Advances in neural information processing systems, pp. 908–918, 2017. ", + "bbox": [ + 173, + 359, + 825, + 402 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Yinlam Chow, Ofir Nachum, Aleksandra Faust, Edgar Duenez-Guzman, and Mohammad Ghavamzadeh. Lyapunov-based safe policy optimization for continuous control. arXiv preprint arXiv:1901.10031, 2019. ", + "bbox": [ + 174, + 412, + 825, + 455 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Alexander I Cowen-Rivers, Daniel Palenicek, Vincent Moens, Mohammed Abdullah, Aivar Sootla, Jun Wang, and Haitham Ammar. Samba: Safe model-based & active reinforcement learning. arXiv preprint arXiv:2006.09436, 2020. ", + "bbox": [ + 173, + 464, + 825, + 507 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Gal Dalal, Krishnamurthy Dvijotham, Matej Vecerik, Todd Hester, Cosmin Paduraru, and Yuval Tassa. Safe exploration in continuous action spaces. arXiv preprint arXiv:1801.08757, 2018. ", + "bbox": [ + 171, + 517, + 825, + 546 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sarah Dean, Stephen Tu, Nikolai Matni, and Benjamin Recht. Safely learning to control the constrained linear quadratic regulator. In 2019 American Control Conference (ACC), pp. 5582–5588. IEEE, 2019. ", + "bbox": [ + 174, + 555, + 825, + 598 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Benjamin Eysenbach, Shixiang Gu, Julian Ibarz, and Sergey Levine. Leave no trace: Learning to reset for safe and autonomous reinforcement learning. arXiv preprint arXiv:1711.06782, 2017. ", + "bbox": [ + 173, + 608, + 821, + 637 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jaime F Fisac, Anayo K Akametalu, Melanie N Zeilinger, Shahab Kaynama, Jeremy Gillula, and Claire J Tomlin. A general safety framework for learning-based control in uncertain robotic systems. IEEE Transactions on Automatic Control, 64(7):2737–2752, 2018. ", + "bbox": [ + 174, + 646, + 825, + 689 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jaime F Fisac, Neil F Lugovoy, Vicenc¸ Rubies-Royo, Shromona Ghosh, and Claire J Tomlin. Bridging hamilton-jacobi safety analysis and reinforcement learning. In 2019 International Conference on Robotics and Automation (ICRA), pp. 8550–8556. IEEE, 2019. ", + "bbox": [ + 173, + 699, + 825, + 742 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Javier Garcıa and Fernando Fernandez. A comprehensive survey on safe reinforcement learning. ´ Journal of Machine Learning Research, 16(1):1437–1480, 2015. ", + "bbox": [ + 168, + 752, + 821, + 781 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Djordje Grbic and Sebastian Risi. Safe reinforcement learning through meta-learned instincts. arXiv preprint arXiv:2005.03233, 2020. ", + "bbox": [ + 169, + 790, + 825, + 819 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Offpolicy maximum entropy deep reinforcement learning with a stochastic actor. arXiv preprint arXiv:1801.01290, 2018. ", + "bbox": [ + 173, + 829, + 825, + 871 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Sylvia L Herbert, Somil Bansal, Shromona Ghosh, and Claire J Tomlin. Reachability-based safety guarantees using efficient initializations. In 2019 IEEE 58th Conference on Decision and Control (CDC), pp. 4810–4816. IEEE, 2019. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Thomas Jaksch, Ronald Ortner, and Peter Auer. Near-optimal regret bounds for reinforcement learning. Journal of Machine Learning Research, 11(4), 2010. ", + "bbox": [ + 171, + 103, + 823, + 133 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sham Kakade and John Langford. Approximately optimal approximate reinforcement learning. In ICML, volume 2, pp. 267–274, 2002. ", + "bbox": [ + 173, + 140, + 823, + 169 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Torsten Koller, Felix Berkenkamp, Matteo Turchetta, and Andreas Krause. Learning-based model predictive control for safe exploration. In 2018 IEEE Conference on Decision and Control (CDC), pp. 6059–6066. IEEE, 2018. ", + "bbox": [ + 174, + 176, + 823, + 219 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative q-learning for offline reinforcement learning. arXiv preprint arXiv:2006.04779, 2020. ", + "bbox": [ + 171, + 227, + 823, + 256 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sascha Lange, Thomas Gabel, and Martin Riedmiller. Batch reinforcement learning. In Reinforcement learning, pp. 45–73. Springer, 2012. ", + "bbox": [ + 176, + 263, + 823, + 292 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Karen Leung, Edward Schmerling, Mo Chen, John Talbot, J Christian Gerdes, and Marco Pavone. On infusing reachability-based safety assurance within probabilistic planning frameworks for human-robot vehicle interactions. In International Symposium on Experimental Robotics, pp. 561–574. Springer, 2018. ", + "bbox": [ + 173, + 300, + 823, + 356 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sergey Levine. Deep reinforcement learning course, 2018. URL http://rail.eecs. berkeley.edu/deeprlcourse-fa18/static/slides/. ", + "bbox": [ + 174, + 363, + 820, + 393 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tutorial, review, and perspectives on open problems. arXiv preprint arXiv:2005.01643, 2020. ", + "bbox": [ + 174, + 401, + 821, + 430 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Santiago Paternain, Miguel Calvo-Fullana, Luiz FO Chamon, and Alejandro Ribeiro. Safe policies for reinforcement learning via primal-dual methods. arXiv preprint arXiv:1911.09101, 2019a. ", + "bbox": [ + 173, + 436, + 823, + 467 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Santiago Paternain, Luiz Chamon, Miguel Calvo-Fullana, and Alejandro Ribeiro. Constrained reinforcement learning has zero duality gap. In Advances in Neural Information Processing Systems, pp. 7555–7565, 2019b. ", + "bbox": [ + 173, + 473, + 823, + 516 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. arXiv preprint arXiv:2004.00784, 2020. ", + "bbox": [ + 173, + 523, + 823, + 554 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Alex Ray, Joshua Achiam, and Dario Amodei. Benchmarking safe exploration in deep reinforcement learning. arXiv preprint arXiv:1910.01708, 2019. ", + "bbox": [ + 174, + 560, + 823, + 590 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Daniel Russo. Worst-case regret bounds for exploration via randomized value functions. In Advances in Neural Information Processing Systems, pp. 14433–14443, 2019. ", + "bbox": [ + 173, + 597, + 823, + 627 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pp. 1889–1897, 2015a. ", + "bbox": [ + 171, + 633, + 820, + 664 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. Highdimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438, 2015b. ", + "bbox": [ + 171, + 670, + 825, + 713 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Krishnan Srinivasan, Benjamin Eysenbach, Sehoon Ha, Jie Tan, and Chelsea Finn. Learning to be safe: Deep rl with a safety critic. arXiv preprint arXiv:2010.14603, 2020. ", + "bbox": [ + 171, + 720, + 823, + 751 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Adam Stooke, Joshua Achiam, and Pieter Abbeel. Responsive safety in reinforcement learning by pid lagrangian methods. arXiv preprint arXiv:2007.03964, 2020. ", + "bbox": [ + 173, + 757, + 821, + 786 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Rich Sutton. Reinforcement learning book, 2020. URL http://incompleteideas.net/ book/first/ebook/node42.html. ", + "bbox": [ + 174, + 794, + 820, + 824 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Chen Tessler, Daniel J Mankowitz, and Shie Mannor. Reward constrained policy optimization. arXiv preprint arXiv:1805.11074, 2018. ", + "bbox": [ + 173, + 830, + 823, + 859 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Brijen Thananjeyan, Ashwin Balakrishna, Ugo Rosolia, Felix Li, Rowan McAllister, Joseph E Gonzalez, Sergey Levine, Francesco Borrelli, and Ken Goldberg. Safety augmented value estimation from demonstrations (saved): Safe deep model-based rl for sparse cost robotic tasks. IEEE Robotics and Automation Letters, 5(2):3612–3619, 2020. ", + "bbox": [ + 174, + 867, + 825, + 924 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Yuke Zhu, Josiah Wong, Ajay Mandlekar, and Roberto Mart´ın-Mart´ın. robosuite: A modular simulation framework and benchmark for robot learning. In arXiv preprint arXiv:2009.12293, 2020. ", + "bbox": [ + 176, + 103, + 823, + 132 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 299, + 118 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "A.1 PROOFS OF ALL THEOREMS AND LEMMAS ", + "text_level": 1, + "bbox": [ + 174, + 136, + 509, + 151 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Note. During policy updates via Equation 5, the $D _ { \\mathrm { K L } }$ constraint is satisfied with high probability if we follow Algorithm 1. ", + "bbox": [ + 171, + 162, + 825, + 193 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Following the steps in the Appendix A.2, we can write the gradient ascent step for $\\phi$ as ", + "bbox": [ + 168, + 198, + 756, + 213 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/20e174e74bf32ba12da247e0ccedaca9e836e3f2040c600302a28ada6175cb17.jpg", + "text": "$$\n\\phi \\phi _ { o l d } + \\beta { \\pmb F } ^ { - 1 } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) \\quad \\beta = \\beta ^ { j } \\sqrt { \\frac { 2 \\delta } { \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } { \\pmb F } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) } }\n$$", + "text_format": "latex", + "bbox": [ + 245, + 213, + 753, + 256 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "$\\pmb { F }$ can be estimated with samples as ", + "bbox": [ + 176, + 253, + 411, + 267 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/a171350c0fdb713b40f271ee6b08fb47c26342cbe3336c7330e548d3718b7449.jpg", + "text": "$$\n\\begin{array} { r } { \\pmb { F } = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ) ^ { T } \\right] \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 290, + 267, + 705, + 294 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Here $\\beta ^ { j }$ is the backtracking coefficient and we perform backtracking line search with exponential decay. $\\nabla _ { \\phi _ { o l d } } \\tilde { J } \\big ( \\phi _ { o l d } \\big )$ is calculated as, ", + "bbox": [ + 173, + 295, + 825, + 324 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/b1ed89bb40f0ced3396fcfcaf6548508e61cf4a453de9a0c4b3b47093c08aa96.jpg", + "text": "$$\n\\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\tilde { A } _ { R } ^ { \\pi _ { \\phi _ { o l d } } } \\right]\n$$", + "text_format": "latex", + "bbox": [ + 287, + 324, + 710, + 351 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "After every update, we check if $\\bar { D } _ { \\mathrm { K L } } ( \\phi | | \\phi _ { o l d } ) \\leq \\delta$ , and if not we decay $\\beta ^ { j } = \\beta ^ { j } ( 1 - \\beta ^ { j } ) ^ { j }$ , set $j j + 1$ and repeat for $L$ steps until $D _ { \\mathrm { K L } } \\leq \\delta$ is satisfied. If this is not satisfied after $L$ steps, we backtrack, and do not update $\\phi$ i.e. set $\\phi \\phi _ { o l d }$ . ", + "bbox": [ + 173, + 357, + 826, + 401 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Lemma 1. If we follow Algorithm $I$ , during policy updates via equation 5, the following is satisfied with high probability $\\geq 1 - \\omega$ ", + "bbox": [ + 173, + 406, + 825, + 434 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/8637d9ec5ff3e6a811968acadac65fe7633c9ec2755d5b4611ed698c330ec4a2.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma }\n$$", + "text_format": "latex", + "bbox": [ + 289, + 434, + 710, + 467 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Here, $\\zeta$ captures sampling error in the estimation of $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)$ and we have $\\zeta \\leq \\frac { C \\sqrt { \\log ( 1 / \\omega ) } } { | N | }$ , where $C$ is a constant and $N$ is the number of samples used in the estimation of $V _ { C }$ . ", + "bbox": [ + 173, + 470, + 823, + 506 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Proof. Based on line 6 of Algorithm 1, for every rollout $\\{ ( s , a ) \\}$ , the following holds: ", + "bbox": [ + 174, + 531, + 736, + 546 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/235df6c75aa93a2ddd0d6de6c2d8ad22c0b329baf8e1a734fd442d8e2ac7c119.jpg", + "text": "$$\n\\begin{array} { r l } & { \\quad Q _ { C } ( s , a ) \\leq ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) ) ) \\quad \\forall ( s , a ) } \\\\ & { \\Longrightarrow \\hat { A } _ { C } ( s , a ) \\leq ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) ) ) \\quad \\forall ( s , a ) } \\\\ & { \\Longrightarrow \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) + \\cfrac { 1 } { 1 - \\gamma } \\hat { A } _ { C } ( s , a ) \\leq \\chi \\quad \\forall ( s , a ) } \\\\ & { \\Longrightarrow \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) + \\cfrac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right] \\leq \\chi } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 313, + 547, + 687, + 654 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "We note that we can only compute a sample estimate $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ instead of the true quantity $V _ { C }$ which can introduce sampling error in practice. In order to ensure that $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ is not much lesser than $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ , we can obtain a bound on their difference. Note that if $\\tilde { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\geq V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ , the Lemma holds directly, so we only need to consider the less than case. ", + "bbox": [ + 174, + 654, + 823, + 717 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Let $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) = V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) - \\zeta$ . With high probability $\\geq 1 - \\omega$ , we can ensure $\\begin{array} { r } { \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | } } \\end{array}$ , where $C ^ { \\prime }$ is a constant independent of $\\omega$ (obtained from union bounds and concentration inequalities) and $N$ is the number of samples used in the estimation of $V _ { C }$ . In addition, our estimate of $\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right]$ is an overestimate of the true $\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right]$ , and we denote their difference by $\\Delta$ . ", + "bbox": [ + 173, + 724, + 826, + 814 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "So, with high probability $\\geq 1 - \\omega$ , we have ", + "bbox": [ + 174, + 819, + 460, + 834 + ], + "page_idx": 12 + }, + { + "type": "equation", + "img_path": "images/53368a2de0a0d4d3e2aa6de9c6bd01d3afd077f072993f0a2ebc0b5d4c946a2d.jpg", + "text": "$$\n\\begin{array} { r l } & { \\quad \\hat { V } _ { C } ^ { \\hat { \\pi _ { \\phi _ { o l d } } } } ( \\mu ) + \\displaystyle \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right] \\leq \\chi } \\\\ & { \\Longrightarrow V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\displaystyle \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 272, + 833, + 728, + 900 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Theorem 1. Consider policy updates that solve the constrained optimization problem defined in equation 5. With high probability $\\geq 1 - \\omega$ , we have the following upper bound on expected probability of failure $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu )$ for $\\pi _ { \\phi _ { n e w } }$ during every policy update iteration ", + "bbox": [ + 171, + 103, + 825, + 147 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/1d6f00221f9072e1b7fe70ec845c18e583d98d7584cb656394919845f6fb789e.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi n e w } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad w h e r e \\quad \\zeta \\leq \\frac { C \\sqrt { \\log ( 1 / \\omega ) } } { | N | }\n$$", + "text_format": "latex", + "bbox": [ + 250, + 146, + 748, + 183 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here, $\\begin{array} { r } { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ( s , a ) | } \\end{array}$ and $\\Delta$ is the overestimation in $\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ A _ { C } ( s , a ) ]$ due to CQL. ", + "bbox": [ + 173, + 200, + 826, + 233 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proof. $C ( s )$ denotes the value of the constraint function from the environment in state $s$ . This is analogous to the task reward function $R ( s , a )$ . In our case $C ( s )$ is a binary indicator of whether a catastrophic failure has occurred, however the analysis we present holds even when $C ( s )$ is a shaped continuous cost function. ", + "bbox": [ + 173, + 306, + 825, + 362 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/2859f99d83e1b0c754fc5beb84f94d027981b402837fa9e88d0955a387c53aca.jpg", + "text": "$$\nC ( s ) = { \\left\\{ \\begin{array} { l l } { 1 , } & { \\mathbb { I } \\{ f a i l u r e \\} = 1 } \\\\ { 0 , } & { { \\mathrm { o t h e r w i s e } } } \\end{array} \\right. }\n$$", + "text_format": "latex", + "bbox": [ + 392, + 361, + 602, + 396 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Let $V _ { R } ^ { \\pi _ { \\phi } } ( \\mu )$ denotes the discounted task rewards obtained in expectation by executing policy $\\pi _ { \\phi }$ for one episode, and let $V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )$ denote the corresponding constraint values. ", + "bbox": [ + 174, + 397, + 826, + 429 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/db3e9a8d1babb77956b9a14014590fd08b0a94f16297c09eca84b14e7257a0bc.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) \\quad s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi\n$$", + "text_format": "latex", + "bbox": [ + 382, + 445, + 616, + 472 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "From the TRPO (Schulman et al., 2015a) and CPO (Achiam et al., 2017) papers, following similar derivations, we obtain the following bounds ", + "bbox": [ + 174, + 478, + 823, + 506 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/a218ecb126ce75ca306a25717c4f1946f018103862a2b6cf3c0560ce1b319221.jpg", + "text": "$$\nV _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) - V _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\geq \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\frac { 2 \\gamma \\epsilon _ { R } } { 1 - \\gamma } \\mathrm { D } _ { T V } ( \\pi _ { \\phi } | | \\pi _ { \\phi _ { o l d } } ) [ s ] \\right]\n$$", + "text_format": "latex", + "bbox": [ + 191, + 506, + 779, + 540 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here, $A _ { R } ^ { \\pi _ { \\phi } }$ is the advantage function corresponding to the task rewards and $\\begin{array} { r l } { \\epsilon _ { R } } & { { } = } \\end{array}$ $\\begin{array} { r } { \\operatorname* { m a x } _ { s } \\vert \\mathbb { E } _ { a \\sim \\pi _ { \\phi } } A _ { R } ^ { \\pi _ { \\phi } } ( s , a ) \\vert } \\end{array}$ . $\\mathbf { D } _ { T V }$ is the total variation distance. We also have, ", + "bbox": [ + 173, + 541, + 825, + 571 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/1c15909c81029537d1068394727e41a6d240f96d08b1b6564c095d5a91cedf69.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) - V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\leq \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) + \\frac { 2 \\gamma \\epsilon _ { C } } { 1 - \\gamma } \\mathrm { D } _ { T V } ( \\pi _ { \\phi } | | \\pi _ { \\phi _ { o l d } } ) [ s ] \\right]\n$$", + "text_format": "latex", + "bbox": [ + 191, + 589, + 781, + 623 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Here, $A _ { C } ^ { \\pi _ { \\phi _ { o l d } } }$ is the advantage function corresponding to the costs and $\\begin{array} { r l } { \\epsilon _ { C } } & { { } = } \\end{array}$ $\\begin{array} { r } { \\operatorname* { m a x } _ { s } \\vert \\mathbb { E } _ { a \\sim \\pi _ { \\phi } } A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\vert } \\end{array}$ . In our case, $A _ { C }$ is defined in terms of the safety $\\mathrm { Q }$ function $Q _ { C } ( s , a )$ , and CQL can bound its expectation directly. To see this, note that, by definition $\\begin{array} { r l r } { { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { = } & { { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ Q _ { \\zeta } ( s , a ) \\right] - { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ Q _ { \\zeta } ( s , a ) \\right] . } \\end{array}$ Here, the RHS is precisely the term in equation 2 of (Kumar et al., 2020) that is bounded by CQL. We get an overstimated advantage $\\hat { A } _ { C } ( s , a )$ from training the safety critic $Q _ { C }$ through updates in equation 2. . Let $\\Delta$ denote the expected magnitude of over-estimate ${ { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right] \\ = \\ { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] + \\Delta }$ , where $\\Delta$ is positive. Note that replacing $A _ { C }$ , by its over-estimate $\\hat { A } _ { C }$ , the inequality in 16 above still holds. ", + "bbox": [ + 173, + 623, + 826, + 779 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Using Pinsker’s inequality, we can convert the bounds in terms of $\\mathrm { D } _ { K L }$ instead of $\\mathrm { D } _ { T V }$ , ", + "bbox": [ + 171, + 784, + 751, + 799 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/a57416cd4f92612230b4fbeacc0c04d95c29cbacad2491b63f970de595191866.jpg", + "text": "$$\n\\begin{array} { r } { \\mathbf { D } _ { T V } ( p | | q ) \\leq \\sqrt { \\mathbf { D } _ { K L } ( p | | q ) / 2 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 398, + 799, + 599, + 818 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "By Jensen’s inequality, ", + "bbox": [ + 174, + 816, + 326, + 832 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/b9ad5dfc3236d36de166096d867695dc45241eafe5c22391721770e6538ba789.jpg", + "text": "$$\n\\mathbb { E } [ \\sqrt { \\mathbf { D } _ { K L } ( p | | q ) / 2 } ] \\le \\sqrt { \\mathbb { E } [ \\mathbf { D } _ { K L } ( p | | q ) ] / 2 }\n$$", + "text_format": "latex", + "bbox": [ + 361, + 832, + 637, + 851 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "So, we can replace the $\\mathbb { E } [ \\mathbf { D } _ { T V } ( p | | q ) ]$ terms in the bounds by $\\sqrt { \\mathbb { E } [ \\mathbf { D } _ { K L } ( \\boldsymbol { p } | | \\boldsymbol { q } ) ] }$ . Then, inequation 16 becomes, ", + "bbox": [ + 174, + 852, + 826, + 881 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/686ef07920eea40b482aae0a0d2b9a9214c000292e8e0bf734b6b4b2474d6a96.jpg", + "text": "$$\n\\tau _ { C } ^ { \\pi \\phi } ( \\mu ) - V _ { C } ^ { \\pi \\phi _ { o l d } } ( \\mu ) \\leq \\frac { 1 } { 1 - \\gamma } \\left[ \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] + \\frac { 2 \\gamma \\epsilon _ { C } } { 1 - \\gamma } \\sqrt { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\mathbf { D } _ { K L } ( \\pi _ { \\phi } | | \\pi _ { \\phi } ) \\right] } \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 181, + 878, + 839, + 912 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Re-visiting our objective in equation 5, ", + "bbox": [ + 174, + 103, + 429, + 118 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/010683a4a2d2afb9595ca5f932ca9db8d59884cff4bd17acda556198bbef87f0.jpg", + "text": "$$\n\\begin{array} { r l } { \\displaystyle \\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { } \\\\ { s . t . \\quad } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . \\quad } & { V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 346, + 117, + 651, + 188 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "From inequation 19 we note that instead of of constraining $V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )$ we can constrain an upper bound on this. Writing the constraint in terms of the current policy iterate $\\pi _ { \\phi _ { o l d } }$ using equation 19, ", + "bbox": [ + 173, + 188, + 825, + 217 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/d0bb7882d86cdefe5ba249e203ce8600ea0482ba9bfddcd1fac872ff2defaad2.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathsf { \\Gamma } _ { \\hat { \\varphi } _ { n \\epsilon w } } = \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\bar { \\mathbb { E } } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } \\\\ & { \\quad s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi o l d } ( \\cdot \\vert s ) \\vert \\vert \\pi _ { \\phi } ( \\cdot \\vert s ) ) ] \\le \\delta } \\\\ & { \\quad s . t . \\quad V _ { G } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\displaystyle \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] + \\beta \\sqrt { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot \\vert s ) \\vert \\vert \\pi _ { \\phi } ( \\cdot \\vert s ) ) ] } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 173, + 215, + 839, + 301 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "As there is already a bound on $D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ]$ , getting rid of the redundant term, we define the following optimization problem, which we actually optimize for ", + "bbox": [ + 173, + 320, + 826, + 351 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/56567d4641b519c63653a24cc666377d1ef8437f0595b1d6f4098f742a6b5922.jpg", + "text": "$$\n\\begin{array} { r l } { \\pi _ { \\phi _ { n e w } } = \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { } \\\\ { s . t . } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . } & { V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\leq \\chi } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 267, + 366, + 728, + 452 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Upper bound on expected probability of failures. If $\\pi _ { \\phi _ { n e w } }$ is updated using equation 5, then we have the following upper bound on $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)$ ", + "bbox": [ + 173, + 455, + 826, + 487 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/3b5772a8d0e54ec82163767ccbe61cf6ae6e4aa8626b038360f2e27fe070381c.jpg", + "text": "$$\n\\gamma _ { C } ^ { \\pi _ { \\partial u _ { c o u l } } } ( \\mu ) \\leq V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\Big ] + \\frac { 2 \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\sqrt { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\big [ \\mathrm { D } _ { K L } \\big ( \\pi _ { \\phi } \\big | \\big | \\pi _ { \\phi } \\big ) \\big | ^ { 2 } \\big ] } .\n$$", + "text_format": "latex", + "bbox": [ + 181, + 486, + 839, + 518 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "If we ensure $\\begin{array} { r } { V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\leq \\chi } \\end{array}$ holds by following Algorithm 1,we have the following upper bound on $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)$ ", + "bbox": [ + 173, + 534, + 826, + 574 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/09c0f4d8645afd77a2138ae33e7473f59db5fc8dd38aad16092fa04ee2ac3c3f.jpg", + "text": "$$\n\\begin{array} { c } { { V _ { C } ^ { \\pi _ { \\theta n e w } } ( \\mu ) \\leq \\chi + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } } } \\\\ { { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ^ { \\pi _ { \\phi _ { n d } } } ( s , a ) | . } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 176, + 574, + 596, + 630 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Now, instead of $A _ { C } ( s , a )$ , we have an over-estimated advantage estimate $\\hat { A } _ { C } ( s , a )$ obtained by training the safety critic $Q _ { C }$ through CQL as in equation 2. Let $\\Delta$ denote the expected magnitude of over-estimate $\\begin{array} { r } { \\dot { \\mathbb { E } } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A _ { C } } ( s , a ) \\right] = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] + \\Delta } \\end{array}$ , where $\\Delta$ is positive. ", + "bbox": [ + 173, + 636, + 825, + 689 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "From Lemma 1, we are able to ensure the following with high probability $\\geq 1 - \\omega$ ", + "bbox": [ + 176, + 694, + 715, + 708 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/05c372d896ca4c7414b07d150ade5edb63f9da732adbc74ffbf49442df2801bc.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma }\n$$", + "text_format": "latex", + "bbox": [ + 287, + 708, + 714, + 741 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "By combining this with the upper bound on $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)$ from inequality 23, we obtain with probability $\\geq 1 - \\omega$ ", + "bbox": [ + 173, + 741, + 821, + 770 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/d9c2090609f469de6ce691aaa314a3d576d1283f3c5641e6820cca9bb41841e0.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad \\mathrm { w h e r e } \\qquad \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }\n$$", + "text_format": "latex", + "bbox": [ + 246, + 768, + 750, + 805 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Since $\\epsilon _ { C }$ depends on the optimized policy $\\pi _ { \\phi _ { n e w } }$ , it can’t be calculated exactly prior to the update. As we cap $Q _ { C } ( s , a )$ to be $\\leq 1$ , therefore, the best bound we can construct for $\\epsilon _ { C }$ is the trivial bound $\\epsilon _ { C } \\leq 2$ . Now, in order to have $V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) < \\chi$ , we require $\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta } \\end{array}$ . To guarantee this, replacing $\\Delta$ by the exact overestimation term from CQL, we have the following condition on ", + "bbox": [ + 173, + 861, + 825, + 925 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "$\\alpha$ : ", + "bbox": [ + 173, + 106, + 192, + 117 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/40145cc844c7ef268947fd422d355f5c96b6fe184eee42f61982b3d14ed7414c.jpg", + "text": "$$\n\\alpha > \\frac { \\mathcal { G } _ { c , T } } { 1 - \\gamma } \\cdot \\operatorname* { m a x } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } } \\left( \\frac { 1 } { \\vert \\sqrt { \\mathcal { D } _ { \\phi _ { o l d ^ { \\prime } } } } \\vert } + \\frac { 2 \\sqrt { 2 \\delta } \\gamma + ( 1 - \\gamma ) ^ { 2 } \\zeta } { \\mathcal { G } _ { c , T } } \\right) \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left( \\frac { \\pi _ { \\phi _ { o l d } } } { \\pi _ { \\phi _ { o l d ^ { \\prime } } } } - 1 \\right) \\right] ^ { - 1 }\n$$", + "text_format": "latex", + "bbox": [ + 187, + 114, + 781, + 156 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Here, $\\mathcal { G } _ { c , T }$ is a constant depending on the concentration properties of the safety constraint function $C ( s , a )$ and the state transition operator $T ( s ^ { \\prime } | s , a )$ (Kumar et al., 2020). $\\phi _ { o l d ^ { \\prime } }$ denotes the parameters of the policy $\\pi$ in the iteration before $\\phi _ { o l d }$ . Now, with probability $\\geq 1 - \\omega$ , we have $\\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }$ . So, if $\\alpha$ is chosen as follows ", + "bbox": [ + 173, + 156, + 826, + 219 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/38d9e3d1b2073eeaeaeb6070a11ee63d48a40b9c84a8a3b71d2e58a325876012.jpg", + "text": "$$\n\\begin{array} { r } { v > \\displaystyle \\frac { \\mathcal { G } _ { c , T } } { 1 - \\gamma } \\cdot \\displaystyle \\operatorname* { m a x } _ { s \\sim \\rho \\phi _ { o l d ^ { \\prime } } } \\left( \\frac { 1 } { | \\sqrt { \\mathcal { D } _ { \\phi _ { o l d ^ { \\prime } } } } | } + \\frac { 2 \\sqrt { 2 \\delta } \\gamma + ( 1 - \\gamma ) ^ { 2 } \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | } } { \\mathcal { G } _ { c , T } } \\right) \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d ^ { \\prime } } } } \\left( \\frac { \\pi _ { \\phi _ { o l d } } } { \\pi _ { \\phi _ { o l d ^ { \\prime } } } } - 1 \\right) \\right] ^ { - 1 } } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 218, + 838, + 277 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Then with probability $\\geq 1 - \\omega$ , we will have, ", + "bbox": [ + 176, + 290, + 472, + 304 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/4fe73cc28ecfbf904b0b60c0f731238724095ee3294fc31d377b2d8826d6c2dc.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi\n$$", + "text_format": "latex", + "bbox": [ + 444, + 303, + 553, + 321 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "In the next theorem, we show that the convergence rate to the optimal solution is not severely affected due to the safety constraint satisfaction guarantee, and gets modified by addition of an extra bounded term. ", + "bbox": [ + 176, + 327, + 825, + 371 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Theorem 2. If we run the policy gradient updates through equation 5, for policy $\\pi _ { \\phi }$ , with $\\mu$ as the starting state distribution, with $\\phi ^ { ( 0 ) } = 0$ , and learning rate $\\eta > 0 _ { : }$ , then for all policy update iterations $T > 0$ we have, with probability $\\geq 1 - \\omega$ , ", + "bbox": [ + 174, + 376, + 825, + 421 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/6039f3ac332c6130a67e1eeafbde64bf38f8dc028245c114583d5840c7a6e4e5.jpg", + "text": "$$\nV _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + \\left( ( 1 - \\chi ) + \\left( 1 - \\frac { 2 \\Delta } { ( 1 - \\gamma ) } \\right) + 2 \\zeta \\right) \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T }\n$$", + "text_format": "latex", + "bbox": [ + 186, + 420, + 808, + 458 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Sin1), of the dual variables is upper-bounded. √ $\\lambda$ strictly decreases durin addition, if we choose radient descent updates (Algorithmas mentioned in the discussion of $\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) }$ $\\alpha$ Theorem 1, we have $\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + \\zeta } \\end{array}$ . Hence, with probability $\\geq 1 - \\omega$ , we can ensure that ", + "bbox": [ + 173, + 468, + 825, + 521 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/c3c79dc7cb34560eff525b8109b22dadc119845a488302f8c1f0b24fb0577f00.jpg", + "text": "$$\nV _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\mathcal { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + K \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T } \\quad \\mathrm { w h e r e } \\quad K \\leq ( 1 - \\chi ) + \\frac { 4 \\sqrt { 2 \\delta } \\gamma } { ( 1 - \\gamma ) ^ { 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 181, + 521, + 818, + 559 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proof. Let superscript $( t )$ denote the $t ^ { \\mathrm { { t h } } }$ policy update iteration. We follow the derivation in Lemma 5.2 of (Agarwal et al., 2019) but replace $A ( s , a )$ with our modified advantage estimator $\\hat { A } ^ { ( t ) } ( s , a ) =$ $A _ { R } ^ { ( t ) } ( s , a ) - \\lambda ^ { ( t ) } A _ { C } ( s , a )$ . The quantity $\\log Z _ { t } ( s )$ is defined in terms of ${ A } _ { R } ^ { ( t ) }$ as ", + "bbox": [ + 174, + 585, + 823, + 636 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/81fdec9ced7ca7a5cc972edd3a31d04ed29ac2a2eb2be55701f42db0ecc41d04.jpg", + "text": "$$\n\\begin{array} { l } { \\log Z _ { t } ( s ) = \\log \\displaystyle \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\exp \\left( \\eta A ^ { ( t ) } / ( 1 - \\gamma ) \\right) } \\\\ { \\mathrm { ~ } \\ge \\displaystyle \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\log \\exp \\eta A ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } \\\\ { \\displaystyle = \\frac { \\eta } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A ^ { ( t ) } ( s , a ) } \\\\ { \\displaystyle = 0 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 315, + 637, + 679, + 756 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We define an equivalent alternate quantity based on $\\hat { A } ^ { ( t ) }$ ", + "bbox": [ + 173, + 756, + 544, + 772 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/64ffe6b42ab2f8049785a23482bb587b1a486c77590c6a07965ba096539d746d.jpg", + "text": "$$\n\\begin{array} { r l } { { \\operatorname { s g } \\hat { Z } _ { t } ( s ) = \\log \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\exp ( \\eta \\hat { A } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } } \\\\ & { = \\log \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\exp ( \\eta ( A _ { R } ^ { ( t ) } ( s , a ) - \\lambda ^ { ( t ) } A _ { C } ( s , a ) ) / ( 1 - \\gamma ) ) } \\\\ & { \\geq \\displaystyle \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\log \\exp ( \\eta A _ { R } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) - \\lambda ^ { ( t ) } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\log \\exp ( \\eta A _ { C } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } \\\\ & { = 0 - \\displaystyle \\frac { \\lambda ^ { ( t ) } \\eta } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 181, + 773, + 839, + 915 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "For simplicity, consider softmax policy parameterization (equivalent results hold under the function approximation regime as shown in (Agarwal et al., 2019)), where we define the policy updates with the modified advantage function $\\hat { A } ^ { ( t ) }$ to take the form: ", + "bbox": [ + 173, + 103, + 825, + 147 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/7a104c874bb24f90ddc3ebec18d8dac69466d8fa44a15afee423a6097ebdc90d.jpg", + "text": "$$\n\\phi ^ { ( t + 1 ) } = \\phi ^ { ( t ) } + \\frac { \\eta } { 1 - \\gamma } \\hat { A } ^ { ( t ) } \\quad \\mathrm { a n d } \\quad \\pi ^ { ( t + 1 ) } ( a | s ) = \\pi ^ { ( t ) } ( a | s ) \\frac { \\exp \\bigl ( \\eta \\hat { A } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) \\bigr ) } { \\hat { Z } _ { t } ( s ) } ,\n$$", + "text_format": "latex", + "bbox": [ + 214, + 147, + 784, + 183 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Here, $\\begin{array} { r } { \\hat { Z } _ { t } ( s ) = \\sum _ { a \\in \\mathcal { A } } \\pi ^ { ( t ) } ( a | s ) \\exp ( \\eta \\hat { A } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } \\end{array}$ . Note that our actual policy updates (with backtracking line search) are almost equivalent to this when $\\eta$ is small. For the sake of notational convenience, we will denote $\\begin{array} { r } { \\log \\hat { Z } _ { t } ( s ) \\dot { + } \\frac { \\lambda ^ { ( t ) } \\eta } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) } \\end{array}$ as $G _ { t } ( s )$ . We have $G _ { t } ( s ) \\geq 0$ from equation 30. ", + "bbox": [ + 173, + 185, + 826, + 250 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We consider the performance improvement lemma (Kakade $\\&$ Langford, 2002) with respect to the task advantage function $A _ { R } ^ { ( t ) } ( \\bar { s } , a )$ and express it in terms of the modified advantage function $\\hat { A } ^ { ( t ) } ( s , a ) = A _ { R } ^ { ( t ) } ( s , a ) - \\lambda ^ { ( t ) } \\ddot { A } _ { C } ( s , a )$ . Let $\\mu$ be the starting state distribution of the MDP, and $\\boldsymbol d ^ { ( t ) }$ denote the stationary distribution of states induced by policy $\\pi$ in the $t ^ { \\mathrm { { t h } } }$ iteration. ", + "bbox": [ + 173, + 256, + 825, + 323 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/140c9722c88918eea2fb3e647d919ea9e5da7ff0a1615cfa788f159a5c60795e.jpg", + "text": "$$\n\\begin{array} { r l } { \\| \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) - \\frac { 1 } { N _ { 1 } } \\sum _ { k = 0 } ^ { N } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } } & { = } \\\\ & { - 1 - \\sum _ { k = 0 } ^ { N } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } } \\\\ & { - \\frac { 1 } { N _ { 1 } ^ { 2 } } \\sum _ { k = 0 } ^ { N } \\exp \\sum _ { j } ^ { k + 1 } \\exp \\frac { \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } \\| \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } \\| _ { \\alpha } ^ { \\prime } \\| _ { \\alpha } ^ { \\prime } } \\\\ & \\phantom { = } \\frac { 1 } { N _ { 1 } } \\sum _ { k = 0 } ^ { N } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } \\\\ & = \\frac { 1 } { N _ { 1 } } \\frac { 1 } { N _ { 1 } ^ { 2 } } \\exp \\sum _ { j } ^ { k + 1 } \\exp ( - \\sum _ { j } ^ \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 192, + 344, + 772, + 666 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "We note that $G _ { t } ( s ) \\geq 0$ from equation 30. We now prove a result upper bounding the difference between the optimal task value for any state distribution $\\rho$ and the task value at the $t ^ { \\mathrm { t h } }$ iteration for the same state distribution. ", + "bbox": [ + 176, + 667, + 823, + 709 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Sub-optimality gap. The difference between the optimal value function and the current value function estimate is upper bounded. ", + "bbox": [ + 171, + 715, + 823, + 746 + ], + "page_idx": 16 + }, + { + "type": "equation", + "img_path": "images/c75a7199ee24945a621c14ca8f099f8aa73b3ed490f235c5374cbf6b50288b73.jpg", + "text": "$$\n\\begin{array} { r l } & { - \\nabla _ { x } ^ { ( k ) } \\cdot \\nabla _ { x } ^ { ( k ) } = \\frac { 1 } { 1 - \\sigma ^ { 2 } } \\mathrm { E } _ { x - \\cdot } \\sum _ { x = \\cdot } \\sum _ { x = \\cdot } \\nabla _ { x } ^ { \\cdot \\gamma } \\cdot \\sigma ( x ) \\hat { A } ^ { ( k ) } \\cdot \\nabla _ { x } \\cdot \\sigma _ { x } + x ^ { ( k ) } \\hat { A } ^ { ( k ) } \\cdot \\sigma _ { x } \\otimes x ) } \\\\ & { = - \\frac { 1 } { \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma _ { x } | \\hat { A } | \\cdot \\frac { \\sigma _ { x } ^ { ( k ) } } { \\sigma _ { x } ^ { ( k ) } } \\hat { A } | \\hat { A } | \\hat { A } | \\cdot } \\\\ & { - \\frac { 1 } { \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots ( \\int _ { 0 } \\mathrm { R } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\rangle - \\frac { 1 } { \\sigma _ { x } \\otimes x } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\rangle + \\frac { 1 } { \\sigma _ { x } } \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\cdot \\hat { A } | \\sigma _ { x } ^ { ( k ) } \\hat { A } | ) } \\\\ & { + \\frac { 1 } { 1 - \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\hat { A } | \\hat { A } | \\cdot \\sigma _ { x } \\otimes x ^ { ( k ) } \\hat { A } | \\cdot } \\\\ & - \\frac { 1 } { \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\hat { A } | \\hat { A } | \\cdot \\sigma _ { x } \\otimes x ^ { ( k ) } \\rangle + \\log _ { 2 } ^ { \\cdot ( k ) } ( \\int _ { 0 } \\mathrm { R } \\langle \\sigma | \\hat { A } | \\cdot \\hat { \\sigma } \\otimes x ^ { ( k ) } \\rangle - \\frac { 1 } { \\sigma _ { x } } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\hat \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 225, + 116, + 883, + 431 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Using equation 31 with $d ^ { \\star }$ as the starting state distribution $\\mu$ , we have: ", + "bbox": [ + 176, + 449, + 635, + 464 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/dcd840b5552e16bef0f53f8a2510c82f50e29c54c6e90c784854071d9819c159.jpg", + "text": "$$\n\\frac { 1 } { \\eta } \\mathbb { E } _ { s \\sim d ^ { \\star } } \\log G _ { t } ( s ) \\le \\frac { 1 } { 1 - \\gamma } \\Big ( V ^ { ( t + 1 ) } ( d ^ { \\star } ) - V ^ { ( t ) } ( d ^ { \\star } ) \\Big )\n$$", + "text_format": "latex", + "bbox": [ + 320, + 463, + 678, + 494 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "which gives us a bound on $\\mathbb { E } _ { s \\sim d ^ { \\star } } \\log G _ { t } ( s )$ . ", + "bbox": [ + 174, + 494, + 464, + 510 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Using the above equation and that $V ^ { ( t + 1 ) } ( \\rho ) \\geq V ^ { ( t ) } ( \\rho )$ (as $V ^ { ( t + 1 ) } ( s ) \\geq V ^ { ( t ) } ( s )$ for all states $s$ ), we have: ", + "bbox": [ + 174, + 515, + 823, + 545 + ], + "page_idx": 17 + }, + { + "type": "equation", + "img_path": "images/2f76fc7f021c93a4c6cbf6901f677637639d184e5c733f53b8b1cebb8fa0a416.jpg", + "text": "$$\n\\begin{array} { r l } { V _ { \\mathrm { G } } ^ { - 1 } ( s ) - V _ { \\mathrm { G } } ^ { \\mathrm { G } } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } \\sum _ { \\ell = 1 } ^ { 1 } ( V _ { \\mathrm { G } } ^ { - 1 } ( \\beta ) - V _ { \\mathrm { G } } ^ { \\mathrm { G } ( 1 ) } ( s ) ) } \\\\ { \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq ( \\operatorname { B a } ( s ) + \\operatorname { B a } ( s ) ) ^ { n } \\setminus S _ { \\mathrm { G } } ( s ) ( s ) + \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) } \\\\ { \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } } ^ { - 1 } ( s ) \\leq ( \\operatorname { B a } ( s ) + \\frac { \\lambda ^ { n } } { T } ) \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } } ^ { \\mathrm { G } ( 1 ) } ( s ) + \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } } ^ { \\mathrm { G } ( 1 ) } ( s ) \\leq \\frac { 1 } { T } } \\\\ { \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } \\frac { 1 } { \\sum _ { \\ell = 1 } ^ { 2 } \\lambda } \\frac { 1 } { ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! } } \\\\ \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { ( 1 ) } ( s ) \\leq ( 1 - s ) ! \\displaystyle \\sum _ { n = 0 } ^ { 1 } \\frac { 1 } { ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! } - \\frac { 1 } { T } \\displaystyle \\sum _ \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 165, + 542, + 816, + 838 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "Here, $\\Delta$ denotes the CQL overestimation penalty, and we have used the fact that each term of $\\begin{array} { r } { \\left( \\frac { 1 } { 1 - \\gamma } \\sum _ { a } \\pi ^ { \\star } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) - \\frac { 1 } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) \\right) } \\end{array}$ is upper bounded by $\\begin{array} { r } { ( \\chi + \\zeta - \\frac { \\Delta } { ( 1 - \\gamma ) } ) } \\end{array}$ from Lemma 1, so the difference is upper-bounded by $\\begin{array} { r } { 2 ( \\chi + \\zeta - \\frac { \\Delta } { ( 1 - \\gamma ) } ) } \\end{array}$ . ", + "bbox": [ + 173, + 843, + 826, + 901 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "By choosing $\\alpha$ as in equation 26, we have $\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta . \\mathrm { S o } , - \\Delta < - \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } - ( 1 - \\gamma ) \\zeta . } \\end{array}$ Hence, we obtain the relation ", + "bbox": [ + 173, + 99, + 823, + 136 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "We also observe that $\\begin{array} { r } { 2 ( \\chi - \\frac { \\Delta } { ( 1 - \\gamma ) } ) + 2 \\zeta = \\chi + \\chi - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } + 2 \\zeta \\leq 2 - \\chi - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } = ( 1 - \\chi ) + } \\end{array}$ $\\begin{array} { r } { 2 \\zeta + ( 1 - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } ) + 2 \\zeta } \\end{array}$ ", + "bbox": [ + 176, + 140, + 823, + 180 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "So, we have the following result for convergence rate ", + "bbox": [ + 173, + 184, + 524, + 200 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/62ea171a5a5265fddfecb880b5af6f2a9331d4c220ee7e3e9f9ee64eedd4604c.jpg", + "text": "$$\nV _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\mathcal { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + \\bigl ( ( 1 - \\chi ) + \\bigl ( 1 - \\frac { 2 \\Delta } { ( 1 - \\gamma ) } \\bigr ) + 2 \\zeta \\bigr ) \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T }\n$$", + "text_format": "latex", + "bbox": [ + 199, + 217, + 799, + 255 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Again, with probability $\\geq 1 - \\omega$ , we can ensure $\\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }$ . Overall, choosing the value of $\\alpha$ from equation 27, we have $\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta } \\end{array}$ . So, − 2 2δγ − (1 − γ)ζ. Hence, with probability $\\geq 1 - \\omega$ , we can ensure that ", + "bbox": [ + 174, + 256, + 825, + 314 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/2cf28189e3b1b24d07fb5107dc8662d4567bc3d1c9f9023a1c698b13a7126d81.jpg", + "text": "$$\nV _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\cal A | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + K \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T }\n$$", + "text_format": "latex", + "bbox": [ + 299, + 313, + 699, + 349 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "where, ", + "bbox": [ + 174, + 356, + 220, + 369 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/bac039ba46a8dc2a906640e153409c69aaf08eabd64ddd723aebd8df26f90ba7.jpg", + "text": "$$\nK \\leq ( 1 - \\chi ) + \\frac { 4 \\sqrt { 2 \\delta } \\gamma } { ( 1 - \\gamma ) ^ { 2 } }\n$$", + "text_format": "latex", + "bbox": [ + 411, + 368, + 584, + 405 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "So far we have demonstrated that the resulting policy iterates from our algorithm all satisfy the desired safety constraint of the CMDP which allows for a maximum safety violation of $\\chi$ for every intermediate policy. While this guarantee ensures that the probability of failures is bounded, it does not elaborate on the total failures incurred by the algorithm. In our next result, we show that the cumulative failures until a certain number of iterations $T$ of the algorithm grows sublinearly when executing Algorithm 1, provided the safety threshold $\\chi$ is set in the right way. ", + "bbox": [ + 173, + 467, + 825, + 553 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Theorem 3. Let $\\chi$ in Algorithm 1 be time-dependent such that $\\chi _ { t } = \\mathcal { O } ( 1 / \\sqrt { t } )$ . Then, the total number of cumulative safety violations until when $T$ transition samples have been collected by Algorithm 1, $R e g _ { C } ( T )$ , scales sub-linearly with $T$ , i.e., $R e g _ { C } ( T ) = \\mathcal { O } ( \\sqrt { | \\cal { S } | | \\mathcal { A } | T } )$ . ", + "bbox": [ + 173, + 560, + 825, + 607 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Proof. From Theorem 1, with probability $\\geq 1 - \\omega$ ", + "bbox": [ + 174, + 647, + 509, + 662 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/dc7774a579cb12e424bccd9bbd8363616ab8a6ad43bf1a8eaed0a3b74c36805d.jpg", + "text": "$$\nV _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad \\mathrm { w h e r e } \\qquad \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }\n$$", + "text_format": "latex", + "bbox": [ + 246, + 662, + 750, + 698 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Instead of using $\\phi _ { n e w }$ , for ease of notation in this analysis, let us write this in terms of subscript $t$ such that $\\pi _ { t }$ denotes the policy at the $t ^ { \\mathrm { t h } }$ iteration. We can write the bound on $\\zeta$ more explicitly as, $\\zeta _ { t } \\le \\mathbb { E } _ { s \\sim d ^ { \\pi _ { t } } , a \\sim \\pi _ { t } } \\left[ \\frac { C ^ { \\prime } \\sqrt { \\log ( { 1 / \\omega } ) } } { \\sqrt { | N _ { t } ( s , a ) | } } \\right]$ Here, $N _ { t } ( s , a )$ denotes the total number of times $( s , a )$ is seen, and can be written as $\\begin{array} { r } { N _ { t } ( s , a ) = \\sum _ { j = 1 } ^ { t } n _ { j } ( s , a ) } \\end{array}$ , where $n _ { j } ( s , a )$ denotes total number of times $( s , a )$ is seen in episode $j$ . $\\begin{array} { r } { T = \\sum _ { s , a } N _ { k } ( s , a ) } \\end{array}$ denotes the total number of samples collected so far. Let $k$ be the number of episodes elapsed when this happens. The safety threshold $\\chi$ can be varied at every iteration, such that it decreases as $\\begin{array} { r } { \\chi _ { t } \\propto \\frac { 1 } { \\sqrt { t } } } \\end{array}$ . So, we have with probability $\\geq 1 - \\omega$ ", + "bbox": [ + 173, + 696, + 826, + 820 + ], + "page_idx": 18 + }, + { + "type": "equation", + "img_path": "images/8031a3dc0640a194d19823afcfafbf43d0057874475975f3424e5b8062bf757f.jpg", + "text": "$$\n\\gamma _ { C } ^ { \\pi _ { t } } ( \\mu ) \\leq \\chi _ { t } + \\zeta _ { t } - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\quad \\mathrm { w h e r e } \\quad \\zeta _ { t } \\leq \\mathbb { E } _ { s \\sim d ^ { \\pi _ { t } } , a \\sim \\pi _ { t } } \\left[ \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { \\sqrt { \\lvert N _ { t } ( s , a ) \\rvert } } \\right] \\quad \\mathrm { a n d } \\ \\chi _ { t } = \\frac { C ^ { \\prime \\prime } } { \\sqrt { t } }\n$$", + "text_format": "latex", + "bbox": [ + 181, + 821, + 831, + 863 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Now, we note that $V _ { C } ^ { \\pi _ { t } }$ denotes the expected probability of episodic failures by executing policy $\\pi _ { t }$ (as described in section 2 of the main paper). Let us consider that a policy is rolled out for one episode after every training iteration to collect data (exploration). ", + "bbox": [ + 173, + 881, + 825, + 924 + ], + "page_idx": 18 + }, + { + "type": "text", + "text": "Hence the total cumulative safety failures when $T$ transition samples have been collected by Algorithm 1 is: ", + "bbox": [ + 169, + 102, + 825, + 131 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/8bcb76b768fb686060b7a8161a01bfeb8eb5bd8a8cc8c1b14ca7699921e32e39.jpg", + "text": "$$\n\\begin{array} { r l } { \\mathbb { E } \\mathbb { E } \\mathbb { E } \\mathbb { E } ( t ) - \\frac { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } \\times } & { 1 } \\\\ & { \\le \\frac { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } \\underbrace { \\sum _ { t = 1 } ^ { K } \\left( \\begin{array} { l l } { \\Delta } & { \\sqrt { \\Delta } \\pi \\kappa \\omega t } \\\\ { 1 - \\gamma } & { ( 1 - \\gamma ) ^ { t } } \\end{array} \\right) } _ { \\le \\alpha } } \\\\ & { \\le \\frac { \\sum _ { t = 1 } ^ { K } \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } } { \\sqrt { \\pi } } + \\sum _ { t = 1 } ^ { K } \\sum _ { s = 1 } ^ { K } \\frac { \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } } \\\\ & \\le \\frac { \\sum _ { t = 1 } ^ { K } \\frac { \\gamma } { \\alpha } + \\mathcal { E } \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\sum _ { s = 1 } ^ { K } \\frac { \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } } \\\\ & \\le C ^ { \\mathcal { N } } \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( \\mathcal { L } \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 295, + 127, + 702, + 431 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Here, we used the concavity of square root function to obtain $\\begin{array} { r } { \\sum _ { s , a } \\sqrt { N _ { k } ( s , a ) } } \\end{array}$ ≤ $\\sqrt { | S | | A | } \\sqrt { \\textstyle \\sum _ { s , a } N _ { k } ( s , a ) }$ and the definition $\\begin{array} { r } { T = \\sum _ { s , a } N _ { k } ( s , a ) } \\end{array}$ ", + "bbox": [ + 174, + 469, + 825, + 513 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "A.2 DERIVATION OF THE POLICY UPDATE EQUATIONS ", + "text_level": 1, + "bbox": [ + 174, + 541, + 555, + 556 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Let $a \\in { \\mathcal { A } }$ denote an action, $s \\in { \\mathcal { S } }$ denote a state, $\\pi _ { \\phi } ( a | s )$ denote a parameterized policy, $r ( s , a )$ denote a reward function for the task being solved, and $\\tau$ denote a trajectory of actions by following policy $\\pi _ { \\phi }$ at each state. To solve the following constrained optimization problem: ", + "bbox": [ + 173, + 569, + 825, + 612 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/847f5562500f839174145bf05f41ee2ff2c41885d271c67becc54a53cc7b581b.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { \\tau \\sim \\pi _ { \\phi } } [ \\sum _ { \\tau } r ( \\cdot ) ] \\quad s . t . \\quad \\mathbb { E } _ { \\tau \\sim \\pi _ { \\phi } } [ \\sum _ { \\tau } \\mathbb { 1 } \\{ f a i l u r e \\} ] = 0\n$$", + "text_format": "latex", + "bbox": [ + 303, + 612, + 694, + 645 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "Here, $\\tau$ is the trajectory corresponding to an episode. The objective is to maximize the cumulative returns while satisfying the constraint. The constraint says that the agent must never fail during every episode. $\\mathbb { 1 } \\{ f a i l u r e \\} = 1$ if there is a failure and $\\mathbb { 1 } \\{ f a i l u r e \\} = 0$ if the agent does not fail. The only way expectation can be 0 for this quantity is if every element is 0, so the constraint essentially is to never fail in any episode. Let’s rewrite the objective, more generally as ", + "bbox": [ + 173, + 650, + 825, + 720 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/2dfea94de165846f367e0b9b0785eed9d382a2b8e96b4d575e96f8049baaadf2.jpg", + "text": "$$\n\\operatorname * { m a x } _ { \\pi _ { \\phi } } V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) \\quad s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\bar { \\mu ) } = 0\n$$", + "text_format": "latex", + "bbox": [ + 382, + 719, + 614, + 746 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "We can relax the constraint slightly, by introducing a tolerance parameter $\\chi \\approx 0$ . The objective below tolerates atmost $\\chi$ failures in expectation. Since the agent can fail only once in an episode, $V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )$ can also be interpreted as the probability of failure, and the constraint $V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi$ says that the probability of failure in expectation must be bounded by $\\chi$ . So, our objective has a very intuitive and practical interpretation. ", + "bbox": [ + 173, + 751, + 825, + 821 + ], + "page_idx": 19 + }, + { + "type": "equation", + "img_path": "images/61785f9e75172d93c55380ca540163fbf68bd5023437395d9a8e09bfdd3c7262.jpg", + "text": "$$\n\\operatorname* { m a x } _ { \\pi _ { \\phi } } V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) \\quad s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi\n$$", + "text_format": "latex", + "bbox": [ + 382, + 820, + 616, + 847 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "We learn one state value function, $V _ { R }$ (corresponding to the task reward), parameterized by $\\theta$ and one state-action value function $Q _ { C }$ (corresponding to the sparse failure indicator), parameterized by $\\zeta$ . We have a task reward function $r ( s , a )$ from the environment which is used to learn $V _ { R }$ . For learning $Q _ { C }$ , we get a signal from the environment indicating whether the agent is dead (1) or alive (0) i.e. $\\mathbb { 1 } \\{ f a i l u r e \\}$ . ", + "bbox": [ + 173, + 853, + 825, + 925 + ], + "page_idx": 19 + }, + { + "type": "text", + "text": "The safety critic $Q _ { C }$ is used to get an estimate of how safe a particular state is, by providing an estimate of probability of failure, that will be used to guide exploration. We desire the estimates to be conservative, in the sense that the probability of failure should be an over-estimate of the actual probability so that the agent can err in the side of caution while exploring. To train such a critic $Q _ { C }$ , we incorporate theoretical insights from $\\mathrm { C Q L }$ , and estimate $Q _ { C }$ through updates similar to those obtained by flipping the sign of $\\alpha$ in equation 2 of the CQL paper (Kumar et al., 2020). The motivation for this is to get an upper bound on $Q _ { C }$ instead of a lower bound, as guaranteed by CQL. ", + "bbox": [ + 173, + 103, + 825, + 202 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We also note that the CQL penalty term (the first two terms of equation 2 of the CQL paper) can be expressed as an estimate for the advantage function of the policy $\\mathbb { E } _ { s \\sim d } { } ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) \\big [ A ( s , a ) \\big ]$ ,where, $A ( s , a )$ is the advantage function. ", + "bbox": [ + 174, + 208, + 823, + 252 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/71237c04bf08477c5636c434f2043177ab77c64900b5875c6c7497462e31a9d6.jpg", + "text": "$$\n\\begin{array} { r l } & { \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ Q ( s , a ) ] - \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi _ { o l d } } ( a \\mid s ) } [ Q ( s , a ) ] } \\\\ & { \\ = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ Q ( s , a ) - \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } ( a \\mid s ) } Q ( s , a ) ] } \\\\ & { \\ = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ Q ( s , a ) - V ( s ) ] = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ A } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 258, + 252, + 738, + 315 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Hence, CQL can help provide an upper bound on the advantage function directly. Although the CQL class of algorithms have been proposed for batch RL, the basic bounds on the value function hold even for online training. ", + "bbox": [ + 176, + 315, + 821, + 357 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "We denote the objective inside arg min as $C Q L ( \\zeta )$ , where $\\zeta$ parameterizes $Q _ { C }$ , and $k$ denotes the $k ^ { \\mathrm { { t h } } }$ update iteration. ", + "bbox": [ + 173, + 363, + 825, + 391 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/ddb2075e29c1639216aba86d5015eb8ff6d6ba07f610e6e57cb18ac62a7bc937.jpg", + "text": "$$\n\\begin{array} { r l } & { \\hat { Q } _ { C } ^ { k + 1 } \\gets \\underset { Q _ { C } } { \\arg \\operatorname* { m i n } } \\alpha \\left( - \\mathbb { E } _ { s \\sim \\mathcal { D } _ { e n v } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q _ { C } ( s , a ) ] + \\mathbb { E } _ { ( s , a ) \\sim \\mathcal { D } _ { e n v } } [ Q _ { C } ( s , a ) ] \\right) } \\\\ & { \\quad \\quad \\quad + \\frac { 1 } { 2 } \\mathbb { E } _ { ( s , a , s ^ { \\prime } , c ) \\sim \\mathcal { D } _ { e n v } } \\left[ \\left( Q _ { C } ( s , a ) - \\hat { B } ^ { \\pi _ { \\phi } } \\hat { Q } _ { C } ^ { k } ( s , a ) \\right) ^ { 2 } \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 238, + 391, + 759, + 457 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "For states sampled from the replay buffer $\\mathcal { D } _ { e n v }$ , the first term seeks to maximize the expectation of $Q _ { C }$ over actions sampled from the current policy, while the second term seeks to minimize the expectation of $Q _ { C }$ over actions sampled from the replay buffer. $\\mathcal { D } _ { e n v }$ can include off-policy data, and also offline-data (if available). Let the over-estimated advantage, corresponding to the overestimated critic $Q _ { C }$ , so obtained from CQL, be denoted as $\\hat { A } _ { C } ( s , \\bar { a } )$ , where the true advantage is $A _ { C } ( s , a )$ . ", + "bbox": [ + 173, + 455, + 825, + 541 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Now, let $\\rho _ { \\phi } ( s )$ denote the stationary distribution of states induced by policy $\\pi _ { \\phi }$ . For policy optimization, we have to solve a constrained optimization problem as described below: ", + "bbox": [ + 174, + 546, + 820, + 575 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/25b67e3a887b722117412dbdb7816cfa4ae8f8812e4c340a3c339dcea249d887.jpg", + "text": "$$\n\\begin{array} { r l } { { \\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } [ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) ] } } \\\\ { s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 346, + 574, + 651, + 645 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "This, as per equation 22 can be rewritten as ", + "bbox": [ + 176, + 643, + 460, + 659 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/76dd911d5909ca0ed824c4fdc2c4fe7da8af47ea175c30911b6b74ef1c120444.jpg", + "text": "$$\n\\begin{array} { r l } { \\pi _ { \\phi _ { n e w } } = \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { } \\\\ { s . t . } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . } & { V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\leq \\chi } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 267, + 675, + 728, + 760 + ], + "page_idx": 20 + }, + { + "type": "text", + "text": "Since we are learning an over-estimate of $A _ { C }$ through the updates in equation 2, we replace $A _ { C }$ by the learned $\\hat { A } _ { C }$ in the constraint above. There are multiple ways to solve this constrained optimization problem, through duality. If we consider the Lagrangian dual of this, then we have the following optimization problem, which we can solve approximately by alternating gradient descent. For now, we keep the KL constraint as is, and later use its second order Taylor expansion in terms of the Fisher Information Matrix. ", + "bbox": [ + 173, + 763, + 826, + 851 + ], + "page_idx": 20 + }, + { + "type": "equation", + "img_path": "images/c2fc6d9dd2b815b6e0383c314030769ff619b361b65ab60b3aa8bb11d0c817fe.jpg", + "text": "$$\n\\begin{array} { r l } { } & { \\underset { \\pi _ { \\phi } } { \\mathrm { a a x } } \\underset { \\lambda \\geq 0 } { \\mathrm { m i n } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\Big ] - \\lambda \\left( V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ \\hat { A } _ { C } ( s , a ) \\Big ] - \\chi \\right) } \\\\ { : \\cdot \\hfill } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 174, + 114, + 826, + 170 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We replace $V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)$ by its sample estimate $\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ and denote $\\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )$ as $\\chi ^ { \\prime }$ . Note that $\\chi ^ { \\prime }$ is independent of parameter $\\phi$ that is being optimized over. So, the objective becomes ", + "bbox": [ + 174, + 190, + 826, + 223 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/c1d81b544da1cec6deb45e1e91517b44494bab7fc8f6288b1508b8c183a37878.jpg", + "text": "$$\n\\begin{array} { r l } & { \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\underset { \\lambda \\geq 0 } { \\operatorname* { m i n } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\frac { \\lambda } { 1 - \\gamma } \\hat { A } _ { C } ( s , a ) \\right] + \\lambda \\chi ^ { \\prime } } \\\\ & { s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 274, + 239, + 705, + 295 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "For notational convenience let $\\lambda ^ { \\prime }$ denote the fraction $\\frac { \\lambda } { 1 - \\gamma }$ . Also, in the expectation, we replace $a \\sim \\pi _ { \\phi }$ by $a \\sim \\pi _ { \\phi _ { o l d } }$ and account for it by importance weighting of the objective. ", + "bbox": [ + 173, + 295, + 821, + 327 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Let us consider $\\operatorname* { m a x } _ { \\pi _ { \\phi } }$ operation and the following gradient necessary for gradient ascent of $\\phi$ ", + "bbox": [ + 173, + 332, + 794, + 348 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/a180fc58d5a090d2d05974b0117c53d3cda96387b41490438709082bc562efd5.jpg", + "text": "$$\n\\begin{array} { r l } & { \\phi \\longleftarrow \\arg \\operatorname* { m a x } _ { \\phi } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\frac { \\pi _ { \\phi } ( a | s ) } { \\pi _ { \\phi _ { o l d } } ( a | s ) } ( A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) ) \\right] \\right] } \\\\ & { s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 243, + 348, + 753, + 404 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/d917c34560e6f99797091e718cc743f460e89dfd175d03bed8cd30b1efa940b8.jpg", + "text": "$$\n\\phi \\arg \\operatorname* { m a x } _ { \\phi } \\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) ^ { T } ( \\phi - \\phi _ { o l d } )\n$$", + "text_format": "latex", + "bbox": [ + 331, + 411, + 593, + 436 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/b386f92aacbcea9abd2a16c8405fe6cb05ee4842ac75afb6b4ecf46c0ed23bba.jpg", + "text": "$$\n\\begin{array} { r } { \\begin{array} { r l } { s . t . } & { { } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array} } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 352, + 438, + 665, + 458 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Here, using slide 20 of Lecture 9 in (Levine, 2018), and the identity $\\nabla _ { \\phi } \\pi _ { \\phi } = \\pi _ { \\phi } \\nabla _ { \\phi } \\log \\pi _ { \\phi }$ we have ", + "bbox": [ + 176, + 457, + 821, + 472 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/65422b3c8de2c4f299739ea1878b6206bb3388280b2981c7dcf3627ec5f338d4.jpg", + "text": "$$\n\\nabla _ { \\phi } \\bar { A } ( \\phi ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\frac { \\pi _ { \\phi } ( a | s ) } { \\pi _ { \\phi _ { o l d } } ( a | s ) } \\nabla _ { \\phi } \\log \\pi _ { \\phi } ( a | s ) ( A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) ) \\right] \\right]\n$$", + "text_format": "latex", + "bbox": [ + 186, + 472, + 784, + 506 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Using slide 24 of Lecture 5 in (Levine, 2018) and estimating locally at $\\phi = \\phi _ { o l d }$ ", + "bbox": [ + 166, + 506, + 696, + 518 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/91f487342952e35cb922c78e435892d1b0289c5b69612b3213f29cbefd950427.jpg", + "text": "$$\n\\begin{array} { r } { \\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) ( A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) ) \\right] \\right] } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 189, + 518, + 784, + 545 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "We note that, Es∼ρφold $\\begin{array} { r l } { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left. \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left. \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right. \\right. = \\nabla _ { \\phi _ { o l d } } J ( \\phi _ { o l d } ) } & { { } } \\end{array}$ ), the original policy gradient corresponding to task rewards. So, we can write equation 48 as ", + "bbox": [ + 174, + 545, + 830, + 582 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/e733d31ba8c059d66efd2ab95a44b5d813a8a267707af819c8e8d760fab60082.jpg", + "text": "$$\n\\nabla _ { \\phi _ { o l d } } a r A ( \\phi _ { o l d } ) = \\overline { { \\nabla _ { \\phi _ { o l d } } J ( \\phi _ { o l d } ) + \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) \\right] \\right] } }\n$$", + "text_format": "latex", + "bbox": [ + 254, + 580, + 740, + 608 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "In practice, we estimate $A _ { R } ^ { \\pi _ { \\phi _ { o l d } } }$ through GAE (Schulman et al., 2015b;a; Levine, 2018) ", + "bbox": [ + 168, + 614, + 745, + 632 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/48a5ce9c42784a733b323a622250d5e848e1513670869ee5a05c56ed071d5e63.jpg", + "text": "$$\n\\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } ( \\gamma ) ^ { t ^ { \\prime } - t } \\Delta _ { t ^ { \\prime } } \\quad \\Delta _ { t ^ { \\prime } } = r ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } ) + \\gamma V _ { R } ( s _ { t ^ { \\prime } + 1 } ) - V _ { R } ( s _ { t ^ { \\prime } } )\n$$", + "text_format": "latex", + "bbox": [ + 267, + 632, + 728, + 672 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Let to e $\\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) = A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } A _ { C } ( s , a )$ denote the modified advantage function corresponding ", + "bbox": [ + 173, + 679, + 826, + 709 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/5036269c407d8270adb68c87ec96bb7f8320fd188b98a5073db7bcc4c51b388c.jpg", + "text": "$$\n\\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } ( \\gamma ) ^ { t ^ { \\prime } - t } \\Delta _ { t ^ { \\prime } } ~ \\Delta _ { t ^ { \\prime } } = r ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } ) + \\gamma V _ { R } ( s _ { t ^ { \\prime } + 1 } ) - V _ { R } ( s _ { t ^ { \\prime } } ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } )\n$$", + "text_format": "latex", + "bbox": [ + 199, + 709, + 772, + 751 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "So, rewriting equations 48 and 53 in terms of $\\tilde { A } ^ { \\pi _ { \\phi _ { o l d } } }$ , we have ", + "bbox": [ + 173, + 751, + 583, + 767 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/77c01311144a6f671abe66391454950ec00aa6e55b1de0c163e6bbebea4f290a.jpg", + "text": "$$\n\\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } \\right] \\right]\n$$", + "text_format": "latex", + "bbox": [ + 272, + 766, + 723, + 792 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/5429af7f0fb538e06aef6ea4e5252045459bb906ef244d6dec27799828eed190.jpg", + "text": "$$\n\\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) = \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } )\n$$", + "text_format": "latex", + "bbox": [ + 395, + 810, + 602, + 830 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "Substituting in equation 46, we have ", + "bbox": [ + 174, + 830, + 413, + 844 + ], + "page_idx": 21 + }, + { + "type": "equation", + "img_path": "images/776ddab6b7e136c28ecad0bbc889646dfe2896828632ee536eed40f747d6c254.jpg", + "text": "$$\n\\begin{array} { r l } & { \\phi \\underset { \\phi } { \\mathrm { a r g } \\mathrm { m a x } } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } ( \\phi - \\phi _ { o l d } ) } \\\\ & { \\quad s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 331, + 842, + 666, + 891 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "As shown in slide 20 of Lecture 9 (Levine, 2018) and (Schulman et al., 2015a), we can approximate $D _ { \\mathrm { K L } }$ in terms of the Fisher Information Matrix $\\mathbf { F }$ (this is the second order term in the Taylor ", + "bbox": [ + 173, + 890, + 821, + 919 + ], + "page_idx": 21 + }, + { + "type": "text", + "text": "expansion of KL; note that around $\\phi = \\phi _ { o l d }$ , both the $\\mathrm { K L }$ term and its gradient are 0), ", + "bbox": [ + 174, + 103, + 733, + 118 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/1eb474a2abe6ac6316917b5fc42e7a37da60f10b0b6c78456d62eb0d806bc648.jpg", + "text": "$$\nD _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) = \\frac { 1 } { 2 } ( \\phi - \\phi _ { o l d } ) ^ { T } \\mathbf { F } ( \\phi - \\phi _ { o l d } )\n$$", + "text_format": "latex", + "bbox": [ + 316, + 117, + 679, + 147 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Where, $\\mathbf { F }$ can be estimated with samples as ", + "bbox": [ + 174, + 145, + 460, + 160 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/c7dcdb4c8df2507ea25cf7c0fa7d5a4be5ef6610df52e6b7c4629dc26f7af3fd.jpg", + "text": "$$\n\\mathbf { F } = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ) ^ { T } \\right] \\right]\n$$", + "text_format": "latex", + "bbox": [ + 292, + 159, + 704, + 185 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "So, finally, we can write the gradient ascent step for $\\phi$ as (natural gradient conversion) ", + "bbox": [ + 166, + 185, + 740, + 199 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/504305fdd6580575aa4c0e3a11a3bb903942e8654b23ef60c2f8aeb3b4515ede.jpg", + "text": "$$\n\\phi \\phi _ { o l d } + \\beta \\mathbf { F } ^ { - 1 } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) \\quad \\beta = \\sqrt { \\frac { 2 \\delta } { \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } \\mathbf { F } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) } }\n$$", + "text_format": "latex", + "bbox": [ + 254, + 199, + 743, + 239 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "In practice, we perform backtracking line search to ensure the $D _ { \\mathrm { K L } }$ constraint satisfaction. So, we have the following update rule ", + "bbox": [ + 176, + 239, + 823, + 267 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/b9487dd803118da957f1e919e4516fbbbcb8e04eaebe59ea98543d8e3dcd44f3.jpg", + "text": "$$\n\\phi \\phi _ { o l d } + \\beta { \\pmb { F } } ^ { - 1 } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) \\quad \\beta = \\beta ^ { j } \\sqrt { \\frac { 2 \\delta } { \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } { \\pmb { F } } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) } }\n$$", + "text_format": "latex", + "bbox": [ + 243, + 267, + 754, + 309 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "After every update, we check if $\\bar { D } _ { \\mathrm { K L } } ( \\phi | | \\phi _ { o l d } ) \\leq \\delta$ , and if not we decay $\\beta ^ { j } = \\beta ^ { j } ( 1 - \\beta ^ { j } ) ^ { j }$ , set $j \\gets j + 1$ and repeat for $L$ steps until $\\bar { D } _ { \\mathrm { K L } } \\leq \\delta$ is satisfied. If this is not satisfied after $L$ steps, we backtrack, and do not update $\\phi$ i.e. set $\\phi \\phi _ { o l d }$ . For gradient descent with respect to the Lagrange multiplier $\\lambda$ we have (from equation 5), ", + "bbox": [ + 173, + 309, + 823, + 364 + ], + "page_idx": 22 + }, + { + "type": "equation", + "img_path": "images/2bcbf6410896bd5c796faaf3885e999d8727e354fdddbd6c6b9a025379d7486b.jpg", + "text": "$$\n\\lambda \\lambda - ( \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ \\hat { A } _ { C } ( s , a ) ] - \\chi ^ { \\prime } )\n$$", + "text_format": "latex", + "bbox": [ + 315, + 364, + 663, + 398 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Note that in the derivations we have ommitted $\\textstyle \\sum _ { t }$ in the outermost loop of all expectations, and subscripts (e.g. $a _ { t } , s _ { t } )$ in order to avoid clutter in notations. ", + "bbox": [ + 174, + 398, + 825, + 426 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "A.3 RELATION TO CPO ", + "text_level": 1, + "bbox": [ + 176, + 443, + 352, + 458 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "The CPO paper (Achiam et al., 2017) considers a very similar overall objective for policy gradient updates, with one major difference. CPO approximates the $V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi$ constraint by replacing $V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )$ with its first order Taylor expansion and enforces the resulting simplified constraint exactly in the dual space. On the other hand, we do not make this simplification, and use primal-dual optimization to optimize an upper bound on $V _ { C }$ through the CQL inspired objective in equation 2. Doing this and not not making the linearity modification allows us to handle sparse (binary) failure indicators from the environment without assuming a continuous safety cost function as done in CPO (Achiam et al., 2017). ", + "bbox": [ + 173, + 468, + 825, + 583 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "A.4 PRACTICAL CONSIDERATIONS ", + "text_level": 1, + "bbox": [ + 178, + 599, + 426, + 613 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "Depending on the value of KL-constraint on successive policies $\\delta$ , the RHS in Theorem 2 can either be a lower or higher rate than the corresponding problem without safety constraint. In particular, let the sampling error $\\zeta = 0$ , then if (1−γ)4(2−χ)22 , the third term is negative. ", + "bbox": [ + 173, + 625, + 825, + 674 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "If we set $\\gamma = 0 . 9 9$ and $\\chi = 0 . 0 5$ , then for any $\\delta > 1 \\mathsf { e } - 8$ , the third term in Theorem 3 will be negative. Also, if $\\alpha$ is chosen to be much greater than that in equation 26, the value of $\\Delta$ can be arbitrarily increased in principle, and we would be overestimating the value of $Q _ { C }$ significantly. While increasing $\\Delta$ significantly will lead to a decrease in the upper bound of $V _ { R } ^ { * } ( \\mu ) \\stackrel { - } { - } V _ { R } ^ { ( T ) } ( \\mu )$ but in practice, we would no longer have a practical algorithm. This is because, when $Q _ { C }$ is overestimated significantly, it would be difficult to guarantee that line 9 of Algorithm 1 is satisfied, and policy execution will stop, resulting in infinite wall clock time for the algorithm. ", + "bbox": [ + 173, + 680, + 825, + 782 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "In order to ensure that the above does not happen, in practice we loop over line 6 of Algorithm 1 for a maximum of 100 iterations. So, in practice the anytime safety constraint satisfaction of Theorem 2 is violated during the early stages of training when the function approximation of $Q _ { C }$ is incorrect. However, as we demonstrate empirically, we are able to ensure the guarantee holds during the majority of the training process. ", + "bbox": [ + 174, + 789, + 825, + 859 + ], + "page_idx": 22 + }, + { + "type": "text", + "text": "A.5 DETAILS ABOUT THE ENVIRONMENTS ", + "text_level": 1, + "bbox": [ + 176, + 103, + 478, + 117 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "In each environment, shown in Figure 2, we define a task objective that the agent must achieve and a criteria for catastrophic failure. The goal is to solve the task without dying. In all the environments, in addition to the task reward, the agent only receives a binary signal indicatin whether it is dead i.e. a catastrophic failure has occurred (1) or alive (0). ", + "bbox": [ + 176, + 128, + 825, + 185 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "• Point agent navigation avoiding traps. Here, a point agent with two independent actuators for turning and moving forward/backward must be controlled in a 2D plane to reach a goal (shown in green in Figure 2) while avoiding traps shown in violet circular regions. The agent has a health counter set to 25 for the episode and it decreases by 1 for every timestep that it resides in a trap. The agent is alive when the health counter is positive, and a catastrophic failure occurs when the counter strikes 0 and the agent dies. Car agent navigation avoiding traps. Similar environment as the above but the agent is a Car with more complex dynamics. It has two independently controllable front wheels and free-rolling rear wheel. We adapt this environment from (Ray et al., 2019). Panda push without toppling. A Franka Emika Panda arm must push a vertically placed block across the table to a goal location without the block toppling over. The workspace dimensions of the table are $2 0 \\mathrm { c m } \\mathrm { x } 4 0 \\mathrm { c m }$ and the dimensions of the block are 5cmx5cmx10cm. The environment is based on Robosuite Zhu et al. (2020) and we use Operational Space Control (OSC) to control the end-effevctor velocities of the robot arm. A catastrophic failure is said to occur is the block topples. \n• Panda push within boundary. A Franka Emika Panda arm must be controlled to push a block across the table to a goal location without the block going outside a rectangular constraint region. Catastrophic failure occurs when the block center of mass $( x , y )$ position) move outside the constraint region on the table with dimensions $1 5 \\mathrm { c m } \\mathrm { x } 3 5 \\mathrm { c m }$ . The dimensions of the block are 5cmx5cmx10cm. The environment is based on Robosuite Zhu et al. (2020) and we use Operational Space Control (OSC) to control the end-effector velocities of the robot arm. \n• Laikago walk without falling, a Laikago quadruped robot must walk without falling. The agent is rewarded for walking as fast as possible (or trotting) and failure occurs when the robot falls. Since this is an extremely challenging task, for all the baselines, we initialize the agent’s policy with a controller that has been trained to keep the agent standing, while not in motion. The environment is implemented in PyBullet and is based on (Peng et al., 2020). ", + "bbox": [ + 214, + 194, + 826, + 613 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "A.6 HYPER-PARAMETER DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 622, + 421, + 637 + ], + "page_idx": 23 + }, + { + "type": "text", + "text": "We chose the learning rate $\\eta _ { Q }$ for the safety-critic $Q _ { C }$ to be $2 e - 4$ after experimenting with $1 e - 4$ and $2 e - 4$ and observing slightly better results with the latter. The value of discount factor $\\gamma$ is set to the usual default value 0.99, the learning rate $\\eta _ { \\lambda }$ of the dual variable $\\lambda$ is set to $4 e - 2$ , the value of $\\delta$ for the $D _ { \\mathrm { K L } }$ constraint on policy updates is set to 0.01, and the value of $\\alpha$ to be 0.5. We experimented with three different $\\alpha$ values 0.05, 0.5, 5 and found nearly same performance across these three values. For policy updates, the backtracking co-efficient $\\beta ^ { ( 0 ) }$ is set to 0.7 and the max. number of line search iterations $L = 2 0$ . For the $Q$ -ensembles baseline, the ensemble size is chosen to be 20 (as mentioned in the LNT paper), with the rest of the common hyper-parameter values consistent with CSC, for a fair comparison.All results are over four random seeds. ", + "bbox": [ + 173, + 648, + 825, + 775 + ], + "page_idx": 23 + }, + { + "type": "image", + "img_path": "images/1cf2b77cc9f60cabe558cbb6c48ab645aaa956ff96da640543e93a3866668c58.jpg", + "image_caption": [ + "Figure 5: Results on the five environments we consider for our experiments. For each environment we plot the average task reward, the average episodic failures, and the cumulative episodic failures. All the plots are for our method with different safety thresholds $\\chi$ . From the plots it is evident that our method can naturally trade-off safety for task performance depending on how strict the safety threshold $\\chi$ is set to. In particular, for a stricter $\\chi$ (i.e. lesser value), the avg. failures decreases, and the task reward plot also has a slower convergence compared to a less strict threshold. " + ], + "image_footnote": [], + "bbox": [ + 176, + 133, + 821, + 558 + ], + "page_idx": 24 + }, + { + "type": "image", + "img_path": "images/5081b8b05b2da95b28152841eec4fa2041a93afcb84904c087340bd15575b32d.jpg", + "image_caption": [ + "Figure 6: Results on the five environments we consider for our experiments. For each environment we plot the average task reward, the average episodic failures, and the cumulative episodic failures. Since Laikago is an extremely challenging task, for all the baselines, we initialize the agent’s policy with a controller that has been trained to keep the agent standing, while not in motion. The task then is to bootstrap learning so that the agent is able to remain standing while walking as well. The safety threshold $\\chi = 0 . 0 5$ for all the baselines in all the environments. " + ], + "image_footnote": [], + "bbox": [ + 176, + 133, + 821, + 558 + ], + "page_idx": 25 + }, + { + "type": "image", + "img_path": "images/5cc9b512d0b1902a706f179dab0c2a78d9a07a12a7cf3c1148b17408c0dc0b16.jpg", + "image_caption": [ + "Figure 7: Comparison between two RL algorithms TRPO (Schulman et al., 2015a), and SAC (Haarnoja et al., 2018) in the Point agent 2D Navigation environment. We see that TRPO has slightly faster convergence in terms of task rewards and also slightly lower average and cumulative failures, and so consider TRPO as the Base RL baseline in Figures 3 and 4. " + ], + "image_footnote": [], + "bbox": [ + 178, + 704, + 818, + 787 + ], + "page_idx": 25 + }, + { + "type": "text", + "text": "A.10 SEEDING THE REPLAY BUFFER WITH VERY FEW SAMPLES ", + "text_level": 1, + "bbox": [ + 179, + 104, + 625, + 117 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "In order to investigate if we can leverage some offline user-specified data to lower the number of failures during training even further, we seed the replay buffer of CSC and the baselines with 1000 tuples in the Car navigation environment. The 1000 tuples are marked as safe or unsafe depending on whether the car is inside a trap location or not in those states. If our method can leverage such manually marked offline data (in small quantity as this marking procedure is not cheap), then we have a more practical method that can be deployed in situations where the cost of visiting an unsafe state is significantly prohibitive. Note that this is different from the setting of offline/batch RL, where the entire training data is assumed to be available offline - in this experimental setting we consider very few tuples (only 1000). Figure 9 shows that our method can successfully leverage this small offline data to bootstrap the learning of the safety critic and significantly lower the average failures. We attribute this to training the safety critic conservatively through CQL, which is an effective method for handling offline data. ", + "bbox": [ + 173, + 130, + 825, + 296 + ], + "page_idx": 26 + }, + { + "type": "image", + "img_path": "images/7ff2b6047f28b88afefe52f9ea4674749980644fcfdd0f42952f8b4e8e9560a9.jpg", + "image_caption": [ + "Figure 8: Results on the Car navigation environment after seeding the replay buffer with 1000 tuples. Although all the baselines improve by seeding, in terms of lower failure rates compared to Figure 3, we observe that CSC is able to particularly leverage the offline seeding data and significantly lower the average and cumulative failures during training. " + ], + "image_footnote": [], + "bbox": [ + 178, + 313, + 820, + 395 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "A.11 CAR NAVIGATION WITH TRAPS USING CONTINUOUS SAFETY SIGNAL ", + "text_level": 1, + "bbox": [ + 176, + 483, + 699, + 496 + ], + "page_idx": 26 + }, + { + "type": "text", + "text": "In this section we consider the case of a continuous safety signal to show that CSC can learn constraints in this setting as well, and minimize failures significantly more compared to the baselines. The car navigation with traps provides a natural setting for this, because every time the agent enters a trap region, it can receive a penalty (and its health counter decreases by 1), and a catastrophic failure occurs when the health counter drops to 0. By training the safety critic with this continuous failure signal instead of on binary failure signals, we can capture a notion of impending failure and hence aim to be more safe. Note that this setting is a strictly easier evaluation setting than what we previously considered with the binary safety signal. ", + "bbox": [ + 174, + 507, + 825, + 619 + ], + "page_idx": 26 + }, + { + "type": "image", + "img_path": "images/ac67f75ee4ec5741cca716c5033ec0fb4a9edffcf58f0a9203b88b038d4b5df7.jpg", + "image_caption": [ + "Figure 9: Results on the Car navigation environment using continuous safety signal. " + ], + "image_footnote": [], + "bbox": [ + 178, + 635, + 818, + 718 + ], + "page_idx": 26 + } +] \ No newline at end of file diff --git a/parse/train/iaO86DUuKi/iaO86DUuKi_model.json b/parse/train/iaO86DUuKi/iaO86DUuKi_model.json new file mode 100644 index 0000000000000000000000000000000000000000..335bcf63e445c4141bbed7c0a694d0eef64e8216 --- /dev/null +++ b/parse/train/iaO86DUuKi/iaO86DUuKi_model.json @@ -0,0 +1,47633 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 633, + 1302, + 633, + 1302, + 1090, + 398, + 1090 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1236, + 1403, + 1236, + 1403, + 1512, + 298, + 1512 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1527, + 1403, + 1527, + 1403, + 1742, + 298, + 1742 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1757, + 1403, + 1757, + 1403, + 1972, + 297, + 1972 + ], + "score": 0.974 + }, + { + "category_id": 0, + "poly": [ + 298, + 218, + 1372, + 218, + 1372, + 269, + 298, + 269 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 328, + 2005, + 1110, + 2005, + 1110, + 2034, + 328, + 2034 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 302, + 1162, + 573, + 1162, + 573, + 1197, + 302, + 1197 + ], + "score": 0.881 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.87 + }, + { + "category_id": 0, + "poly": [ + 773, + 559, + 927, + 559, + 927, + 592, + 773, + 592 + ], + "score": 0.853 + }, + { + "category_id": 1, + "poly": [ + 315, + 387, + 754, + 387, + 754, + 477, + 315, + 477 + ], + "score": 0.815 + }, + { + "category_id": 1, + "poly": [ + 313, + 318, + 1056, + 318, + 1056, + 385, + 313, + 385 + ], + "score": 0.723 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.701 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 216.0, + 1380.0, + 216.0, + 1380.0, + 275.0, + 295.0, + 275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 329.0, + 1999.0, + 1112.0, + 1999.0, + 1112.0, + 2040.0, + 329.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 579.0, + 1159.0, + 579.0, + 1206.0, + 294.0, + 1206.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 555.0, + 934.0, + 555.0, + 934.0, + 598.0, + 768.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 633.0, + 1304.0, + 633.0, + 1304.0, + 668.0, + 395.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 665.0, + 1306.0, + 665.0, + 1306.0, + 700.0, + 394.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 695.0, + 1306.0, + 695.0, + 1306.0, + 729.0, + 394.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 724.0, + 1307.0, + 724.0, + 1307.0, + 762.0, + 393.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 756.0, + 1306.0, + 756.0, + 1306.0, + 791.0, + 394.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 787.0, + 1306.0, + 787.0, + 1306.0, + 822.0, + 394.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 815.0, + 1304.0, + 815.0, + 1304.0, + 852.0, + 393.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 848.0, + 1306.0, + 848.0, + 1306.0, + 881.0, + 393.0, + 881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 878.0, + 1306.0, + 878.0, + 1306.0, + 913.0, + 393.0, + 913.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 907.0, + 1304.0, + 907.0, + 1304.0, + 943.0, + 393.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 937.0, + 1306.0, + 937.0, + 1306.0, + 974.0, + 394.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 969.0, + 1306.0, + 969.0, + 1306.0, + 1004.0, + 394.0, + 1004.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1001.0, + 1303.0, + 1001.0, + 1303.0, + 1032.0, + 394.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1029.0, + 1303.0, + 1029.0, + 1303.0, + 1064.0, + 395.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 1062.0, + 1264.0, + 1062.0, + 1264.0, + 1094.0, + 399.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1238.0, + 1404.0, + 1238.0, + 1404.0, + 1271.0, + 296.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1267.0, + 1403.0, + 1267.0, + 1403.0, + 1300.0, + 296.0, + 1300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1297.0, + 1405.0, + 1297.0, + 1405.0, + 1336.0, + 292.0, + 1336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1330.0, + 1404.0, + 1330.0, + 1404.0, + 1362.0, + 294.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1358.0, + 1407.0, + 1358.0, + 1407.0, + 1395.0, + 294.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1387.0, + 1405.0, + 1387.0, + 1405.0, + 1426.0, + 293.0, + 1426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1420.0, + 1405.0, + 1420.0, + 1405.0, + 1456.0, + 294.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 1405.0, + 1450.0, + 1405.0, + 1486.0, + 294.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1482.0, + 1220.0, + 1482.0, + 1220.0, + 1515.0, + 296.0, + 1515.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1524.0, + 1405.0, + 1524.0, + 1405.0, + 1563.0, + 292.0, + 1563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1557.0, + 1405.0, + 1557.0, + 1405.0, + 1593.0, + 293.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1585.0, + 1405.0, + 1585.0, + 1405.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1683.0, + 293.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1678.0, + 1405.0, + 1678.0, + 1405.0, + 1713.0, + 293.0, + 1713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1711.0, + 1192.0, + 1711.0, + 1192.0, + 1745.0, + 296.0, + 1745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1756.0, + 1407.0, + 1756.0, + 1407.0, + 1794.0, + 293.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1787.0, + 1403.0, + 1787.0, + 1403.0, + 1821.0, + 295.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1817.0, + 1404.0, + 1817.0, + 1404.0, + 1853.0, + 291.0, + 1853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1846.0, + 1404.0, + 1846.0, + 1404.0, + 1887.0, + 292.0, + 1887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1878.0, + 1405.0, + 1878.0, + 1405.0, + 1914.0, + 292.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1909.0, + 1405.0, + 1909.0, + 1405.0, + 1944.0, + 294.0, + 1944.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1940.0, + 1402.0, + 1940.0, + 1402.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 379.0, + 758.0, + 379.0, + 758.0, + 422.0, + 311.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 411.0, + 703.0, + 411.0, + 703.0, + 454.0, + 310.0, + 454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 448.0, + 684.0, + 448.0, + 684.0, + 481.0, + 312.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 311.0, + 317.0, + 1058.0, + 317.0, + 1058.0, + 356.0, + 311.0, + 356.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 310.0, + 348.0, + 907.0, + 348.0, + 907.0, + 391.0, + 310.0, + 391.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 710, + 1404, + 710, + 1404, + 1168, + 298, + 1168 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 308, + 1403, + 308, + 1403, + 521, + 298, + 521 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1476, + 1404, + 1476, + 1404, + 1723, + 297, + 1723 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1182, + 1404, + 1182, + 1404, + 1462, + 297, + 1462 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1819, + 1404, + 1819, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 633, + 1401, + 633, + 1401, + 695, + 300, + 695 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1401, + 229, + 1401, + 293, + 294, + 293 + ], + "score": 0.952 + }, + { + "category_id": 0, + "poly": [ + 300, + 1765, + 1140, + 1765, + 1140, + 1801, + 300, + 1801 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.898 + }, + { + "category_id": 0, + "poly": [ + 300, + 566, + 578, + 566, + 578, + 600, + 300, + 600 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.7 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.18 + }, + { + "category_id": 13, + "poly": [ + 361, + 803, + 586, + 803, + 586, + 835, + 361, + 835 + ], + "score": 0.93, + "latex": "( \\mathcal S , \\mathcal A , P , R , \\gamma , \\mu , \\mathcal C )" + }, + { + "category_id": 13, + "poly": [ + 860, + 955, + 1028, + 955, + 1028, + 987, + 860, + 987 + ], + "score": 0.93, + "latex": "\\pi : S { \\mathcal { P } } ( A )" + }, + { + "category_id": 13, + "poly": [ + 1141, + 1599, + 1217, + 1599, + 1217, + 1633, + 1141, + 1633 + ], + "score": 0.93, + "latex": "\\mu ( a | s )" + }, + { + "category_id": 13, + "poly": [ + 798, + 1306, + 876, + 1306, + 876, + 1340, + 798, + 1340 + ], + "score": 0.93, + "latex": "V _ { C } ^ { \\pi } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 1206, + 1015, + 1393, + 1015, + 1393, + 1049, + 1206, + 1049 + ], + "score": 0.92, + "latex": "{ \\mathsf { \\bar { C } } } : { \\mathcal { S } } \\to \\{ 0 , 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 940, + 1182, + 1395, + 1182, + 1395, + 1219, + 940, + 1219 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { d ^ { \\pi } ( s ) = ( 1 - \\gamma ) \\sum _ { t = 0 } ^ { \\infty } \\gamma ^ { t } P ( s _ { t } = s | \\pi ) . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1122, + 1245, + 1403, + 1245, + 1403, + 1278, + 1122, + 1278 + ], + "score": 0.92, + "latex": "{ \\cal A } _ { R } ^ { \\pi } ( s , a ) = { \\cal Q } _ { R } ^ { \\pi } ( s , a ) -" + }, + { + "category_id": 13, + "poly": [ + 915, + 1630, + 991, + 1630, + 991, + 1663, + 915, + 1663 + ], + "score": 0.92, + "latex": "\\pi ( a | s )" + }, + { + "category_id": 13, + "poly": [ + 1286, + 1850, + 1402, + 1850, + 1402, + 1884, + 1286, + 1884 + ], + "score": 0.92, + "latex": "C ( s ) = 0" + }, + { + "category_id": 13, + "poly": [ + 1059, + 1077, + 1117, + 1077, + 1117, + 1110, + 1059, + 1110 + ], + "score": 0.92, + "latex": "C ( s )" + }, + { + "category_id": 13, + "poly": [ + 1275, + 1402, + 1353, + 1402, + 1353, + 1435, + 1275, + 1435 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 1070, + 1973, + 1181, + 1973, + 1181, + 2005, + 1070, + 2005 + ], + "score": 0.91, + "latex": "\\chi \\in [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 715, + 863, + 1188, + 863, + 1188, + 897, + 715, + 897 + ], + "score": 0.91, + "latex": "{ \\mathcal { C } } = \\{ ( c _ { i } : { \\mathcal { S } } \\{ 0 , 1 \\} , \\chi _ { i } \\in \\mathbb { R } ) | i \\in \\mathbb { Z } \\}" + }, + { + "category_id": 13, + "poly": [ + 595, + 1972, + 986, + 1972, + 986, + 2009, + 595, + 2009 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi } ( \\mu ) = \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ \\sum _ { t = 0 } ^ { \\infty } C ( s _ { t } ) \\right] \\le \\chi" + }, + { + "category_id": 13, + "poly": [ + 842, + 1850, + 957, + 1850, + 957, + 1883, + 842, + 1883 + ], + "score": 0.91, + "latex": "C ( s ) = 1" + }, + { + "category_id": 13, + "poly": [ + 1222, + 1820, + 1280, + 1820, + 1280, + 1853, + 1222, + 1853 + ], + "score": 0.91, + "latex": "C ( s )" + }, + { + "category_id": 13, + "poly": [ + 545, + 834, + 747, + 834, + 747, + 863, + 545, + 863 + ], + "score": 0.91, + "latex": "R : S \\times \\mathcal { A } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 296, + 1369, + 711, + 1369, + 711, + 1403, + 296, + 1403 + ], + "score": 0.91, + "latex": "\\mathbb { E } _ { \\tau \\sim \\pi } [ \\mathbb { 1 } \\{ f a i l u r e \\} ] = \\mathbb { P } ( f a i l u r e | \\mu )" + }, + { + "category_id": 13, + "poly": [ + 1019, + 1338, + 1403, + 1338, + 1403, + 1373, + 1019, + 1373 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { V _ { C } ^ { \\pi } ( \\mu ) = \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ \\sum _ { t = 0 } ^ { \\infty } C ( s _ { t } ) \\right] = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1048, + 834, + 1169, + 834, + 1169, + 865, + 1048, + 865 + ], + "score": 0.91, + "latex": "\\gamma \\in ( 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 515, + 1403, + 703, + 1403, + 703, + 1434, + 515, + 1434 + ], + "score": 0.9, + "latex": "C : \\mathcal { S } \\{ 0 , 1 \\}" + }, + { + "category_id": 13, + "poly": [ + 358, + 1305, + 742, + 1305, + 742, + 1341, + 358, + 1341 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { V _ { R } ^ { \\pi } ( \\mu ) = \\mathbb E _ { \\tau \\sim \\pi } \\left[ \\sum _ { t = 0 } ^ { \\infty } R ( s _ { t } , a _ { t } ) \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1278, + 371, + 1278, + 371, + 1310, + 297, + 1310 + ], + "score": 0.9, + "latex": "V _ { R } ^ { \\pi } ( s )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1245, + 761, + 1245, + 761, + 1279, + 296, + 1279 + ], + "score": 0.9, + "latex": "Q _ { R } ^ { \\pi } ( s , a ) = \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ R ( \\tau ) | s _ { 0 } = \\dot { s } , a _ { 0 } = a \\right]" + }, + { + "category_id": 13, + "poly": [ + 626, + 1215, + 986, + 1215, + 986, + 1248, + 626, + 1248 + ], + "score": 0.9, + "latex": "V _ { R } ^ { \\pi } ( s ) ~ = ~ \\mathbb { E } _ { \\tau \\sim \\pi } \\left[ R ( \\tau ) | s _ { 0 } = s \\right]" + }, + { + "category_id": 13, + "poly": [ + 734, + 1882, + 950, + 1882, + 950, + 1915, + 734, + 1915 + ], + "score": 0.9, + "latex": "C ( s ) \\ = \\ 0 \\ \\forall s \\in \\mathcal { S }" + }, + { + "category_id": 13, + "poly": [ + 1136, + 802, + 1401, + 802, + 1401, + 835, + 1136, + 835 + ], + "score": 0.89, + "latex": "P : \\bar { \\mathcal { S } } \\times \\mathcal { A } \\times \\mathcal { S } [ 0 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1223, + 1373, + 1289, + 1373, + 1289, + 1403, + 1223, + 1403 + ], + "score": 0.89, + "latex": "d ^ { \\pi } ( s )" + }, + { + "category_id": 13, + "poly": [ + 1327, + 1373, + 1393, + 1373, + 1393, + 1403, + 1327, + 1403 + ], + "score": 0.88, + "latex": "\\rho _ { \\phi } ( s )" + }, + { + "category_id": 13, + "poly": [ + 820, + 1510, + 845, + 1510, + 845, + 1540, + 820, + 1540 + ], + "score": 0.87, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 1052, + 1377, + 1085, + 1377, + 1085, + 1404, + 1052, + 1404 + ], + "score": 0.87, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 301, + 1046, + 519, + 1046, + 519, + 1080, + 301, + 1080 + ], + "score": 0.86, + "latex": "( C \\equiv \\mathbb { 1 } \\{ f a i l u r e \\} )" + }, + { + "category_id": 13, + "poly": [ + 1249, + 1278, + 1291, + 1278, + 1291, + 1306, + 1249, + 1306 + ], + "score": 0.86, + "latex": "A _ { C }" + }, + { + "category_id": 13, + "poly": [ + 568, + 1600, + 593, + 1600, + 593, + 1632, + 568, + 1632 + ], + "score": 0.86, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 568, + 928, + 598, + 928, + 598, + 956, + 568, + 956 + ], + "score": 0.86, + "latex": "\\chi _ { i }" + }, + { + "category_id": 13, + "poly": [ + 497, + 1540, + 523, + 1540, + 523, + 1571, + 497, + 1571 + ], + "score": 0.85, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 1289, + 461, + 1344, + 461, + 1344, + 490, + 1289, + 490 + ], + "score": 0.85, + "latex": "50 \\%" + }, + { + "category_id": 13, + "poly": [ + 1037, + 901, + 1062, + 901, + 1062, + 925, + 1037, + 925 + ], + "score": 0.84, + "latex": "c _ { i }" + }, + { + "category_id": 13, + "poly": [ + 671, + 804, + 693, + 804, + 693, + 830, + 671, + 830 + ], + "score": 0.82, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1276, + 1121, + 1276, + 1121, + 1306, + 1084, + 1306 + ], + "score": 0.82, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1278, + 1181, + 1278, + 1181, + 1308, + 1138, + 1308 + ], + "score": 0.82, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 894, + 804, + 919, + 804, + 919, + 831, + 894, + 831 + ], + "score": 0.79, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 1148, + 1051, + 1168, + 1051, + 1168, + 1078, + 1148, + 1078 + ], + "score": 0.78, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 297, + 869, + 318, + 869, + 318, + 896, + 297, + 896 + ], + "score": 0.77, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 694, + 1017, + 715, + 1017, + 715, + 1044, + 694, + 1044 + ], + "score": 0.77, + "latex": "\\mathcal { C }" + }, + { + "category_id": 13, + "poly": [ + 671, + 1437, + 691, + 1437, + 691, + 1458, + 671, + 1458 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 879, + 1190, + 901, + 1190, + 901, + 1212, + 879, + 1212 + ], + "score": 0.75, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 1109, + 1857, + 1126, + 1857, + 1126, + 1878, + 1109, + 1878 + ], + "score": 0.72, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 990, + 988, + 1014, + 988, + 1014, + 1013, + 990, + 1013 + ], + "score": 0.51, + "latex": "\\Pi" + }, + { + "category_id": 13, + "poly": [ + 446, + 1632, + 471, + 1632, + 471, + 1663, + 446, + 1663 + ], + "score": 0.3, + "latex": "\\mathrm { Q }" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1760.0, + 1145.0, + 1760.0, + 1145.0, + 1807.0, + 292.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 561.0, + 583.0, + 561.0, + 583.0, + 607.0, + 292.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 710.0, + 1404.0, + 710.0, + 1404.0, + 745.0, + 297.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 741.0, + 1405.0, + 741.0, + 1405.0, + 776.0, + 293.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 768.0, + 1406.0, + 768.0, + 1406.0, + 808.0, + 293.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 799.0, + 360.0, + 799.0, + 360.0, + 839.0, + 292.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 799.0, + 670.0, + 799.0, + 670.0, + 839.0, + 587.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 694.0, + 799.0, + 893.0, + 799.0, + 893.0, + 839.0, + 694.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 920.0, + 799.0, + 1135.0, + 799.0, + 1135.0, + 839.0, + 920.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 839.0, + 1402.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 833.0, + 544.0, + 833.0, + 544.0, + 868.0, + 293.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 833.0, + 1047.0, + 833.0, + 1047.0, + 868.0, + 748.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 833.0, + 1405.0, + 833.0, + 1405.0, + 868.0, + 1170.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 862.0, + 296.0, + 862.0, + 296.0, + 900.0, + 291.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 862.0, + 714.0, + 862.0, + 714.0, + 900.0, + 319.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 862.0, + 1406.0, + 862.0, + 1406.0, + 900.0, + 1189.0, + 900.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 894.0, + 1036.0, + 894.0, + 1036.0, + 928.0, + 292.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 894.0, + 1405.0, + 894.0, + 1405.0, + 928.0, + 1063.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 923.0, + 567.0, + 923.0, + 567.0, + 960.0, + 292.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 599.0, + 923.0, + 1407.0, + 923.0, + 1407.0, + 960.0, + 599.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 953.0, + 859.0, + 953.0, + 859.0, + 990.0, + 292.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 953.0, + 1406.0, + 953.0, + 1406.0, + 990.0, + 1029.0, + 990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 985.0, + 989.0, + 985.0, + 989.0, + 1020.0, + 294.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1015.0, + 985.0, + 1405.0, + 985.0, + 1405.0, + 1020.0, + 1015.0, + 1020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1012.0, + 693.0, + 1012.0, + 693.0, + 1052.0, + 292.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1012.0, + 1205.0, + 1012.0, + 1205.0, + 1052.0, + 716.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1012.0, + 1405.0, + 1012.0, + 1405.0, + 1052.0, + 1394.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1045.0, + 300.0, + 1045.0, + 300.0, + 1082.0, + 293.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 520.0, + 1045.0, + 1147.0, + 1045.0, + 1147.0, + 1082.0, + 520.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 1045.0, + 1405.0, + 1045.0, + 1405.0, + 1082.0, + 1169.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1073.0, + 1058.0, + 1073.0, + 1058.0, + 1115.0, + 291.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 1073.0, + 1405.0, + 1073.0, + 1405.0, + 1115.0, + 1118.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1108.0, + 1404.0, + 1108.0, + 1404.0, + 1140.0, + 296.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1138.0, + 1350.0, + 1138.0, + 1350.0, + 1169.0, + 296.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 306.0, + 1407.0, + 306.0, + 1407.0, + 346.0, + 293.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 337.0, + 1405.0, + 337.0, + 1405.0, + 373.0, + 294.0, + 373.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 369.0, + 1405.0, + 369.0, + 1405.0, + 403.0, + 294.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 401.0, + 1407.0, + 401.0, + 1407.0, + 433.0, + 293.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 428.0, + 1408.0, + 428.0, + 1408.0, + 465.0, + 292.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 460.0, + 1288.0, + 460.0, + 1288.0, + 495.0, + 293.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 460.0, + 1407.0, + 460.0, + 1407.0, + 495.0, + 1345.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 492.0, + 649.0, + 492.0, + 649.0, + 523.0, + 293.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1477.0, + 1402.0, + 1477.0, + 1402.0, + 1510.0, + 296.0, + 1510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1507.0, + 819.0, + 1507.0, + 819.0, + 1542.0, + 294.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1507.0, + 1404.0, + 1507.0, + 1404.0, + 1542.0, + 846.0, + 1542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1538.0, + 496.0, + 1538.0, + 496.0, + 1572.0, + 295.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1538.0, + 1404.0, + 1538.0, + 1404.0, + 1572.0, + 524.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1569.0, + 1402.0, + 1569.0, + 1402.0, + 1603.0, + 295.0, + 1603.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1598.0, + 567.0, + 1598.0, + 567.0, + 1634.0, + 292.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1598.0, + 1140.0, + 1598.0, + 1140.0, + 1634.0, + 594.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1598.0, + 1406.0, + 1598.0, + 1406.0, + 1634.0, + 1218.0, + 1634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1631.0, + 445.0, + 1631.0, + 445.0, + 1665.0, + 295.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 472.0, + 1631.0, + 914.0, + 1631.0, + 914.0, + 1665.0, + 472.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1631.0, + 1406.0, + 1631.0, + 1406.0, + 1665.0, + 992.0, + 1665.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1662.0, + 1404.0, + 1662.0, + 1404.0, + 1692.0, + 296.0, + 1692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1693.0, + 771.0, + 1693.0, + 771.0, + 1724.0, + 296.0, + 1724.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1168.0, + 878.0, + 1168.0, + 878.0, + 1234.0, + 287.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 902.0, + 1168.0, + 939.0, + 1168.0, + 939.0, + 1234.0, + 902.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1396.0, + 1168.0, + 1414.0, + 1168.0, + 1414.0, + 1234.0, + 1396.0, + 1234.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1211.0, + 625.0, + 1211.0, + 625.0, + 1252.0, + 292.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1211.0, + 1409.0, + 1211.0, + 1409.0, + 1252.0, + 987.0, + 1252.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1240.0, + 295.0, + 1240.0, + 295.0, + 1284.0, + 291.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 1240.0, + 1121.0, + 1240.0, + 1121.0, + 1284.0, + 762.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1271.0, + 296.0, + 1271.0, + 296.0, + 1313.0, + 291.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1271.0, + 1083.0, + 1271.0, + 1083.0, + 1313.0, + 372.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1122.0, + 1271.0, + 1137.0, + 1271.0, + 1137.0, + 1313.0, + 1122.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1271.0, + 1248.0, + 1271.0, + 1248.0, + 1313.0, + 1182.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1292.0, + 1271.0, + 1409.0, + 1271.0, + 1409.0, + 1313.0, + 1292.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1296.0, + 357.0, + 1296.0, + 357.0, + 1350.0, + 288.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 1296.0, + 797.0, + 1296.0, + 797.0, + 1350.0, + 743.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 877.0, + 1296.0, + 1411.0, + 1296.0, + 1411.0, + 1350.0, + 877.0, + 1350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1333.0, + 1018.0, + 1333.0, + 1018.0, + 1381.0, + 290.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1333.0, + 1410.0, + 1333.0, + 1410.0, + 1381.0, + 1404.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1366.0, + 295.0, + 1366.0, + 295.0, + 1409.0, + 291.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 1366.0, + 1051.0, + 1366.0, + 1051.0, + 1409.0, + 712.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1366.0, + 1222.0, + 1366.0, + 1222.0, + 1409.0, + 1086.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 1366.0, + 1326.0, + 1366.0, + 1326.0, + 1409.0, + 1290.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1366.0, + 1405.0, + 1366.0, + 1405.0, + 1409.0, + 1394.0, + 1409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1397.0, + 514.0, + 1397.0, + 514.0, + 1438.0, + 292.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 704.0, + 1397.0, + 1274.0, + 1397.0, + 1274.0, + 1438.0, + 704.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1397.0, + 1409.0, + 1397.0, + 1409.0, + 1438.0, + 1354.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1429.0, + 670.0, + 1429.0, + 670.0, + 1467.0, + 292.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 1429.0, + 703.0, + 1429.0, + 703.0, + 1467.0, + 692.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1818.0, + 1221.0, + 1818.0, + 1221.0, + 1855.0, + 293.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 1818.0, + 1404.0, + 1818.0, + 1404.0, + 1855.0, + 1281.0, + 1855.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1848.0, + 841.0, + 1848.0, + 841.0, + 1886.0, + 293.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1848.0, + 1108.0, + 1848.0, + 1108.0, + 1886.0, + 958.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1127.0, + 1848.0, + 1285.0, + 1848.0, + 1285.0, + 1886.0, + 1127.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1881.0, + 733.0, + 1881.0, + 733.0, + 1916.0, + 296.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1881.0, + 1406.0, + 1881.0, + 1406.0, + 1916.0, + 951.0, + 1916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1947.0, + 294.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1406.0, + 1943.0, + 1406.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1958.0, + 594.0, + 1958.0, + 594.0, + 2020.0, + 286.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 987.0, + 1958.0, + 1069.0, + 1958.0, + 1069.0, + 2020.0, + 987.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 1958.0, + 1413.0, + 1958.0, + 1413.0, + 2020.0, + 1182.0, + 2020.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2001.0, + 1405.0, + 2001.0, + 1405.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 630.0, + 1405.0, + 630.0, + 1405.0, + 668.0, + 295.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 665.0, + 1342.0, + 665.0, + 1342.0, + 697.0, + 298.0, + 697.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1406.0, + 228.0, + 1406.0, + 267.0, + 294.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 261.0, + 1073.0, + 261.0, + 1073.0, + 297.0, + 296.0, + 297.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1664, + 1405, + 1664, + 1405, + 1912, + 297, + 1912 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 296, + 1190, + 1405, + 1190, + 1405, + 1444, + 296, + 1444 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 296, + 720, + 1403, + 720, + 1403, + 907, + 296, + 907 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 297, + 919, + 1405, + 919, + 1405, + 1044, + 297, + 1044 + ], + "score": 0.974 + }, + { + "category_id": 3, + "poly": [ + 427, + 244, + 1282, + 244, + 1282, + 558, + 427, + 558 + ], + "score": 0.972 + }, + { + "category_id": 8, + "poly": [ + 393, + 1523, + 1306, + 1523, + 1306, + 1660, + 393, + 1660 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 296, + 642, + 1400, + 642, + 1400, + 706, + 296, + 706 + ], + "score": 0.948 + }, + { + "category_id": 1, + "poly": [ + 297, + 1925, + 1398, + 1925, + 1398, + 1987, + 297, + 1987 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 293, + 1058, + 1400, + 1058, + 1400, + 1120, + 293, + 1120 + ], + "score": 0.944 + }, + { + "category_id": 8, + "poly": [ + 485, + 1121, + 1213, + 1121, + 1213, + 1176, + 485, + 1176 + ], + "score": 0.932 + }, + { + "category_id": 1, + "poly": [ + 294, + 1456, + 1402, + 1456, + 1402, + 1520, + 294, + 1520 + ], + "score": 0.929 + }, + { + "category_id": 4, + "poly": [ + 292, + 566, + 1402, + 566, + 1402, + 625, + 292, + 625 + ], + "score": 0.929 + }, + { + "category_id": 8, + "poly": [ + 482, + 1989, + 1221, + 1989, + 1221, + 2035, + 482, + 2035 + ], + "score": 0.911 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 812, + 76, + 812, + 104, + 300, + 104 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1124, + 1400, + 1124, + 1400, + 1154, + 1366, + 1154 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1575, + 1400, + 1575, + 1400, + 1604, + 1366, + 1604 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1365, + 2000, + 1400, + 2000, + 1400, + 2029, + 1365, + 2029 + ], + "score": 0.845 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.694 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.385 + }, + { + "category_id": 13, + "poly": [ + 586, + 595, + 695, + 595, + 695, + 625, + 586, + 625 + ], + "score": 0.93, + "latex": "C ( s ^ { \\prime } ) = 1" + }, + { + "category_id": 13, + "poly": [ + 420, + 874, + 523, + 874, + 523, + 907, + 420, + 907 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 672, + 752, + 776, + 752, + 776, + 785, + 672, + 785 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1300, + 874, + 1376, + 874, + 1376, + 907, + 1300, + 907 + ], + "score": 0.93, + "latex": "\\pi ( a | s )" + }, + { + "category_id": 13, + "poly": [ + 361, + 843, + 466, + 843, + 466, + 876, + 361, + 876 + ], + "score": 0.92, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1278, + 567, + 1357, + 567, + 1357, + 598, + 1278, + 598 + ], + "score": 0.92, + "latex": "R ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 650, + 1818, + 712, + 1818, + 712, + 1849, + 650, + 1849 + ], + "score": 0.92, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 450, + 1757, + 512, + 1757, + 512, + 1788, + 450, + 1788 + ], + "score": 0.91, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 1115, + 1406, + 1156, + 1406, + 1156, + 1442, + 1115, + 1442 + ], + "score": 0.91, + "latex": "\\hat { A } _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1757, + 1154, + 1757, + 1154, + 1789, + 1111, + 1789 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 824, + 1193, + 868, + 1193, + 868, + 1225, + 824, + 1225 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 298, + 594, + 359, + 594, + 359, + 625, + 298, + 625 + ], + "score": 0.9, + "latex": "C \\bar { ( } s ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 723, + 1316, + 766, + 1316, + 766, + 1346, + 723, + 1346 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 772, + 1376, + 816, + 1376, + 816, + 1408, + 772, + 1408 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 365, + 1660, + 413, + 1660, + 413, + 1694, + 365, + 1694 + ], + "score": 0.89, + "latex": "\\hat { B } ^ { \\pi _ { \\phi } }" + }, + { + "category_id": 13, + "poly": [ + 689, + 1488, + 732, + 1488, + 732, + 1520, + 689, + 1520 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 14, + "poly": [ + 388, + 1522, + 1305, + 1522, + 1305, + 1663, + 388, + 1663 + ], + "score": 0.89, + "latex": "\\begin{array} { r l } & { \\hat { Q } _ { C } ^ { k + 1 } \\gets \\underset { Q _ { C } } { \\mathrm { a r g m i n } } \\quad \\cdot \\left( - \\mathbb { E } _ { s \\sim \\mathcal { D } _ { e n v } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q _ { C } ( s , a ) ] + \\mathbb { E } _ { ( s , a ) \\sim \\mathcal { D } _ { e n v } } [ Q _ { C } ( s , a ) ] \\right) } \\\\ & { \\quad \\quad \\quad + \\frac { 1 } { 2 } \\mathbb { E } _ { ( s , a , s ^ { \\prime } , c ) \\sim \\mathcal { D } _ { e n v } } \\left[ \\left( Q _ { C } ( s , a ) - \\hat { \\mathcal { B } } ^ { \\pi _ { \\phi } } \\hat { Q } _ { C } ^ { k } ( s , a ) \\right) ^ { 2 } \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1218, + 1788, + 1261, + 1788, + 1261, + 1819, + 1218, + 1819 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1263, + 1315, + 1307, + 1315, + 1307, + 1346, + 1263, + 1346 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 924, + 1850, + 966, + 1850, + 966, + 1880, + 924, + 1880 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 438, + 952, + 481, + 952, + 481, + 984, + 438, + 984 + ], + "score": 0.88, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 945, + 1487, + 980, + 1487, + 980, + 1515, + 945, + 1515 + ], + "score": 0.88, + "latex": "k ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 992, + 596, + 1032, + 596, + 1032, + 624, + 992, + 624 + ], + "score": 0.88, + "latex": "Q _ { C }" + }, + { + "category_id": 14, + "poly": [ + 486, + 1119, + 1214, + 1119, + 1214, + 1179, + 486, + 1179 + ], + "score": 0.87, + "latex": "\\pi ^ { * } = \\underset { \\pi \\in \\Pi _ { C } } { \\operatorname { a r g m a x } } V _ { R } ^ { \\pi } ( \\mu ) \\quad \\mathrm { ~ w h e r e ~ } \\quad \\Pi _ { C } = \\{ \\pi \\in \\Pi : V _ { C } ^ { \\pi } ( \\mu ) \\leq \\chi \\}" + }, + { + "category_id": 13, + "poly": [ + 396, + 1487, + 496, + 1487, + 496, + 1521, + 396, + 1521 + ], + "score": 0.87, + "latex": "C Q L ( \\zeta )" + }, + { + "category_id": 13, + "poly": [ + 978, + 1062, + 1009, + 1062, + 1009, + 1087, + 978, + 1087 + ], + "score": 0.87, + "latex": "\\pi ^ { * }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1883, + 331, + 1883, + 331, + 1913, + 297, + 1913 + ], + "score": 0.85, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 567, + 1138, + 567, + 1138, + 593, + 1116, + 593 + ], + "score": 0.83, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 791, + 1489, + 810, + 1489, + 810, + 1515, + 791, + 1515 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 928, + 1882, + 946, + 1882, + 946, + 1906, + 928, + 1906 + ], + "score": 0.82, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 929, + 1349, + 950, + 1349, + 950, + 1372, + 929, + 1372 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1133, + 440, + 1185, + 440, + 1185, + 466, + 1133, + 466 + ], + "score": 0.79, + "latex": "( s , a )" + }, + { + "category_id": 14, + "poly": [ + 480, + 1986, + 1222, + 1986, + 1222, + 2043, + 480, + 2043 + ], + "score": 0.77, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi } , a \\sim \\pi _ { \\phi } } \\big [ \\bar { A } _ { R } ^ { \\pi _ { \\phi } } ( s , a ) \\big ] \\ \\quad \\mathrm { s . t . } \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi } , a \\sim \\pi _ { \\phi } } Q _ { C } ( s , a ) \\leq \\chi" + }, + { + "category_id": 13, + "poly": [ + 507, + 1489, + 526, + 1489, + 526, + 1520, + 507, + 1520 + ], + "score": 0.76, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 1355, + 849, + 1372, + 849, + 1372, + 871, + 1355, + 871 + ], + "score": 0.75, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 394, + 1702, + 416, + 1702, + 416, + 1723, + 394, + 1723 + ], + "score": 0.74, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 355, + 880, + 371, + 880, + 371, + 901, + 355, + 901 + ], + "score": 0.71, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1016, + 879, + 1034, + 879, + 1034, + 901, + 1016, + 901 + ], + "score": 0.68, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 759, + 879, + 775, + 879, + 775, + 901, + 759, + 901 + ], + "score": 0.61, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1136, + 1346, + 1237, + 1346, + 1237, + 1376, + 1136, + 1376 + ], + "score": 0.56, + "latex": "\\operatorname { C Q L } ( { \\mathcal { H } } )" + }, + { + "category_id": 13, + "poly": [ + 620, + 568, + 748, + 568, + 748, + 597, + 620, + 597 + ], + "score": 0.4, + "latex": "e n v . s t e p ( a )" + }, + { + "category_id": 13, + "poly": [ + 1042, + 525, + 1058, + 525, + 1058, + 548, + 1042, + 548 + ], + "score": 0.33, + "latex": "\\phi" + }, + { + "category_id": 14, + "poly": [ + 875, + 1119, + 1212, + 1119, + 1212, + 1158, + 875, + 1158 + ], + "score": 0.28, + "latex": "\\Pi _ { C } = \\{ \\pi \\in \\Pi : V _ { C } ^ { \\pi } ( \\mu ) \\leq \\chi \\}" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 243.0, + 721.0, + 243.0, + 721.0, + 271.0, + 591.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 996.0, + 259.0, + 1214.0, + 259.0, + 1214.0, + 288.0, + 996.0, + 288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 659.0, + 298.0, + 693.0, + 298.0, + 693.0, + 324.0, + 659.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 816.0, + 297.0, + 856.0, + 297.0, + 856.0, + 326.0, + 816.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 312.0, + 1061.0, + 312.0, + 1061.0, + 328.0, + 1042.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 294.0, + 1229.0, + 294.0, + 1229.0, + 315.0, + 1088.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1013.0, + 328.0, + 1047.0, + 328.0, + 1047.0, + 420.0, + 1013.0, + 420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 350.0, + 1059.0, + 350.0, + 1059.0, + 365.0, + 1042.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 567.0, + 368.0, + 692.0, + 368.0, + 692.0, + 401.0, + 567.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 371.0, + 820.0, + 371.0, + 820.0, + 404.0, + 696.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 371.0, + 927.0, + 371.0, + 927.0, + 397.0, + 842.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 389.0, + 1058.0, + 389.0, + 1058.0, + 401.0, + 1043.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 425.0, + 1060.0, + 425.0, + 1060.0, + 440.0, + 1042.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 444.0, + 556.0, + 444.0, + 556.0, + 467.0, + 449.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 673.0, + 443.0, + 768.0, + 443.0, + 768.0, + 472.0, + 673.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 452.0, + 1122.0, + 452.0, + 1122.0, + 474.0, + 1102.0, + 474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1129.0, + 433.0, + 1132.0, + 433.0, + 1132.0, + 472.0, + 1129.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 433.0, + 1192.0, + 433.0, + 1192.0, + 472.0, + 1186.0, + 472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 472.0, + 1123.0, + 472.0, + 1123.0, + 501.0, + 1101.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 564.0, + 519.0, + 830.0, + 519.0, + 830.0, + 547.0, + 564.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 931.0, + 521.0, + 1041.0, + 521.0, + 1041.0, + 550.0, + 931.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 521.0, + 1207.0, + 521.0, + 1207.0, + 550.0, + 1059.0, + 550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 567.0, + 619.0, + 567.0, + 619.0, + 598.0, + 296.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 567.0, + 1115.0, + 567.0, + 1115.0, + 598.0, + 749.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1139.0, + 567.0, + 1277.0, + 567.0, + 1277.0, + 598.0, + 1139.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 567.0, + 1404.0, + 567.0, + 1404.0, + 598.0, + 1358.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 594.0, + 585.0, + 594.0, + 585.0, + 625.0, + 360.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 696.0, + 594.0, + 991.0, + 594.0, + 991.0, + 625.0, + 696.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 594.0, + 1303.0, + 594.0, + 1303.0, + 625.0, + 1033.0, + 625.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1660.0, + 364.0, + 1660.0, + 364.0, + 1701.0, + 292.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1660.0, + 1405.0, + 1660.0, + 1405.0, + 1701.0, + 414.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1694.0, + 393.0, + 1694.0, + 393.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 417.0, + 1694.0, + 1405.0, + 1694.0, + 1405.0, + 1730.0, + 417.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1726.0, + 1405.0, + 1726.0, + 1405.0, + 1760.0, + 295.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1755.0, + 449.0, + 1755.0, + 449.0, + 1793.0, + 292.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 1755.0, + 1110.0, + 1755.0, + 1110.0, + 1793.0, + 513.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1755.0, + 1407.0, + 1755.0, + 1407.0, + 1793.0, + 1155.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1789.0, + 1217.0, + 1789.0, + 1217.0, + 1819.0, + 296.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1789.0, + 1403.0, + 1789.0, + 1403.0, + 1819.0, + 1262.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1817.0, + 649.0, + 1817.0, + 649.0, + 1851.0, + 294.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 713.0, + 1817.0, + 1403.0, + 1817.0, + 1403.0, + 1851.0, + 713.0, + 1851.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1845.0, + 923.0, + 1845.0, + 923.0, + 1886.0, + 294.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1845.0, + 1405.0, + 1845.0, + 1405.0, + 1886.0, + 967.0, + 1886.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1879.0, + 296.0, + 1879.0, + 296.0, + 1914.0, + 292.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 1879.0, + 927.0, + 1879.0, + 927.0, + 1914.0, + 332.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 1879.0, + 1231.0, + 1879.0, + 1231.0, + 1914.0, + 947.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1192.0, + 823.0, + 1192.0, + 823.0, + 1226.0, + 295.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 1192.0, + 1407.0, + 1192.0, + 1407.0, + 1226.0, + 869.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1224.0, + 1405.0, + 1224.0, + 1405.0, + 1258.0, + 294.0, + 1258.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1253.0, + 1405.0, + 1253.0, + 1405.0, + 1287.0, + 296.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1285.0, + 1405.0, + 1285.0, + 1405.0, + 1319.0, + 295.0, + 1319.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1315.0, + 722.0, + 1315.0, + 722.0, + 1349.0, + 295.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 767.0, + 1315.0, + 1262.0, + 1315.0, + 1262.0, + 1349.0, + 767.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1315.0, + 1405.0, + 1315.0, + 1405.0, + 1349.0, + 1308.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1344.0, + 928.0, + 1344.0, + 928.0, + 1377.0, + 294.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1344.0, + 1135.0, + 1344.0, + 1135.0, + 1377.0, + 951.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1238.0, + 1344.0, + 1403.0, + 1344.0, + 1403.0, + 1377.0, + 1238.0, + 1377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1372.0, + 771.0, + 1372.0, + 771.0, + 1412.0, + 291.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 817.0, + 1372.0, + 1406.0, + 1372.0, + 1406.0, + 1412.0, + 817.0, + 1412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1408.0, + 1114.0, + 1408.0, + 1114.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1408.0, + 1168.0, + 1408.0, + 1168.0, + 1446.0, + 1157.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 719.0, + 1404.0, + 719.0, + 1404.0, + 755.0, + 294.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 753.0, + 671.0, + 753.0, + 671.0, + 785.0, + 295.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 753.0, + 1404.0, + 753.0, + 1404.0, + 785.0, + 777.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 780.0, + 1406.0, + 780.0, + 1406.0, + 820.0, + 293.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 1401.0, + 812.0, + 1401.0, + 848.0, + 295.0, + 848.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 844.0, + 360.0, + 844.0, + 360.0, + 876.0, + 296.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 844.0, + 1354.0, + 844.0, + 1354.0, + 876.0, + 467.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 844.0, + 1404.0, + 844.0, + 1404.0, + 876.0, + 1373.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 874.0, + 354.0, + 874.0, + 354.0, + 910.0, + 294.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 874.0, + 419.0, + 874.0, + 419.0, + 910.0, + 372.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 874.0, + 758.0, + 874.0, + 758.0, + 910.0, + 524.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 874.0, + 1015.0, + 874.0, + 1015.0, + 910.0, + 776.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 874.0, + 1299.0, + 874.0, + 1299.0, + 910.0, + 1035.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 874.0, + 1390.0, + 874.0, + 1390.0, + 910.0, + 1377.0, + 910.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 918.0, + 1403.0, + 918.0, + 1403.0, + 953.0, + 295.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 952.0, + 437.0, + 952.0, + 437.0, + 985.0, + 295.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 952.0, + 1403.0, + 952.0, + 1403.0, + 985.0, + 482.0, + 985.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 980.0, + 1405.0, + 980.0, + 1405.0, + 1015.0, + 294.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1014.0, + 449.0, + 1014.0, + 449.0, + 1042.0, + 295.0, + 1042.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 641.0, + 1404.0, + 641.0, + 1404.0, + 679.0, + 295.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 674.0, + 569.0, + 674.0, + 569.0, + 707.0, + 294.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1921.0, + 1403.0, + 1921.0, + 1403.0, + 1962.0, + 292.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1956.0, + 865.0, + 1956.0, + 865.0, + 1990.0, + 296.0, + 1990.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1059.0, + 977.0, + 1059.0, + 977.0, + 1094.0, + 297.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1059.0, + 1403.0, + 1059.0, + 1403.0, + 1094.0, + 1010.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1090.0, + 970.0, + 1090.0, + 970.0, + 1123.0, + 293.0, + 1123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1455.0, + 1403.0, + 1455.0, + 1403.0, + 1491.0, + 296.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1484.0, + 395.0, + 1484.0, + 395.0, + 1524.0, + 293.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 497.0, + 1484.0, + 506.0, + 1484.0, + 506.0, + 1524.0, + 497.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1484.0, + 688.0, + 1484.0, + 688.0, + 1524.0, + 527.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 733.0, + 1484.0, + 790.0, + 1484.0, + 790.0, + 1524.0, + 733.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1484.0, + 944.0, + 1484.0, + 944.0, + 1524.0, + 811.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 1484.0, + 1171.0, + 1484.0, + 1171.0, + 1524.0, + 981.0, + 1524.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1031, + 1404, + 1031, + 1404, + 1189, + 296, + 1189 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1602, + 1405, + 1602, + 1405, + 1727, + 297, + 1727 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1200, + 1404, + 1200, + 1404, + 1356, + 297, + 1356 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1372, + 1405, + 1372, + 1405, + 1482, + 297, + 1482 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 344, + 1403, + 344, + 1403, + 437, + 299, + 437 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 300, + 510, + 1405, + 510, + 1405, + 1001, + 300, + 1001 + ], + "score": 0.966 + }, + { + "category_id": 1, + "poly": [ + 298, + 1914, + 1405, + 1914, + 1405, + 2035, + 298, + 2035 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 301, + 1494, + 1402, + 1494, + 1402, + 1590, + 301, + 1590 + ], + "score": 0.961 + }, + { + "category_id": 1, + "poly": [ + 292, + 1740, + 1398, + 1740, + 1398, + 1804, + 292, + 1804 + ], + "score": 0.943 + }, + { + "category_id": 8, + "poly": [ + 527, + 290, + 1141, + 290, + 1141, + 339, + 527, + 339 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 293, + 229, + 1400, + 229, + 1400, + 289, + 293, + 289 + ], + "score": 0.93 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.913 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1844, + 1400, + 1844, + 1400, + 1873, + 1366, + 1873 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.774 + }, + { + "category_id": 8, + "poly": [ + 470, + 1807, + 1229, + 1807, + 1229, + 1913, + 470, + 1913 + ], + "score": 0.751 + }, + { + "category_id": 0, + "poly": [ + 298, + 468, + 1045, + 468, + 1045, + 500, + 298, + 500 + ], + "score": 0.616 + }, + { + "category_id": 8, + "poly": [ + 476, + 1807, + 880, + 1807, + 880, + 1864, + 476, + 1864 + ], + "score": 0.402 + }, + { + "category_id": 6, + "poly": [ + 298, + 468, + 1045, + 468, + 1045, + 500, + 298, + 500 + ], + "score": 0.23 + }, + { + "category_id": 8, + "poly": [ + 464, + 1871, + 1250, + 1871, + 1250, + 1913, + 464, + 1913 + ], + "score": 0.158 + }, + { + "category_id": 13, + "poly": [ + 643, + 1154, + 789, + 1154, + 789, + 1189, + 643, + 1189 + ], + "score": 0.94, + "latex": "a \\sim \\pi _ { \\phi _ { o l d } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 1113, + 1262, + 1217, + 1262, + 1217, + 1295, + 1113, + 1295 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 944, + 1154, + 1097, + 1154, + 1097, + 1188, + 944, + 1188 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a ) \\leq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1267, + 1292, + 1372, + 1292, + 1372, + 1326, + 1267, + 1326 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 635, + 1201, + 731, + 1201, + 731, + 1236, + 635, + 1236 + ], + "score": 0.93, + "latex": "\\pi _ { \\phi _ { o l d } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 1062, + 1369, + 1400, + 1369, + 1400, + 1410, + 1062, + 1410 + ], + "score": 0.93, + "latex": "\\epsilon = ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) )" + }, + { + "category_id": 13, + "poly": [ + 754, + 1633, + 907, + 1633, + 907, + 1667, + 754, + 1667 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a ) \\leq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 379, + 1406, + 497, + 1406, + 497, + 1445, + 379, + 1445 + ], + "score": 0.92, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 1278, + 1525, + 1392, + 1525, + 1392, + 1559, + 1278, + 1559 + ], + "score": 0.92, + "latex": "c = { C } ( s ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 899, + 1495, + 1042, + 1495, + 1042, + 1528, + 899, + 1528 + ], + "score": 0.92, + "latex": "( s , a , s ^ { \\prime } , r , c )" + }, + { + "category_id": 13, + "poly": [ + 505, + 802, + 617, + 802, + 617, + 832, + 505, + 832 + ], + "score": 0.91, + "latex": "\\{ 1 , . . . , \\Nu \\}" + }, + { + "category_id": 13, + "poly": [ + 419, + 1446, + 536, + 1446, + 536, + 1484, + 419, + 1484 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 690, + 1973, + 967, + 1973, + 967, + 2007, + 690, + 2007 + ], + "score": 0.91, + "latex": "D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) )" + }, + { + "category_id": 13, + "poly": [ + 767, + 1294, + 932, + 1294, + 932, + 1326, + 767, + 1326 + ], + "score": 0.91, + "latex": "Q _ { C } ( s , a ) \\ \\leq \\ \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 757, + 1262, + 918, + 1262, + 918, + 1294, + 757, + 1294 + ], + "score": 0.91, + "latex": "Q _ { C } ( s , a ) \\leq \\epsilon" + }, + { + "category_id": 14, + "poly": [ + 526, + 288, + 1143, + 288, + 1143, + 344, + 526, + 344 + ], + "score": 0.9, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\operatorname* { m i n } _ { \\lambda \\geq 0 } \\mathbb { E } _ { s \\sim \\rho _ { \\phi } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi } } ( s , a ) - \\lambda \\left( Q _ { C } ( s , a ) - \\chi \\right) \\right]" + }, + { + "category_id": 13, + "poly": [ + 781, + 857, + 1035, + 857, + 1035, + 887, + 781, + 887 + ], + "score": 0.9, + "latex": "\\zeta : = \\zeta - \\eta _ { Q } \\nabla _ { \\zeta } { \\bar { C } } Q L ( \\zeta )" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1665, + 1194, + 1665, + 1194, + 1697, + 1150, + 1697 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1665, + 435, + 1665, + 435, + 1696, + 374, + 1696 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 1252, + 1094, + 1295, + 1094, + 1295, + 1126, + 1252, + 1126 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 519, + 1497, + 579, + 1497, + 579, + 1526, + 519, + 1526 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 483, + 375, + 527, + 375, + 527, + 407, + 483, + 407 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 614, + 1915, + 667, + 1915, + 667, + 1943, + 614, + 1943 + ], + "score": 0.89, + "latex": "D _ { \\mathrm { K L } }" + }, + { + "category_id": 13, + "poly": [ + 924, + 720, + 981, + 720, + 981, + 746, + 924, + 746 + ], + "score": 0.89, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 542, + 602, + 657, + 602, + 657, + 632, + 542, + 632 + ], + "score": 0.89, + "latex": "\\{ 1 , . . . , \\mathbf { M } \\}" + }, + { + "category_id": 13, + "poly": [ + 379, + 542, + 547, + 542, + 547, + 577, + 379, + 577 + ], + "score": 0.89, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\chi" + }, + { + "category_id": 13, + "poly": [ + 728, + 768, + 982, + 768, + 982, + 806, + 728, + 806 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\sum _ { e = 1 } ^ { M } \\hat { V } _ { C } ^ { e } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 629, + 511, + 663, + 511, + 663, + 544, + 629, + 544 + ], + "score": 0.87, + "latex": "Q _ { \\zeta } ^ { s }" + }, + { + "category_id": 13, + "poly": [ + 1342, + 803, + 1399, + 803, + 1399, + 829, + 1342, + 829 + ], + "score": 0.87, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 781, + 664, + 923, + 664, + 923, + 693, + 781, + 693 + ], + "score": 0.87, + "latex": "Q _ { C } ( s , a ) \\leq \\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1126, + 512, + 1188, + 512, + 1188, + 541, + 1126, + 541 + ], + "score": 0.86, + "latex": "\\epsilon , \\delta , \\chi" + }, + { + "category_id": 13, + "poly": [ + 452, + 632, + 761, + 632, + 761, + 666, + 452, + 666 + ], + "score": 0.86, + "latex": "\\epsilon ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) )" + }, + { + "category_id": 13, + "poly": [ + 496, + 667, + 634, + 667, + 634, + 693, + 496, + 693 + ], + "score": 0.86, + "latex": "a \\sim \\pi _ { \\phi _ { o l d } } ( s )" + }, + { + "category_id": 13, + "poly": [ + 609, + 832, + 628, + 832, + 628, + 858, + 609, + 858 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 549, + 1528, + 572, + 1528, + 572, + 1553, + 549, + 1553 + ], + "score": 0.84, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 838, + 542, + 963, + 542, + 963, + 578, + 838, + 578 + ], + "score": 0.84, + "latex": "> \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 435, + 511, + 469, + 511, + 469, + 541, + 435, + 541 + ], + "score": 0.83, + "latex": "V _ { \\theta } ^ { r }" + }, + { + "category_id": 13, + "poly": [ + 757, + 692, + 862, + 692, + 862, + 721, + 757, + 721 + ], + "score": 0.82, + "latex": "c = { C } ( { s } ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 732, + 1696, + 752, + 1696, + 752, + 1722, + 732, + 1722 + ], + "score": 0.8, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 413, + 719, + 755, + 719, + 755, + 749, + 413, + 749 + ], + "score": 0.8, + "latex": "\\mathcal { D } _ { e n v } \\gets \\mathcal { D } _ { e n v } \\cup \\{ ( s , a , s ^ { \\prime } , r , c ) \\}" + }, + { + "category_id": 13, + "poly": [ + 949, + 511, + 1006, + 511, + 1006, + 539, + 949, + 539 + ], + "score": 0.79, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 881, + 515, + 911, + 515, + 911, + 541, + 881, + 541 + ], + "score": 0.78, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 959, + 2006, + 985, + 2006, + 985, + 2030, + 959, + 2030 + ], + "score": 0.77, + "latex": "\\pmb { F }" + }, + { + "category_id": 13, + "poly": [ + 536, + 1873, + 984, + 1873, + 984, + 1913, + 536, + 1913 + ], + "score": 0.77, + "latex": "\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta" + }, + { + "category_id": 13, + "poly": [ + 873, + 888, + 892, + 888, + 892, + 910, + 873, + 910 + ], + "score": 0.77, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 497, + 607, + 512, + 607, + 512, + 627, + 497, + 627 + ], + "score": 0.75, + "latex": "e" + }, + { + "category_id": 13, + "poly": [ + 732, + 669, + 748, + 669, + 748, + 689, + 732, + 689 + ], + "score": 0.75, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 602, + 691, + 746, + 691, + 746, + 719, + 602, + 719 + ], + "score": 0.75, + "latex": "{ \\bf \\omega } ^ { \\prime } , \\dot { r } = R ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 702, + 1454, + 717, + 1454, + 717, + 1476, + 702, + 1476 + ], + "score": 0.74, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1123, + 1300, + 1142, + 1300, + 1142, + 1320, + 1123, + 1320 + ], + "score": 0.74, + "latex": "a" + }, + { + "category_id": 14, + "poly": [ + 475, + 1803, + 884, + 1803, + 884, + 1865, + 475, + 1865 + ], + "score": 0.74, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 690, + 1563, + 706, + 1563, + 706, + 1584, + 690, + 1584 + ], + "score": 0.73, + "latex": "c" + }, + { + "category_id": 13, + "poly": [ + 759, + 1531, + 776, + 1531, + 776, + 1553, + 759, + 1553 + ], + "score": 0.73, + "latex": "r" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1501, + 1146, + 1501, + 1146, + 1523, + 1129, + 1523 + ], + "score": 0.72, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 367, + 1381, + 383, + 1381, + 383, + 1402, + 367, + 1402 + ], + "score": 0.7, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 473, + 1299, + 491, + 1299, + 491, + 1320, + 473, + 1320 + ], + "score": 0.69, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 618, + 1268, + 635, + 1268, + 635, + 1290, + 618, + 1290 + ], + "score": 0.69, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 298, + 1269, + 316, + 1269, + 316, + 1290, + 298, + 1290 + ], + "score": 0.68, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 463, + 804, + 476, + 804, + 476, + 827, + 463, + 827 + ], + "score": 0.68, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 1383, + 1503, + 1401, + 1503, + 1401, + 1523, + 1383, + 1523 + ], + "score": 0.67, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 1085, + 670, + 1100, + 670, + 1100, + 688, + 1085, + 688 + ], + "score": 0.65, + "latex": "a" + }, + { + "category_id": 13, + "poly": [ + 922, + 513, + 940, + 513, + 940, + 536, + 922, + 536 + ], + "score": 0.58, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 590, + 693, + 610, + 693, + 610, + 716, + 590, + 716 + ], + "score": 0.52, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 378, + 939, + 484, + 939, + 484, + 970, + 378, + 970 + ], + "score": 0.36, + "latex": "\\phi _ { o l d } \\phi" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 464.0, + 1049.0, + 464.0, + 1049.0, + 507.0, + 296.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 464.0, + 1049.0, + 464.0, + 1049.0, + 507.0, + 296.0, + 507.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1032.0, + 1405.0, + 1032.0, + 1405.0, + 1066.0, + 296.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1060.0, + 1404.0, + 1060.0, + 1404.0, + 1101.0, + 293.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1093.0, + 1251.0, + 1093.0, + 1251.0, + 1130.0, + 294.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1093.0, + 1406.0, + 1093.0, + 1406.0, + 1130.0, + 1296.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1118.0, + 1405.0, + 1118.0, + 1405.0, + 1164.0, + 291.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1152.0, + 642.0, + 1152.0, + 642.0, + 1192.0, + 293.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 1152.0, + 943.0, + 1152.0, + 943.0, + 1192.0, + 790.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.0, + 1152.0, + 1109.0, + 1152.0, + 1109.0, + 1192.0, + 1098.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1604.0, + 1403.0, + 1604.0, + 1403.0, + 1636.0, + 296.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1633.0, + 753.0, + 1633.0, + 753.0, + 1669.0, + 294.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 908.0, + 1633.0, + 1405.0, + 1633.0, + 1405.0, + 1669.0, + 908.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1662.0, + 373.0, + 1662.0, + 373.0, + 1701.0, + 292.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1662.0, + 1149.0, + 1662.0, + 1149.0, + 1701.0, + 436.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1195.0, + 1662.0, + 1406.0, + 1662.0, + 1406.0, + 1701.0, + 1195.0, + 1701.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1694.0, + 731.0, + 1694.0, + 731.0, + 1730.0, + 295.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 1694.0, + 1323.0, + 1694.0, + 1323.0, + 1730.0, + 753.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1198.0, + 634.0, + 1198.0, + 634.0, + 1238.0, + 295.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 732.0, + 1198.0, + 1406.0, + 1198.0, + 1406.0, + 1238.0, + 732.0, + 1238.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1230.0, + 1405.0, + 1230.0, + 1405.0, + 1266.0, + 294.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1259.0, + 297.0, + 1259.0, + 297.0, + 1297.0, + 292.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 1259.0, + 617.0, + 1259.0, + 617.0, + 1297.0, + 317.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 636.0, + 1259.0, + 756.0, + 1259.0, + 756.0, + 1297.0, + 636.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1259.0, + 1112.0, + 1259.0, + 1112.0, + 1297.0, + 919.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1259.0, + 1406.0, + 1259.0, + 1406.0, + 1297.0, + 1218.0, + 1297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1291.0, + 472.0, + 1291.0, + 472.0, + 1329.0, + 294.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 492.0, + 1291.0, + 766.0, + 1291.0, + 766.0, + 1329.0, + 492.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1291.0, + 1122.0, + 1291.0, + 1122.0, + 1329.0, + 933.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1291.0, + 1266.0, + 1291.0, + 1266.0, + 1329.0, + 1143.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1373.0, + 1291.0, + 1406.0, + 1291.0, + 1406.0, + 1329.0, + 1373.0, + 1329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1322.0, + 762.0, + 1322.0, + 762.0, + 1359.0, + 295.0, + 1359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1363.0, + 366.0, + 1363.0, + 366.0, + 1416.0, + 290.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 1363.0, + 1061.0, + 1363.0, + 1061.0, + 1416.0, + 384.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1363.0, + 1408.0, + 1363.0, + 1408.0, + 1416.0, + 1401.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 281.0, + 1389.0, + 378.0, + 1389.0, + 378.0, + 1458.0, + 281.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 1389.0, + 1417.0, + 1389.0, + 1417.0, + 1458.0, + 498.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 281.0, + 1436.0, + 418.0, + 1436.0, + 418.0, + 1494.0, + 281.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 537.0, + 1436.0, + 701.0, + 1436.0, + 701.0, + 1494.0, + 537.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 1436.0, + 1286.0, + 1436.0, + 1286.0, + 1494.0, + 718.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 342.0, + 1404.0, + 342.0, + 1404.0, + 380.0, + 295.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 375.0, + 482.0, + 375.0, + 482.0, + 409.0, + 295.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 528.0, + 375.0, + 1405.0, + 375.0, + 1405.0, + 409.0, + 528.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 405.0, + 1272.0, + 405.0, + 1272.0, + 440.0, + 293.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 504.0, + 434.0, + 504.0, + 434.0, + 545.0, + 303.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 504.0, + 628.0, + 504.0, + 628.0, + 545.0, + 470.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 504.0, + 880.0, + 504.0, + 880.0, + 545.0, + 664.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 504.0, + 921.0, + 504.0, + 921.0, + 545.0, + 912.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 504.0, + 948.0, + 504.0, + 948.0, + 545.0, + 941.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1007.0, + 504.0, + 1125.0, + 504.0, + 1125.0, + 545.0, + 1007.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 504.0, + 1199.0, + 504.0, + 1199.0, + 545.0, + 1189.0, + 545.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 532.0, + 378.0, + 532.0, + 378.0, + 586.0, + 299.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 548.0, + 532.0, + 563.0, + 532.0, + 563.0, + 586.0, + 548.0, + 586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 538.0, + 837.0, + 538.0, + 837.0, + 581.0, + 824.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 538.0, + 1406.0, + 538.0, + 1406.0, + 581.0, + 964.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 568.0, + 670.0, + 568.0, + 670.0, + 607.0, + 301.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 743.0, + 572.0, + 1402.0, + 572.0, + 1402.0, + 607.0, + 743.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 599.0, + 337.0, + 599.0, + 337.0, + 631.0, + 304.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 597.0, + 496.0, + 597.0, + 496.0, + 635.0, + 370.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 513.0, + 597.0, + 541.0, + 597.0, + 541.0, + 635.0, + 513.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 658.0, + 597.0, + 698.0, + 597.0, + 698.0, + 635.0, + 658.0, + 635.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 632.0, + 335.0, + 632.0, + 335.0, + 666.0, + 304.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 614.0, + 451.0, + 614.0, + 451.0, + 684.0, + 401.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 762.0, + 614.0, + 778.0, + 614.0, + 778.0, + 684.0, + 762.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 662.0, + 335.0, + 662.0, + 335.0, + 694.0, + 304.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 659.0, + 495.0, + 659.0, + 495.0, + 698.0, + 409.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 635.0, + 659.0, + 731.0, + 659.0, + 731.0, + 698.0, + 635.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 659.0, + 780.0, + 659.0, + 780.0, + 698.0, + 749.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 659.0, + 1084.0, + 659.0, + 1084.0, + 698.0, + 924.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1101.0, + 659.0, + 1112.0, + 659.0, + 1112.0, + 698.0, + 1101.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 688.0, + 335.0, + 688.0, + 335.0, + 721.0, + 304.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 688.0, + 589.0, + 688.0, + 589.0, + 721.0, + 412.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 688.0, + 756.0, + 688.0, + 756.0, + 721.0, + 747.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 863.0, + 688.0, + 874.0, + 688.0, + 874.0, + 721.0, + 863.0, + 721.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 304.0, + 716.0, + 335.0, + 716.0, + 335.0, + 749.0, + 304.0, + 749.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 712.0, + 412.0, + 712.0, + 412.0, + 753.0, + 408.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 756.0, + 712.0, + 923.0, + 712.0, + 923.0, + 753.0, + 756.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 982.0, + 712.0, + 1405.0, + 712.0, + 1405.0, + 753.0, + 982.0, + 753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 744.0, + 338.0, + 744.0, + 338.0, + 804.0, + 295.0, + 804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 744.0, + 462.0, + 744.0, + 462.0, + 824.0, + 363.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 744.0, + 727.0, + 744.0, + 727.0, + 824.0, + 477.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 800.0, + 338.0, + 800.0, + 338.0, + 833.0, + 294.0, + 833.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 800.0, + 462.0, + 800.0, + 462.0, + 832.0, + 375.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 477.0, + 800.0, + 504.0, + 800.0, + 504.0, + 832.0, + 477.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 618.0, + 800.0, + 655.0, + 800.0, + 655.0, + 832.0, + 618.0, + 832.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 797.0, + 1341.0, + 797.0, + 1341.0, + 836.0, + 949.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 797.0, + 1404.0, + 797.0, + 1404.0, + 836.0, + 1400.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 826.0, + 339.0, + 826.0, + 339.0, + 863.0, + 293.0, + 863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 826.0, + 608.0, + 826.0, + 608.0, + 861.0, + 413.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 629.0, + 826.0, + 1242.0, + 826.0, + 1242.0, + 861.0, + 629.0, + 861.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 854.0, + 338.0, + 854.0, + 338.0, + 887.0, + 294.0, + 887.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 412.0, + 853.0, + 780.0, + 853.0, + 780.0, + 890.0, + 412.0, + 890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 880.0, + 339.0, + 880.0, + 339.0, + 917.0, + 291.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 880.0, + 872.0, + 880.0, + 872.0, + 918.0, + 411.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 893.0, + 880.0, + 1058.0, + 880.0, + 1058.0, + 918.0, + 893.0, + 918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 909.0, + 339.0, + 909.0, + 339.0, + 941.0, + 293.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 909.0, + 465.0, + 909.0, + 465.0, + 943.0, + 374.0, + 943.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 938.0, + 339.0, + 938.0, + 339.0, + 970.0, + 293.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 938.0, + 377.0, + 938.0, + 377.0, + 972.0, + 371.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 938.0, + 488.0, + 938.0, + 488.0, + 972.0, + 485.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 964.0, + 428.0, + 964.0, + 428.0, + 999.0, + 294.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1908.0, + 613.0, + 1908.0, + 613.0, + 1947.0, + 295.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1908.0, + 1403.0, + 1908.0, + 1403.0, + 1947.0, + 668.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1944.0, + 1403.0, + 1944.0, + 1403.0, + 1974.0, + 297.0, + 1974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1969.0, + 689.0, + 1969.0, + 689.0, + 2012.0, + 293.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1969.0, + 1405.0, + 1969.0, + 1405.0, + 2012.0, + 968.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 2002.0, + 958.0, + 2002.0, + 958.0, + 2037.0, + 296.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 986.0, + 2002.0, + 1403.0, + 2002.0, + 1403.0, + 2037.0, + 986.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1493.0, + 518.0, + 1493.0, + 518.0, + 1531.0, + 294.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1493.0, + 898.0, + 1493.0, + 898.0, + 1531.0, + 580.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 1493.0, + 1128.0, + 1493.0, + 1128.0, + 1531.0, + 1043.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1493.0, + 1382.0, + 1493.0, + 1382.0, + 1531.0, + 1147.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1493.0, + 1408.0, + 1493.0, + 1408.0, + 1531.0, + 1402.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1524.0, + 548.0, + 1524.0, + 548.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 573.0, + 1524.0, + 758.0, + 1524.0, + 758.0, + 1562.0, + 573.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1524.0, + 1277.0, + 1524.0, + 1277.0, + 1562.0, + 777.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1524.0, + 1406.0, + 1524.0, + 1406.0, + 1562.0, + 1393.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1554.0, + 689.0, + 1554.0, + 689.0, + 1591.0, + 295.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1554.0, + 1391.0, + 1554.0, + 1391.0, + 1591.0, + 707.0, + 1591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1737.0, + 1402.0, + 1737.0, + 1402.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1772.0, + 1145.0, + 1772.0, + 1145.0, + 1808.0, + 295.0, + 1808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 1404.0, + 225.0, + 1404.0, + 268.0, + 293.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 258.0, + 452.0, + 258.0, + 452.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1092, + 1406, + 1092, + 1406, + 1378, + 295, + 1378 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 955, + 1404, + 955, + 1404, + 1080, + 297, + 1080 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 296, + 817, + 1404, + 817, + 1404, + 941, + 296, + 941 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 1870, + 1405, + 1870, + 1405, + 1967, + 297, + 1967 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 297, + 1751, + 1404, + 1751, + 1404, + 1854, + 297, + 1854 + ], + "score": 0.964 + }, + { + "category_id": 1, + "poly": [ + 296, + 1532, + 1403, + 1532, + 1403, + 1641, + 296, + 1641 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 298, + 580, + 1404, + 580, + 1404, + 673, + 298, + 673 + ], + "score": 0.945 + }, + { + "category_id": 1, + "poly": [ + 298, + 1673, + 1401, + 1673, + 1401, + 1737, + 298, + 1737 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 422, + 1969, + 1278, + 1969, + 1278, + 2045, + 422, + 2045 + ], + "score": 0.937 + }, + { + "category_id": 1, + "poly": [ + 301, + 377, + 1404, + 377, + 1404, + 442, + 301, + 442 + ], + "score": 0.915 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 814, + 74, + 814, + 106, + 299, + 106 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 300, + 736, + 711, + 736, + 711, + 772, + 300, + 772 + ], + "score": 0.902 + }, + { + "category_id": 1, + "poly": [ + 296, + 1390, + 1402, + 1390, + 1402, + 1452, + 296, + 1452 + ], + "score": 0.89 + }, + { + "category_id": 8, + "poly": [ + 299, + 445, + 1396, + 445, + 1396, + 515, + 299, + 515 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1365, + 338, + 1400, + 338, + 1400, + 367, + 1365, + 367 + ], + "score": 0.866 + }, + { + "category_id": 9, + "poly": [ + 1365, + 537, + 1400, + 537, + 1400, + 567, + 1365, + 567 + ], + "score": 0.865 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1992, + 1401, + 1992, + 1401, + 2022, + 1365, + 2022 + ], + "score": 0.854 + }, + { + "category_id": 8, + "poly": [ + 491, + 1455, + 1207, + 1455, + 1207, + 1523, + 491, + 1523 + ], + "score": 0.835 + }, + { + "category_id": 1, + "poly": [ + 296, + 228, + 1212, + 228, + 1212, + 263, + 296, + 263 + ], + "score": 0.824 + }, + { + "category_id": 2, + "poly": [ + 840, + 2087, + 859, + 2087, + 859, + 2112, + 840, + 2112 + ], + "score": 0.76 + }, + { + "category_id": 8, + "poly": [ + 301, + 265, + 1364, + 265, + 1364, + 327, + 301, + 327 + ], + "score": 0.742 + }, + { + "category_id": 1, + "poly": [ + 301, + 334, + 811, + 334, + 811, + 374, + 301, + 374 + ], + "score": 0.628 + }, + { + "category_id": 8, + "poly": [ + 300, + 520, + 721, + 520, + 721, + 580, + 300, + 580 + ], + "score": 0.267 + }, + { + "category_id": 1, + "poly": [ + 301, + 265, + 1364, + 265, + 1364, + 327, + 301, + 327 + ], + "score": 0.171 + }, + { + "category_id": 1, + "poly": [ + 300, + 520, + 721, + 520, + 721, + 580, + 300, + 580 + ], + "score": 0.099 + }, + { + "category_id": 13, + "poly": [ + 910, + 1534, + 1028, + 1534, + 1028, + 1574, + 910, + 1574 + ], + "score": 0.93, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)" + }, + { + "category_id": 13, + "poly": [ + 679, + 1751, + 796, + 1751, + 796, + 1791, + 679, + 1791 + ], + "score": 0.93, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 945, + 1128, + 1260, + 1128, + 1260, + 1175, + 945, + 1175 + ], + "score": 0.93, + "latex": "\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ \\hat { A } _ { C } ( s , a ) ]" + }, + { + "category_id": 13, + "poly": [ + 1192, + 1524, + 1392, + 1524, + 1392, + 1580, + 1192, + 1580 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 494, + 1217, + 612, + 1217, + 612, + 1254, + 494, + 1254 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)" + }, + { + "category_id": 13, + "poly": [ + 873, + 1216, + 991, + 1216, + 991, + 1254, + 873, + 1254 + ], + "score": 0.92, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 898, + 374, + 940, + 374, + 940, + 410, + 898, + 410 + ], + "score": 0.91, + "latex": "\\hat { A } _ { C }" + }, + { + "category_id": 13, + "poly": [ + 476, + 1095, + 856, + 1095, + 856, + 1132, + 476, + 1132 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ( s , a ) | } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 516, + 1818, + 644, + 1818, + 644, + 1855, + 516, + 1855 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)" + }, + { + "category_id": 14, + "poly": [ + 491, + 1452, + 1210, + 1452, + 1210, + 1523, + 491, + 1523 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma }" + }, + { + "category_id": 14, + "poly": [ + 424, + 1965, + 1276, + 1965, + 1276, + 2045, + 424, + 2045 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad w h e r e \\quad \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 13, + "poly": [ + 1051, + 1214, + 1401, + 1214, + 1401, + 1254, + 1051, + 1254 + ], + "score": 0.91, + "latex": "\\zeta = | \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) - V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) | )" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1253, + 1142, + 1253, + 1142, + 1284, + 1036, + 1284 + ], + "score": 0.9, + "latex": "\\mathrm { R e g } _ { C } ( T )" + }, + { + "category_id": 13, + "poly": [ + 1161, + 1791, + 1230, + 1791, + 1230, + 1821, + 1161, + 1821 + ], + "score": 0.9, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 518, + 381, + 560, + 381, + 560, + 410, + 518, + 410 + ], + "score": 0.89, + "latex": "A _ { C }" + }, + { + "category_id": 13, + "poly": [ + 546, + 1347, + 606, + 1347, + 606, + 1377, + 546, + 1377 + ], + "score": 0.89, + "latex": "\\pi _ { \\phi _ { o l d } }" + }, + { + "category_id": 13, + "poly": [ + 998, + 1760, + 1059, + 1760, + 1059, + 1790, + 998, + 1790 + ], + "score": 0.89, + "latex": "\\pi _ { \\phi _ { o l d } }" + }, + { + "category_id": 13, + "poly": [ + 484, + 1171, + 1000, + 1171, + 1000, + 1217, + 484, + 1217 + ], + "score": 0.89, + "latex": "\\Delta = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ \\hat { A } _ { C } ( s , a ) - A _ { C } ( s , a ) ]" + }, + { + "category_id": 13, + "poly": [ + 943, + 1254, + 979, + 1254, + 979, + 1281, + 943, + 1281 + ], + "score": 0.89, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1217, + 1608, + 1252, + 1608, + 1252, + 1638, + 1217, + 1638 + ], + "score": 0.88, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 641, + 1936, + 710, + 1936, + 710, + 1967, + 641, + 1967 + ], + "score": 0.86, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 470, + 1931, + 596, + 1931, + 596, + 1967, + 470, + 1967 + ], + "score": 0.86, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 1060, + 1179, + 1078, + 1179, + 1078, + 1210, + 1060, + 1210 + ], + "score": 0.85, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 536, + 1424, + 636, + 1424, + 636, + 1453, + 536, + 1453 + ], + "score": 0.85, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 298, + 1791, + 366, + 1791, + 366, + 1821, + 298, + 1821 + ], + "score": 0.85, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 680, + 1903, + 776, + 1903, + 776, + 1932, + 680, + 1932 + ], + "score": 0.85, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 372, + 1576, + 405, + 1576, + 405, + 1605, + 372, + 1605 + ], + "score": 0.84, + "latex": "C ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1706, + 1067, + 1706, + 1067, + 1732, + 1040, + 1732 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 914, + 1097, + 941, + 1097, + 941, + 1123, + 914, + 1123 + ], + "score": 0.82, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 344, + 1252, + 372, + 1252, + 372, + 1278, + 344, + 1278 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 823, + 1282, + 847, + 1282, + 847, + 1308, + 823, + 1308 + ], + "score": 0.82, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 369, + 1540, + 387, + 1540, + 387, + 1571, + 369, + 1571 + ], + "score": 0.82, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 625, + 1610, + 652, + 1610, + 652, + 1635, + 625, + 1635 + ], + "score": 0.77, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 364, + 520, + 722, + 520, + 722, + 581, + 364, + 581 + ], + "score": 0.68, + "latex": "\\frac { 1 } { 2 } ( \\phi - \\phi _ { o l d } ) ^ { T } { \\cal F } ( \\phi - \\phi _ { o l d } ) \\leq \\delta" + }, + { + "category_id": 14, + "poly": [ + 363, + 334, + 811, + 334, + 811, + 374, + 363, + 374 + ], + "score": 0.67, + "latex": "\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta" + }, + { + "category_id": 14, + "poly": [ + 294, + 262, + 1368, + 262, + 1368, + 333, + 294, + 333 + ], + "score": 0.66, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\quad \\mathrm { s . t . } \\quad V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi" + }, + { + "category_id": 14, + "poly": [ + 364, + 521, + 722, + 521, + 722, + 582, + 364, + 582 + ], + "score": 0.61, + "latex": "\\frac { 1 } { 2 } ( \\phi - \\phi _ { o l d } ) ^ { T } { \\cal F } ( \\phi - \\phi _ { o l d } ) \\leq \\delta" + }, + { + "category_id": 13, + "poly": [ + 683, + 1396, + 697, + 1396, + 697, + 1419, + 683, + 1419 + ], + "score": 0.57, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 735, + 1582, + 756, + 1582, + 756, + 1604, + 735, + 1604 + ], + "score": 0.49, + "latex": "\\omega" + }, + { + "category_id": 14, + "poly": [ + 310, + 440, + 1406, + 440, + 1406, + 585, + 310, + 585 + ], + "score": 0.45, + "latex": "\\begin{array} { r l r } { { \\operatorname* { m a x } \\operatorname* { m i n } _ { \\lambda \\geq 0 } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } \\big ( s , a \\big ) \\Big ] - \\lambda ( V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\big ( \\mu \\big ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ \\hat { A } _ { C } ( s , a ) \\Big ] - \\chi ) } } \\\\ { \\cdot \\ L _ { \\delta } } & { \\frac { 1 } { 2 } ( \\phi - \\phi _ { o l d } ) ^ { T } F ( \\phi - \\phi _ { o l d } ) \\leq \\delta } & { ( 6 ) } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 732.0, + 715.0, + 732.0, + 715.0, + 777.0, + 292.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1090.0, + 475.0, + 1090.0, + 475.0, + 1136.0, + 293.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 1090.0, + 913.0, + 1090.0, + 913.0, + 1136.0, + 857.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1090.0, + 1407.0, + 1090.0, + 1407.0, + 1136.0, + 942.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1126.0, + 944.0, + 1126.0, + 944.0, + 1183.0, + 289.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1261.0, + 1126.0, + 1408.0, + 1126.0, + 1408.0, + 1183.0, + 1261.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1171.0, + 483.0, + 1171.0, + 483.0, + 1224.0, + 290.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1001.0, + 1171.0, + 1059.0, + 1171.0, + 1059.0, + 1224.0, + 1001.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1171.0, + 1409.0, + 1171.0, + 1409.0, + 1224.0, + 1079.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1204.0, + 493.0, + 1204.0, + 493.0, + 1264.0, + 288.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 613.0, + 1204.0, + 872.0, + 1204.0, + 872.0, + 1264.0, + 613.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1204.0, + 1050.0, + 1204.0, + 1050.0, + 1264.0, + 992.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1204.0, + 1412.0, + 1204.0, + 1412.0, + 1264.0, + 1402.0, + 1264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1248.0, + 343.0, + 1248.0, + 343.0, + 1286.0, + 294.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1248.0, + 942.0, + 1248.0, + 942.0, + 1286.0, + 373.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 1248.0, + 1035.0, + 1248.0, + 1035.0, + 1286.0, + 980.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 1248.0, + 1407.0, + 1248.0, + 1407.0, + 1286.0, + 1143.0, + 1286.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1280.0, + 822.0, + 1280.0, + 822.0, + 1313.0, + 294.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 1280.0, + 1406.0, + 1280.0, + 1406.0, + 1313.0, + 848.0, + 1313.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 1404.0, + 1309.0, + 1404.0, + 1347.0, + 293.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1340.5, + 618.0, + 1340.5, + 618.0, + 1379.5, + 294.0, + 1379.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 957.0, + 1405.0, + 957.0, + 1405.0, + 989.0, + 296.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 988.0, + 1404.0, + 988.0, + 1404.0, + 1021.0, + 295.0, + 1021.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1017.0, + 1405.0, + 1017.0, + 1405.0, + 1054.0, + 294.0, + 1054.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1049.0, + 801.0, + 1049.0, + 801.0, + 1082.0, + 295.0, + 1082.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 817.0, + 1404.0, + 817.0, + 1404.0, + 852.0, + 293.0, + 852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 846.0, + 1406.0, + 846.0, + 1406.0, + 884.0, + 294.0, + 884.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 878.0, + 1404.0, + 878.0, + 1404.0, + 915.0, + 293.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 913.0, + 419.0, + 913.0, + 419.0, + 941.0, + 292.0, + 941.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1870.0, + 1403.0, + 1870.0, + 1403.0, + 1905.0, + 297.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1901.0, + 679.0, + 1901.0, + 679.0, + 1936.0, + 296.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 1901.0, + 1403.0, + 1901.0, + 1403.0, + 1936.0, + 777.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1924.0, + 469.0, + 1924.0, + 469.0, + 1977.0, + 284.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 597.0, + 1924.0, + 640.0, + 1924.0, + 640.0, + 1977.0, + 597.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1924.0, + 1130.0, + 1924.0, + 1130.0, + 1977.0, + 711.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1740.0, + 678.0, + 1740.0, + 678.0, + 1801.0, + 287.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 1740.0, + 997.0, + 1740.0, + 997.0, + 1801.0, + 797.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1060.0, + 1740.0, + 1409.0, + 1740.0, + 1409.0, + 1801.0, + 1060.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1782.0, + 297.0, + 1782.0, + 297.0, + 1829.0, + 291.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1782.0, + 1160.0, + 1782.0, + 1160.0, + 1829.0, + 367.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1231.0, + 1782.0, + 1407.0, + 1782.0, + 1407.0, + 1829.0, + 1231.0, + 1829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1815.0, + 515.0, + 1815.0, + 515.0, + 1860.0, + 294.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 1815.0, + 1042.0, + 1815.0, + 1042.0, + 1860.0, + 645.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1522.0, + 368.0, + 1522.0, + 368.0, + 1585.0, + 289.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 388.0, + 1522.0, + 909.0, + 1522.0, + 909.0, + 1585.0, + 388.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1522.0, + 1191.0, + 1522.0, + 1191.0, + 1585.0, + 1029.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1393.0, + 1522.0, + 1399.0, + 1522.0, + 1399.0, + 1585.0, + 1393.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1575.0, + 371.0, + 1575.0, + 371.0, + 1611.0, + 295.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1575.0, + 734.0, + 1575.0, + 734.0, + 1611.0, + 406.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 1575.0, + 1404.0, + 1575.0, + 1404.0, + 1611.0, + 757.0, + 1611.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1605.0, + 624.0, + 1605.0, + 624.0, + 1644.0, + 294.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 653.0, + 1605.0, + 1216.0, + 1605.0, + 1216.0, + 1644.0, + 653.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1605.0, + 1262.0, + 1605.0, + 1262.0, + 1644.0, + 1253.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 578.0, + 1404.0, + 578.0, + 1404.0, + 616.0, + 294.0, + 616.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 608.0, + 1405.0, + 608.0, + 1405.0, + 646.0, + 293.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 640.0, + 668.0, + 640.0, + 668.0, + 674.0, + 293.0, + 674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1673.0, + 1406.0, + 1673.0, + 1406.0, + 1709.0, + 296.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1702.0, + 1039.0, + 1702.0, + 1039.0, + 1742.0, + 293.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1702.0, + 1244.0, + 1702.0, + 1244.0, + 1742.0, + 1068.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 377.0, + 517.0, + 377.0, + 517.0, + 413.0, + 297.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 561.0, + 377.0, + 897.0, + 377.0, + 897.0, + 413.0, + 561.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 941.0, + 377.0, + 1404.0, + 377.0, + 1404.0, + 413.0, + 941.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 408.0, + 1295.0, + 408.0, + 1295.0, + 444.0, + 296.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1392.0, + 682.0, + 1392.0, + 682.0, + 1427.0, + 295.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 698.0, + 1392.0, + 1404.0, + 1392.0, + 1404.0, + 1427.0, + 698.0, + 1427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1422.0, + 535.0, + 1422.0, + 535.0, + 1458.0, + 295.0, + 1458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 224.0, + 1217.0, + 224.0, + 1217.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 328.0, + 362.0, + 328.0, + 362.0, + 380.0, + 292.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 328.0, + 818.0, + 328.0, + 818.0, + 380.0, + 812.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1369.0, + 261.0, + 1377.0, + 261.0, + 1377.0, + 334.0, + 1369.0, + 334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 294.0, + 1030.0, + 294.0, + 1030.0, + 332.0, + 953.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 532.0, + 309.0, + 532.0, + 309.0, + 569.0, + 293.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 375.0, + 551.0, + 388.0, + 551.0, + 388.0, + 578.0, + 375.0, + 578.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1055, + 1403, + 1055, + 1403, + 1241, + 297, + 1241 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 228, + 1403, + 228, + 1403, + 414, + 298, + 414 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 561, + 1403, + 561, + 1403, + 689, + 297, + 689 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 298, + 428, + 1403, + 428, + 1403, + 554, + 298, + 554 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 298, + 1911, + 1404, + 1911, + 1404, + 2034, + 298, + 2034 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 299, + 1402, + 1403, + 1402, + 1403, + 1531, + 299, + 1531 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 300, + 1644, + 1401, + 1644, + 1401, + 1738, + 300, + 1738 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 298, + 1249, + 1403, + 1249, + 1403, + 1382, + 298, + 1382 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 313, + 691, + 1384, + 691, + 1384, + 770, + 313, + 770 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 362, + 1764, + 1324, + 1764, + 1324, + 1835, + 362, + 1835 + ], + "score": 0.927 + }, + { + "category_id": 0, + "poly": [ + 300, + 1575, + 558, + 1575, + 558, + 1609, + 300, + 1609 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 299, + 1869, + 642, + 1869, + 642, + 1901, + 299, + 1901 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 296, + 788, + 1405, + 788, + 1405, + 1041, + 296, + 1041 + ], + "score": 0.847 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 859, + 2089, + 859, + 2112, + 840, + 2112 + ], + "score": 0.785 + }, + { + "category_id": 13, + "poly": [ + 509, + 1494, + 599, + 1494, + 599, + 1532, + 509, + 1532 + ], + "score": 0.94, + "latex": "\\mathcal { O } ( \\sqrt { T } )" + }, + { + "category_id": 13, + "poly": [ + 532, + 1278, + 708, + 1278, + 708, + 1313, + 532, + 1313 + ], + "score": 0.93, + "latex": "\\chi _ { t } \\stackrel { \\cdot } { = } \\mathcal { O } ( 1 / \\sqrt { t } )" + }, + { + "category_id": 13, + "poly": [ + 335, + 820, + 460, + 820, + 460, + 862, + 335, + 862 + ], + "score": 0.92, + "latex": "\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 352, + 624, + 451, + 624, + 451, + 660, + 352, + 660 + ], + "score": 0.92, + "latex": "\\phi ^ { ( 0 ) } = 0" + }, + { + "category_id": 13, + "poly": [ + 613, + 628, + 681, + 628, + 681, + 660, + 613, + 660 + ], + "score": 0.91, + "latex": "\\eta > 0" + }, + { + "category_id": 13, + "poly": [ + 378, + 1342, + 680, + 1342, + 680, + 1382, + 378, + 1382 + ], + "score": 0.91, + "latex": "R e g _ { C } ( T ) = \\mathcal { O } ( \\sqrt { | \\mathcal { S } | | \\mathcal { A } | T } )" + }, + { + "category_id": 13, + "poly": [ + 1015, + 1311, + 1119, + 1311, + 1119, + 1344, + 1015, + 1344 + ], + "score": 0.91, + "latex": "R e g _ { C } ( T )" + }, + { + "category_id": 13, + "poly": [ + 472, + 296, + 541, + 296, + 541, + 326, + 472, + 326 + ], + "score": 0.9, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 1330, + 358, + 1399, + 358, + 1399, + 387, + 1330, + 387 + ], + "score": 0.9, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 298, + 327, + 366, + 327, + 366, + 356, + 298, + 356 + ], + "score": 0.89, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 643, + 660, + 717, + 660, + 717, + 687, + 643, + 687 + ], + "score": 0.89, + "latex": "T > 0" + }, + { + "category_id": 14, + "poly": [ + 311, + 689, + 1389, + 689, + 1389, + 770, + 311, + 770 + ], + "score": 0.85, + "latex": "V _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\mathcal { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + K \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T } w h e r e K \\leq ( 1 - \\chi ) + \\frac { 4 \\sqrt { 2 \\delta } \\gamma } { ( 1 - \\gamma ) ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 909, + 597, + 943, + 597, + 943, + 627, + 909, + 627 + ], + "score": 0.84, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 1273, + 1090, + 1295, + 1090, + 1295, + 1119, + 1273, + 1119 + ], + "score": 0.83, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 889, + 1212, + 912, + 1212, + 912, + 1241, + 889, + 1241 + ], + "score": 0.83, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 664, + 920, + 686, + 920, + 686, + 950, + 664, + 950 + ], + "score": 0.82, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1246, + 827, + 1275, + 827, + 1275, + 854, + 1246, + 854 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 661, + 356, + 682, + 356, + 682, + 385, + 661, + 385 + ], + "score": 0.8, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 329, + 949, + 359, + 949, + 359, + 976, + 329, + 976 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1100, + 950, + 1127, + 950, + 1127, + 976, + 1100, + 976 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 699, + 792, + 719, + 792, + 719, + 818, + 699, + 818 + ], + "score": 0.8, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 442, + 953, + 466, + 953, + 466, + 981, + 442, + 981 + ], + "score": 0.8, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 909, + 1180, + 933, + 1180, + 933, + 1206, + 909, + 1206 + ], + "score": 0.79, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1079, + 893, + 1101, + 893, + 1101, + 916, + 1079, + 916 + ], + "score": 0.75, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1005, + 659, + 1105, + 659, + 1105, + 689, + 1005, + 689 + ], + "score": 0.73, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 364, + 1312, + 388, + 1312, + 388, + 1338, + 364, + 1338 + ], + "score": 0.65, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1119, + 1253, + 1141, + 1253, + 1141, + 1282, + 1119, + 1282 + ], + "score": 0.56, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1010, + 598, + 1031, + 598, + 1031, + 625, + 1010, + 625 + ], + "score": 0.52, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 297, + 1347, + 320, + 1347, + 320, + 1375, + 297, + 1375 + ], + "score": 0.47, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1323, + 630, + 1338, + 630, + 1338, + 655, + 1323, + 655 + ], + "score": 0.3, + "latex": "I" + }, + { + "category_id": 13, + "poly": [ + 823, + 634, + 845, + 634, + 845, + 655, + 823, + 655 + ], + "score": 0.28, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1570.0, + 561.0, + 1570.0, + 561.0, + 1615.0, + 291.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1866.0, + 643.0, + 1866.0, + 643.0, + 1906.0, + 293.0, + 1906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 862.0, + 2087.0, + 862.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1091.0, + 294.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1082.0, + 1272.0, + 1082.0, + 1272.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1296.0, + 1082.0, + 1403.0, + 1082.0, + 1403.0, + 1125.0, + 1296.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1119.0, + 1404.0, + 1119.0, + 1404.0, + 1151.0, + 296.0, + 1151.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1145.0, + 1404.0, + 1145.0, + 1404.0, + 1182.0, + 295.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1180.0, + 908.0, + 1180.0, + 908.0, + 1212.0, + 296.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 934.0, + 1180.0, + 1404.0, + 1180.0, + 1404.0, + 1212.0, + 934.0, + 1212.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1209.0, + 888.0, + 1209.0, + 888.0, + 1245.0, + 295.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 1209.0, + 1164.0, + 1209.0, + 1164.0, + 1245.0, + 913.0, + 1245.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 228.0, + 1405.0, + 228.0, + 1405.0, + 266.0, + 293.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 261.0, + 1405.0, + 261.0, + 1405.0, + 297.0, + 294.0, + 297.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 289.0, + 471.0, + 289.0, + 471.0, + 333.0, + 293.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 542.0, + 289.0, + 1405.0, + 289.0, + 1405.0, + 333.0, + 542.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 318.0, + 297.0, + 318.0, + 297.0, + 362.0, + 294.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 318.0, + 1407.0, + 318.0, + 1407.0, + 362.0, + 367.0, + 362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 343.0, + 660.0, + 343.0, + 660.0, + 399.0, + 288.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 343.0, + 1329.0, + 343.0, + 1329.0, + 399.0, + 683.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 343.0, + 1407.0, + 343.0, + 1407.0, + 399.0, + 1400.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 382.0, + 1071.0, + 382.0, + 1071.0, + 414.0, + 294.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 561.0, + 1404.0, + 561.0, + 1404.0, + 598.0, + 296.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 594.0, + 908.0, + 594.0, + 908.0, + 627.0, + 293.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 594.0, + 1009.0, + 594.0, + 1009.0, + 627.0, + 944.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 594.0, + 1403.0, + 594.0, + 1403.0, + 627.0, + 1032.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 621.0, + 351.0, + 621.0, + 351.0, + 667.0, + 291.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 452.0, + 621.0, + 612.0, + 621.0, + 612.0, + 667.0, + 452.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 682.0, + 621.0, + 822.0, + 621.0, + 822.0, + 667.0, + 682.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 621.0, + 1322.0, + 621.0, + 1322.0, + 667.0, + 846.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1339.0, + 621.0, + 1407.0, + 621.0, + 1407.0, + 667.0, + 1339.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 657.0, + 642.0, + 657.0, + 642.0, + 695.0, + 291.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 657.0, + 1004.0, + 657.0, + 1004.0, + 695.0, + 718.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 657.0, + 1116.0, + 657.0, + 1116.0, + 695.0, + 1106.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 426.0, + 1407.0, + 426.0, + 1407.0, + 465.0, + 293.0, + 465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 459.0, + 1407.0, + 459.0, + 1407.0, + 495.0, + 292.0, + 495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 490.0, + 1405.0, + 490.0, + 1405.0, + 525.0, + 292.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 520.0, + 1130.0, + 520.0, + 1130.0, + 557.0, + 293.0, + 557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1910.0, + 1405.0, + 1910.0, + 1405.0, + 1945.0, + 295.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1943.0, + 1404.0, + 1943.0, + 1404.0, + 1976.0, + 294.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1975.0, + 1402.0, + 1975.0, + 1402.0, + 2007.0, + 294.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2038.0, + 293.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1401.0, + 1407.0, + 1401.0, + 1407.0, + 1438.0, + 294.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1432.0, + 1403.0, + 1432.0, + 1403.0, + 1468.0, + 294.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1463.0, + 1405.0, + 1463.0, + 1405.0, + 1499.0, + 294.0, + 1499.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1495.0, + 508.0, + 1495.0, + 508.0, + 1532.0, + 294.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 1495.0, + 658.0, + 1495.0, + 658.0, + 1532.0, + 600.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1642.0, + 1402.0, + 1642.0, + 1402.0, + 1679.0, + 295.0, + 1679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1675.0, + 1403.0, + 1675.0, + 1403.0, + 1709.0, + 296.0, + 1709.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1707.0, + 670.0, + 1707.0, + 670.0, + 1741.0, + 296.0, + 1741.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1248.0, + 1118.0, + 1248.0, + 1118.0, + 1285.0, + 294.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1142.0, + 1248.0, + 1403.0, + 1248.0, + 1403.0, + 1285.0, + 1142.0, + 1285.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1278.0, + 531.0, + 1278.0, + 531.0, + 1314.0, + 293.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 1278.0, + 1407.0, + 1278.0, + 1407.0, + 1314.0, + 709.0, + 1314.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1309.0, + 363.0, + 1309.0, + 363.0, + 1347.0, + 293.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1309.0, + 1014.0, + 1309.0, + 1014.0, + 1347.0, + 389.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 1309.0, + 1405.0, + 1309.0, + 1405.0, + 1347.0, + 1120.0, + 1347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1345.0, + 296.0, + 1345.0, + 296.0, + 1381.0, + 292.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1345.0, + 377.0, + 1345.0, + 377.0, + 1381.0, + 321.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 681.0, + 1345.0, + 690.0, + 1345.0, + 690.0, + 1381.0, + 681.0, + 1381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 366.0, + 1762.0, + 1154.0, + 1762.0, + 1154.0, + 1801.0, + 366.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 1798.0, + 1329.0, + 1798.0, + 1329.0, + 1838.0, + 367.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 787.0, + 698.0, + 787.0, + 698.0, + 826.0, + 293.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 787.0, + 1405.0, + 787.0, + 1405.0, + 826.0, + 720.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 856.0, + 1405.0, + 856.0, + 1405.0, + 894.0, + 291.0, + 894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 885.0, + 1078.0, + 885.0, + 1078.0, + 923.0, + 293.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 885.0, + 1403.0, + 885.0, + 1403.0, + 923.0, + 1102.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 918.0, + 663.0, + 918.0, + 663.0, + 952.0, + 294.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 687.0, + 918.0, + 1405.0, + 918.0, + 1405.0, + 952.0, + 687.0, + 952.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 946.0, + 328.0, + 946.0, + 328.0, + 983.0, + 291.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 946.0, + 441.0, + 946.0, + 441.0, + 983.0, + 360.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 946.0, + 1099.0, + 946.0, + 1099.0, + 983.0, + 467.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1128.0, + 946.0, + 1406.0, + 946.0, + 1406.0, + 983.0, + 1128.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 977.0, + 1407.0, + 977.0, + 1407.0, + 1015.0, + 293.0, + 1015.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1009.0, + 469.0, + 1009.0, + 469.0, + 1044.0, + 294.0, + 1044.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 819.5, + 415.0, + 819.5, + 415.0, + 868.5, + 292.0, + 868.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.75, + 812.5, + 1411.75, + 812.5, + 1411.75, + 870.5, + 345.75, + 870.5 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 846, + 1403, + 846, + 1403, + 1152, + 297, + 1152 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 1167, + 1404, + 1167, + 1404, + 1382, + 297, + 1382 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 525, + 1404, + 525, + 1404, + 830, + 297, + 830 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1500, + 1404, + 1500, + 1404, + 1715, + 297, + 1715 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1729, + 1404, + 1729, + 1404, + 2005, + 298, + 2005 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 331, + 234, + 1367, + 234, + 1367, + 431, + 331, + 431 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 299, + 436, + 1403, + 436, + 1403, + 519, + 299, + 519 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 300, + 1436, + 616, + 1436, + 616, + 1466, + 300, + 1466 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 858, + 2088, + 858, + 2111, + 842, + 2111 + ], + "score": 0.649 + }, + { + "category_id": 2, + "poly": [ + 842, + 2088, + 859, + 2088, + 859, + 2111, + 842, + 2111 + ], + "score": 0.215 + }, + { + "category_id": 13, + "poly": [ + 1085, + 938, + 1170, + 938, + 1170, + 970, + 1085, + 970 + ], + "score": 0.94, + "latex": "R ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1205, + 937, + 1400, + 937, + 1400, + 971, + 1205, + 971 + ], + "score": 0.92, + "latex": "R ( s , a ) { \\bar { } } - P C ( s )" + }, + { + "category_id": 13, + "poly": [ + 372, + 1913, + 469, + 1913, + 469, + 1944, + 372, + 1944 + ], + "score": 0.91, + "latex": "\\chi = 0 . 0" + }, + { + "category_id": 13, + "poly": [ + 372, + 969, + 466, + 969, + 466, + 997, + 372, + 997 + ], + "score": 0.91, + "latex": "P = 1 0" + }, + { + "category_id": 13, + "poly": [ + 517, + 968, + 576, + 968, + 576, + 1001, + 517, + 1001 + ], + "score": 0.9, + "latex": "C ( s )" + }, + { + "category_id": 13, + "poly": [ + 300, + 678, + 373, + 678, + 373, + 711, + 300, + 711 + ], + "score": 0.89, + "latex": "( ( x , y )" + }, + { + "category_id": 13, + "poly": [ + 913, + 1943, + 955, + 1943, + 955, + 1975, + 913, + 1975 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1272, + 1765, + 1294, + 1765, + 1294, + 1793, + 1272, + 1793 + ], + "score": 0.8, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 527, + 1654, + 551, + 1654, + 551, + 1684, + 527, + 1684 + ], + "score": 0.79, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 297, + 1765, + 318, + 1765, + 318, + 1793, + 297, + 1793 + ], + "score": 0.76, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1330, + 1000, + 1355, + 1000, + 1355, + 1031, + 1330, + 1031 + ], + "score": 0.72, + "latex": "\\varrho" + }, + { + "category_id": 13, + "poly": [ + 416, + 1653, + 475, + 1653, + 475, + 1682, + 416, + 1682 + ], + "score": 0.42, + "latex": "C P O" + }, + { + "category_id": 13, + "poly": [ + 422, + 877, + 497, + 877, + 497, + 908, + 422, + 908 + ], + "score": 0.36, + "latex": "( C P O )" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 402.0, + 463.0, + 402.0, + 463.0, + 436.0, + 331.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 547.0, + 402.0, + 685.0, + 402.0, + 685.0, + 436.0, + 547.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 397.0, + 1320.0, + 397.0, + 1320.0, + 439.0, + 760.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 432.0, + 1404.0, + 432.0, + 1404.0, + 468.0, + 293.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 460.0, + 1404.0, + 460.0, + 1404.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 488.0, + 598.0, + 488.0, + 598.0, + 523.0, + 293.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1435.0, + 619.0, + 1435.0, + 619.0, + 1470.0, + 295.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2119.0, + 840.0, + 2119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 843.0, + 1402.0, + 843.0, + 1402.0, + 882.0, + 292.0, + 882.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 877.0, + 421.0, + 877.0, + 421.0, + 909.0, + 295.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 877.0, + 1401.0, + 877.0, + 1401.0, + 909.0, + 498.0, + 909.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 906.0, + 1405.0, + 906.0, + 1405.0, + 942.0, + 293.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 937.0, + 1084.0, + 937.0, + 1084.0, + 972.0, + 295.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 937.0, + 1204.0, + 937.0, + 1204.0, + 972.0, + 1171.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 937.0, + 1404.0, + 937.0, + 1404.0, + 972.0, + 1401.0, + 972.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 967.0, + 371.0, + 967.0, + 371.0, + 1003.0, + 295.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 467.0, + 967.0, + 516.0, + 967.0, + 516.0, + 1003.0, + 467.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 967.0, + 1405.0, + 967.0, + 1405.0, + 1003.0, + 577.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 998.0, + 1329.0, + 998.0, + 1329.0, + 1032.0, + 295.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1356.0, + 998.0, + 1404.0, + 998.0, + 1404.0, + 1032.0, + 1356.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1027.0, + 1408.0, + 1027.0, + 1408.0, + 1065.0, + 292.0, + 1065.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1057.0, + 1405.0, + 1057.0, + 1405.0, + 1094.0, + 292.0, + 1094.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1088.0, + 1407.0, + 1088.0, + 1407.0, + 1127.0, + 292.0, + 1127.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1120.0, + 442.0, + 1120.0, + 442.0, + 1153.0, + 293.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1166.0, + 1404.0, + 1166.0, + 1404.0, + 1201.0, + 295.0, + 1201.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1198.0, + 1404.0, + 1198.0, + 1404.0, + 1229.0, + 296.0, + 1229.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1228.0, + 1402.0, + 1228.0, + 1402.0, + 1261.0, + 292.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1260.0, + 1401.0, + 1260.0, + 1401.0, + 1290.0, + 296.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1287.0, + 1404.0, + 1287.0, + 1404.0, + 1323.0, + 294.0, + 1323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1318.0, + 1402.0, + 1318.0, + 1402.0, + 1355.0, + 294.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1350.0, + 1035.0, + 1350.0, + 1035.0, + 1384.0, + 295.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 527.0, + 1402.0, + 527.0, + 1402.0, + 559.0, + 295.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 553.0, + 1404.0, + 553.0, + 1404.0, + 590.0, + 292.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 583.0, + 1405.0, + 583.0, + 1405.0, + 624.0, + 291.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 614.0, + 1404.0, + 614.0, + 1404.0, + 651.0, + 292.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 643.0, + 1408.0, + 643.0, + 1408.0, + 685.0, + 291.0, + 685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 676.0, + 299.0, + 676.0, + 299.0, + 715.0, + 292.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 676.0, + 1407.0, + 676.0, + 1407.0, + 715.0, + 374.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 710.0, + 1405.0, + 710.0, + 1405.0, + 742.0, + 294.0, + 742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 737.0, + 1405.0, + 737.0, + 1405.0, + 776.0, + 292.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 770.0, + 1405.0, + 770.0, + 1405.0, + 805.0, + 294.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 798.0, + 1064.0, + 798.0, + 1064.0, + 834.0, + 295.0, + 834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1497.0, + 1405.0, + 1497.0, + 1405.0, + 1537.0, + 294.0, + 1537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1529.0, + 1406.0, + 1529.0, + 1406.0, + 1566.0, + 294.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1598.0, + 294.0, + 1598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1590.0, + 1405.0, + 1590.0, + 1405.0, + 1627.0, + 292.0, + 1627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1622.0, + 1402.0, + 1622.0, + 1402.0, + 1657.0, + 295.0, + 1657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1651.0, + 415.0, + 1651.0, + 415.0, + 1687.0, + 294.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 476.0, + 1651.0, + 526.0, + 1651.0, + 526.0, + 1687.0, + 476.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1651.0, + 1404.0, + 1651.0, + 1404.0, + 1687.0, + 552.0, + 1687.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1681.0, + 883.0, + 1681.0, + 883.0, + 1720.0, + 294.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1729.0, + 1405.0, + 1729.0, + 1405.0, + 1765.0, + 296.0, + 1765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1759.0, + 296.0, + 1759.0, + 296.0, + 1798.0, + 291.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1759.0, + 1271.0, + 1759.0, + 1271.0, + 1798.0, + 319.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1759.0, + 1407.0, + 1759.0, + 1407.0, + 1798.0, + 1295.0, + 1798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1789.0, + 1406.0, + 1789.0, + 1406.0, + 1827.0, + 292.0, + 1827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1857.0, + 292.0, + 1857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1853.0, + 1402.0, + 1853.0, + 1402.0, + 1885.0, + 294.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1880.0, + 1406.0, + 1880.0, + 1406.0, + 1918.0, + 292.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1911.0, + 371.0, + 1911.0, + 371.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1911.0, + 1406.0, + 1911.0, + 1406.0, + 1948.0, + 470.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1939.0, + 912.0, + 1939.0, + 912.0, + 1982.0, + 291.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 956.0, + 1939.0, + 1405.0, + 1939.0, + 1405.0, + 1982.0, + 956.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1970.0, + 1370.0, + 1970.0, + 1370.0, + 2013.0, + 292.0, + 2013.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1009, + 1403, + 1009, + 1403, + 1315, + 298, + 1315 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1499, + 1403, + 1499, + 1403, + 1775, + 298, + 1775 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1788, + 1403, + 1788, + 1403, + 2035, + 298, + 2035 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 1330, + 1402, + 1330, + 1402, + 1483, + 299, + 1483 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 323, + 225, + 1348, + 225, + 1348, + 654, + 323, + 654 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 297, + 681, + 1404, + 681, + 1404, + 849, + 297, + 849 + ], + "score": 0.956 + }, + { + "category_id": 0, + "poly": [ + 300, + 875, + 587, + 875, + 587, + 910, + 300, + 910 + ], + "score": 0.896 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.796 + }, + { + "category_id": 1, + "poly": [ + 301, + 961, + 1139, + 961, + 1139, + 993, + 301, + 993 + ], + "score": 0.771 + }, + { + "category_id": 0, + "poly": [ + 301, + 961, + 1139, + 961, + 1139, + 993, + 301, + 993 + ], + "score": 0.186 + }, + { + "category_id": 13, + "poly": [ + 998, + 766, + 1102, + 766, + 1102, + 794, + 998, + 794 + ], + "score": 0.9, + "latex": "\\chi = 0 . 0 3" + }, + { + "category_id": 13, + "poly": [ + 882, + 1591, + 974, + 1591, + 974, + 1621, + 882, + 1621 + ], + "score": 0.76, + "latex": "\\mathrm { C V A R } _ { \\alpha }" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 229.0, + 586.0, + 229.0, + 586.0, + 264.0, + 396.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 655.0, + 232.0, + 811.0, + 232.0, + 811.0, + 262.0, + 655.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 232.0, + 1057.0, + 232.0, + 1057.0, + 261.0, + 888.0, + 261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 229.0, + 1261.0, + 229.0, + 1261.0, + 264.0, + 1167.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 327.0, + 246.0, + 387.0, + 246.0, + 387.0, + 411.0, + 327.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 593.0, + 253.0, + 630.0, + 253.0, + 630.0, + 287.0, + 593.0, + 287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 268.0, + 1112.0, + 268.0, + 1112.0, + 307.0, + 1073.0, + 307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 367.0, + 320.0, + 386.0, + 320.0, + 386.0, + 345.0, + 367.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 308.0, + 630.0, + 308.0, + 630.0, + 344.0, + 594.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 299.0, + 869.0, + 299.0, + 869.0, + 327.0, + 835.0, + 327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1076.0, + 316.0, + 1111.0, + 316.0, + 1111.0, + 350.0, + 1076.0, + 350.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 378.0, + 386.0, + 378.0, + 386.0, + 398.0, + 368.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 607.0, + 376.0, + 629.0, + 376.0, + 629.0, + 402.0, + 607.0, + 402.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 369.0, + 868.0, + 369.0, + 868.0, + 389.0, + 850.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 370.0, + 1111.0, + 370.0, + 1111.0, + 397.0, + 1089.0, + 397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 404.0, + 410.0, + 404.0, + 410.0, + 427.0, + 390.0, + 427.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 399.0, + 525.0, + 399.0, + 525.0, + 431.0, + 463.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 399.0, + 622.0, + 399.0, + 622.0, + 431.0, + 555.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 630.0, + 404.0, + 650.0, + 404.0, + 650.0, + 425.0, + 630.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 399.0, + 766.0, + 399.0, + 766.0, + 431.0, + 703.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 796.0, + 399.0, + 865.0, + 399.0, + 865.0, + 431.0, + 796.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 403.0, + 892.0, + 403.0, + 892.0, + 428.0, + 868.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 944.0, + 399.0, + 1007.0, + 399.0, + 1007.0, + 431.0, + 944.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1037.0, + 399.0, + 1102.0, + 399.0, + 1102.0, + 431.0, + 1037.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 403.0, + 1134.0, + 403.0, + 1134.0, + 428.0, + 1111.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 401.0, + 1253.0, + 401.0, + 1253.0, + 429.0, + 1181.0, + 429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1271.0, + 399.0, + 1348.0, + 399.0, + 1348.0, + 431.0, + 1271.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 418.0, + 355.0, + 418.0, + 355.0, + 575.0, + 316.0, + 575.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 422.0, + 871.0, + 422.0, + 871.0, + 452.0, + 825.0, + 452.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 442.0, + 388.0, + 442.0, + 388.0, + 508.0, + 346.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 584.0, + 445.0, + 633.0, + 445.0, + 633.0, + 517.0, + 584.0, + 517.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 647.0, + 463.0, + 661.0, + 463.0, + 661.0, + 480.0, + 647.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 476.0, + 872.0, + 476.0, + 872.0, + 512.0, + 826.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1066.0, + 467.0, + 1113.0, + 467.0, + 1113.0, + 501.0, + 1066.0, + 501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 502.0, + 1133.0, + 502.0, + 1133.0, + 516.0, + 1118.0, + 516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1135.0, + 449.0, + 1173.0, + 449.0, + 1173.0, + 491.0, + 1135.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 534.0, + 388.0, + 534.0, + 388.0, + 565.0, + 346.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 534.0, + 630.0, + 534.0, + 630.0, + 566.0, + 587.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 535.0, + 871.0, + 535.0, + 871.0, + 566.0, + 827.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 540.0, + 1111.0, + 540.0, + 1111.0, + 569.0, + 1068.0, + 569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 572.0, + 411.0, + 572.0, + 411.0, + 598.0, + 389.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 569.0, + 525.0, + 569.0, + 525.0, + 601.0, + 463.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 568.0, + 621.0, + 568.0, + 621.0, + 600.0, + 555.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 573.0, + 653.0, + 573.0, + 653.0, + 598.0, + 628.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 569.0, + 766.0, + 569.0, + 766.0, + 601.0, + 703.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 797.0, + 571.0, + 860.0, + 571.0, + 860.0, + 599.0, + 797.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 573.0, + 893.0, + 573.0, + 893.0, + 598.0, + 868.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 569.0, + 1007.0, + 569.0, + 1007.0, + 601.0, + 943.0, + 601.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 571.0, + 1102.0, + 571.0, + 1102.0, + 599.0, + 1038.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1111.0, + 573.0, + 1134.0, + 573.0, + 1134.0, + 598.0, + 1111.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 569.0, + 1254.0, + 569.0, + 1254.0, + 598.0, + 1181.0, + 598.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 571.0, + 1345.0, + 571.0, + 1345.0, + 599.0, + 1273.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 410.0, + 615.0, + 599.0, + 615.0, + 599.0, + 646.0, + 410.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 684.0, + 615.0, + 824.0, + 615.0, + 824.0, + 646.0, + 684.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 907.0, + 614.0, + 962.0, + 614.0, + 962.0, + 649.0, + 907.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 613.0, + 1106.0, + 613.0, + 1106.0, + 652.0, + 1044.0, + 652.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1186.0, + 614.0, + 1323.0, + 614.0, + 1323.0, + 649.0, + 1186.0, + 649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 627.0, + 334.0, + 736.0, + 334.0, + 736.0, + 369.0, + 627.0, + 369.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 292.5, + 1048.0, + 292.5, + 1048.0, + 354.5, + 839.0, + 354.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 290.0, + 1317.0, + 290.0, + 1317.0, + 355.5, + 1077.0, + 355.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 335.5, + 509.0, + 335.5, + 509.0, + 366.0, + 381.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 901.0, + 456.5, + 905.0, + 456.5, + 905.0, + 487.5, + 901.0, + 487.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1213.0, + 515.5, + 1286.0, + 515.5, + 1286.0, + 547.5, + 1213.0, + 547.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 682.0, + 1405.0, + 682.0, + 1405.0, + 712.0, + 296.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 710.0, + 1405.0, + 710.0, + 1405.0, + 740.0, + 295.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 740.0, + 1404.0, + 740.0, + 1404.0, + 766.0, + 297.0, + 766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 765.0, + 997.0, + 765.0, + 997.0, + 795.0, + 295.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 765.0, + 1402.0, + 765.0, + 1402.0, + 795.0, + 1103.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 793.0, + 1404.0, + 793.0, + 1404.0, + 822.0, + 296.0, + 822.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 820.0, + 560.0, + 820.0, + 560.0, + 850.0, + 295.0, + 850.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 872.0, + 593.0, + 872.0, + 593.0, + 916.0, + 293.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2117.0, + 838.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 1143.0, + 957.0, + 1143.0, + 1002.0, + 295.0, + 1002.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1008.0, + 1407.0, + 1008.0, + 1407.0, + 1043.0, + 294.0, + 1043.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1038.0, + 1404.0, + 1038.0, + 1404.0, + 1074.0, + 293.0, + 1074.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1069.0, + 1405.0, + 1069.0, + 1405.0, + 1105.0, + 293.0, + 1105.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1099.0, + 1405.0, + 1099.0, + 1405.0, + 1135.0, + 293.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1133.0, + 1404.0, + 1133.0, + 1404.0, + 1165.0, + 296.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1163.0, + 1404.0, + 1163.0, + 1404.0, + 1195.0, + 296.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1194.0, + 1404.0, + 1194.0, + 1404.0, + 1225.0, + 294.0, + 1225.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1222.0, + 1406.0, + 1222.0, + 1406.0, + 1257.0, + 292.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1252.0, + 1407.0, + 1252.0, + 1407.0, + 1288.0, + 293.0, + 1288.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 1290.0, + 1285.0, + 1290.0, + 1317.0, + 294.0, + 1317.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1500.0, + 1404.0, + 1500.0, + 1404.0, + 1533.0, + 297.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1530.0, + 1404.0, + 1530.0, + 1404.0, + 1561.0, + 294.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1560.0, + 1405.0, + 1560.0, + 1405.0, + 1593.0, + 296.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1590.0, + 881.0, + 1590.0, + 881.0, + 1623.0, + 294.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 975.0, + 1590.0, + 1403.0, + 1590.0, + 1403.0, + 1623.0, + 975.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1619.0, + 1404.0, + 1619.0, + 1404.0, + 1654.0, + 293.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1647.0, + 1405.0, + 1647.0, + 1405.0, + 1688.0, + 294.0, + 1688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1681.0, + 1403.0, + 1681.0, + 1403.0, + 1714.0, + 296.0, + 1714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1710.0, + 1405.0, + 1710.0, + 1405.0, + 1748.0, + 292.0, + 1748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1741.0, + 982.0, + 1741.0, + 982.0, + 1777.0, + 295.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1787.0, + 1405.0, + 1787.0, + 1405.0, + 1824.0, + 293.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1820.0, + 1405.0, + 1820.0, + 1405.0, + 1854.0, + 293.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1851.0, + 1405.0, + 1851.0, + 1405.0, + 1885.0, + 293.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1403.0, + 1881.0, + 1403.0, + 1915.0, + 294.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1407.0, + 1910.0, + 1407.0, + 1948.0, + 294.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1942.0, + 1406.0, + 1942.0, + 1406.0, + 1977.0, + 292.0, + 1977.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2002.0, + 1001.0, + 2002.0, + 1001.0, + 2037.0, + 293.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1330.0, + 1404.0, + 1330.0, + 1404.0, + 1363.0, + 297.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1362.0, + 1406.0, + 1362.0, + 1406.0, + 1395.0, + 295.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1392.0, + 1404.0, + 1392.0, + 1404.0, + 1425.0, + 295.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1422.0, + 1403.0, + 1422.0, + 1403.0, + 1455.0, + 295.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1450.0, + 552.0, + 1450.0, + 552.0, + 1487.0, + 294.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 957.0, + 1143.0, + 957.0, + 1143.0, + 1002.0, + 295.0, + 1002.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1328, + 1403, + 1328, + 1403, + 1573, + 298, + 1573 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 299, + 1128, + 1403, + 1128, + 1403, + 1313, + 299, + 1313 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 298, + 1588, + 1404, + 1588, + 1404, + 1771, + 298, + 1771 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1850, + 1403, + 1850, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 299, + 875, + 1403, + 875, + 1403, + 1029, + 299, + 1029 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 339, + 247, + 1287, + 247, + 1287, + 644, + 339, + 644 + ], + "score": 0.973 + }, + { + "category_id": 4, + "poly": [ + 297, + 672, + 1403, + 672, + 1403, + 869, + 297, + 869 + ], + "score": 0.954 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 812, + 76, + 812, + 104, + 299, + 104 + ], + "score": 0.888 + }, + { + "category_id": 0, + "poly": [ + 300, + 1808, + 558, + 1808, + 558, + 1835, + 300, + 1835 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 300, + 1068, + 1017, + 1068, + 1017, + 1103, + 300, + 1103 + ], + "score": 0.848 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2111, + 841, + 2111 + ], + "score": 0.789 + }, + { + "category_id": 13, + "poly": [ + 340, + 1393, + 364, + 1393, + 364, + 1422, + 340, + 1422 + ], + "score": 0.82, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 405, + 789, + 425, + 789, + 425, + 814, + 405, + 814 + ], + "score": 0.75, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 401, + 705, + 418, + 705, + 418, + 726, + 401, + 726 + ], + "score": 0.29, + "latex": "\\mathbf { X }" + }, + { + "category_id": 15, + "poly": [ + 405.0, + 248.0, + 584.0, + 248.0, + 584.0, + 284.0, + 405.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 249.0, + 791.0, + 249.0, + 791.0, + 284.0, + 644.0, + 284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 861.0, + 251.0, + 1019.0, + 251.0, + 1019.0, + 283.0, + 861.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 250.0, + 1206.0, + 250.0, + 1206.0, + 283.0, + 1118.0, + 283.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 339.0, + 264.0, + 397.0, + 264.0, + 397.0, + 417.0, + 339.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 292.0, + 624.0, + 292.0, + 624.0, + 330.0, + 589.0, + 330.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 297.0, + 971.0, + 297.0, + 971.0, + 305.0, + 957.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 286.0, + 1070.0, + 286.0, + 1070.0, + 332.0, + 1032.0, + 332.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 335.0, + 396.0, + 335.0, + 396.0, + 357.0, + 380.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 308.0, + 483.0, + 308.0, + 483.0, + 380.0, + 443.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 600.0, + 342.0, + 621.0, + 342.0, + 621.0, + 368.0, + 600.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 316.0, + 846.0, + 316.0, + 846.0, + 347.0, + 811.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 326.0, + 1071.0, + 326.0, + 1071.0, + 367.0, + 1036.0, + 367.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 384.0, + 397.0, + 384.0, + 397.0, + 403.0, + 381.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 603.0, + 389.0, + 621.0, + 389.0, + 621.0, + 409.0, + 603.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 826.0, + 384.0, + 844.0, + 384.0, + 844.0, + 404.0, + 826.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 380.0, + 1067.0, + 380.0, + 1067.0, + 406.0, + 1047.0, + 406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 336.0, + 419.0, + 399.0, + 419.0, + 399.0, + 570.0, + 336.0, + 570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 412.0, + 419.0, + 412.0, + 419.0, + 431.0, + 403.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 407.0, + 527.0, + 407.0, + 527.0, + 437.0, + 468.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 409.0, + 623.0, + 409.0, + 623.0, + 458.0, + 555.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 412.0, + 641.0, + 412.0, + 641.0, + 431.0, + 624.0, + 431.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 407.0, + 749.0, + 407.0, + 749.0, + 437.0, + 691.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 407.0, + 865.0, + 407.0, + 865.0, + 467.0, + 776.0, + 467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 407.0, + 971.0, + 407.0, + 971.0, + 437.0, + 913.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 407.0, + 1059.0, + 407.0, + 1059.0, + 437.0, + 999.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 410.0, + 1089.0, + 410.0, + 1089.0, + 433.0, + 1067.0, + 433.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1131.0, + 407.0, + 1199.0, + 407.0, + 1199.0, + 437.0, + 1131.0, + 437.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 409.0, + 1284.0, + 409.0, + 1284.0, + 436.0, + 1216.0, + 436.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 467.0, + 398.0, + 467.0, + 398.0, + 491.0, + 362.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 469.0, + 623.0, + 469.0, + 623.0, + 498.0, + 583.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 803.0, + 479.0, + 847.0, + 479.0, + 847.0, + 518.0, + 803.0, + 518.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1026.0, + 474.0, + 1070.0, + 474.0, + 1070.0, + 508.0, + 1026.0, + 508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1083.0, + 465.0, + 1119.0, + 465.0, + 1119.0, + 503.0, + 1083.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 535.0, + 397.0, + 535.0, + 397.0, + 559.0, + 362.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 583.0, + 532.0, + 623.0, + 532.0, + 623.0, + 561.0, + 583.0, + 561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 805.0, + 530.0, + 846.0, + 530.0, + 846.0, + 560.0, + 805.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1027.0, + 531.0, + 1068.0, + 531.0, + 1068.0, + 560.0, + 1027.0, + 560.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 568.0, + 420.0, + 568.0, + 420.0, + 589.0, + 401.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 468.0, + 563.0, + 526.0, + 563.0, + 526.0, + 594.0, + 468.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 564.0, + 615.0, + 564.0, + 615.0, + 594.0, + 554.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 623.0, + 569.0, + 642.0, + 569.0, + 642.0, + 589.0, + 623.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 564.0, + 749.0, + 564.0, + 749.0, + 594.0, + 691.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 564.0, + 837.0, + 564.0, + 837.0, + 594.0, + 776.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 566.0, + 866.0, + 566.0, + 866.0, + 591.0, + 844.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 913.0, + 564.0, + 972.0, + 564.0, + 972.0, + 594.0, + 913.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 564.0, + 1060.0, + 564.0, + 1060.0, + 594.0, + 999.0, + 594.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1067.0, + 568.0, + 1089.0, + 568.0, + 1089.0, + 591.0, + 1067.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1132.0, + 565.0, + 1199.0, + 565.0, + 1199.0, + 592.0, + 1132.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1216.0, + 564.0, + 1284.0, + 564.0, + 1284.0, + 591.0, + 1216.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 498.0, + 607.0, + 631.0, + 607.0, + 631.0, + 636.0, + 498.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 705.0, + 607.0, + 806.0, + 607.0, + 806.0, + 636.0, + 705.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 612.0, + 879.0, + 612.0, + 879.0, + 631.0, + 838.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 607.0, + 972.0, + 607.0, + 972.0, + 636.0, + 883.0, + 636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 604.0, + 1143.0, + 604.0, + 1143.0, + 641.0, + 1047.0, + 641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 674.25, + 309.0, + 683.25, + 309.0, + 683.25, + 396.0, + 674.25, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 321.0, + 980.0, + 321.0, + 980.0, + 366.0, + 837.0, + 366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 322.5, + 1220.0, + 322.5, + 1220.0, + 377.5, + 1044.0, + 377.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.25, + 455.5, + 658.25, + 455.5, + 658.25, + 479.5, + 645.25, + 479.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 424.0, + 458.0, + 427.0, + 458.0, + 427.0, + 469.0, + 424.0, + 469.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 451.0, + 899.0, + 451.0, + 899.0, + 497.0, + 847.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 487.5, + 691.0, + 487.5, + 691.0, + 503.5, + 676.0, + 503.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 674.0, + 1405.0, + 674.0, + 1405.0, + 707.0, + 295.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 700.0, + 400.0, + 700.0, + 400.0, + 733.0, + 293.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 419.0, + 700.0, + 1405.0, + 700.0, + 1405.0, + 733.0, + 419.0, + 733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 728.0, + 1404.0, + 728.0, + 1404.0, + 761.0, + 295.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 755.0, + 1405.0, + 755.0, + 1405.0, + 792.0, + 293.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 781.0, + 404.0, + 781.0, + 404.0, + 820.0, + 294.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 426.0, + 781.0, + 1404.0, + 781.0, + 1404.0, + 820.0, + 426.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 812.0, + 1404.0, + 812.0, + 1404.0, + 845.0, + 295.0, + 845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 839.0, + 1138.0, + 839.0, + 1138.0, + 873.0, + 296.0, + 873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 1809.0, + 560.0, + 1809.0, + 560.0, + 1836.0, + 299.0, + 1836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1065.0, + 1022.0, + 1065.0, + 1022.0, + 1111.0, + 292.0, + 1111.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1328.0, + 1403.0, + 1328.0, + 1403.0, + 1362.0, + 296.0, + 1362.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1359.0, + 1403.0, + 1359.0, + 1403.0, + 1392.0, + 293.0, + 1392.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1391.0, + 339.0, + 1391.0, + 339.0, + 1424.0, + 294.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1391.0, + 1405.0, + 1391.0, + 1405.0, + 1424.0, + 365.0, + 1424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1421.0, + 1403.0, + 1421.0, + 1403.0, + 1454.0, + 294.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1450.0, + 1403.0, + 1450.0, + 1403.0, + 1480.0, + 296.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1480.0, + 1405.0, + 1480.0, + 1405.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1509.0, + 1407.0, + 1509.0, + 1407.0, + 1546.0, + 293.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1544.0, + 1242.0, + 1544.0, + 1242.0, + 1574.0, + 296.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1126.0, + 1405.0, + 1126.0, + 1405.0, + 1163.0, + 295.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1159.0, + 1404.0, + 1159.0, + 1404.0, + 1195.0, + 294.0, + 1195.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1189.0, + 1405.0, + 1189.0, + 1405.0, + 1224.0, + 294.0, + 1224.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1220.0, + 1403.0, + 1220.0, + 1403.0, + 1256.0, + 294.0, + 1256.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1248.0, + 1405.0, + 1248.0, + 1405.0, + 1287.0, + 293.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1278.0, + 1156.0, + 1278.0, + 1156.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1586.0, + 1405.0, + 1586.0, + 1405.0, + 1623.0, + 293.0, + 1623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1620.0, + 1405.0, + 1620.0, + 1405.0, + 1649.0, + 293.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1649.0, + 1405.0, + 1649.0, + 1405.0, + 1685.0, + 294.0, + 1685.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1680.0, + 1406.0, + 1680.0, + 1406.0, + 1715.0, + 294.0, + 1715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1712.0, + 1404.0, + 1712.0, + 1404.0, + 1744.0, + 294.0, + 1744.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1738.0, + 1387.0, + 1738.0, + 1387.0, + 1777.0, + 293.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1947.0, + 292.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1405.0, + 1942.0, + 1405.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 293.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1998.0, + 1077.0, + 1998.0, + 1077.0, + 2040.0, + 292.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 874.0, + 1402.0, + 874.0, + 1402.0, + 907.0, + 294.0, + 907.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 904.0, + 1404.0, + 904.0, + 1404.0, + 940.0, + 294.0, + 940.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 935.0, + 1405.0, + 935.0, + 1405.0, + 970.0, + 294.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 967.0, + 1404.0, + 967.0, + 1404.0, + 1000.0, + 295.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 997.0, + 1137.0, + 997.0, + 1137.0, + 1034.0, + 295.0, + 1034.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 532, + 1405, + 532, + 1405, + 656, + 298, + 656 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 298, + 679, + 1405, + 679, + 1405, + 774, + 298, + 774 + ], + "score": 0.922 + }, + { + "category_id": 1, + "poly": [ + 298, + 910, + 1404, + 910, + 1404, + 1003, + 298, + 1003 + ], + "score": 0.911 + }, + { + "category_id": 1, + "poly": [ + 299, + 1941, + 1403, + 1941, + 1403, + 2034, + 299, + 2034 + ], + "score": 0.91 + }, + { + "category_id": 1, + "poly": [ + 297, + 794, + 1405, + 794, + 1405, + 888, + 297, + 888 + ], + "score": 0.908 + }, + { + "category_id": 1, + "poly": [ + 297, + 1024, + 1403, + 1024, + 1403, + 1119, + 297, + 1119 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 294, + 280, + 1401, + 280, + 1401, + 342, + 294, + 342 + ], + "score": 0.905 + }, + { + "category_id": 1, + "poly": [ + 300, + 364, + 1400, + 364, + 1400, + 457, + 300, + 457 + ], + "score": 0.904 + }, + { + "category_id": 1, + "poly": [ + 298, + 1224, + 1404, + 1224, + 1404, + 1318, + 298, + 1318 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 295, + 1340, + 1399, + 1340, + 1399, + 1405, + 295, + 1405 + ], + "score": 0.897 + }, + { + "category_id": 1, + "poly": [ + 290, + 1740, + 1403, + 1740, + 1403, + 1804, + 290, + 1804 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 297, + 1540, + 1405, + 1540, + 1405, + 1635, + 297, + 1635 + ], + "score": 0.893 + }, + { + "category_id": 1, + "poly": [ + 298, + 1423, + 1405, + 1423, + 1405, + 1519, + 298, + 1519 + ], + "score": 0.891 + }, + { + "category_id": 1, + "poly": [ + 287, + 1656, + 1399, + 1656, + 1399, + 1720, + 287, + 1720 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 813, + 75, + 813, + 105, + 300, + 105 + ], + "score": 0.886 + }, + { + "category_id": 1, + "poly": [ + 297, + 1825, + 1404, + 1825, + 1404, + 1919, + 297, + 1919 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 293, + 1139, + 1403, + 1139, + 1403, + 1205, + 293, + 1205 + ], + "score": 0.883 + }, + { + "category_id": 1, + "poly": [ + 299, + 478, + 1237, + 478, + 1237, + 513, + 299, + 513 + ], + "score": 0.875 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 489, + 227, + 489, + 262, + 300, + 262 + ], + "score": 0.824 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.816 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 534.0, + 1403.0, + 534.0, + 1403.0, + 567.0, + 295.0, + 567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 566.0, + 1403.0, + 566.0, + 1403.0, + 599.0, + 322.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 595.0, + 1406.0, + 595.0, + 1406.0, + 631.0, + 320.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 624.0, + 401.0, + 624.0, + 401.0, + 659.0, + 320.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 680.0, + 1407.0, + 680.0, + 1407.0, + 714.0, + 296.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 711.0, + 1406.0, + 711.0, + 1406.0, + 745.0, + 323.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 740.0, + 740.0, + 740.0, + 740.0, + 778.0, + 323.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 907.0, + 1404.0, + 907.0, + 1404.0, + 946.0, + 296.0, + 946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 942.0, + 1404.0, + 942.0, + 1404.0, + 976.0, + 325.0, + 976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 972.0, + 611.0, + 972.0, + 611.0, + 1001.0, + 323.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1939.0, + 1404.0, + 1939.0, + 1404.0, + 1979.0, + 293.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1972.0, + 1407.0, + 1972.0, + 1407.0, + 2007.0, + 320.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2002.0, + 738.0, + 2002.0, + 738.0, + 2036.0, + 323.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 795.0, + 1405.0, + 795.0, + 1405.0, + 829.0, + 295.0, + 829.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 824.0, + 1405.0, + 824.0, + 1405.0, + 864.0, + 320.0, + 864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 858.0, + 645.0, + 858.0, + 645.0, + 888.0, + 322.0, + 888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1024.0, + 1405.0, + 1024.0, + 1405.0, + 1059.0, + 295.0, + 1059.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1052.0, + 1407.0, + 1052.0, + 1407.0, + 1095.0, + 321.0, + 1095.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1087.0, + 775.0, + 1087.0, + 775.0, + 1120.0, + 322.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 279.0, + 1403.0, + 279.0, + 1403.0, + 315.0, + 294.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 312.0, + 708.0, + 312.0, + 708.0, + 341.0, + 319.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 362.0, + 1405.0, + 362.0, + 1405.0, + 401.0, + 294.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 396.0, + 1405.0, + 396.0, + 1405.0, + 430.0, + 324.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 421.0, + 399.0, + 421.0, + 399.0, + 462.0, + 320.0, + 462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1224.0, + 1403.0, + 1224.0, + 1403.0, + 1261.0, + 295.0, + 1261.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1256.0, + 1402.0, + 1256.0, + 1402.0, + 1290.0, + 323.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1285.0, + 469.0, + 1285.0, + 469.0, + 1321.0, + 323.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1340.0, + 1406.0, + 1340.0, + 1406.0, + 1376.0, + 295.0, + 1376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1371.0, + 1384.0, + 1371.0, + 1384.0, + 1406.0, + 320.0, + 1406.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1740.0, + 1405.0, + 1740.0, + 1405.0, + 1776.0, + 296.0, + 1776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1773.0, + 706.0, + 1773.0, + 706.0, + 1804.0, + 319.0, + 1804.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1538.0, + 1403.0, + 1538.0, + 1403.0, + 1579.0, + 292.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1573.0, + 1403.0, + 1573.0, + 1403.0, + 1607.0, + 322.0, + 1607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1602.0, + 1064.0, + 1602.0, + 1064.0, + 1636.0, + 322.0, + 1636.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1423.0, + 1406.0, + 1423.0, + 1406.0, + 1462.0, + 292.0, + 1462.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1455.0, + 1405.0, + 1455.0, + 1405.0, + 1491.0, + 322.0, + 1491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1489.0, + 1176.0, + 1489.0, + 1176.0, + 1520.0, + 322.0, + 1520.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1652.0, + 1405.0, + 1652.0, + 1405.0, + 1696.0, + 291.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1687.0, + 1049.0, + 1687.0, + 1049.0, + 1720.0, + 322.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1824.0, + 1404.0, + 1824.0, + 1404.0, + 1860.0, + 294.0, + 1860.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1857.0, + 1405.0, + 1857.0, + 1405.0, + 1891.0, + 322.0, + 1891.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1888.0, + 611.0, + 1888.0, + 611.0, + 1917.0, + 322.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1142.0, + 1404.0, + 1142.0, + 1404.0, + 1175.0, + 296.0, + 1175.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1171.0, + 1360.0, + 1171.0, + 1360.0, + 1207.0, + 324.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 478.0, + 1243.0, + 478.0, + 1243.0, + 517.0, + 293.0, + 517.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.824 + }, + { + "category_id": 1, + "poly": [ + 298, + 884, + 1399, + 884, + 1399, + 949, + 298, + 949 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 299, + 802, + 1396, + 802, + 1396, + 869, + 299, + 869 + ], + "score": 0.685 + }, + { + "category_id": 1, + "poly": [ + 297, + 963, + 1401, + 963, + 1401, + 1030, + 297, + 1030 + ], + "score": 0.659 + }, + { + "category_id": 1, + "poly": [ + 297, + 1044, + 1402, + 1044, + 1402, + 1138, + 297, + 1138 + ], + "score": 0.655 + }, + { + "category_id": 1, + "poly": [ + 294, + 1587, + 1401, + 1587, + 1401, + 1653, + 294, + 1653 + ], + "score": 0.647 + }, + { + "category_id": 1, + "poly": [ + 294, + 1396, + 1397, + 1396, + 1397, + 1462, + 294, + 1462 + ], + "score": 0.644 + }, + { + "category_id": 1, + "poly": [ + 298, + 1236, + 1400, + 1236, + 1400, + 1301, + 298, + 1301 + ], + "score": 0.641 + }, + { + "category_id": 1, + "poly": [ + 300, + 581, + 1401, + 581, + 1401, + 647, + 300, + 647 + ], + "score": 0.64 + }, + { + "category_id": 1, + "poly": [ + 294, + 500, + 1402, + 500, + 1402, + 565, + 294, + 565 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 296, + 1155, + 1401, + 1155, + 1401, + 1221, + 296, + 1221 + ], + "score": 0.611 + }, + { + "category_id": 1, + "poly": [ + 296, + 1669, + 1398, + 1669, + 1398, + 1733, + 296, + 1733 + ], + "score": 0.606 + }, + { + "category_id": 1, + "poly": [ + 298, + 1749, + 1397, + 1749, + 1397, + 1814, + 298, + 1814 + ], + "score": 0.584 + }, + { + "category_id": 1, + "poly": [ + 295, + 1830, + 1401, + 1830, + 1401, + 1894, + 295, + 1894 + ], + "score": 0.57 + }, + { + "category_id": 1, + "poly": [ + 299, + 390, + 1400, + 390, + 1400, + 484, + 299, + 484 + ], + "score": 0.561 + }, + { + "category_id": 1, + "poly": [ + 294, + 1476, + 1403, + 1476, + 1403, + 1572, + 294, + 1572 + ], + "score": 0.547 + }, + { + "category_id": 1, + "poly": [ + 297, + 1316, + 1400, + 1316, + 1400, + 1382, + 297, + 1382 + ], + "score": 0.534 + }, + { + "category_id": 1, + "poly": [ + 296, + 310, + 1401, + 310, + 1401, + 374, + 296, + 374 + ], + "score": 0.52 + }, + { + "category_id": 1, + "poly": [ + 294, + 229, + 1402, + 229, + 1402, + 295, + 294, + 295 + ], + "score": 0.486 + }, + { + "category_id": 1, + "poly": [ + 299, + 1911, + 1404, + 1911, + 1404, + 2034, + 299, + 2034 + ], + "score": 0.416 + }, + { + "category_id": 1, + "poly": [ + 297, + 662, + 1402, + 662, + 1402, + 786, + 297, + 786 + ], + "score": 0.408 + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 1404.0, + 882.0, + 1404.0, + 923.0, + 294.0, + 923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 915.0, + 1316.0, + 915.0, + 1316.0, + 951.0, + 322.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 796.0, + 1401.0, + 796.0, + 1401.0, + 841.0, + 294.0, + 841.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 833.0, + 1087.0, + 833.0, + 1087.0, + 869.0, + 320.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 963.0, + 1403.0, + 963.0, + 1403.0, + 999.0, + 296.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 995.0, + 1373.0, + 995.0, + 1373.0, + 1032.0, + 324.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1046.0, + 1402.0, + 1046.0, + 1402.0, + 1080.0, + 296.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1074.0, + 1406.0, + 1074.0, + 1406.0, + 1113.0, + 322.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1108.0, + 587.0, + 1108.0, + 587.0, + 1137.0, + 321.0, + 1137.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1588.0, + 1403.0, + 1588.0, + 1403.0, + 1624.0, + 296.0, + 1624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1620.0, + 1147.0, + 1620.0, + 1147.0, + 1654.0, + 322.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1395.0, + 1403.0, + 1395.0, + 1403.0, + 1434.0, + 293.0, + 1434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1429.0, + 1371.0, + 1429.0, + 1371.0, + 1463.0, + 322.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1232.0, + 1405.0, + 1232.0, + 1405.0, + 1275.0, + 293.0, + 1275.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1266.0, + 883.0, + 1266.0, + 883.0, + 1302.0, + 322.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 578.0, + 1403.0, + 578.0, + 1403.0, + 619.0, + 295.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 613.0, + 795.0, + 613.0, + 795.0, + 646.0, + 321.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 502.0, + 1404.0, + 502.0, + 1404.0, + 538.0, + 297.0, + 538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 531.0, + 1043.0, + 531.0, + 1043.0, + 566.0, + 321.0, + 566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1156.0, + 1403.0, + 1156.0, + 1403.0, + 1192.0, + 296.0, + 1192.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1187.0, + 1403.0, + 1187.0, + 1403.0, + 1223.0, + 324.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1665.0, + 1403.0, + 1665.0, + 1403.0, + 1707.0, + 294.0, + 1707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1700.0, + 1051.0, + 1700.0, + 1051.0, + 1734.0, + 322.0, + 1734.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1744.0, + 1402.0, + 1744.0, + 1402.0, + 1787.0, + 292.0, + 1787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1781.0, + 803.0, + 1781.0, + 803.0, + 1814.0, + 324.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1830.0, + 1403.0, + 1830.0, + 1403.0, + 1866.0, + 297.0, + 1866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1862.0, + 707.0, + 1862.0, + 707.0, + 1895.0, + 320.0, + 1895.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 391.0, + 1404.0, + 391.0, + 1404.0, + 425.0, + 295.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 420.0, + 1405.0, + 420.0, + 1405.0, + 458.0, + 322.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 452.0, + 649.0, + 452.0, + 649.0, + 484.0, + 320.0, + 484.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1476.0, + 1404.0, + 1476.0, + 1404.0, + 1514.0, + 294.0, + 1514.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 1509.0, + 1405.0, + 1509.0, + 1405.0, + 1544.0, + 325.0, + 1544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1540.0, + 623.0, + 1540.0, + 623.0, + 1569.0, + 322.0, + 1569.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1317.0, + 1401.0, + 1317.0, + 1401.0, + 1349.0, + 297.0, + 1349.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1348.0, + 1164.0, + 1348.0, + 1164.0, + 1383.0, + 322.0, + 1383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 310.0, + 1405.0, + 310.0, + 1405.0, + 346.0, + 296.0, + 346.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 339.0, + 743.0, + 339.0, + 743.0, + 374.0, + 321.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 1405.0, + 230.0, + 1405.0, + 266.0, + 296.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 261.0, + 1023.0, + 261.0, + 1023.0, + 296.0, + 321.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1910.0, + 1404.0, + 1910.0, + 1404.0, + 1946.0, + 294.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1943.0, + 1403.0, + 1943.0, + 1403.0, + 1976.0, + 322.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1974.0, + 1404.0, + 1974.0, + 1404.0, + 2006.0, + 323.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 2003.0, + 964.0, + 2003.0, + 964.0, + 2035.0, + 323.0, + 2035.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 660.0, + 1406.0, + 660.0, + 1406.0, + 700.0, + 294.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 691.0, + 1404.0, + 691.0, + 1404.0, + 729.0, + 322.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 723.0, + 1406.0, + 723.0, + 1406.0, + 763.0, + 320.0, + 763.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 756.0, + 614.0, + 756.0, + 614.0, + 788.0, + 324.0, + 788.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 301, + 229, + 1400, + 229, + 1400, + 293, + 301, + 293 + ], + "score": 0.9 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.827 + }, + { + "category_id": 2, + "poly": [ + 300, + 75, + 814, + 75, + 814, + 105, + 300, + 105 + ], + "score": 0.825 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 816.0, + 73.0, + 816.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 229.0, + 1403.0, + 229.0, + 1403.0, + 265.0, + 297.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 262.0, + 1391.0, + 262.0, + 1391.0, + 294.0, + 323.0, + 294.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1596, + 1407, + 1596, + 1407, + 1793, + 295, + 1793 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1439, + 1402, + 1439, + 1402, + 1579, + 299, + 1579 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 469, + 1840, + 1237, + 1840, + 1237, + 1980, + 469, + 1980 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 788, + 1406, + 788, + 1406, + 886, + 297, + 886 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 537, + 1207, + 1167, + 1207, + 1167, + 1437, + 537, + 1437 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 294, + 359, + 1405, + 359, + 1405, + 425, + 294, + 425 + ], + "score": 0.941 + }, + { + "category_id": 8, + "poly": [ + 489, + 718, + 1210, + 718, + 1210, + 776, + 489, + 776 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 296, + 650, + 1404, + 650, + 1404, + 714, + 296, + 714 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 493, + 592, + 1201, + 592, + 1201, + 648, + 493, + 648 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 300, + 559, + 700, + 559, + 700, + 590, + 300, + 590 + ], + "score": 0.927 + }, + { + "category_id": 1, + "poly": [ + 296, + 1038, + 1402, + 1038, + 1402, + 1115, + 296, + 1115 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 295, + 896, + 1403, + 896, + 1403, + 958, + 295, + 958 + ], + "score": 0.921 + }, + { + "category_id": 1, + "poly": [ + 298, + 1170, + 1253, + 1170, + 1253, + 1205, + 298, + 1205 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.917 + }, + { + "category_id": 1, + "poly": [ + 288, + 437, + 1287, + 437, + 1287, + 472, + 288, + 472 + ], + "score": 0.912 + }, + { + "category_id": 1, + "poly": [ + 299, + 1804, + 786, + 1804, + 786, + 1838, + 299, + 1838 + ], + "score": 0.91 + }, + { + "category_id": 8, + "poly": [ + 416, + 477, + 1284, + 477, + 1284, + 560, + 416, + 560 + ], + "score": 0.91 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1305, + 1401, + 1305, + 1401, + 1336, + 1350, + 1336 + ], + "score": 0.901 + }, + { + "category_id": 8, + "poly": [ + 492, + 961, + 1208, + 961, + 1208, + 1029, + 492, + 1029 + ], + "score": 0.893 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1890, + 1401, + 1890, + 1401, + 1922, + 1351, + 1922 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1351, + 728, + 1400, + 728, + 1400, + 760, + 1351, + 760 + ], + "score": 0.883 + }, + { + "category_id": 9, + "poly": [ + 1365, + 603, + 1400, + 603, + 1400, + 633, + 1365, + 633 + ], + "score": 0.883 + }, + { + "category_id": 0, + "poly": [ + 300, + 226, + 509, + 226, + 509, + 262, + 300, + 262 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1365, + 501, + 1401, + 501, + 1401, + 531, + 1365, + 531 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1975, + 1403, + 1975, + 1403, + 2004, + 1374, + 2004 + ], + "score": 0.833 + }, + { + "category_id": 0, + "poly": [ + 298, + 300, + 868, + 300, + 868, + 334, + 298, + 334 + ], + "score": 0.619 + }, + { + "category_id": 1, + "poly": [ + 298, + 300, + 868, + 300, + 868, + 334, + 298, + 334 + ], + "score": 0.263 + }, + { + "category_id": 13, + "poly": [ + 1133, + 789, + 1353, + 789, + 1353, + 825, + 1133, + 825 + ], + "score": 0.94, + "latex": "\\beta ^ { j } = \\beta ^ { j } ( 1 - \\beta ^ { j } ) ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 376, + 679, + 531, + 679, + 531, + 718, + 376, + 718 + ], + "score": 0.94, + "latex": "\\nabla _ { \\phi _ { o l d } } \\tilde { J } \\big ( \\phi _ { o l d } \\big )" + }, + { + "category_id": 13, + "poly": [ + 1138, + 1030, + 1319, + 1030, + 1319, + 1085, + 1138, + 1085 + ], + "score": 0.93, + "latex": "\\zeta \\leq \\frac { C \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 14, + "poly": [ + 417, + 470, + 1283, + 470, + 1283, + 564, + 417, + 564 + ], + "score": 0.93, + "latex": "\\phi \\phi _ { o l d } + \\beta { \\pmb F } ^ { - 1 } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) \\quad \\beta = \\beta ^ { j } \\sqrt { \\frac { 2 \\delta } { \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } { \\pmb F } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) } }" + }, + { + "category_id": 13, + "poly": [ + 1202, + 1594, + 1393, + 1594, + 1393, + 1651, + 1202, + 1651 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 926, + 1172, + 1017, + 1172, + 1017, + 1206, + 926, + 1206 + ], + "score": 0.93, + "latex": "\\{ ( s , a ) \\}" + }, + { + "category_id": 13, + "poly": [ + 919, + 1438, + 1037, + 1438, + 1037, + 1477, + 919, + 1477 + ], + "score": 0.93, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 343, + 1605, + 663, + 1605, + 663, + 1644, + 343, + 1644 + ], + "score": 0.92, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) = V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) - \\zeta" + }, + { + "category_id": 13, + "poly": [ + 873, + 1040, + 991, + 1040, + 991, + 1079, + 873, + 1079 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)" + }, + { + "category_id": 14, + "poly": [ + 492, + 958, + 1210, + 958, + 1210, + 1030, + 492, + 1030 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 298, + 823, + 419, + 823, + 419, + 855, + 298, + 855 + ], + "score": 0.92, + "latex": "j j + 1" + }, + { + "category_id": 14, + "poly": [ + 497, + 590, + 1201, + 590, + 1201, + 648, + 497, + 648 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\pmb { F } = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ) ^ { T } \\right] \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 352, + 1512, + 469, + 1512, + 469, + 1550, + 352, + 1550 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 14, + "poly": [ + 466, + 1835, + 1240, + 1835, + 1240, + 1983, + 466, + 1983 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\quad \\hat { V } _ { C } ^ { \\hat { \\pi _ { \\phi _ { o l d } } } } ( \\mu ) + \\displaystyle \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right] \\leq \\chi } \\\\ & { \\Longrightarrow V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\displaystyle \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1708, + 593, + 1708, + 593, + 1763, + 297, + 1763 + ], + "score": 0.91, + "latex": "\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right]" + }, + { + "category_id": 14, + "poly": [ + 489, + 716, + 1209, + 716, + 1209, + 774, + 489, + 774 + ], + "score": 0.91, + "latex": "\\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\tilde { A } _ { R } ^ { \\pi _ { \\phi _ { o l d } } } \\right]" + }, + { + "category_id": 13, + "poly": [ + 669, + 788, + 882, + 788, + 882, + 824, + 669, + 824 + ], + "score": 0.91, + "latex": "\\bar { D } _ { \\mathrm { K L } } ( \\phi | | \\phi _ { o l d } ) \\leq \\delta" + }, + { + "category_id": 13, + "poly": [ + 1078, + 1474, + 1196, + 1474, + 1196, + 1513, + 1078, + 1513 + ], + "score": 0.91, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 14, + "poly": [ + 534, + 1206, + 1170, + 1206, + 1170, + 1440, + 534, + 1440 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\quad Q _ { C } ( s , a ) \\leq ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) ) ) \\quad \\forall ( s , a ) } \\\\ & { \\Longrightarrow \\hat { A } _ { C } ( s , a ) \\leq ( 1 - \\gamma ) ( \\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) ) ) \\quad \\forall ( s , a ) } \\\\ & { \\Longrightarrow \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) + \\cfrac { 1 } { 1 - \\gamma } \\hat { A } _ { C } ( s , a ) \\leq \\chi \\quad \\forall ( s , a ) } \\\\ & { \\Longrightarrow \\hat { V } _ { C } ^ { \\pi _ { \\phi o l d } } ( \\mu ) + \\cfrac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right] \\leq \\chi } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1082, + 1511, + 1354, + 1511, + 1354, + 1551, + 1082, + 1551 + ], + "score": 0.9, + "latex": "\\tilde { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\geq V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 841, + 363, + 895, + 363, + 895, + 393, + 841, + 393 + ], + "score": 0.9, + "latex": "D _ { \\mathrm { K L } }" + }, + { + "category_id": 13, + "poly": [ + 727, + 855, + 836, + 855, + 836, + 885, + 727, + 885 + ], + "score": 0.9, + "latex": "\\phi \\phi _ { o l d }" + }, + { + "category_id": 13, + "poly": [ + 936, + 1716, + 1221, + 1716, + 1221, + 1756, + 936, + 1756 + ], + "score": 0.9, + "latex": "\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 1031, + 1679, + 1067, + 1679, + 1067, + 1710, + 1031, + 1710 + ], + "score": 0.89, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 924, + 1610, + 1024, + 1610, + 1024, + 1641, + 924, + 1641 + ], + "score": 0.89, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 732, + 825, + 835, + 825, + 835, + 853, + 732, + 853 + ], + "score": 0.89, + "latex": "D _ { \\mathrm { K L } } \\leq \\delta" + }, + { + "category_id": 13, + "poly": [ + 360, + 648, + 391, + 648, + 391, + 681, + 360, + 681 + ], + "score": 0.89, + "latex": "\\beta ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 581, + 1807, + 679, + 1807, + 679, + 1836, + 581, + 1836 + ], + "score": 0.89, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 1114, + 1082, + 1149, + 1082, + 1149, + 1112, + 1114, + 1112 + ], + "score": 0.87, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1362, + 1443, + 1401, + 1443, + 1401, + 1474, + 1362, + 1474 + ], + "score": 0.86, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 374, + 1648, + 405, + 1648, + 405, + 1675, + 374, + 1675 + ], + "score": 0.86, + "latex": "C ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 536, + 930, + 636, + 930, + 636, + 959, + 536, + 959 + ], + "score": 0.85, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 623, + 854, + 643, + 854, + 643, + 884, + 623, + 884 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1233, + 443, + 1252, + 443, + 1252, + 472, + 1233, + 472 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 1270, + 826, + 1292, + 826, + 1292, + 850, + 1270, + 850 + ], + "score": 0.84, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 365, + 1046, + 382, + 1046, + 382, + 1077, + 365, + 1077 + ], + "score": 0.83, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 411, + 1680, + 439, + 1680, + 439, + 1706, + 411, + 1706 + ], + "score": 0.82, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 585, + 824, + 607, + 824, + 607, + 850, + 585, + 850 + ], + "score": 0.82, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 299, + 561, + 324, + 561, + 324, + 587, + 299, + 587 + ], + "score": 0.79, + "latex": "\\pmb { F }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1081, + 323, + 1081, + 323, + 1108, + 298, + 1108 + ], + "score": 0.79, + "latex": "C" + }, + { + "category_id": 13, + "poly": [ + 522, + 1081, + 550, + 1081, + 550, + 1108, + 522, + 1108 + ], + "score": 0.78, + "latex": "N" + }, + { + "category_id": 13, + "poly": [ + 509, + 1762, + 534, + 1762, + 534, + 1786, + 509, + 1786 + ], + "score": 0.76, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 736, + 1652, + 759, + 1652, + 759, + 1675, + 736, + 1675 + ], + "score": 0.76, + "latex": "\\omega" + }, + { + "category_id": 13, + "poly": [ + 685, + 901, + 699, + 901, + 699, + 925, + 685, + 925 + ], + "score": 0.48, + "latex": "I" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 220.0, + 514.0, + 220.0, + 514.0, + 271.0, + 294.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 831.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 831.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1979.0, + 1403.0, + 1979.0, + 1403.0, + 2006.0, + 1377.0, + 2006.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 300.0, + 868.0, + 300.0, + 868.0, + 336.0, + 296.0, + 336.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1592.0, + 342.0, + 1592.0, + 342.0, + 1653.0, + 288.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1592.0, + 923.0, + 1592.0, + 923.0, + 1653.0, + 664.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 1592.0, + 1201.0, + 1592.0, + 1201.0, + 1653.0, + 1025.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 1592.0, + 1397.0, + 1592.0, + 1397.0, + 1653.0, + 1394.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1641.0, + 373.0, + 1641.0, + 373.0, + 1683.0, + 296.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 406.0, + 1641.0, + 735.0, + 1641.0, + 735.0, + 1683.0, + 406.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 760.0, + 1641.0, + 1406.0, + 1641.0, + 1406.0, + 1683.0, + 760.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1675.0, + 410.0, + 1675.0, + 410.0, + 1716.0, + 293.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1675.0, + 1030.0, + 1675.0, + 1030.0, + 1716.0, + 440.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1068.0, + 1675.0, + 1408.0, + 1675.0, + 1408.0, + 1716.0, + 1068.0, + 1716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1705.0, + 296.0, + 1705.0, + 296.0, + 1769.0, + 290.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1705.0, + 935.0, + 1705.0, + 935.0, + 1769.0, + 594.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 1705.0, + 1412.0, + 1705.0, + 1412.0, + 1769.0, + 1222.0, + 1769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1750.0, + 508.0, + 1750.0, + 508.0, + 1795.0, + 293.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 535.0, + 1750.0, + 547.0, + 1750.0, + 547.0, + 1795.0, + 535.0, + 1795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1429.0, + 918.0, + 1429.0, + 918.0, + 1486.0, + 286.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1429.0, + 1361.0, + 1429.0, + 1361.0, + 1486.0, + 1038.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1429.0, + 1412.0, + 1429.0, + 1412.0, + 1486.0, + 1402.0, + 1486.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1462.0, + 1077.0, + 1462.0, + 1077.0, + 1528.0, + 285.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1197.0, + 1462.0, + 1416.0, + 1462.0, + 1416.0, + 1528.0, + 1197.0, + 1528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 1498.0, + 351.0, + 1498.0, + 351.0, + 1566.0, + 280.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 470.0, + 1498.0, + 1081.0, + 1498.0, + 1081.0, + 1566.0, + 470.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1355.0, + 1498.0, + 1418.0, + 1498.0, + 1418.0, + 1566.0, + 1355.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1548.0, + 1072.0, + 1548.0, + 1072.0, + 1579.0, + 295.0, + 1579.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 788.0, + 668.0, + 788.0, + 668.0, + 826.0, + 292.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 883.0, + 788.0, + 1132.0, + 788.0, + 1132.0, + 826.0, + 883.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 788.0, + 1406.0, + 788.0, + 1406.0, + 826.0, + 1354.0, + 826.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 820.0, + 297.0, + 820.0, + 297.0, + 858.0, + 291.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 820.0, + 584.0, + 820.0, + 584.0, + 858.0, + 420.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 820.0, + 731.0, + 820.0, + 731.0, + 858.0, + 608.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 820.0, + 1269.0, + 820.0, + 1269.0, + 858.0, + 836.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 820.0, + 1406.0, + 820.0, + 1406.0, + 858.0, + 1293.0, + 858.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 847.0, + 622.0, + 847.0, + 622.0, + 892.0, + 291.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 847.0, + 726.0, + 847.0, + 726.0, + 892.0, + 644.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 837.0, + 847.0, + 849.0, + 847.0, + 849.0, + 892.0, + 837.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 360.0, + 840.0, + 360.0, + 840.0, + 396.0, + 294.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 896.0, + 360.0, + 1406.0, + 360.0, + 1406.0, + 396.0, + 896.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 392.0, + 567.0, + 392.0, + 567.0, + 426.0, + 293.0, + 426.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 646.0, + 359.0, + 646.0, + 359.0, + 686.0, + 294.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 646.0, + 1405.0, + 646.0, + 1405.0, + 686.0, + 392.0, + 686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 681.0, + 375.0, + 681.0, + 375.0, + 720.0, + 295.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 681.0, + 718.0, + 681.0, + 718.0, + 720.0, + 532.0, + 720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 554.0, + 298.0, + 554.0, + 298.0, + 596.0, + 295.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 554.0, + 702.0, + 554.0, + 702.0, + 596.0, + 325.0, + 596.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1027.0, + 364.0, + 1027.0, + 364.0, + 1096.0, + 291.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1027.0, + 872.0, + 1027.0, + 872.0, + 1096.0, + 383.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 992.0, + 1027.0, + 1137.0, + 1027.0, + 1137.0, + 1096.0, + 992.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1027.0, + 1409.0, + 1027.0, + 1409.0, + 1096.0, + 1320.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1074.0, + 297.0, + 1074.0, + 297.0, + 1119.0, + 294.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 1074.0, + 521.0, + 1074.0, + 521.0, + 1119.0, + 324.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 551.0, + 1074.0, + 1113.0, + 1074.0, + 1113.0, + 1119.0, + 551.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1074.0, + 1161.0, + 1074.0, + 1161.0, + 1119.0, + 1150.0, + 1119.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 894.0, + 684.0, + 894.0, + 684.0, + 936.0, + 292.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 700.0, + 894.0, + 1406.0, + 894.0, + 1406.0, + 936.0, + 700.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 928.0, + 535.0, + 928.0, + 535.0, + 964.0, + 295.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 637.0, + 928.0, + 640.0, + 928.0, + 640.0, + 964.0, + 637.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1166.0, + 925.0, + 1166.0, + 925.0, + 1211.0, + 292.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1018.0, + 1166.0, + 1259.0, + 1166.0, + 1259.0, + 1211.0, + 1018.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 435.0, + 1232.0, + 435.0, + 1232.0, + 477.0, + 292.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 435.0, + 1288.0, + 435.0, + 1288.0, + 477.0, + 1253.0, + 477.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1802.0, + 580.0, + 1802.0, + 580.0, + 1842.0, + 296.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1802.0, + 788.0, + 1802.0, + 788.0, + 1842.0, + 680.0, + 1842.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 300.0, + 868.0, + 300.0, + 868.0, + 336.0, + 296.0, + 336.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 12, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 676, + 1405, + 676, + 1405, + 799, + 297, + 799 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 295, + 1373, + 1406, + 1373, + 1406, + 1715, + 295, + 1715 + ], + "score": 0.97 + }, + { + "category_id": 8, + "poly": [ + 670, + 800, + 1026, + 800, + 1026, + 871, + 670, + 871 + ], + "score": 0.954 + }, + { + "category_id": 1, + "poly": [ + 293, + 228, + 1404, + 228, + 1404, + 325, + 293, + 325 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 298, + 876, + 1406, + 876, + 1406, + 946, + 298, + 946 + ], + "score": 0.947 + }, + { + "category_id": 1, + "poly": [ + 298, + 1053, + 1402, + 1053, + 1402, + 1116, + 298, + 1116 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 295, + 1193, + 1404, + 1193, + 1404, + 1260, + 295, + 1260 + ], + "score": 0.94 + }, + { + "category_id": 8, + "poly": [ + 424, + 328, + 1276, + 328, + 1276, + 406, + 424, + 406 + ], + "score": 0.939 + }, + { + "category_id": 1, + "poly": [ + 296, + 444, + 1406, + 444, + 1406, + 516, + 296, + 516 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 652, + 982, + 1048, + 982, + 1048, + 1036, + 652, + 1036 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 327, + 1119, + 1321, + 1119, + 1321, + 1190, + 327, + 1190 + ], + "score": 0.925 + }, + { + "category_id": 1, + "poly": [ + 298, + 1876, + 1406, + 1876, + 1406, + 1939, + 298, + 1939 + ], + "score": 0.923 + }, + { + "category_id": 8, + "poly": [ + 616, + 1833, + 1084, + 1833, + 1084, + 1874, + 616, + 1874 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.918 + }, + { + "category_id": 9, + "poly": [ + 1351, + 996, + 1400, + 996, + 1400, + 1027, + 1351, + 1027 + ], + "score": 0.899 + }, + { + "category_id": 1, + "poly": [ + 298, + 1799, + 556, + 1799, + 556, + 1831, + 298, + 1831 + ], + "score": 0.895 + }, + { + "category_id": 8, + "poly": [ + 681, + 1763, + 1021, + 1763, + 1021, + 1803, + 681, + 1803 + ], + "score": 0.891 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1839, + 1401, + 1839, + 1401, + 1869, + 1351, + 1869 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1767, + 1400, + 1767, + 1400, + 1798, + 1351, + 1798 + ], + "score": 0.884 + }, + { + "category_id": 1, + "poly": [ + 294, + 1727, + 1278, + 1727, + 1278, + 1760, + 294, + 1760 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1351, + 350, + 1400, + 350, + 1400, + 382, + 1351, + 382 + ], + "score": 0.878 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2113, + 836, + 2113 + ], + "score": 0.864 + }, + { + "category_id": 8, + "poly": [ + 298, + 1940, + 1529, + 1940, + 1529, + 2011, + 298, + 2011 + ], + "score": 0.858 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1138, + 1400, + 1138, + 1400, + 1169, + 1352, + 1169 + ], + "score": 0.852 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1320, + 1400, + 1320, + 1400, + 1351, + 1352, + 1351 + ], + "score": 0.836 + }, + { + "category_id": 8, + "poly": [ + 319, + 1297, + 1323, + 1297, + 1323, + 1372, + 319, + 1372 + ], + "score": 0.828 + }, + { + "category_id": 9, + "poly": [ + 1351, + 2009, + 1401, + 2009, + 1401, + 2038, + 1351, + 2038 + ], + "score": 0.724 + }, + { + "category_id": 14, + "poly": [ + 669, + 796, + 1027, + 796, + 1027, + 874, + 669, + 874 + ], + "score": 0.94, + "latex": "C ( s ) = { \\left\\{ \\begin{array} { l l } { 1 , } & { \\mathbb { I } \\{ f a i l u r e \\} = 1 } \\\\ { 0 , } & { { \\mathrm { o t h e r w i s e } } } \\end{array} \\right. }" + }, + { + "category_id": 13, + "poly": [ + 856, + 1557, + 961, + 1557, + 961, + 1596, + 856, + 1596 + ], + "score": 0.93, + "latex": "\\hat { A } _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 524, + 910, + 615, + 910, + 615, + 947, + 524, + 947 + ], + "score": 0.93, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 341, + 875, + 432, + 875, + 432, + 912, + 341, + 912 + ], + "score": 0.93, + "latex": "V _ { R } ^ { \\pi _ { \\phi } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 726, + 709, + 811, + 709, + 811, + 742, + 726, + 742 + ], + "score": 0.92, + "latex": "R ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 651, + 981, + 1049, + 981, + 1049, + 1040, + 651, + 1040 + ], + "score": 0.92, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) \\quad s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi" + }, + { + "category_id": 14, + "poly": [ + 327, + 1116, + 1325, + 1116, + 1325, + 1191, + 327, + 1191 + ], + "score": 0.92, + "latex": "V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) - V _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\geq \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\frac { 2 \\gamma \\epsilon _ { R } } { 1 - \\gamma } \\mathrm { D } _ { T V } ( \\pi _ { \\phi } | | \\pi _ { \\phi _ { o l d } } ) [ s ] \\right]" + }, + { + "category_id": 13, + "poly": [ + 554, + 1877, + 710, + 1877, + 710, + 1911, + 554, + 1911 + ], + "score": 0.92, + "latex": "\\mathbb { E } [ \\mathbf { D } _ { T V } ( p | | q ) ]" + }, + { + "category_id": 13, + "poly": [ + 977, + 1874, + 1160, + 1874, + 1160, + 1912, + 977, + 1912 + ], + "score": 0.92, + "latex": "\\sqrt { \\mathbb { E } [ \\mathbf { D } _ { K L } ( \\boldsymbol { p } | | \\boldsymbol { q } ) ] }" + }, + { + "category_id": 13, + "poly": [ + 1216, + 739, + 1274, + 739, + 1274, + 773, + 1216, + 773 + ], + "score": 0.92, + "latex": "C ( s )" + }, + { + "category_id": 14, + "poly": [ + 327, + 1298, + 1328, + 1298, + 1328, + 1374, + 327, + 1374 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) - V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) \\leq \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) + \\frac { 2 \\gamma \\epsilon _ { C } } { 1 - \\gamma } \\mathrm { D } _ { T V } ( \\pi _ { \\phi } | | \\pi _ { \\phi _ { o l d } } ) [ s ] \\right]" + }, + { + "category_id": 13, + "poly": [ + 380, + 678, + 439, + 678, + 439, + 712, + 380, + 712 + ], + "score": 0.91, + "latex": "C ( s )" + }, + { + "category_id": 13, + "poly": [ + 960, + 709, + 1019, + 709, + 1019, + 743, + 960, + 743 + ], + "score": 0.91, + "latex": "C ( s )" + }, + { + "category_id": 13, + "poly": [ + 685, + 1676, + 726, + 1676, + 726, + 1712, + 685, + 1712 + ], + "score": 0.91, + "latex": "\\hat { A } _ { C }" + }, + { + "category_id": 14, + "poly": [ + 426, + 324, + 1273, + 324, + 1273, + 405, + 426, + 405 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi n e w } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad w h e r e \\quad \\zeta \\leq \\frac { C \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1447, + 401, + 1447, + 401, + 1480, + 297, + 1480 + ], + "score": 0.91, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1410, + 595, + 1410, + 595, + 1449, + 297, + 1449 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { \\operatorname* { m a x } _ { s } \\vert \\mathbb { E } _ { a \\sim \\pi _ { \\phi } } A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\vert } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1084, + 447, + 1400, + 447, + 1400, + 488, + 1084, + 488 + ], + "score": 0.9, + "latex": "\\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ A _ { C } ( s , a ) ]" + }, + { + "category_id": 14, + "poly": [ + 310, + 1935, + 1428, + 1935, + 1428, + 2011, + 310, + 2011 + ], + "score": 0.89, + "latex": "\\tau _ { C } ^ { \\pi \\phi } ( \\mu ) - V _ { C } ^ { \\pi \\phi _ { o l d } } ( \\mu ) \\leq \\frac { 1 } { 1 - \\gamma } \\left[ \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] + \\frac { 2 \\gamma \\epsilon _ { C } } { 1 - \\gamma } \\sqrt { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\mathbf { D } _ { K L } ( \\pi _ { \\phi } | | \\pi _ { \\phi } ) \\right] } \\right] ." + }, + { + "category_id": 13, + "poly": [ + 366, + 447, + 734, + 447, + 734, + 484, + 366, + 484 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ( s , a ) | } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1207, + 1729, + 1265, + 1729, + 1265, + 1760, + 1207, + 1760 + ], + "score": 0.88, + "latex": "\\mathrm { D } _ { T V }" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1730, + 1087, + 1730, + 1087, + 1759, + 1028, + 1759 + ], + "score": 0.88, + "latex": "\\mathrm { D } _ { K L }" + }, + { + "category_id": 13, + "poly": [ + 641, + 295, + 710, + 295, + 710, + 325, + 641, + 325 + ], + "score": 0.88, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 408, + 1682, + 449, + 1682, + 449, + 1713, + 408, + 1713 + ], + "score": 0.88, + "latex": "A _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1356, + 1562, + 1401, + 1562, + 1401, + 1595, + 1356, + 1595 + ], + "score": 0.88, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1224, + 569, + 1224, + 569, + 1263, + 298, + 1263 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\operatorname* { m a x } _ { s } \\vert \\mathbb { E } _ { a \\sim \\pi _ { \\phi } } A _ { R } ^ { \\pi _ { \\phi } } ( s , a ) \\vert } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 798, + 1414, + 842, + 1414, + 842, + 1445, + 798, + 1445 + ], + "score": 0.88, + "latex": "A _ { C }" + }, + { + "category_id": 13, + "poly": [ + 383, + 1190, + 435, + 1190, + 435, + 1226, + 383, + 1226 + ], + "score": 0.87, + "latex": "A _ { R } ^ { \\pi _ { \\phi } }" + }, + { + "category_id": 14, + "poly": [ + 679, + 1759, + 1020, + 1759, + 1020, + 1802, + 679, + 1802 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { \\mathbf { D } _ { T V } ( p | | q ) \\leq \\sqrt { \\mathbf { D } _ { K L } ( p | | q ) / 2 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 470, + 291, + 597, + 291, + 597, + 326, + 470, + 326 + ], + "score": 0.86, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1622, + 1003, + 1622, + 1003, + 1679, + 296, + 1679 + ], + "score": 0.86, + "latex": "{ { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } _ { C } ( s , a ) \\right] \\ = \\ { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] + \\Delta }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 883, + 1361, + 883, + 1361, + 913, + 1328, + 913 + ], + "score": 0.86, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 14, + "poly": [ + 616, + 1831, + 1085, + 1831, + 1085, + 1873, + 616, + 1873 + ], + "score": 0.84, + "latex": "\\mathbb { E } [ \\sqrt { \\mathbf { D } _ { K L } ( p | | q ) / 2 } ] \\le \\sqrt { \\mathbb { E } [ \\mathbf { D } _ { K L } ( p | | q ) ] / 2 }" + }, + { + "category_id": 13, + "poly": [ + 674, + 263, + 775, + 263, + 775, + 292, + 674, + 292 + ], + "score": 0.84, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 786, + 449, + 813, + 449, + 813, + 475, + 786, + 475 + ], + "score": 0.83, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 1299, + 1380, + 1403, + 1380, + 1403, + 1411, + 1299, + 1411 + ], + "score": 0.82, + "latex": "\\begin{array} { r l } { \\epsilon _ { C } } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 389, + 1373, + 470, + 1373, + 470, + 1411, + 389, + 1411 + ], + "score": 0.8, + "latex": "A _ { C } ^ { \\pi _ { \\phi _ { o l d } } }" + }, + { + "category_id": 13, + "poly": [ + 1101, + 1633, + 1129, + 1633, + 1129, + 1661, + 1101, + 1661 + ], + "score": 0.79, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 790, + 1594, + 818, + 1594, + 818, + 1622, + 790, + 1622 + ], + "score": 0.78, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 1309, + 1196, + 1403, + 1196, + 1403, + 1226, + 1309, + 1226 + ], + "score": 0.78, + "latex": "\\begin{array} { r l } { \\epsilon _ { R } } & { { } = } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 579, + 1226, + 639, + 1226, + 639, + 1259, + 579, + 1259 + ], + "score": 0.74, + "latex": "\\mathbf { D } _ { T V }" + }, + { + "category_id": 13, + "poly": [ + 355, + 1477, + 1398, + 1477, + 1398, + 1533, + 355, + 1533 + ], + "score": 0.73, + "latex": "\\begin{array} { r l r } { { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { = } & { { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ Q _ { \\zeta } ( s , a ) \\right] - { \\mathbb E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ Q _ { \\zeta } ( s , a ) \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1287, + 685, + 1302, + 685, + 1302, + 706, + 1287, + 706 + ], + "score": 0.61, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1268, + 1414, + 1296, + 1414, + 1296, + 1445, + 1268, + 1445 + ], + "score": 0.5, + "latex": "\\mathrm { Q }" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 871.0, + 2085.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 677.0, + 379.0, + 677.0, + 379.0, + 713.0, + 295.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 677.0, + 1286.0, + 677.0, + 1286.0, + 713.0, + 440.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 677.0, + 1406.0, + 677.0, + 1406.0, + 713.0, + 1303.0, + 713.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 708.0, + 725.0, + 708.0, + 725.0, + 743.0, + 292.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 812.0, + 708.0, + 959.0, + 708.0, + 959.0, + 743.0, + 812.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 708.0, + 1408.0, + 708.0, + 1408.0, + 743.0, + 1020.0, + 743.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 739.0, + 1215.0, + 739.0, + 1215.0, + 775.0, + 294.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1275.0, + 739.0, + 1406.0, + 739.0, + 1406.0, + 775.0, + 1275.0, + 775.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 770.0, + 583.0, + 770.0, + 583.0, + 803.0, + 295.0, + 803.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 280.0, + 1356.0, + 388.0, + 1356.0, + 388.0, + 1425.0, + 280.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 471.0, + 1356.0, + 1298.0, + 1356.0, + 1298.0, + 1425.0, + 471.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1404.0, + 1356.0, + 1418.0, + 1356.0, + 1418.0, + 1425.0, + 1404.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1398.0, + 296.0, + 1398.0, + 296.0, + 1454.0, + 284.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 1398.0, + 797.0, + 1398.0, + 797.0, + 1454.0, + 596.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 843.0, + 1398.0, + 1267.0, + 1398.0, + 1267.0, + 1454.0, + 843.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1297.0, + 1398.0, + 1412.0, + 1398.0, + 1412.0, + 1454.0, + 1297.0, + 1454.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1441.0, + 296.0, + 1441.0, + 296.0, + 1487.0, + 292.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1441.0, + 1408.0, + 1441.0, + 1408.0, + 1487.0, + 402.0, + 1487.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1471.0, + 354.0, + 1471.0, + 354.0, + 1538.0, + 284.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1399.0, + 1471.0, + 1413.0, + 1471.0, + 1413.0, + 1538.0, + 1399.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1523.0, + 1408.0, + 1523.0, + 1408.0, + 1565.0, + 292.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1558.0, + 855.0, + 1558.0, + 855.0, + 1599.0, + 293.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 962.0, + 1558.0, + 1355.0, + 1558.0, + 1355.0, + 1599.0, + 962.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1558.0, + 1406.0, + 1558.0, + 1406.0, + 1599.0, + 1402.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1589.0, + 789.0, + 1589.0, + 789.0, + 1631.0, + 293.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 1589.0, + 1406.0, + 1589.0, + 1406.0, + 1631.0, + 819.0, + 1631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1626.0, + 295.0, + 1626.0, + 295.0, + 1677.0, + 289.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1626.0, + 1100.0, + 1626.0, + 1100.0, + 1677.0, + 1004.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1626.0, + 1412.0, + 1626.0, + 1412.0, + 1677.0, + 1130.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1673.0, + 407.0, + 1673.0, + 407.0, + 1717.0, + 292.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 450.0, + 1673.0, + 684.0, + 1673.0, + 684.0, + 1717.0, + 450.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1673.0, + 1156.0, + 1673.0, + 1156.0, + 1717.0, + 727.0, + 1717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 267.0, + 295.0, + 267.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 262.0, + 673.0, + 262.0, + 673.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 296.0, + 776.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 283.0, + 469.0, + 283.0, + 469.0, + 333.0, + 290.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 283.0, + 640.0, + 283.0, + 640.0, + 333.0, + 598.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 283.0, + 1125.0, + 283.0, + 1125.0, + 333.0, + 711.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 870.0, + 340.0, + 870.0, + 340.0, + 920.0, + 292.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 433.0, + 870.0, + 1327.0, + 870.0, + 1327.0, + 920.0, + 433.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1362.0, + 870.0, + 1407.0, + 870.0, + 1407.0, + 920.0, + 1362.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 906.0, + 523.0, + 906.0, + 523.0, + 951.0, + 293.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 906.0, + 1106.0, + 906.0, + 1106.0, + 951.0, + 616.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1049.0, + 1406.0, + 1049.0, + 1406.0, + 1091.0, + 293.0, + 1091.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1086.0, + 789.0, + 1086.0, + 789.0, + 1118.0, + 297.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1176.0, + 382.0, + 1176.0, + 382.0, + 1241.0, + 285.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 436.0, + 1176.0, + 1308.0, + 1176.0, + 1308.0, + 1241.0, + 436.0, + 1241.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1220.0, + 297.0, + 1220.0, + 297.0, + 1266.0, + 293.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.0, + 1220.0, + 578.0, + 1220.0, + 578.0, + 1266.0, + 570.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1220.0, + 1139.0, + 1220.0, + 1139.0, + 1266.0, + 640.0, + 1266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 436.0, + 365.0, + 436.0, + 365.0, + 497.0, + 288.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 436.0, + 785.0, + 436.0, + 785.0, + 497.0, + 735.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 814.0, + 436.0, + 1083.0, + 436.0, + 1083.0, + 497.0, + 814.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 436.0, + 1409.0, + 436.0, + 1409.0, + 497.0, + 1401.0, + 497.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 477.0, + 443.0, + 477.0, + 443.0, + 519.0, + 293.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1872.0, + 553.0, + 1872.0, + 553.0, + 1914.0, + 291.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1872.0, + 976.0, + 1872.0, + 976.0, + 1914.0, + 711.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1161.0, + 1872.0, + 1407.0, + 1872.0, + 1407.0, + 1914.0, + 1161.0, + 1914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1903.0, + 412.0, + 1903.0, + 412.0, + 1945.0, + 292.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1792.0, + 559.0, + 1792.0, + 559.0, + 1840.0, + 293.0, + 1840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1721.0, + 1027.0, + 1721.0, + 1027.0, + 1770.0, + 292.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 1721.0, + 1206.0, + 1721.0, + 1206.0, + 1770.0, + 1088.0, + 1770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1266.0, + 1721.0, + 1281.0, + 1721.0, + 1281.0, + 1770.0, + 1266.0, + 1770.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 13, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1895, + 1404, + 1895, + 1404, + 2037, + 296, + 2037 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 297, + 1400, + 1403, + 1400, + 1403, + 1517, + 297, + 1517 + ], + "score": 0.959 + }, + { + "category_id": 1, + "poly": [ + 297, + 1175, + 1406, + 1175, + 1406, + 1266, + 297, + 1266 + ], + "score": 0.957 + }, + { + "category_id": 8, + "poly": [ + 460, + 808, + 1244, + 808, + 1244, + 992, + 460, + 992 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 689, + 1270, + 1010, + 1270, + 1010, + 1343, + 689, + 1343 + ], + "score": 0.952 + }, + { + "category_id": 8, + "poly": [ + 590, + 262, + 1111, + 262, + 1111, + 413, + 590, + 413 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 296, + 707, + 1407, + 707, + 1407, + 773, + 296, + 773 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 295, + 1005, + 1408, + 1005, + 1408, + 1074, + 295, + 1074 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 297, + 1631, + 1398, + 1631, + 1398, + 1695, + 297, + 1695 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 295, + 414, + 1404, + 414, + 1404, + 480, + 295, + 480 + ], + "score": 0.93 + }, + { + "category_id": 8, + "poly": [ + 422, + 1697, + 1279, + 1697, + 1279, + 1775, + 422, + 1775 + ], + "score": 0.927 + }, + { + "category_id": 8, + "poly": [ + 390, + 546, + 1544, + 546, + 1544, + 660, + 390, + 660 + ], + "score": 0.926 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 733, + 229, + 733, + 262, + 298, + 262 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 297, + 73, + 814, + 73, + 814, + 106, + 297, + 106 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 300, + 1342, + 778, + 1342, + 778, + 1385, + 300, + 1385 + ], + "score": 0.915 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1287, + 1401, + 1287, + 1401, + 1319, + 1351, + 1319 + ], + "score": 0.892 + }, + { + "category_id": 8, + "poly": [ + 300, + 1076, + 1507, + 1076, + 1507, + 1145, + 300, + 1145 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1351, + 322, + 1401, + 322, + 1401, + 353, + 1351, + 353 + ], + "score": 0.877 + }, + { + "category_id": 8, + "poly": [ + 490, + 1564, + 1212, + 1564, + 1212, + 1629, + 490, + 1629 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1719, + 1400, + 1719, + 1400, + 1751, + 1351, + 1751 + ], + "score": 0.872 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.867 + }, + { + "category_id": 9, + "poly": [ + 1351, + 886, + 1401, + 886, + 1401, + 917, + 1351, + 917 + ], + "score": 0.864 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1141, + 1401, + 1141, + 1401, + 1172, + 1351, + 1172 + ], + "score": 0.862 + }, + { + "category_id": 9, + "poly": [ + 1350, + 658, + 1402, + 658, + 1402, + 689, + 1350, + 689 + ], + "score": 0.836 + }, + { + "category_id": 1, + "poly": [ + 302, + 1528, + 1219, + 1528, + 1219, + 1561, + 302, + 1561 + ], + "score": 0.807 + }, + { + "category_id": 2, + "poly": [ + 1374, + 1786, + 1403, + 1786, + 1403, + 1816, + 1374, + 1816 + ], + "score": 0.802 + }, + { + "category_id": 1, + "poly": [ + 297, + 482, + 788, + 482, + 788, + 539, + 297, + 539 + ], + "score": 0.438 + }, + { + "category_id": 13, + "poly": [ + 1152, + 1398, + 1256, + 1398, + 1256, + 1436, + 1152, + 1436 + ], + "score": 0.95, + "latex": "\\hat { A } _ { C } ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 458, + 808, + 1241, + 808, + 1241, + 996, + 458, + 996 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } { \\pi _ { \\phi _ { n e w } } = \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { } \\\\ { s . t . } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . } & { V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\leq \\chi } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 969, + 1958, + 1240, + 1958, + 1240, + 2007, + 969, + 2007 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 418, + 1928, + 523, + 1928, + 523, + 1961, + 418, + 1961 + ], + "score": 0.93, + "latex": "Q _ { C } ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 422, + 1692, + 1277, + 1692, + 1277, + 1774, + 422, + 1774 + ], + "score": 0.93, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad \\mathrm { w h e r e } \\qquad \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 13, + "poly": [ + 491, + 1402, + 593, + 1402, + 593, + 1435, + 491, + 1435 + ], + "score": 0.92, + "latex": "A _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 944, + 413, + 1034, + 413, + 1034, + 451, + 944, + 451 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 648, + 1965, + 831, + 1965, + 831, + 2003, + 648, + 2003 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) < \\chi" + }, + { + "category_id": 13, + "poly": [ + 667, + 709, + 952, + 709, + 952, + 744, + 667, + 744 + ], + "score": 0.92, + "latex": "D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ]" + }, + { + "category_id": 14, + "poly": [ + 310, + 476, + 1428, + 476, + 1428, + 664, + 310, + 664 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } & { \\mathsf { \\Gamma } _ { \\hat { \\varphi } _ { n \\epsilon w } } = \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\bar { \\mathbb { E } } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } \\\\ & { \\quad s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi o l d } ( \\cdot \\vert s ) \\vert \\vert \\pi _ { \\phi } ( \\cdot \\vert s ) ) ] \\le \\delta } \\\\ & { \\quad s . t . \\quad V _ { G } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\displaystyle \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] + \\beta \\sqrt { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot \\vert s ) \\vert \\vert \\pi _ { \\phi } ( \\cdot \\vert s ) ) ] } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 590, + 260, + 1111, + 260, + 1111, + 415, + 590, + 415 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { \\displaystyle \\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { } \\\\ { s . t . \\quad } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . \\quad } & { V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1967, + 384, + 1967, + 384, + 2000, + 297, + 2000 + ], + "score": 0.9, + "latex": "\\epsilon _ { C } \\leq 2" + }, + { + "category_id": 13, + "poly": [ + 450, + 1176, + 1053, + 1176, + 1053, + 1231, + 450, + 1231 + ], + "score": 0.9, + "latex": "\\begin{array} { r } { V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\leq \\chi } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 691, + 1038, + 819, + 1038, + 819, + 1075, + 691, + 1075 + ], + "score": 0.89, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)" + }, + { + "category_id": 13, + "poly": [ + 1042, + 452, + 1102, + 452, + 1102, + 480, + 1042, + 480 + ], + "score": 0.89, + "latex": "\\pi _ { \\phi _ { o l d } }" + }, + { + "category_id": 13, + "poly": [ + 914, + 1010, + 983, + 1010, + 983, + 1041, + 914, + 1041 + ], + "score": 0.89, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 13, + "poly": [ + 572, + 1435, + 615, + 1435, + 615, + 1464, + 572, + 1464 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 795, + 1631, + 922, + 1631, + 922, + 1667, + 795, + 1667 + ], + "score": 0.89, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)" + }, + { + "category_id": 13, + "poly": [ + 362, + 1664, + 462, + 1664, + 462, + 1694, + 362, + 1694 + ], + "score": 0.88, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 775, + 1903, + 844, + 1903, + 844, + 1931, + 775, + 1931 + ], + "score": 0.88, + "latex": "\\pi _ { \\phi _ { n e w } }" + }, + { + "category_id": 14, + "poly": [ + 491, + 1561, + 1214, + 1561, + 1214, + 1631, + 491, + 1631 + ], + "score": 0.86, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 366, + 1902, + 400, + 1902, + 400, + 1928, + 366, + 1928 + ], + "score": 0.86, + "latex": "\\epsilon _ { C }" + }, + { + "category_id": 13, + "poly": [ + 587, + 1929, + 633, + 1929, + 633, + 1959, + 587, + 1959 + ], + "score": 0.85, + "latex": "\\leq 1" + }, + { + "category_id": 13, + "poly": [ + 1155, + 1932, + 1187, + 1932, + 1187, + 1958, + 1155, + 1958 + ], + "score": 0.83, + "latex": "\\epsilon _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1243, + 1474, + 1269, + 1474, + 1269, + 1501, + 1243, + 1501 + ], + "score": 0.83, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 482, + 1462, + 1157, + 1462, + 1157, + 1519, + 482, + 1519 + ], + "score": 0.83, + "latex": "\\begin{array} { r } { \\dot { \\mathbb { E } } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A _ { C } } ( s , a ) \\right] = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ( s , a ) \\right] + \\Delta } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 310, + 1072, + 1428, + 1072, + 1428, + 1144, + 310, + 1144 + ], + "score": 0.82, + "latex": "\\gamma _ { C } ^ { \\pi _ { \\partial u _ { c o u l } } } ( \\mu ) \\leq V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\Big ] + \\frac { 2 \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\sqrt { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\big [ \\mathrm { D } _ { K L } \\big ( \\pi _ { \\phi } \\big | \\big | \\pi _ { \\phi } \\big ) \\big | ^ { 2 } \\big ] } ." + }, + { + "category_id": 13, + "poly": [ + 752, + 1230, + 879, + 1230, + 879, + 1266, + 752, + 1266 + ], + "score": 0.82, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } \\left( \\mu \\right)" + }, + { + "category_id": 13, + "poly": [ + 1020, + 1434, + 1047, + 1434, + 1047, + 1460, + 1020, + 1460 + ], + "score": 0.81, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 464, + 2003, + 490, + 2003, + 490, + 2030, + 464, + 2030 + ], + "score": 0.81, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 365, + 1342, + 769, + 1342, + 769, + 1384, + 365, + 1384 + ], + "score": 0.74, + "latex": "\\begin{array} { r } { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) | } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1119, + 1531, + 1218, + 1531, + 1218, + 1560, + 1119, + 1560 + ], + "score": 0.63, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 14, + "poly": [ + 364, + 1266, + 1014, + 1266, + 1014, + 1388, + 364, + 1388 + ], + "score": 0.39, + "latex": "\\begin{array} { c } { { V _ { C } ^ { \\pi _ { \\theta n e w } } ( \\mu ) \\leq \\chi + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } } } \\\\ { { \\epsilon _ { C } = \\operatorname* { m a x } _ { s } | \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { n e w } } } A _ { C } ^ { \\pi _ { \\phi _ { n d } } } ( s , a ) | . } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 687, + 1265, + 1012, + 1265, + 1012, + 1341, + 687, + 1341 + ], + "score": 0.39, + "latex": "V _ { C } ^ { { \\pi } _ { \\phi n e w } } ( \\mu ) \\leq \\chi + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 71.0, + 815.0, + 71.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1377.0, + 1790.0, + 1403.0, + 1790.0, + 1403.0, + 1818.0, + 1377.0, + 1818.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1895.0, + 365.0, + 1895.0, + 365.0, + 1934.0, + 295.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 401.0, + 1895.0, + 774.0, + 1895.0, + 774.0, + 1934.0, + 401.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 845.0, + 1895.0, + 1404.0, + 1895.0, + 1404.0, + 1934.0, + 845.0, + 1934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1927.0, + 417.0, + 1927.0, + 417.0, + 1962.0, + 296.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1927.0, + 586.0, + 1927.0, + 586.0, + 1962.0, + 524.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 634.0, + 1927.0, + 1154.0, + 1927.0, + 1154.0, + 1962.0, + 634.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1927.0, + 1404.0, + 1927.0, + 1404.0, + 1962.0, + 1188.0, + 1962.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1950.0, + 296.0, + 1950.0, + 296.0, + 2013.0, + 288.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 385.0, + 1950.0, + 647.0, + 1950.0, + 647.0, + 2013.0, + 385.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 1950.0, + 968.0, + 1950.0, + 968.0, + 2013.0, + 832.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 1950.0, + 1412.0, + 1950.0, + 1412.0, + 2013.0, + 1241.0, + 2013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2000.0, + 463.0, + 2000.0, + 463.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 491.0, + 2000.0, + 1405.0, + 2000.0, + 1405.0, + 2039.0, + 491.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1045.0, + 1988.5, + 1089.0, + 1988.5, + 1089.0, + 2005.5, + 1045.0, + 2005.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1397.0, + 490.0, + 1397.0, + 490.0, + 1440.0, + 293.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 1397.0, + 1151.0, + 1397.0, + 1151.0, + 1440.0, + 594.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1257.0, + 1397.0, + 1404.0, + 1397.0, + 1404.0, + 1440.0, + 1257.0, + 1440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1429.0, + 571.0, + 1429.0, + 571.0, + 1471.0, + 293.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 616.0, + 1429.0, + 1019.0, + 1429.0, + 1019.0, + 1471.0, + 616.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1048.0, + 1429.0, + 1404.0, + 1429.0, + 1404.0, + 1471.0, + 1048.0, + 1471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1465.0, + 481.0, + 1465.0, + 481.0, + 1521.0, + 290.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1158.0, + 1465.0, + 1242.0, + 1465.0, + 1242.0, + 1521.0, + 1158.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1270.0, + 1465.0, + 1403.0, + 1465.0, + 1403.0, + 1521.0, + 1270.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1173.0, + 449.0, + 1173.0, + 449.0, + 1231.0, + 285.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 1173.0, + 1410.0, + 1173.0, + 1410.0, + 1231.0, + 1054.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1214.0, + 751.0, + 1214.0, + 751.0, + 1276.0, + 290.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 880.0, + 1214.0, + 889.0, + 1214.0, + 889.0, + 1276.0, + 880.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 706.0, + 666.0, + 706.0, + 666.0, + 746.0, + 294.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 706.0, + 1408.0, + 706.0, + 1408.0, + 746.0, + 953.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 740.0, + 1130.0, + 740.0, + 1130.0, + 772.0, + 295.0, + 772.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1001.0, + 913.0, + 1001.0, + 913.0, + 1050.0, + 293.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 984.0, + 1001.0, + 1409.0, + 1001.0, + 1409.0, + 1050.0, + 984.0, + 1050.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1025.0, + 690.0, + 1025.0, + 690.0, + 1085.0, + 288.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 1025.0, + 828.0, + 1025.0, + 828.0, + 1085.0, + 820.0, + 1085.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1623.0, + 794.0, + 1623.0, + 794.0, + 1671.0, + 289.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 1623.0, + 1407.0, + 1623.0, + 1407.0, + 1671.0, + 923.0, + 1671.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1656.0, + 361.0, + 1656.0, + 361.0, + 1700.0, + 294.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 463.0, + 1656.0, + 467.0, + 1656.0, + 467.0, + 1700.0, + 463.0, + 1700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 409.0, + 943.0, + 409.0, + 943.0, + 453.0, + 293.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 409.0, + 1407.0, + 409.0, + 1407.0, + 453.0, + 1035.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 444.0, + 1041.0, + 444.0, + 1041.0, + 485.0, + 294.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1103.0, + 444.0, + 1319.0, + 444.0, + 1319.0, + 485.0, + 1103.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 225.0, + 738.0, + 225.0, + 738.0, + 269.0, + 293.0, + 269.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1329.0, + 363.0, + 1329.0, + 363.0, + 1395.0, + 292.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1525.0, + 1118.0, + 1525.0, + 1118.0, + 1566.0, + 293.0, + 1566.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 268.0, + 473.0, + 309.0, + 473.0, + 309.0, + 556.0, + 268.0, + 556.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 14, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 299, + 1290, + 1402, + 1290, + 1402, + 1402, + 299, + 1402 + ], + "score": 0.972 + }, + { + "category_id": 1, + "poly": [ + 297, + 345, + 1408, + 345, + 1408, + 486, + 297, + 486 + ], + "score": 0.968 + }, + { + "category_id": 1, + "poly": [ + 297, + 1032, + 1404, + 1032, + 1404, + 1148, + 297, + 1148 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 538, + 1405, + 1160, + 1405, + 1160, + 1659, + 538, + 1659 + ], + "score": 0.963 + }, + { + "category_id": 8, + "poly": [ + 296, + 1705, + 1492, + 1705, + 1492, + 2008, + 296, + 2008 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 300, + 722, + 1404, + 722, + 1404, + 817, + 300, + 817 + ], + "score": 0.958 + }, + { + "category_id": 1, + "poly": [ + 299, + 828, + 1403, + 828, + 1403, + 928, + 299, + 928 + ], + "score": 0.956 + }, + { + "category_id": 8, + "poly": [ + 295, + 490, + 1426, + 490, + 1426, + 611, + 295, + 611 + ], + "score": 0.931 + }, + { + "category_id": 8, + "poly": [ + 314, + 933, + 1382, + 933, + 1382, + 1010, + 314, + 1010 + ], + "score": 0.929 + }, + { + "category_id": 1, + "poly": [ + 300, + 639, + 804, + 639, + 804, + 672, + 300, + 672 + ], + "score": 0.921 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 316, + 255, + 1328, + 255, + 1328, + 343, + 316, + 343 + ], + "score": 0.917 + }, + { + "category_id": 8, + "poly": [ + 756, + 671, + 943, + 671, + 943, + 711, + 756, + 711 + ], + "score": 0.916 + }, + { + "category_id": 1, + "poly": [ + 296, + 1666, + 926, + 1666, + 926, + 1702, + 296, + 1702 + ], + "score": 0.905 + }, + { + "category_id": 9, + "poly": [ + 1351, + 673, + 1401, + 673, + 1401, + 704, + 1351, + 704 + ], + "score": 0.889 + }, + { + "category_id": 8, + "poly": [ + 307, + 1151, + 1389, + 1151, + 1389, + 1231, + 307, + 1231 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1350, + 608, + 1402, + 608, + 1402, + 639, + 1350, + 639 + ], + "score": 0.866 + }, + { + "category_id": 9, + "poly": [ + 1351, + 286, + 1401, + 286, + 1401, + 315, + 1351, + 315 + ], + "score": 0.862 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1518, + 1402, + 1518, + 1402, + 1549, + 1350, + 1549 + ], + "score": 0.862 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 866, + 2087, + 866, + 2114, + 836, + 2114 + ], + "score": 0.851 + }, + { + "category_id": 9, + "poly": [ + 1351, + 2008, + 1402, + 2008, + 1402, + 2038, + 1351, + 2038 + ], + "score": 0.844 + }, + { + "category_id": 1, + "poly": [ + 296, + 235, + 328, + 235, + 328, + 260, + 296, + 260 + ], + "score": 0.827 + }, + { + "category_id": 14, + "poly": [ + 538, + 1403, + 1158, + 1403, + 1158, + 1666, + 538, + 1666 + ], + "score": 0.94, + "latex": "\\begin{array} { l } { \\log Z _ { t } ( s ) = \\log \\displaystyle \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\exp \\left( \\eta A ^ { ( t ) } / ( 1 - \\gamma ) \\right) } \\\\ { \\mathrm { ~ } \\ge \\displaystyle \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\log \\exp \\eta A ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } \\\\ { \\displaystyle = \\frac { \\eta } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A ^ { ( t ) } ( s , a ) } \\\\ { \\displaystyle = 0 } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 530, + 1102, + 710, + 1102, + 710, + 1152, + 530, + 1152 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + \\zeta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 742, + 1329, + 826, + 1329, + 826, + 1363, + 742, + 1363 + ], + "score": 0.94, + "latex": "A ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1203, + 405, + 1393, + 405, + 1393, + 462, + 1203, + 462 + ], + "score": 0.93, + "latex": "\\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 13, + "poly": [ + 735, + 374, + 848, + 374, + 848, + 408, + 735, + 408 + ], + "score": 0.93, + "latex": "T ( s ^ { \\prime } | s , a )" + }, + { + "category_id": 14, + "poly": [ + 320, + 255, + 1329, + 255, + 1329, + 347, + 320, + 347 + ], + "score": 0.93, + "latex": "\\alpha > \\frac { \\mathcal { G } _ { c , T } } { 1 - \\gamma } \\cdot \\operatorname* { m a x } _ { s \\sim \\rho _ { \\phi _ { o l d ^ { \\prime } } } } \\left( \\frac { 1 } { \\vert \\sqrt { \\mathcal { D } _ { \\phi _ { o l d ^ { \\prime } } } } \\vert } + \\frac { 2 \\sqrt { 2 \\delta } \\gamma + ( 1 - \\gamma ) ^ { 2 } \\zeta } { \\mathcal { G } _ { c , T } } \\right) \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left( \\frac { \\pi _ { \\phi _ { o l d } } } { \\pi _ { \\phi _ { o l d ^ { \\prime } } } } - 1 \\right) \\right] ^ { - 1 }" + }, + { + "category_id": 13, + "poly": [ + 697, + 862, + 803, + 862, + 803, + 897, + 697, + 897 + ], + "score": 0.93, + "latex": "\\phi ^ { ( 0 ) } = 0" + }, + { + "category_id": 14, + "poly": [ + 310, + 483, + 1427, + 483, + 1427, + 612, + 310, + 612 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { v > \\displaystyle \\frac { \\mathcal { G } _ { c , T } } { 1 - \\gamma } \\cdot \\displaystyle \\operatorname* { m a x } _ { s \\sim \\rho \\phi _ { o l d ^ { \\prime } } } \\left( \\frac { 1 } { | \\sqrt { \\mathcal { D } _ { \\phi _ { o l d ^ { \\prime } } } } | } + \\frac { 2 \\sqrt { 2 \\delta } \\gamma + ( 1 - \\gamma ) ^ { 2 } \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | } } { \\mathcal { G } _ { c , T } } \\right) \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d ^ { \\prime } } } } \\left( \\frac { \\pi _ { \\phi _ { o l d } } } { \\pi _ { \\phi _ { o l d ^ { \\prime } } } } - 1 \\right) \\right] ^ { - 1 } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1099, + 1360, + 1149, + 1360, + 1149, + 1403, + 1099, + 1403 + ], + "score": 0.93, + "latex": "{ A } _ { R } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 337, + 1064, + 463, + 1064, + 463, + 1107, + 337, + 1107 + ], + "score": 0.92, + "latex": "\\textstyle \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) }" + }, + { + "category_id": 14, + "poly": [ + 310, + 1703, + 1428, + 1703, + 1428, + 2015, + 310, + 2015 + ], + "score": 0.92, + "latex": "\\begin{array} { r l } { { \\operatorname { s g } \\hat { Z } _ { t } ( s ) = \\log \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\exp ( \\eta \\hat { A } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } } \\\\ & { = \\log \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\exp ( \\eta ( A _ { R } ^ { ( t ) } ( s , a ) - \\lambda ^ { ( t ) } A _ { C } ( s , a ) ) / ( 1 - \\gamma ) ) } \\\\ & { \\geq \\displaystyle \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\log \\exp ( \\eta A _ { R } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) - \\lambda ^ { ( t ) } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) \\log \\exp ( \\eta A _ { C } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } \\\\ & { = 0 - \\displaystyle \\frac { \\lambda ^ { ( t ) } \\eta } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1360, + 588, + 1360, + 588, + 1403, + 297, + 1403 + ], + "score": 0.92, + "latex": "A _ { R } ^ { ( t ) } ( s , a ) - \\lambda ^ { ( t ) } A _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1260, + 1324, + 1406, + 1324, + 1406, + 1363, + 1260, + 1363 + ], + "score": 0.92, + "latex": "\\hat { A } ^ { ( t ) } ( s , a ) =" + }, + { + "category_id": 13, + "poly": [ + 298, + 374, + 383, + 374, + 383, + 408, + 298, + 408 + ], + "score": 0.92, + "latex": "C ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 746, + 1367, + 853, + 1367, + 853, + 1402, + 746, + 1402 + ], + "score": 0.91, + "latex": "\\log Z _ { t } ( s )" + }, + { + "category_id": 14, + "poly": [ + 319, + 926, + 1377, + 926, + 1377, + 1009, + 319, + 1009 + ], + "score": 0.91, + "latex": "V _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + \\left( ( 1 - \\chi ) + \\left( 1 - \\frac { 2 \\Delta } { ( 1 - \\gamma ) } \\right) + 2 \\zeta \\right) \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T }" + }, + { + "category_id": 13, + "poly": [ + 703, + 423, + 752, + 423, + 752, + 454, + 703, + 454 + ], + "score": 0.91, + "latex": "\\phi _ { o l d }" + }, + { + "category_id": 14, + "poly": [ + 757, + 669, + 944, + 669, + 944, + 711, + 757, + 711 + ], + "score": 0.9, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi" + }, + { + "category_id": 14, + "poly": [ + 311, + 1149, + 1392, + 1149, + 1392, + 1231, + 311, + 1231 + ], + "score": 0.9, + "latex": "V _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\mathcal { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + K \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T } \\quad \\mathrm { w h e r e } \\quad K \\leq ( 1 - \\chi ) + \\frac { 4 \\sqrt { 2 \\delta } \\gamma } { ( 1 - \\gamma ) ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 991, + 1112, + 1090, + 1112, + 1090, + 1143, + 991, + 1143 + ], + "score": 0.9, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 1022, + 866, + 1098, + 866, + 1098, + 898, + 1022, + 898 + ], + "score": 0.89, + "latex": "\\eta > 0 _ { : }" + }, + { + "category_id": 13, + "poly": [ + 410, + 897, + 484, + 897, + 484, + 925, + 410, + 925 + ], + "score": 0.89, + "latex": "T > 0" + }, + { + "category_id": 13, + "poly": [ + 545, + 642, + 644, + 642, + 644, + 671, + 545, + 671 + ], + "score": 0.89, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 874, + 1665, + 923, + 1665, + 923, + 1697, + 874, + 1697 + ], + "score": 0.89, + "latex": "\\hat { A } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 365, + 345, + 419, + 345, + 419, + 377, + 365, + 377 + ], + "score": 0.89, + "latex": "\\mathcal { G } _ { c , T }" + }, + { + "category_id": 13, + "poly": [ + 1086, + 375, + 1143, + 375, + 1143, + 407, + 1086, + 407 + ], + "score": 0.89, + "latex": "\\phi _ { o l d ^ { \\prime } }" + }, + { + "category_id": 13, + "poly": [ + 1010, + 422, + 1098, + 422, + 1098, + 453, + 1010, + 453 + ], + "score": 0.89, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 710, + 1292, + 740, + 1292, + 740, + 1322, + 710, + 1322 + ], + "score": 0.86, + "latex": "t ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 1240, + 836, + 1273, + 836, + 1273, + 865, + 1240, + 865 + ], + "score": 0.84, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 774, + 899, + 871, + 899, + 871, + 928, + 774, + 928 + ], + "score": 0.82, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 699, + 1036, + 719, + 1036, + 719, + 1062, + 699, + 1062 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 364, + 462, + 386, + 462, + 386, + 485, + 364, + 485 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1346, + 837, + 1367, + 837, + 1367, + 863, + 1346, + 863 + ], + "score": 0.78, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 553, + 1294, + 586, + 1294, + 586, + 1327, + 553, + 1327 + ], + "score": 0.76, + "latex": "( t )" + }, + { + "category_id": 13, + "poly": [ + 439, + 427, + 460, + 427, + 460, + 449, + 439, + 449 + ], + "score": 0.76, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 987, + 1076, + 1009, + 1076, + 1009, + 1098, + 987, + 1098 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 298, + 239, + 318, + 239, + 318, + 258, + 298, + 258 + ], + "score": 0.52, + "latex": "\\alpha" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1290.0, + 552.0, + 1290.0, + 552.0, + 1327.0, + 295.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 587.0, + 1290.0, + 709.0, + 1290.0, + 709.0, + 1327.0, + 587.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 1290.0, + 1404.0, + 1290.0, + 1404.0, + 1327.0, + 741.0, + 1327.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1325.0, + 741.0, + 1325.0, + 741.0, + 1366.0, + 293.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 827.0, + 1325.0, + 1259.0, + 1325.0, + 1259.0, + 1366.0, + 827.0, + 1366.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1348.0, + 296.0, + 1348.0, + 296.0, + 1420.0, + 283.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 589.0, + 1348.0, + 745.0, + 1348.0, + 745.0, + 1420.0, + 589.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1348.0, + 1098.0, + 1348.0, + 1098.0, + 1420.0, + 854.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1150.0, + 1348.0, + 1183.0, + 1348.0, + 1183.0, + 1420.0, + 1150.0, + 1420.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 339.0, + 364.0, + 339.0, + 364.0, + 379.0, + 292.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 339.0, + 1408.0, + 339.0, + 1408.0, + 379.0, + 420.0, + 379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 374.0, + 734.0, + 374.0, + 734.0, + 409.0, + 384.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 849.0, + 374.0, + 1085.0, + 374.0, + 1085.0, + 409.0, + 849.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 374.0, + 1405.0, + 374.0, + 1405.0, + 409.0, + 1144.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 404.0, + 438.0, + 404.0, + 438.0, + 468.0, + 291.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 461.0, + 404.0, + 702.0, + 404.0, + 702.0, + 468.0, + 461.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 753.0, + 404.0, + 1009.0, + 404.0, + 1009.0, + 468.0, + 753.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 404.0, + 1202.0, + 404.0, + 1202.0, + 468.0, + 1099.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1394.0, + 404.0, + 1400.0, + 404.0, + 1400.0, + 468.0, + 1394.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 458.0, + 363.0, + 458.0, + 363.0, + 490.0, + 295.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 387.0, + 458.0, + 618.0, + 458.0, + 618.0, + 490.0, + 387.0, + 490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1033.0, + 336.0, + 1033.0, + 336.0, + 1122.0, + 285.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 464.0, + 1033.0, + 698.0, + 1033.0, + 698.0, + 1122.0, + 464.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 1033.0, + 986.0, + 1033.0, + 986.0, + 1122.0, + 720.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 1033.0, + 1418.0, + 1033.0, + 1418.0, + 1122.0, + 1010.0, + 1122.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1098.0, + 529.0, + 1098.0, + 529.0, + 1148.0, + 293.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1098.0, + 990.0, + 1098.0, + 990.0, + 1148.0, + 711.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1098.0, + 1315.0, + 1098.0, + 1315.0, + 1148.0, + 1091.0, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 722.0, + 1405.0, + 722.0, + 1405.0, + 760.0, + 294.0, + 760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 755.0, + 1405.0, + 755.0, + 1405.0, + 789.0, + 295.0, + 789.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 788.0, + 363.0, + 788.0, + 363.0, + 817.0, + 294.0, + 817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 828.0, + 1239.0, + 828.0, + 1239.0, + 867.0, + 295.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1274.0, + 828.0, + 1345.0, + 828.0, + 1345.0, + 867.0, + 1274.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1368.0, + 828.0, + 1404.0, + 828.0, + 1404.0, + 867.0, + 1368.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 858.0, + 696.0, + 858.0, + 696.0, + 904.0, + 292.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 804.0, + 858.0, + 1021.0, + 858.0, + 1021.0, + 904.0, + 804.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1099.0, + 858.0, + 1408.0, + 858.0, + 1408.0, + 904.0, + 1099.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 890.0, + 409.0, + 890.0, + 409.0, + 936.0, + 293.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 890.0, + 773.0, + 890.0, + 773.0, + 936.0, + 485.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 872.0, + 890.0, + 885.0, + 890.0, + 885.0, + 936.0, + 872.0, + 936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 634.0, + 544.0, + 634.0, + 544.0, + 679.0, + 294.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 645.0, + 634.0, + 811.0, + 634.0, + 811.0, + 679.0, + 645.0, + 679.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1665.0, + 873.0, + 1665.0, + 873.0, + 1706.0, + 293.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 924.0, + 1665.0, + 927.0, + 1665.0, + 927.0, + 1706.0, + 924.0, + 1706.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 235.0, + 297.0, + 235.0, + 297.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 235.0, + 331.0, + 235.0, + 331.0, + 266.0, + 319.0, + 266.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 15, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 564, + 1404, + 564, + 1404, + 712, + 297, + 712 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 296, + 410, + 1406, + 410, + 1406, + 552, + 296, + 552 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 334, + 746, + 1313, + 746, + 1313, + 1469, + 334, + 1469 + ], + "score": 0.967 + }, + { + "category_id": 1, + "poly": [ + 300, + 1471, + 1402, + 1471, + 1402, + 1563, + 300, + 1563 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 293, + 1577, + 1401, + 1577, + 1401, + 1642, + 293, + 1642 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 296, + 229, + 1403, + 229, + 1403, + 326, + 296, + 326 + ], + "score": 0.941 + }, + { + "category_id": 2, + "poly": [ + 299, + 74, + 813, + 74, + 813, + 105, + 299, + 105 + ], + "score": 0.914 + }, + { + "category_id": 8, + "poly": [ + 362, + 330, + 1334, + 330, + 1334, + 406, + 362, + 406 + ], + "score": 0.907 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1095, + 1402, + 1095, + 1402, + 1127, + 1351, + 1127 + ], + "score": 0.883 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2112, + 836, + 2112 + ], + "score": 0.852 + }, + { + "category_id": 14, + "poly": [ + 330, + 760, + 1316, + 760, + 1316, + 1468, + 330, + 1468 + ], + "score": 0.94, + "latex": "\\begin{array} { r l } { \\| \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) - \\frac { 1 } { N _ { 1 } } \\sum _ { k = 0 } ^ { N } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } } & { = } \\\\ & { - 1 - \\sum _ { k = 0 } ^ { N } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } } \\\\ & { - \\frac { 1 } { N _ { 1 } ^ { 2 } } \\sum _ { k = 0 } ^ { N } \\exp \\sum _ { j } ^ { k + 1 } \\exp \\frac { \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } { N _ { 1 } ^ { 2 } } \\| \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } \\| _ { \\alpha } ^ { \\prime } \\| _ { \\alpha } ^ { \\prime } } \\\\ & \\phantom { = } \\frac { 1 } { N _ { 1 } } \\sum _ { k = 0 } ^ { N } \\exp \\frac { - \\sum _ { j } ^ { k + 1 } \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\mathrm { H G } } ( \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } , \\hat { \\mathbf { r } } _ { \\alpha } ^ { \\prime } ) } \\\\ & = \\frac { 1 } { N _ { 1 } } \\frac { 1 } { N _ { 1 } ^ { 2 } } \\exp \\sum _ { j } ^ { k + 1 } \\exp ( - \\sum _ { j } ^ \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 616, + 596, + 728, + 596, + 728, + 639, + 616, + 639 + ], + "score": 0.93, + "latex": "A _ { R } ^ { ( t ) } ( \\bar { s } , a )" + }, + { + "category_id": 13, + "poly": [ + 1281, + 482, + 1402, + 482, + 1402, + 518, + 1281, + 518 + ], + "score": 0.93, + "latex": "G _ { t } ( s ) \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 297, + 635, + 734, + 635, + 734, + 679, + 297, + 679 + ], + "score": 0.93, + "latex": "\\hat { A } ^ { ( t ) } ( s , a ) = A _ { R } ^ { ( t ) } ( s , a ) - \\lambda ^ { ( t ) } \\ddot { A } _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 449, + 1470, + 576, + 1470, + 576, + 1504, + 449, + 1504 + ], + "score": 0.92, + "latex": "G _ { t } ( s ) \\geq 0" + }, + { + "category_id": 13, + "poly": [ + 661, + 291, + 711, + 291, + 711, + 323, + 661, + 323 + ], + "score": 0.92, + "latex": "\\hat { A } ^ { ( t ) }" + }, + { + "category_id": 13, + "poly": [ + 615, + 474, + 1068, + 474, + 1068, + 525, + 615, + 525 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { \\log \\hat { Z } _ { t } ( s ) \\dot { + } \\frac { \\lambda ^ { ( t ) } \\eta } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1100, + 482, + 1168, + 482, + 1168, + 518, + 1100, + 518 + ], + "score": 0.91, + "latex": "G _ { t } ( s )" + }, + { + "category_id": 13, + "poly": [ + 1356, + 639, + 1399, + 639, + 1399, + 672, + 1356, + 672 + ], + "score": 0.88, + "latex": "\\boldsymbol d ^ { ( t ) }" + }, + { + "category_id": 14, + "poly": [ + 364, + 326, + 1334, + 326, + 1334, + 405, + 364, + 405 + ], + "score": 0.88, + "latex": "\\phi ^ { ( t + 1 ) } = \\phi ^ { ( t ) } + \\frac { \\eta } { 1 - \\gamma } \\hat { A } ^ { ( t ) } \\quad \\mathrm { a n d } \\quad \\pi ^ { ( t + 1 ) } ( a | s ) = \\pi ^ { ( t ) } ( a | s ) \\frac { \\exp \\bigl ( \\eta \\hat { A } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) \\bigr ) } { \\hat { Z } _ { t } ( s ) } ," + }, + { + "category_id": 13, + "poly": [ + 1065, + 677, + 1095, + 677, + 1095, + 706, + 1065, + 706 + ], + "score": 0.88, + "latex": "t ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 13, + "poly": [ + 1231, + 1500, + 1261, + 1500, + 1261, + 1528, + 1231, + 1528 + ], + "score": 0.86, + "latex": "t ^ { \\mathrm { t h } }" + }, + { + "category_id": 13, + "poly": [ + 931, + 1506, + 949, + 1506, + 949, + 1533, + 931, + 1533 + ], + "score": 0.83, + "latex": "\\rho" + }, + { + "category_id": 13, + "poly": [ + 788, + 648, + 810, + 648, + 810, + 677, + 788, + 677 + ], + "score": 0.81, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 366, + 410, + 935, + 410, + 935, + 450, + 366, + 450 + ], + "score": 0.79, + "latex": "\\begin{array} { r } { \\hat { Z } _ { t } ( s ) = \\sum _ { a \\in \\mathcal { A } } \\pi ^ { ( t ) } ( a | s ) \\exp ( \\eta \\hat { A } ^ { ( t ) } ( s , a ) / ( 1 - \\gamma ) ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 980, + 450, + 998, + 450, + 998, + 476, + 980, + 476 + ], + "score": 0.79, + "latex": "\\eta" + }, + { + "category_id": 13, + "poly": [ + 972, + 683, + 993, + 683, + 993, + 706, + 972, + 706 + ], + "score": 0.77, + "latex": "\\pi" + }, + { + "category_id": 13, + "poly": [ + 995, + 568, + 1022, + 568, + 1022, + 595, + 995, + 595 + ], + "score": 0.51, + "latex": "\\&" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 562.0, + 994.0, + 562.0, + 994.0, + 602.0, + 295.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1023.0, + 562.0, + 1405.0, + 562.0, + 1405.0, + 602.0, + 1023.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 590.0, + 615.0, + 590.0, + 615.0, + 651.0, + 288.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 729.0, + 590.0, + 1411.0, + 590.0, + 1411.0, + 651.0, + 729.0, + 651.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 735.0, + 632.0, + 787.0, + 632.0, + 787.0, + 684.0, + 735.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 632.0, + 1355.0, + 632.0, + 1355.0, + 684.0, + 811.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1400.0, + 632.0, + 1407.0, + 632.0, + 1407.0, + 684.0, + 1400.0, + 684.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 674.0, + 971.0, + 674.0, + 971.0, + 714.0, + 294.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 674.0, + 1064.0, + 674.0, + 1064.0, + 714.0, + 994.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1096.0, + 674.0, + 1206.0, + 674.0, + 1206.0, + 714.0, + 1096.0, + 714.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 407.0, + 365.0, + 407.0, + 365.0, + 453.0, + 291.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 407.0, + 1407.0, + 407.0, + 1407.0, + 453.0, + 936.0, + 453.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 443.0, + 979.0, + 443.0, + 979.0, + 478.0, + 295.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 999.0, + 443.0, + 1406.0, + 443.0, + 1406.0, + 478.0, + 999.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 468.0, + 614.0, + 468.0, + 614.0, + 533.0, + 290.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1069.0, + 468.0, + 1099.0, + 468.0, + 1099.0, + 533.0, + 1069.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1169.0, + 468.0, + 1280.0, + 468.0, + 1280.0, + 533.0, + 1169.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 468.0, + 1414.0, + 468.0, + 1414.0, + 533.0, + 1403.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 518.0, + 501.0, + 518.0, + 501.0, + 554.0, + 295.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 448.0, + 1468.0, + 448.0, + 1506.0, + 295.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 577.0, + 1468.0, + 1404.0, + 1468.0, + 1404.0, + 1506.0, + 577.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1500.0, + 930.0, + 1500.0, + 930.0, + 1533.0, + 294.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 950.0, + 1500.0, + 1230.0, + 1500.0, + 1230.0, + 1533.0, + 950.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1500.0, + 1404.0, + 1500.0, + 1404.0, + 1533.0, + 1262.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1529.0, + 598.0, + 1529.0, + 598.0, + 1565.0, + 297.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1575.0, + 1404.0, + 1575.0, + 1404.0, + 1615.0, + 295.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1607.0, + 651.0, + 1607.0, + 651.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 228.0, + 1404.0, + 228.0, + 1404.0, + 266.0, + 294.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 262.0, + 1404.0, + 262.0, + 1404.0, + 295.0, + 292.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 291.0, + 660.0, + 291.0, + 660.0, + 329.0, + 294.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 291.0, + 907.0, + 291.0, + 907.0, + 329.0, + 712.0, + 329.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 16, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1857, + 1408, + 1857, + 1408, + 1986, + 296, + 1986 + ], + "score": 0.971 + }, + { + "category_id": 8, + "poly": [ + 300, + 250, + 1565, + 250, + 1565, + 948, + 300, + 948 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 297, + 1203, + 1429, + 1203, + 1429, + 1847, + 297, + 1847 + ], + "score": 0.966 + }, + { + "category_id": 8, + "poly": [ + 546, + 1025, + 1153, + 1025, + 1153, + 1090, + 546, + 1090 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 300, + 990, + 1083, + 990, + 1083, + 1023, + 300, + 1023 + ], + "score": 0.928 + }, + { + "category_id": 1, + "poly": [ + 299, + 1134, + 1402, + 1134, + 1402, + 1202, + 299, + 1202 + ], + "score": 0.927 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 299, + 1090, + 789, + 1090, + 789, + 1123, + 299, + 1123 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.863 + }, + { + "category_id": 9, + "poly": [ + 1351, + 941, + 1401, + 941, + 1401, + 972, + 1351, + 972 + ], + "score": 0.701 + }, + { + "category_id": 14, + "poly": [ + 385, + 256, + 1503, + 256, + 1503, + 951, + 385, + 951 + ], + "score": 0.98, + "latex": "\\begin{array} { r l } & { - \\nabla _ { x } ^ { ( k ) } \\cdot \\nabla _ { x } ^ { ( k ) } = \\frac { 1 } { 1 - \\sigma ^ { 2 } } \\mathrm { E } _ { x - \\cdot } \\sum _ { x = \\cdot } \\sum _ { x = \\cdot } \\nabla _ { x } ^ { \\cdot \\gamma } \\cdot \\sigma ( x ) \\hat { A } ^ { ( k ) } \\cdot \\nabla _ { x } \\cdot \\sigma _ { x } + x ^ { ( k ) } \\hat { A } ^ { ( k ) } \\cdot \\sigma _ { x } \\otimes x ) } \\\\ & { = - \\frac { 1 } { \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma _ { x } | \\hat { A } | \\cdot \\frac { \\sigma _ { x } ^ { ( k ) } } { \\sigma _ { x } ^ { ( k ) } } \\hat { A } | \\hat { A } | \\hat { A } | \\cdot } \\\\ & { - \\frac { 1 } { \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots ( \\int _ { 0 } \\mathrm { R } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\rangle - \\frac { 1 } { \\sigma _ { x } \\otimes x } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\rangle + \\frac { 1 } { \\sigma _ { x } } \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\cdot \\hat { A } | \\sigma _ { x } ^ { ( k ) } \\hat { A } | ) } \\\\ & { + \\frac { 1 } { 1 - \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\hat { A } | \\hat { A } | \\cdot \\sigma _ { x } \\otimes x ^ { ( k ) } \\hat { A } | \\cdot } \\\\ & - \\frac { 1 } { \\sigma ^ { 2 } } \\cdot \\sigma _ { x } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\cdot \\nabla _ { x } ^ { ( k ) } \\hat { A } | \\hat { A } | \\cdot \\sigma _ { x } \\otimes x ^ { ( k ) } \\rangle + \\log _ { 2 } ^ { \\cdot ( k ) } ( \\int _ { 0 } \\mathrm { R } \\langle \\sigma | \\hat { A } | \\cdot \\hat { \\sigma } \\otimes x ^ { ( k ) } \\rangle - \\frac { 1 } { \\sigma _ { x } } \\otimes \\cdots \\sum _ { x = \\cdot } ^ { \\cdot \\gamma + 1 } \\langle \\sigma | \\hat { A } | \\hat \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 689, + 1136, + 932, + 1136, + 932, + 1175, + 689, + 1175 + ], + "score": 0.93, + "latex": "V ^ { ( t + 1 ) } ( \\rho ) \\geq V ^ { ( t ) } ( \\rho )" + }, + { + "category_id": 13, + "poly": [ + 978, + 1136, + 1217, + 1136, + 1217, + 1175, + 978, + 1175 + ], + "score": 0.93, + "latex": "V ^ { ( t + 1 ) } ( s ) \\geq V ^ { ( t ) } ( s )" + }, + { + "category_id": 13, + "poly": [ + 1212, + 1894, + 1400, + 1894, + 1400, + 1940, + 1212, + 1940 + ], + "score": 0.92, + "latex": "\\begin{array} { r } { ( \\chi + \\zeta - \\frac { \\Delta } { ( 1 - \\gamma ) } ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 281, + 1196, + 1391, + 1196, + 1391, + 1847, + 281, + 1847 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { V _ { \\mathrm { G } } ^ { - 1 } ( s ) - V _ { \\mathrm { G } } ^ { \\mathrm { G } } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } \\sum _ { \\ell = 1 } ^ { 1 } ( V _ { \\mathrm { G } } ^ { - 1 } ( \\beta ) - V _ { \\mathrm { G } } ^ { \\mathrm { G } ( 1 ) } ( s ) ) } \\\\ { \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq ( \\operatorname { B a } ( s ) + \\operatorname { B a } ( s ) ) ^ { n } \\setminus S _ { \\mathrm { G } } ( s ) ( s ) + \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) } \\\\ { \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } } ^ { - 1 } ( s ) \\leq ( \\operatorname { B a } ( s ) + \\frac { \\lambda ^ { n } } { T } ) \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } } ^ { \\mathrm { G } ( 1 ) } ( s ) + \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } } ^ { \\mathrm { G } ( 1 ) } ( s ) \\leq \\frac { 1 } { T } } \\\\ { \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { - 1 } ( s ) \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } \\frac { 1 } { \\sum _ { \\ell = 1 } ^ { 2 } \\lambda } \\frac { 1 } { ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! } } \\\\ \\leq \\frac { 1 } { T } \\displaystyle \\sum _ { n = 0 } ^ { 1 } S _ { \\mathrm { G } , n } ^ { ( 1 ) } ( s ) \\leq ( 1 - s ) ! \\displaystyle \\sum _ { n = 0 } ^ { 1 } \\frac { 1 } { ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! ( 1 - s ) ! } - \\frac { 1 } { T } \\displaystyle \\sum _ \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 903, + 1940, + 1101, + 1940, + 1101, + 1985, + 903, + 1985 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { 2 ( \\chi + \\zeta - \\frac { \\Delta } { ( 1 - \\gamma ) } ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 598, + 1091, + 782, + 1091, + 782, + 1123, + 598, + 1123 + ], + "score": 0.89, + "latex": "\\mathbb { E } _ { s \\sim d ^ { \\star } } \\log G _ { t } ( s )" + }, + { + "category_id": 14, + "poly": [ + 546, + 1022, + 1153, + 1022, + 1153, + 1090, + 546, + 1090 + ], + "score": 0.88, + "latex": "\\frac { 1 } { \\eta } \\mathbb { E } _ { s \\sim d ^ { \\star } } \\log G _ { t } ( s ) \\le \\frac { 1 } { 1 - \\gamma } \\Big ( V ^ { ( t + 1 ) } ( d ^ { \\star } ) - V ^ { ( t ) } ( d ^ { \\star } ) \\Big )" + }, + { + "category_id": 13, + "poly": [ + 563, + 993, + 593, + 993, + 593, + 1019, + 563, + 1019 + ], + "score": 0.88, + "latex": "d ^ { \\star }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1890, + 971, + 1890, + 971, + 1941, + 298, + 1941 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { \\left( \\frac { 1 } { 1 - \\gamma } \\sum _ { a } \\pi ^ { \\star } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) - \\frac { 1 } { 1 - \\gamma } \\sum _ { a } \\pi ^ { ( t ) } ( a | s ) A _ { C } ^ { ( t ) } ( s , a ) \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 370, + 1859, + 397, + 1859, + 397, + 1886, + 370, + 1886 + ], + "score": 0.82, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 950, + 998, + 969, + 998, + 969, + 1023, + 950, + 1023 + ], + "score": 0.77, + "latex": "\\mu" + }, + { + "category_id": 13, + "poly": [ + 1369, + 1149, + 1384, + 1149, + 1384, + 1168, + 1369, + 1168 + ], + "score": 0.62, + "latex": "s" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1854.0, + 369.0, + 1854.0, + 369.0, + 1894.0, + 294.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 398.0, + 1854.0, + 1407.0, + 1854.0, + 1407.0, + 1894.0, + 398.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 972.0, + 1886.0, + 1211.0, + 1886.0, + 1211.0, + 1948.0, + 972.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1886.0, + 1410.0, + 1886.0, + 1410.0, + 1948.0, + 1401.0, + 1948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1935.0, + 902.0, + 1935.0, + 902.0, + 1991.0, + 288.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 1935.0, + 1118.0, + 1935.0, + 1118.0, + 1991.0, + 1102.0, + 1991.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 989.0, + 562.0, + 989.0, + 562.0, + 1028.0, + 295.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 594.0, + 989.0, + 949.0, + 989.0, + 949.0, + 1028.0, + 594.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 989.0, + 1086.0, + 989.0, + 1086.0, + 1028.0, + 970.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1125.0, + 688.0, + 1125.0, + 688.0, + 1184.0, + 289.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 933.0, + 1125.0, + 977.0, + 1125.0, + 977.0, + 1184.0, + 933.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 1125.0, + 1368.0, + 1125.0, + 1368.0, + 1184.0, + 1218.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1385.0, + 1125.0, + 1408.0, + 1125.0, + 1408.0, + 1184.0, + 1385.0, + 1184.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1169.0, + 406.0, + 1169.0, + 406.0, + 1207.0, + 293.0, + 1207.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1083.0, + 597.0, + 1083.0, + 597.0, + 1128.0, + 294.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 783.0, + 1083.0, + 792.0, + 1083.0, + 792.0, + 1128.0, + 783.0, + 1128.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 17, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 296, + 1029, + 1405, + 1029, + 1405, + 1217, + 296, + 1217 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 295, + 1536, + 1406, + 1536, + 1406, + 1808, + 295, + 1808 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1940, + 1404, + 1940, + 1404, + 2036, + 297, + 2036 + ], + "score": 0.971 + }, + { + "category_id": 1, + "poly": [ + 299, + 566, + 1404, + 566, + 1404, + 692, + 299, + 692 + ], + "score": 0.969 + }, + { + "category_id": 1, + "poly": [ + 297, + 1234, + 1405, + 1234, + 1405, + 1338, + 297, + 1338 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 703, + 816, + 995, + 816, + 995, + 892, + 703, + 892 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 510, + 696, + 1186, + 696, + 1186, + 772, + 510, + 772 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 298, + 785, + 376, + 785, + 376, + 816, + 298, + 816 + ], + "score": 0.922 + }, + { + "category_id": 8, + "poly": [ + 423, + 1462, + 1276, + 1462, + 1276, + 1536, + 423, + 1536 + ], + "score": 0.922 + }, + { + "category_id": 8, + "poly": [ + 341, + 484, + 1360, + 484, + 1360, + 561, + 341, + 561 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 297, + 408, + 894, + 408, + 894, + 442, + 297, + 442 + ], + "score": 0.908 + }, + { + "category_id": 8, + "poly": [ + 297, + 1812, + 1406, + 1812, + 1406, + 1900, + 297, + 1900 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1485, + 1400, + 1485, + 1400, + 1516, + 1352, + 1516 + ], + "score": 0.889 + }, + { + "category_id": 1, + "poly": [ + 299, + 1427, + 867, + 1427, + 867, + 1460, + 299, + 1460 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1896, + 1400, + 1896, + 1400, + 1925, + 1351, + 1925 + ], + "score": 0.87 + }, + { + "category_id": 2, + "poly": [ + 836, + 2087, + 865, + 2087, + 865, + 2113, + 836, + 2113 + ], + "score": 0.869 + }, + { + "category_id": 1, + "poly": [ + 296, + 222, + 1401, + 222, + 1401, + 300, + 296, + 300 + ], + "score": 0.847 + }, + { + "category_id": 2, + "poly": [ + 1374, + 905, + 1404, + 905, + 1404, + 935, + 1374, + 935 + ], + "score": 0.832 + }, + { + "category_id": 1, + "poly": [ + 300, + 309, + 1402, + 309, + 1402, + 399, + 300, + 399 + ], + "score": 0.822 + }, + { + "category_id": 8, + "poly": [ + 297, + 353, + 574, + 353, + 574, + 397, + 297, + 397 + ], + "score": 0.13 + }, + { + "category_id": 13, + "poly": [ + 837, + 562, + 1027, + 562, + 1027, + 618, + 837, + 618 + ], + "score": 0.93, + "latex": "\\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 13, + "poly": [ + 934, + 615, + 1248, + 615, + 1248, + 665, + 934, + 665 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { - \\Delta < - \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } - ( 1 - \\gamma ) \\zeta } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 702, + 812, + 997, + 812, + 997, + 892, + 702, + 892 + ], + "score": 0.93, + "latex": "K \\leq ( 1 - \\chi ) + \\frac { 4 \\sqrt { 2 \\delta } \\gamma } { ( 1 - \\gamma ) ^ { 2 } }" + }, + { + "category_id": 13, + "poly": [ + 604, + 615, + 874, + 615, + 874, + 665, + 604, + 665 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 911, + 1669, + 1003, + 1669, + 1003, + 1705, + 911, + 1705 + ], + "score": 0.93, + "latex": "n _ { j } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1028, + 1234, + 1209, + 1234, + 1209, + 1270, + 1028, + 1270 + ], + "score": 0.92, + "latex": "\\chi _ { t } = \\mathcal { O } ( 1 / \\sqrt { t } )" + }, + { + "category_id": 13, + "poly": [ + 910, + 1299, + 1211, + 1299, + 1211, + 1338, + 910, + 1338 + ], + "score": 0.92, + "latex": "R e g _ { C } ( T ) = \\mathcal { O } ( \\sqrt { | \\cal { S } | | \\mathcal { A } | T } )" + }, + { + "category_id": 13, + "poly": [ + 675, + 1768, + 777, + 1768, + 777, + 1811, + 675, + 1811 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { \\chi _ { t } \\propto \\frac { 1 } { \\sqrt { t } } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 511, + 689, + 1190, + 689, + 1190, + 771, + 511, + 771 + ], + "score": 0.91, + "latex": "V _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\cal A | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + K \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T }" + }, + { + "category_id": 14, + "poly": [ + 341, + 480, + 1360, + 480, + 1360, + 562, + 341, + 562 + ], + "score": 0.91, + "latex": "V _ { R } ^ { * } ( \\mu ) - V _ { R } ^ { ( T ) } ( \\mu ) \\leq \\frac { \\log | \\mathcal { A } | } { \\eta T } + \\frac { 1 } { ( 1 - \\gamma ) ^ { 2 } T } + \\bigl ( ( 1 - \\chi ) + \\bigl ( 1 - \\frac { 2 \\Delta } { ( 1 - \\gamma ) } \\bigr ) + 2 \\zeta \\bigr ) \\frac { \\sum _ { t = 0 } ^ { T - 1 } \\lambda ^ { ( t ) } } { \\eta T }" + }, + { + "category_id": 13, + "poly": [ + 1245, + 1615, + 1307, + 1615, + 1307, + 1649, + 1245, + 1649 + ], + "score": 0.91, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 507, + 1942, + 554, + 1942, + 554, + 1978, + 507, + 1978 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { t } }" + }, + { + "category_id": 13, + "poly": [ + 533, + 1667, + 828, + 1667, + 828, + 1708, + 533, + 1708 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { N _ { t } ( s , a ) = \\sum _ { j = 1 } ^ { t } n _ { j } ( s , a ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1338, + 1669, + 1400, + 1669, + 1400, + 1703, + 1338, + 1703 + ], + "score": 0.91, + "latex": "( s , a )" + }, + { + "category_id": 13, + "poly": [ + 485, + 1538, + 545, + 1538, + 545, + 1568, + 485, + 1568 + ], + "score": 0.9, + "latex": "\\phi _ { n e w }" + }, + { + "category_id": 13, + "poly": [ + 1118, + 1772, + 1219, + 1772, + 1219, + 1801, + 1118, + 1801 + ], + "score": 0.9, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 443, + 1302, + 549, + 1302, + 549, + 1336, + 443, + 1336 + ], + "score": 0.89, + "latex": "R e g _ { C } ( T )" + }, + { + "category_id": 13, + "poly": [ + 425, + 660, + 524, + 660, + 524, + 691, + 425, + 691 + ], + "score": 0.89, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 562, + 578, + 661, + 578, + 661, + 609, + 562, + 609 + ], + "score": 0.89, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 767, + 223, + 1031, + 223, + 1031, + 274, + 767, + 274 + ], + "score": 0.89, + "latex": "\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 535, + 310, + 1399, + 310, + 1399, + 359, + 535, + 359 + ], + "score": 0.87, + "latex": "\\begin{array} { r } { 2 ( \\chi - \\frac { \\Delta } { ( 1 - \\gamma ) } ) + 2 \\zeta = \\chi + \\chi - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } + 2 \\zeta \\leq 2 - \\chi - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } = ( 1 - \\chi ) + } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 717, + 1565, + 747, + 1565, + 747, + 1594, + 717, + 1594 + ], + "score": 0.86, + "latex": "t ^ { \\mathrm { t h } }" + }, + { + "category_id": 13, + "poly": [ + 405, + 1570, + 434, + 1570, + 434, + 1597, + 405, + 1597 + ], + "score": 0.86, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 528, + 1706, + 748, + 1706, + 748, + 1743, + 528, + 1743 + ], + "score": 0.86, + "latex": "\\begin{array} { r } { T = \\sum _ { s , a } N _ { k } ( s , a ) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 766, + 1430, + 865, + 1430, + 865, + 1459, + 766, + 1459 + ], + "score": 0.86, + "latex": "\\geq 1 - \\omega" + }, + { + "category_id": 13, + "poly": [ + 753, + 1614, + 849, + 1614, + 849, + 1649, + 753, + 1649 + ], + "score": 0.85, + "latex": "N _ { t } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 889, + 1187, + 911, + 1187, + 911, + 1216, + 889, + 1216 + ], + "score": 0.84, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 296, + 1596, + 676, + 1596, + 676, + 1669, + 296, + 1669 + ], + "score": 0.84, + "latex": "\\zeta _ { t } \\le \\mathbb { E } _ { s \\sim d ^ { \\pi _ { t } } , a \\sim \\pi _ { t } } \\left[ \\frac { C ^ { \\prime } \\sqrt { \\log ( { 1 / \\omega } ) } } { \\sqrt { | N _ { t } ( s , a ) | } } \\right]" + }, + { + "category_id": 14, + "poly": [ + 310, + 1809, + 1414, + 1809, + 1414, + 1902, + 310, + 1902 + ], + "score": 0.84, + "latex": "\\gamma _ { C } ^ { \\pi _ { t } } ( \\mu ) \\leq \\chi _ { t } + \\zeta _ { t } - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\quad \\mathrm { w h e r e } \\quad \\zeta _ { t } \\leq \\mathbb { E } _ { s \\sim d ^ { \\pi _ { t } } , a \\sim \\pi _ { t } } \\left[ \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { \\sqrt { \\lvert N _ { t } ( s , a ) \\rvert } } \\right] \\quad \\mathrm { a n d } \\ \\chi _ { t } = \\frac { C ^ { \\prime \\prime } } { \\sqrt { t } }" + }, + { + "category_id": 13, + "poly": [ + 1169, + 1568, + 1187, + 1568, + 1187, + 1599, + 1169, + 1599 + ], + "score": 0.84, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 1273, + 1067, + 1294, + 1067, + 1294, + 1094, + 1273, + 1094 + ], + "score": 0.83, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1086, + 229, + 1395, + 229, + 1395, + 274, + 1086, + 274 + ], + "score": 0.83, + "latex": "\\begin{array} { r } { - \\Delta < - \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } - ( 1 - \\gamma ) \\zeta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1372, + 1949, + 1400, + 1949, + 1400, + 1973, + 1372, + 1973 + ], + "score": 0.82, + "latex": "\\pi _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1133, + 1743, + 1155, + 1743, + 1155, + 1771, + 1133, + 1771 + ], + "score": 0.81, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 915, + 1155, + 939, + 1155, + 939, + 1181, + 915, + 1181 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1382, + 1707, + 1401, + 1707, + 1401, + 1733, + 1382, + 1733 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 494, + 1241, + 518, + 1241, + 518, + 1269, + 494, + 1269 + ], + "score": 0.79, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1380, + 584, + 1401, + 584, + 1401, + 605, + 1380, + 605 + ], + "score": 0.77, + "latex": "\\alpha" + }, + { + "category_id": 14, + "poly": [ + 533, + 309, + 1400, + 309, + 1400, + 359, + 533, + 359 + ], + "score": 0.77, + "latex": "\\begin{array} { r } { 2 ( \\chi - \\frac { \\Delta } { ( 1 - \\gamma ) } ) + 2 \\zeta = \\chi + \\chi - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } + 2 \\zeta \\leq 2 - \\chi - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } = ( 1 - \\chi ) + } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 443, + 240, + 465, + 240, + 465, + 261, + 443, + 261 + ], + "score": 0.76, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 884, + 1269, + 909, + 1269, + 909, + 1295, + 884, + 1295 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1388, + 1539, + 1401, + 1539, + 1401, + 1563, + 1388, + 1563 + ], + "score": 0.75, + "latex": "t" + }, + { + "category_id": 13, + "poly": [ + 297, + 352, + 574, + 352, + 574, + 399, + 297, + 399 + ], + "score": 0.73, + "latex": "\\begin{array} { r } { 2 \\zeta + ( 1 - 2 \\frac { \\Delta } { ( 1 - \\gamma ) } ) + 2 \\zeta } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 421, + 1460, + 1276, + 1460, + 1276, + 1538, + 421, + 1538 + ], + "score": 0.72, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { n e w } } } ( \\mu ) \\leq \\chi + \\zeta - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } } \\qquad \\mathrm { w h e r e } \\qquad \\zeta \\leq \\frac { C ^ { \\prime } \\sqrt { \\log ( 1 / \\omega ) } } { | N | }" + }, + { + "category_id": 13, + "poly": [ + 830, + 1303, + 853, + 1303, + 853, + 1331, + 830, + 1331 + ], + "score": 0.68, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 498, + 1707, + 514, + 1707, + 514, + 1738, + 498, + 1738 + ], + "score": 0.65, + "latex": "j" + }, + { + "category_id": 13, + "poly": [ + 296, + 1813, + 708, + 1813, + 708, + 1892, + 296, + 1892 + ], + "score": 0.44, + "latex": "V _ { C } ^ { \\pi _ { t } } ( \\mu ) \\leq \\chi _ { t } + \\zeta _ { t } - \\frac { \\Delta } { 1 - \\gamma } + \\frac { \\sqrt { 2 \\delta } \\gamma \\epsilon _ { C } } { ( 1 - \\gamma ) ^ { 2 } }" + }, + { + "category_id": 14, + "poly": [ + 1295, + 1814, + 1408, + 1814, + 1408, + 1893, + 1295, + 1893 + ], + "score": 0.44, + "latex": "\\chi _ { t } = { \\frac { C ^ { \\prime \\prime } } { \\sqrt { t } } }" + }, + { + "category_id": 13, + "poly": [ + 1295, + 1813, + 1409, + 1813, + 1409, + 1894, + 1295, + 1894 + ], + "score": 0.31, + "latex": "\\chi _ { t } = { \\frac { C ^ { \\prime \\prime } } { \\sqrt { t } } }" + }, + { + "category_id": 13, + "poly": [ + 769, + 223, + 1397, + 223, + 1397, + 275, + 769, + 275 + ], + "score": 0.27, + "latex": "\\begin{array} { r } { \\Delta > \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } + ( 1 - \\gamma ) \\zeta . \\mathrm { S o } , - \\Delta < - \\frac { 2 \\sqrt { 2 \\delta } \\gamma } { 1 - \\gamma } - ( 1 - \\gamma ) \\zeta . } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 910.0, + 1403.0, + 910.0, + 1403.0, + 937.0, + 1378.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1030.0, + 1405.0, + 1030.0, + 1405.0, + 1066.0, + 295.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1059.0, + 1272.0, + 1059.0, + 1272.0, + 1100.0, + 293.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 1059.0, + 1403.0, + 1059.0, + 1403.0, + 1100.0, + 1295.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1093.0, + 1403.0, + 1093.0, + 1403.0, + 1126.0, + 295.0, + 1126.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1120.0, + 1403.0, + 1120.0, + 1403.0, + 1158.0, + 294.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1155.0, + 914.0, + 1155.0, + 914.0, + 1187.0, + 296.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 1155.0, + 1403.0, + 1155.0, + 1403.0, + 1187.0, + 940.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1183.0, + 888.0, + 1183.0, + 888.0, + 1219.0, + 295.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 912.0, + 1183.0, + 1163.0, + 1183.0, + 1163.0, + 1219.0, + 912.0, + 1219.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1534.0, + 484.0, + 1534.0, + 484.0, + 1570.0, + 295.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 1534.0, + 1387.0, + 1534.0, + 1387.0, + 1570.0, + 546.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1534.0, + 1406.0, + 1534.0, + 1406.0, + 1570.0, + 1402.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1561.0, + 404.0, + 1561.0, + 404.0, + 1604.0, + 293.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 435.0, + 1561.0, + 716.0, + 1561.0, + 716.0, + 1604.0, + 435.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 748.0, + 1561.0, + 1168.0, + 1561.0, + 1168.0, + 1604.0, + 748.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1561.0, + 1406.0, + 1561.0, + 1406.0, + 1604.0, + 1188.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1611.0, + 752.0, + 1611.0, + 752.0, + 1650.0, + 680.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 1611.0, + 1244.0, + 1611.0, + 1244.0, + 1650.0, + 850.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1308.0, + 1611.0, + 1404.0, + 1611.0, + 1404.0, + 1650.0, + 1308.0, + 1650.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 1651.0, + 532.0, + 1651.0, + 532.0, + 1720.0, + 285.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 1651.0, + 910.0, + 1651.0, + 910.0, + 1720.0, + 829.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 1651.0, + 1337.0, + 1651.0, + 1337.0, + 1720.0, + 1004.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1651.0, + 1413.0, + 1651.0, + 1413.0, + 1720.0, + 1401.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1702.0, + 497.0, + 1702.0, + 497.0, + 1742.0, + 290.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 515.0, + 1702.0, + 527.0, + 1702.0, + 527.0, + 1742.0, + 515.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 749.0, + 1702.0, + 1381.0, + 1702.0, + 1381.0, + 1742.0, + 749.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1702.0, + 1408.0, + 1702.0, + 1408.0, + 1742.0, + 1402.0, + 1742.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1737.0, + 1132.0, + 1737.0, + 1132.0, + 1777.0, + 294.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1156.0, + 1737.0, + 1404.0, + 1737.0, + 1404.0, + 1777.0, + 1156.0, + 1777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1765.0, + 674.0, + 1765.0, + 674.0, + 1812.0, + 290.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 778.0, + 1765.0, + 1117.0, + 1765.0, + 1117.0, + 1812.0, + 778.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1220.0, + 1765.0, + 1226.0, + 1765.0, + 1226.0, + 1812.0, + 1220.0, + 1812.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1614.5, + 507.0, + 1614.5, + 507.0, + 1656.0, + 294.0, + 1656.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1937.0, + 506.0, + 1937.0, + 506.0, + 1982.0, + 291.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 1937.0, + 1371.0, + 1937.0, + 1371.0, + 1982.0, + 555.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1937.0, + 1405.0, + 1937.0, + 1405.0, + 1982.0, + 1401.0, + 1982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2008.0, + 295.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2003.0, + 1025.0, + 2003.0, + 1025.0, + 2038.0, + 292.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 561.0, + 561.0, + 561.0, + 561.0, + 622.0, + 291.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 561.0, + 836.0, + 561.0, + 836.0, + 622.0, + 662.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 561.0, + 1379.0, + 561.0, + 1379.0, + 622.0, + 1028.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 561.0, + 1405.0, + 561.0, + 1405.0, + 622.0, + 1402.0, + 622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 613.0, + 603.0, + 613.0, + 603.0, + 670.0, + 293.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 613.0, + 933.0, + 613.0, + 933.0, + 670.0, + 875.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 661.0, + 424.0, + 661.0, + 424.0, + 693.0, + 295.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 525.0, + 661.0, + 747.0, + 661.0, + 747.0, + 693.0, + 525.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1025.0, + 611.5, + 1406.0, + 611.5, + 1406.0, + 663.0, + 1025.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1234.0, + 493.0, + 1234.0, + 493.0, + 1273.0, + 296.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 519.0, + 1234.0, + 1027.0, + 1234.0, + 1027.0, + 1273.0, + 519.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1234.0, + 1406.0, + 1234.0, + 1406.0, + 1273.0, + 1210.0, + 1273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1267.0, + 883.0, + 1267.0, + 883.0, + 1303.0, + 295.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 910.0, + 1267.0, + 1403.0, + 1267.0, + 1403.0, + 1303.0, + 910.0, + 1303.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1299.0, + 442.0, + 1299.0, + 442.0, + 1338.0, + 292.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1299.0, + 829.0, + 1299.0, + 829.0, + 1338.0, + 550.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 854.0, + 1299.0, + 909.0, + 1299.0, + 909.0, + 1338.0, + 854.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1212.0, + 1299.0, + 1221.0, + 1299.0, + 1221.0, + 1338.0, + 1212.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 777.0, + 383.0, + 777.0, + 383.0, + 823.0, + 295.0, + 823.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 405.0, + 894.0, + 405.0, + 894.0, + 447.0, + 295.0, + 447.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1422.0, + 765.0, + 1422.0, + 765.0, + 1467.0, + 293.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 1422.0, + 869.0, + 1422.0, + 869.0, + 1467.0, + 866.0, + 1467.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 210.0, + 442.0, + 210.0, + 442.0, + 279.0, + 286.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 466.0, + 210.0, + 766.0, + 210.0, + 766.0, + 279.0, + 466.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1398.0, + 210.0, + 1412.0, + 210.0, + 1412.0, + 279.0, + 1398.0, + 279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 265.0, + 628.0, + 265.0, + 628.0, + 300.0, + 295.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 302.0, + 532.0, + 302.0, + 532.0, + 364.0, + 289.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 302.0, + 1408.0, + 302.0, + 1408.0, + 364.0, + 1401.0, + 364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 344.0, + 296.0, + 344.0, + 296.0, + 405.0, + 289.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 575.0, + 344.0, + 586.0, + 344.0, + 586.0, + 405.0, + 575.0, + 405.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 18, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 295, + 1655, + 1404, + 1655, + 1404, + 1810, + 295, + 1810 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 295, + 1433, + 1405, + 1433, + 1405, + 1588, + 295, + 1588 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1879, + 1403, + 1879, + 1403, + 2037, + 297, + 2037 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 296, + 1255, + 1405, + 1255, + 1405, + 1349, + 296, + 1349 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1036, + 1403, + 1036, + 1403, + 1132, + 299, + 1132 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 506, + 293, + 1195, + 293, + 1195, + 948, + 506, + 948 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 515, + 1351, + 1184, + 1351, + 1184, + 1420, + 515, + 1420 + ], + "score": 0.948 + }, + { + "category_id": 8, + "poly": [ + 655, + 1811, + 1047, + 1811, + 1047, + 1863, + 655, + 1863 + ], + "score": 0.938 + }, + { + "category_id": 1, + "poly": [ + 290, + 227, + 1403, + 227, + 1403, + 291, + 290, + 291 + ], + "score": 0.936 + }, + { + "category_id": 8, + "poly": [ + 657, + 1588, + 1045, + 1588, + 1045, + 1637, + 657, + 1637 + ], + "score": 0.925 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.917 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1369, + 1400, + 1369, + 1400, + 1400, + 1351, + 1400 + ], + "score": 0.901 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1599, + 1400, + 1599, + 1400, + 1630, + 1352, + 1630 + ], + "score": 0.895 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1822, + 1400, + 1822, + 1400, + 1853, + 1352, + 1853 + ], + "score": 0.888 + }, + { + "category_id": 9, + "poly": [ + 1350, + 599, + 1402, + 599, + 1402, + 632, + 1350, + 632 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 866, + 2087, + 866, + 2113, + 834, + 2113 + ], + "score": 0.874 + }, + { + "category_id": 2, + "poly": [ + 1374, + 942, + 1404, + 942, + 1404, + 973, + 1374, + 973 + ], + "score": 0.833 + }, + { + "category_id": 0, + "poly": [ + 298, + 1193, + 947, + 1193, + 947, + 1226, + 298, + 1226 + ], + "score": 0.718 + }, + { + "category_id": 14, + "poly": [ + 505, + 281, + 1196, + 281, + 1196, + 952, + 505, + 952 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } { \\mathbb { E } \\mathbb { E } \\mathbb { E } \\mathbb { E } ( t ) - \\frac { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } \\times } & { 1 } \\\\ & { \\le \\frac { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } ( t ) } \\underbrace { \\sum _ { t = 1 } ^ { K } \\left( \\begin{array} { l l } { \\Delta } & { \\sqrt { \\Delta } \\pi \\kappa \\omega t } \\\\ { 1 - \\gamma } & { ( 1 - \\gamma ) ^ { t } } \\end{array} \\right) } _ { \\le \\alpha } } \\\\ & { \\le \\frac { \\sum _ { t = 1 } ^ { K } \\sum _ { t = 1 } ^ { K } \\hat { \\psi } _ { t } ^ { t } } { \\sqrt { \\pi } } + \\sum _ { t = 1 } ^ { K } \\sum _ { s = 1 } ^ { K } \\frac { \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } } \\\\ & \\le \\frac { \\sum _ { t = 1 } ^ { K } \\frac { \\gamma } { \\alpha } + \\mathcal { E } \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\sum _ { s = 1 } ^ { K } \\frac { \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } { \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) \\hat { \\psi } _ { t } ^ { t } ( t ) } } \\\\ & \\le C ^ { \\mathcal { N } } \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( \\mathcal { L } \\sqrt { \\pi } \\hat { \\psi } _ { t } ^ { t } ( t ) \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 958, + 1496, + 1156, + 1496, + 1156, + 1530, + 958, + 1530 + ], + "score": 0.94, + "latex": "\\mathbb { 1 } \\{ f a i l u r e \\} = 0" + }, + { + "category_id": 13, + "poly": [ + 876, + 1256, + 966, + 1256, + 966, + 1290, + 876, + 1290 + ], + "score": 0.93, + "latex": "\\pi _ { \\phi } ( a | s )" + }, + { + "category_id": 13, + "poly": [ + 298, + 1076, + 589, + 1076, + 589, + 1135, + 298, + 1135 + ], + "score": 0.93, + "latex": "\\sqrt { | S | | A | } \\sqrt { \\textstyle \\sum _ { s , a } N _ { k } ( s , a ) }" + }, + { + "category_id": 13, + "poly": [ + 704, + 1943, + 780, + 1943, + 780, + 1976, + 704, + 1976 + ], + "score": 0.92, + "latex": "r ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 1193, + 1717, + 1345, + 1717, + 1345, + 1753, + 1193, + 1753 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi" + }, + { + "category_id": 13, + "poly": [ + 477, + 1496, + 676, + 1496, + 676, + 1530, + 477, + 1530 + ], + "score": 0.92, + "latex": "\\mathbb { 1 } \\{ f a i l u r e \\} = 1" + }, + { + "category_id": 13, + "poly": [ + 298, + 1716, + 388, + 1716, + 388, + 1753, + 298, + 1753 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 1323, + 1256, + 1400, + 1256, + 1400, + 1290, + 1323, + 1290 + ], + "score": 0.92, + "latex": "r ( s , a )" + }, + { + "category_id": 14, + "poly": [ + 651, + 1807, + 1048, + 1807, + 1048, + 1866, + 651, + 1866 + ], + "score": 0.91, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) \\quad s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi" + }, + { + "category_id": 13, + "poly": [ + 794, + 1085, + 1015, + 1085, + 1015, + 1127, + 794, + 1127 + ], + "score": 0.91, + "latex": "\\begin{array} { r } { T = \\sum _ { s , a } N _ { k } ( s , a ) } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 517, + 1349, + 1183, + 1349, + 1183, + 1422, + 517, + 1422 + ], + "score": 0.91, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { \\tau \\sim \\pi _ { \\phi } } [ \\sum _ { \\tau } r ( \\cdot ) ] \\quad s . t . \\quad \\mathbb { E } _ { \\tau \\sim \\pi _ { \\phi } } [ \\sum _ { \\tau } \\mathbb { 1 } \\{ f a i l u r e \\} ] = 0" + }, + { + "category_id": 14, + "poly": [ + 652, + 1586, + 1045, + 1586, + 1045, + 1644, + 652, + 1644 + ], + "score": 0.9, + "latex": "\\operatorname * { m a x } _ { \\pi _ { \\phi } } V _ { R } ^ { \\pi _ { \\phi } } ( \\mu ) \\quad s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\bar { \\mu ) } = 0" + }, + { + "category_id": 13, + "poly": [ + 1143, + 1659, + 1224, + 1659, + 1224, + 1690, + 1143, + 1690 + ], + "score": 0.9, + "latex": "\\chi \\approx 0" + }, + { + "category_id": 13, + "poly": [ + 627, + 1258, + 702, + 1258, + 702, + 1285, + 627, + 1285 + ], + "score": 0.9, + "latex": "s \\in { \\mathcal { S } }" + }, + { + "category_id": 13, + "poly": [ + 343, + 1257, + 422, + 1257, + 422, + 1286, + 343, + 1286 + ], + "score": 0.9, + "latex": "a \\in { \\mathcal { A } }" + }, + { + "category_id": 13, + "poly": [ + 643, + 1913, + 685, + 1913, + 685, + 1943, + 643, + 1943 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 395, + 1973, + 438, + 1973, + 438, + 2004, + 395, + 2004 + ], + "score": 0.88, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 689, + 1882, + 727, + 1882, + 727, + 1912, + 689, + 1912 + ], + "score": 0.88, + "latex": "V _ { R }" + }, + { + "category_id": 13, + "poly": [ + 1302, + 1943, + 1339, + 1943, + 1339, + 1973, + 1302, + 1973 + ], + "score": 0.87, + "latex": "V _ { R }" + }, + { + "category_id": 13, + "poly": [ + 373, + 1322, + 407, + 1322, + 407, + 1352, + 373, + 1352 + ], + "score": 0.87, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 382, + 2004, + 517, + 2004, + 517, + 2036, + 382, + 2036 + ], + "score": 0.82, + "latex": "\\mathbb { 1 } \\{ f a i l u r e \\}" + }, + { + "category_id": 13, + "poly": [ + 834, + 232, + 859, + 232, + 859, + 258, + 834, + 258 + ], + "score": 0.81, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 557, + 1693, + 579, + 1693, + 579, + 1720, + 557, + 1720 + ], + "score": 0.81, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1149, + 1037, + 1345, + 1037, + 1345, + 1080, + 1149, + 1080 + ], + "score": 0.8, + "latex": "\\begin{array} { r } { \\sum _ { s , a } \\sqrt { N _ { k } ( s , a ) } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 1040, + 1755, + 1062, + 1755, + 1062, + 1782, + 1040, + 1782 + ], + "score": 0.79, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 1335, + 1883, + 1353, + 1883, + 1353, + 1909, + 1335, + 1909 + ], + "score": 0.79, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 297, + 1944, + 314, + 1944, + 314, + 1975, + 297, + 1975 + ], + "score": 0.79, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 908, + 1293, + 927, + 1293, + 927, + 1314, + 908, + 1314 + ], + "score": 0.78, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 366, + 1442, + 385, + 1442, + 385, + 1463, + 366, + 1463 + ], + "score": 0.75, + "latex": "\\tau" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2083.0, + 872.0, + 2083.0, + 872.0, + 2124.0, + 828.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1378.0, + 947.0, + 1403.0, + 947.0, + 1403.0, + 974.0, + 1378.0, + 974.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1193.0, + 953.0, + 1193.0, + 953.0, + 1231.0, + 295.0, + 1231.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1655.0, + 1142.0, + 1655.0, + 1142.0, + 1693.0, + 295.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1225.0, + 1655.0, + 1406.0, + 1655.0, + 1406.0, + 1693.0, + 1225.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1685.0, + 556.0, + 1685.0, + 556.0, + 1725.0, + 293.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 580.0, + 1685.0, + 1404.0, + 1685.0, + 1404.0, + 1725.0, + 580.0, + 1725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 297.0, + 1711.0, + 297.0, + 1760.0, + 293.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1711.0, + 1192.0, + 1711.0, + 1192.0, + 1760.0, + 389.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1711.0, + 1410.0, + 1711.0, + 1410.0, + 1760.0, + 1346.0, + 1760.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1744.0, + 1039.0, + 1744.0, + 1039.0, + 1790.0, + 292.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1063.0, + 1744.0, + 1405.0, + 1744.0, + 1405.0, + 1790.0, + 1063.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1779.0, + 707.0, + 1779.0, + 707.0, + 1816.0, + 294.0, + 1816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1435.0, + 365.0, + 1435.0, + 365.0, + 1468.0, + 295.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 386.0, + 1435.0, + 1403.0, + 1435.0, + 1403.0, + 1468.0, + 386.0, + 1468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1463.0, + 1405.0, + 1463.0, + 1405.0, + 1503.0, + 292.0, + 1503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1495.0, + 476.0, + 1495.0, + 476.0, + 1532.0, + 292.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1495.0, + 957.0, + 1495.0, + 957.0, + 1532.0, + 677.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1157.0, + 1495.0, + 1406.0, + 1495.0, + 1406.0, + 1532.0, + 1157.0, + 1532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1525.0, + 1406.0, + 1525.0, + 1406.0, + 1562.0, + 294.0, + 1562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1554.0, + 1267.0, + 1554.0, + 1267.0, + 1593.0, + 293.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1881.0, + 688.0, + 1881.0, + 688.0, + 1915.0, + 297.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 728.0, + 1881.0, + 1334.0, + 1881.0, + 1334.0, + 1915.0, + 728.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1354.0, + 1881.0, + 1404.0, + 1881.0, + 1404.0, + 1915.0, + 1354.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1908.0, + 642.0, + 1908.0, + 642.0, + 1949.0, + 292.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 1908.0, + 1405.0, + 1908.0, + 1405.0, + 1949.0, + 686.0, + 1949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1941.0, + 703.0, + 1941.0, + 703.0, + 1978.0, + 315.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 781.0, + 1941.0, + 1301.0, + 1941.0, + 1301.0, + 1978.0, + 781.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1340.0, + 1941.0, + 1407.0, + 1941.0, + 1407.0, + 1978.0, + 1340.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1971.0, + 394.0, + 1971.0, + 394.0, + 2009.0, + 293.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 439.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2009.0, + 439.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2001.0, + 381.0, + 2001.0, + 381.0, + 2040.0, + 295.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 2001.0, + 530.0, + 2001.0, + 530.0, + 2040.0, + 518.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1252.0, + 342.0, + 1252.0, + 342.0, + 1292.0, + 291.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 423.0, + 1252.0, + 626.0, + 1252.0, + 626.0, + 1292.0, + 423.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1252.0, + 875.0, + 1252.0, + 875.0, + 1292.0, + 703.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 967.0, + 1252.0, + 1322.0, + 1252.0, + 1322.0, + 1292.0, + 967.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1252.0, + 1405.0, + 1252.0, + 1405.0, + 1292.0, + 1401.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1280.0, + 907.0, + 1280.0, + 907.0, + 1326.0, + 291.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 928.0, + 1280.0, + 1406.0, + 1280.0, + 1406.0, + 1326.0, + 928.0, + 1326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1317.0, + 372.0, + 1317.0, + 372.0, + 1353.0, + 293.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1317.0, + 1204.0, + 1317.0, + 1204.0, + 1353.0, + 408.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1030.0, + 1148.0, + 1030.0, + 1148.0, + 1083.0, + 290.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1346.0, + 1030.0, + 1412.0, + 1030.0, + 1412.0, + 1083.0, + 1346.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1078.0, + 297.0, + 1078.0, + 297.0, + 1130.0, + 294.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1078.0, + 793.0, + 1078.0, + 793.0, + 1130.0, + 590.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1016.0, + 1078.0, + 1025.0, + 1078.0, + 1025.0, + 1130.0, + 1016.0, + 1130.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 222.0, + 833.0, + 222.0, + 833.0, + 271.0, + 291.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 860.0, + 222.0, + 1404.0, + 222.0, + 1404.0, + 271.0, + 860.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 256.0, + 421.0, + 256.0, + 421.0, + 296.0, + 292.0, + 296.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 19, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 228, + 1404, + 228, + 1404, + 447, + 297, + 447 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1003, + 1405, + 1003, + 1405, + 1193, + 297, + 1193 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 296, + 1683, + 1406, + 1683, + 1406, + 1873, + 296, + 1873 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 459, + 1402, + 459, + 1402, + 557, + 299, + 557 + ], + "score": 0.969 + }, + { + "category_id": 8, + "poly": [ + 444, + 560, + 1257, + 560, + 1257, + 694, + 444, + 694 + ], + "score": 0.968 + }, + { + "category_id": 8, + "poly": [ + 405, + 867, + 1293, + 867, + 1293, + 1002, + 405, + 1002 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 591, + 1270, + 1110, + 1270, + 1110, + 1418, + 591, + 1418 + ], + "score": 0.958 + }, + { + "category_id": 8, + "poly": [ + 459, + 1487, + 1244, + 1487, + 1244, + 1672, + 459, + 1672 + ], + "score": 0.956 + }, + { + "category_id": 1, + "poly": [ + 302, + 696, + 1399, + 696, + 1399, + 788, + 302, + 788 + ], + "score": 0.949 + }, + { + "category_id": 1, + "poly": [ + 298, + 1205, + 1397, + 1205, + 1397, + 1267, + 298, + 1267 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 296, + 800, + 1403, + 800, + 1403, + 863, + 296, + 863 + ], + "score": 0.935 + }, + { + "category_id": 1, + "poly": [ + 300, + 1418, + 785, + 1418, + 785, + 1450, + 300, + 1450 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 299, + 75, + 813, + 75, + 813, + 105, + 299, + 105 + ], + "score": 0.906 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1327, + 1400, + 1327, + 1400, + 1359, + 1351, + 1359 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1564, + 1400, + 1564, + 1400, + 1594, + 1351, + 1594 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1351, + 919, + 1400, + 919, + 1400, + 949, + 1351, + 949 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1352, + 612, + 1400, + 612, + 1400, + 643, + 1352, + 643 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2112, + 834, + 2112 + ], + "score": 0.432 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 863, + 2088, + 863, + 2113, + 834, + 2113 + ], + "score": 0.426 + }, + { + "category_id": 14, + "poly": [ + 458, + 1487, + 1239, + 1487, + 1239, + 1674, + 458, + 1674 + ], + "score": 0.95, + "latex": "\\begin{array} { r l } { \\pi _ { \\phi _ { n e w } } = \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] } & { } \\\\ { s . t . } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . } & { V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ A _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right] \\leq \\chi } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 969, + 1123, + 1072, + 1123, + 1072, + 1162, + 969, + 1162 + ], + "score": 0.94, + "latex": "\\hat { A } _ { C } ( s , \\bar { a } )" + }, + { + "category_id": 13, + "poly": [ + 401, + 1206, + 468, + 1206, + 468, + 1240, + 401, + 1240 + ], + "score": 0.92, + "latex": "\\rho _ { \\phi } ( s )" + }, + { + "category_id": 13, + "poly": [ + 705, + 521, + 789, + 521, + 789, + 555, + 705, + 555 + ], + "score": 0.92, + "latex": "A ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1159, + 401, + 1159, + 401, + 1193, + 297, + 1193 + ], + "score": 0.92, + "latex": "A _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 467, + 1715, + 510, + 1715, + 510, + 1750, + 467, + 1750 + ], + "score": 0.91, + "latex": "\\hat { A } _ { C }" + }, + { + "category_id": 13, + "poly": [ + 775, + 802, + 874, + 802, + 874, + 836, + 775, + 836 + ], + "score": 0.91, + "latex": "C Q L ( \\zeta )" + }, + { + "category_id": 13, + "poly": [ + 1022, + 1064, + 1085, + 1064, + 1085, + 1095, + 1022, + 1095 + ], + "score": 0.91, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 14, + "poly": [ + 590, + 1266, + 1109, + 1266, + 1109, + 1420, + 590, + 1420 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } { { \\operatorname* { m a x } _ { \\pi _ { \\phi } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } [ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) ] } } \\\\ { s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\\\ { s . t . \\quad V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 814, + 414, + 856, + 414, + 856, + 446, + 814, + 446 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 329, + 1033, + 372, + 1033, + 372, + 1065, + 329, + 1065 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 464, + 1064, + 508, + 1064, + 508, + 1095, + 464, + 1095 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 777, + 1005, + 838, + 1005, + 838, + 1033, + 777, + 1033 + ], + "score": 0.9, + "latex": "\\mathcal { D } _ { e n v }" + }, + { + "category_id": 13, + "poly": [ + 780, + 1685, + 822, + 1685, + 822, + 1716, + 780, + 1716 + ], + "score": 0.89, + "latex": "A _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1046, + 353, + 1091, + 353, + 1091, + 385, + 1046, + 385 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 478, + 1128, + 521, + 1128, + 521, + 1161, + 478, + 1161 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 491, + 231, + 535, + 231, + 535, + 262, + 491, + 262 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1146, + 804, + 1188, + 804, + 1188, + 834, + 1146, + 834 + ], + "score": 0.89, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 297, + 353, + 340, + 353, + 340, + 385, + 297, + 385 + ], + "score": 0.88, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1358, + 1685, + 1401, + 1685, + 1401, + 1716, + 1358, + 1716 + ], + "score": 0.87, + "latex": "A _ { C }" + }, + { + "category_id": 13, + "poly": [ + 1168, + 1212, + 1201, + 1212, + 1201, + 1240, + 1168, + 1240 + ], + "score": 0.87, + "latex": "\\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 298, + 831, + 332, + 831, + 332, + 860, + 298, + 860 + ], + "score": 0.86, + "latex": "k ^ { \\mathrm { { t h } } }" + }, + { + "category_id": 14, + "poly": [ + 407, + 863, + 1292, + 863, + 1292, + 1007, + 407, + 1007 + ], + "score": 0.86, + "latex": "\\begin{array} { r l } & { \\hat { Q } _ { C } ^ { k + 1 } \\gets \\underset { Q _ { C } } { \\arg \\operatorname* { m i n } } \\alpha \\left( - \\mathbb { E } _ { s \\sim \\mathcal { D } _ { e n v } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q _ { C } ( s , a ) ] + \\mathbb { E } _ { ( s , a ) \\sim \\mathcal { D } _ { e n v } } [ Q _ { C } ( s , a ) ] \\right) } \\\\ & { \\quad \\quad \\quad + \\frac { 1 } { 2 } \\mathbb { E } _ { ( s , a , s ^ { \\prime } , c ) \\sim \\mathcal { D } _ { e n v } } \\left[ \\left( Q _ { C } ( s , a ) - \\hat { B } ^ { \\pi _ { \\phi } } \\hat { Q } _ { C } ^ { k } ( s , a ) \\right) ^ { 2 } \\right] } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 962, + 804, + 980, + 804, + 980, + 835, + 962, + 835 + ], + "score": 0.83, + "latex": "\\zeta" + }, + { + "category_id": 13, + "poly": [ + 1249, + 804, + 1267, + 804, + 1267, + 830, + 1249, + 830 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 723, + 388, + 745, + 388, + 745, + 410, + 723, + 410 + ], + "score": 0.78, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1049, + 655, + 1251, + 655, + 1251, + 696, + 1049, + 696 + ], + "score": 0.58, + "latex": "a \\sim \\pi _ { \\phi } ( a | s ) \\left[ A ( s , a ) \\right]" + }, + { + "category_id": 13, + "poly": [ + 446, + 657, + 1252, + 657, + 1252, + 697, + 446, + 697 + ], + "score": 0.49, + "latex": "= \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q ( s , a ) - V ( s ) ] = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) } [ A ( s , a ) ]" + }, + { + "category_id": 13, + "poly": [ + 297, + 521, + 619, + 521, + 619, + 562, + 297, + 562 + ], + "score": 0.49, + "latex": "\\mathbb { E } _ { s \\sim d } { } ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) \\big [ A ( s , a ) \\big ]" + }, + { + "category_id": 14, + "poly": [ + 442, + 560, + 1152, + 560, + 1152, + 603, + 442, + 603 + ], + "score": 0.46, + "latex": "\\begin{array} { r } { { \\mathbb E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q ( s , a ) ] - { \\mathbb E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi _ { o l d } } ( a | s ) } [ Q ( s , a ) ] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 440, + 558, + 1177, + 558, + 1177, + 695, + 440, + 695 + ], + "score": 0.44, + "latex": "\\begin{array} { r l } & { \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ Q ( s , a ) ] - \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi _ { o l d } } ( a \\mid s ) } [ Q ( s , a ) ] } \\\\ & { \\ = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ Q ( s , a ) - \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } ( a \\mid s ) } Q ( s , a ) ] } \\\\ & { \\ = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ Q ( s , a ) - V ( s ) ] = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a \\mid s ) } [ A } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 451, + 652, + 1258, + 652, + 1258, + 697, + 451, + 697 + ], + "score": 0.29, + "latex": "= \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) } [ Q ( s , a ) - V ( s ) ] = \\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) } [ A ( s , a ) ]" + }, + { + "category_id": 14, + "poly": [ + 506, + 609, + 1064, + 609, + 1064, + 650, + 506, + 650 + ], + "score": 0.29, + "latex": "s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } , a \\sim \\pi _ { \\phi } ( a | s ) [ Q ( s , a ) - \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } ( a | s ) } Q ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 297, + 523, + 408, + 523, + 408, + 558, + 297, + 558 + ], + "score": 0.27, + "latex": "\\mathbb { E } _ { s \\sim d ^ { \\pi _ { \\phi _ { o l d } } } }" + }, + { + "category_id": 13, + "poly": [ + 821, + 353, + 880, + 353, + 880, + 383, + 821, + 383 + ], + "score": 0.25, + "latex": "\\mathrm { C Q L }" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 73.0, + 815.0, + 73.0, + 815.0, + 108.0, + 295.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 868.0, + 2084.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 490.0, + 227.0, + 490.0, + 268.0, + 292.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 536.0, + 227.0, + 1405.0, + 227.0, + 1405.0, + 268.0, + 536.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 261.0, + 1406.0, + 261.0, + 1406.0, + 296.0, + 295.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 293.0, + 1404.0, + 293.0, + 1404.0, + 324.0, + 296.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 321.0, + 1405.0, + 321.0, + 1405.0, + 355.0, + 292.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 351.0, + 296.0, + 351.0, + 296.0, + 389.0, + 292.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 351.0, + 820.0, + 351.0, + 820.0, + 389.0, + 341.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 351.0, + 1045.0, + 351.0, + 1045.0, + 389.0, + 881.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1092.0, + 351.0, + 1407.0, + 351.0, + 1407.0, + 389.0, + 1092.0, + 389.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 382.0, + 722.0, + 382.0, + 722.0, + 417.0, + 295.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 746.0, + 382.0, + 1405.0, + 382.0, + 1405.0, + 417.0, + 746.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 411.0, + 813.0, + 411.0, + 813.0, + 451.0, + 291.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 411.0, + 1404.0, + 411.0, + 1404.0, + 451.0, + 857.0, + 451.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1003.0, + 776.0, + 1003.0, + 776.0, + 1036.0, + 296.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 1003.0, + 1403.0, + 1003.0, + 1403.0, + 1036.0, + 839.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1028.0, + 328.0, + 1028.0, + 328.0, + 1071.0, + 292.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1028.0, + 1406.0, + 1028.0, + 1406.0, + 1071.0, + 373.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1062.0, + 463.0, + 1062.0, + 463.0, + 1102.0, + 292.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 509.0, + 1062.0, + 1021.0, + 1062.0, + 1021.0, + 1102.0, + 509.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1086.0, + 1062.0, + 1406.0, + 1062.0, + 1406.0, + 1102.0, + 1086.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1093.0, + 1403.0, + 1093.0, + 1403.0, + 1128.0, + 294.0, + 1128.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1125.0, + 477.0, + 1125.0, + 477.0, + 1164.0, + 292.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1125.0, + 968.0, + 1125.0, + 968.0, + 1164.0, + 522.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1125.0, + 1405.0, + 1125.0, + 1405.0, + 1164.0, + 1073.0, + 1164.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1153.0, + 296.0, + 1153.0, + 296.0, + 1199.0, + 292.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 402.0, + 1153.0, + 416.0, + 1153.0, + 416.0, + 1199.0, + 402.0, + 1199.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1683.0, + 779.0, + 1683.0, + 779.0, + 1720.0, + 295.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 1683.0, + 1357.0, + 1683.0, + 1357.0, + 1720.0, + 823.0, + 1720.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1717.0, + 466.0, + 1717.0, + 466.0, + 1756.0, + 294.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 511.0, + 1717.0, + 1404.0, + 1717.0, + 1404.0, + 1756.0, + 511.0, + 1756.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1748.0, + 1406.0, + 1748.0, + 1406.0, + 1784.0, + 294.0, + 1784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1778.0, + 1404.0, + 1778.0, + 1404.0, + 1817.0, + 293.0, + 1817.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1808.0, + 1406.0, + 1808.0, + 1406.0, + 1847.0, + 293.0, + 1847.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1840.0, + 672.0, + 1840.0, + 672.0, + 1873.0, + 295.0, + 1873.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 457.0, + 1407.0, + 457.0, + 1407.0, + 496.0, + 295.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 487.0, + 1404.0, + 487.0, + 1404.0, + 528.0, + 294.0, + 528.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 519.0, + 296.0, + 519.0, + 296.0, + 562.0, + 290.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 620.0, + 519.0, + 704.0, + 519.0, + 704.0, + 562.0, + 620.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 790.0, + 519.0, + 1087.0, + 519.0, + 1087.0, + 562.0, + 790.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 694.0, + 1403.0, + 694.0, + 1403.0, + 728.0, + 297.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 725.0, + 1404.0, + 725.0, + 1404.0, + 759.0, + 296.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 753.0, + 570.0, + 753.0, + 570.0, + 793.0, + 295.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1202.0, + 400.0, + 1202.0, + 400.0, + 1244.0, + 293.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 469.0, + 1202.0, + 1167.0, + 1202.0, + 1167.0, + 1244.0, + 469.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1202.0, + 1202.0, + 1401.0, + 1202.0, + 1401.0, + 1244.0, + 1202.0, + 1244.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1235.0, + 1223.0, + 1235.0, + 1223.0, + 1271.0, + 293.0, + 1271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 797.0, + 774.0, + 797.0, + 774.0, + 839.0, + 294.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 797.0, + 961.0, + 797.0, + 961.0, + 839.0, + 875.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 981.0, + 797.0, + 1145.0, + 797.0, + 1145.0, + 839.0, + 981.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1189.0, + 797.0, + 1248.0, + 797.0, + 1248.0, + 839.0, + 1189.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1268.0, + 797.0, + 1405.0, + 797.0, + 1405.0, + 839.0, + 1268.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 825.0, + 297.0, + 825.0, + 297.0, + 870.0, + 293.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 825.0, + 523.0, + 825.0, + 523.0, + 870.0, + 333.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1416.0, + 784.0, + 1416.0, + 784.0, + 1453.0, + 297.0, + 1453.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 20, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 8, + "poly": [ + 469, + 528, + 1201, + 528, + 1201, + 647, + 469, + 647 + ], + "score": 0.962 + }, + { + "category_id": 1, + "poly": [ + 298, + 422, + 1408, + 422, + 1408, + 493, + 298, + 493 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 295, + 650, + 1398, + 650, + 1398, + 721, + 295, + 721 + ], + "score": 0.94 + }, + { + "category_id": 1, + "poly": [ + 297, + 1960, + 1399, + 1960, + 1399, + 2024, + 297, + 2024 + ], + "score": 0.938 + }, + { + "category_id": 8, + "poly": [ + 439, + 1284, + 1258, + 1284, + 1258, + 1340, + 439, + 1340 + ], + "score": 0.936 + }, + { + "category_id": 1, + "poly": [ + 297, + 1497, + 1406, + 1497, + 1406, + 1563, + 297, + 1563 + ], + "score": 0.926 + }, + { + "category_id": 8, + "poly": [ + 469, + 1690, + 1229, + 1690, + 1229, + 1749, + 469, + 1749 + ], + "score": 0.922 + }, + { + "category_id": 2, + "poly": [ + 298, + 73, + 814, + 73, + 814, + 106, + 298, + 106 + ], + "score": 0.918 + }, + { + "category_id": 8, + "poly": [ + 306, + 1043, + 1324, + 1043, + 1324, + 1112, + 306, + 1112 + ], + "score": 0.903 + }, + { + "category_id": 8, + "poly": [ + 675, + 1788, + 1027, + 1788, + 1027, + 1829, + 675, + 1829 + ], + "score": 0.901 + }, + { + "category_id": 1, + "poly": [ + 295, + 1655, + 993, + 1655, + 993, + 1689, + 295, + 1689 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1351, + 573, + 1401, + 573, + 1401, + 605, + 1351, + 605 + ], + "score": 0.897 + }, + { + "category_id": 8, + "poly": [ + 567, + 1861, + 1010, + 1861, + 1010, + 1913, + 567, + 1913 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1895, + 1401, + 1895, + 1401, + 1926, + 1351, + 1926 + ], + "score": 0.894 + }, + { + "category_id": 1, + "poly": [ + 296, + 733, + 1352, + 733, + 1352, + 767, + 296, + 767 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1351, + 371, + 1401, + 371, + 1401, + 402, + 1351, + 402 + ], + "score": 0.887 + }, + { + "category_id": 8, + "poly": [ + 459, + 1397, + 1245, + 1397, + 1245, + 1480, + 459, + 1480 + ], + "score": 0.883 + }, + { + "category_id": 8, + "poly": [ + 416, + 770, + 1278, + 770, + 1278, + 845, + 416, + 845 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1793, + 1401, + 1793, + 1401, + 1824, + 1351, + 1824 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1351, + 811, + 1401, + 811, + 1401, + 843, + 1351, + 843 + ], + "score": 0.879 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1420, + 1401, + 1420, + 1401, + 1452, + 1350, + 1452 + ], + "score": 0.878 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1700, + 1401, + 1700, + 1401, + 1732, + 1351, + 1732 + ], + "score": 0.878 + }, + { + "category_id": 1, + "poly": [ + 301, + 1007, + 1398, + 1007, + 1398, + 1040, + 301, + 1040 + ], + "score": 0.876 + }, + { + "category_id": 9, + "poly": [ + 1350, + 1294, + 1401, + 1294, + 1401, + 1325, + 1350, + 1325 + ], + "score": 0.873 + }, + { + "category_id": 9, + "poly": [ + 1351, + 1592, + 1401, + 1592, + 1401, + 1624, + 1351, + 1624 + ], + "score": 0.871 + }, + { + "category_id": 2, + "poly": [ + 834, + 2087, + 865, + 2087, + 865, + 2113, + 834, + 2113 + ], + "score": 0.858 + }, + { + "category_id": 8, + "poly": [ + 314, + 1146, + 1314, + 1146, + 1314, + 1199, + 314, + 1199 + ], + "score": 0.851 + }, + { + "category_id": 1, + "poly": [ + 299, + 1829, + 705, + 1829, + 705, + 1861, + 299, + 1861 + ], + "score": 0.846 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1155, + 1401, + 1155, + 1401, + 1187, + 1352, + 1187 + ], + "score": 0.843 + }, + { + "category_id": 1, + "poly": [ + 298, + 1201, + 1412, + 1201, + 1412, + 1282, + 298, + 1282 + ], + "score": 0.842 + }, + { + "category_id": 9, + "poly": [ + 1351, + 943, + 1400, + 943, + 1400, + 973, + 1351, + 973 + ], + "score": 0.833 + }, + { + "category_id": 9, + "poly": [ + 1352, + 1061, + 1401, + 1061, + 1401, + 1092, + 1352, + 1092 + ], + "score": 0.828 + }, + { + "category_id": 8, + "poly": [ + 335, + 1572, + 1315, + 1572, + 1315, + 1651, + 335, + 1651 + ], + "score": 0.803 + }, + { + "category_id": 1, + "poly": [ + 288, + 1355, + 1269, + 1355, + 1269, + 1393, + 288, + 1393 + ], + "score": 0.781 + }, + { + "category_id": 1, + "poly": [ + 298, + 332, + 812, + 332, + 812, + 373, + 298, + 373 + ], + "score": 0.744 + }, + { + "category_id": 1, + "poly": [ + 286, + 1114, + 1185, + 1114, + 1185, + 1142, + 286, + 1142 + ], + "score": 0.721 + }, + { + "category_id": 8, + "poly": [ + 566, + 907, + 1138, + 907, + 1138, + 1006, + 566, + 1006 + ], + "score": 0.715 + }, + { + "category_id": 8, + "poly": [ + 619, + 1921, + 1137, + 1921, + 1137, + 1960, + 619, + 1960 + ], + "score": 0.616 + }, + { + "category_id": 8, + "poly": [ + 293, + 255, + 1403, + 255, + 1403, + 327, + 293, + 327 + ], + "score": 0.561 + }, + { + "category_id": 8, + "poly": [ + 469, + 847, + 989, + 847, + 989, + 889, + 469, + 889 + ], + "score": 0.403 + }, + { + "category_id": 8, + "poly": [ + 566, + 907, + 1013, + 907, + 1013, + 959, + 566, + 959 + ], + "score": 0.225 + }, + { + "category_id": 8, + "poly": [ + 573, + 1861, + 1136, + 1861, + 1136, + 1959, + 573, + 1959 + ], + "score": 0.175 + }, + { + "category_id": 9, + "poly": [ + 1351, + 943, + 1401, + 943, + 1401, + 974, + 1351, + 974 + ], + "score": 0.1 + }, + { + "category_id": 13, + "poly": [ + 341, + 1497, + 801, + 1497, + 801, + 1538, + 341, + 1538 + ], + "score": 0.93, + "latex": "\\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) = A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } A _ { C } ( s , a )" + }, + { + "category_id": 13, + "poly": [ + 911, + 649, + 960, + 649, + 960, + 692, + 911, + 692 + ], + "score": 0.92, + "latex": "\\frac { \\lambda } { 1 - \\gamma }" + }, + { + "category_id": 13, + "poly": [ + 802, + 424, + 920, + 424, + 920, + 463, + 802, + 463 + ], + "score": 0.92, + "latex": "\\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 14, + "poly": [ + 673, + 1786, + 1027, + 1786, + 1027, + 1829, + 673, + 1829 + ], + "score": 0.91, + "latex": "\\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) = \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } )" + }, + { + "category_id": 13, + "poly": [ + 428, + 426, + 545, + 426, + 545, + 463, + 428, + 463 + ], + "score": 0.91, + "latex": "V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } \\left( \\mu \\right)" + }, + { + "category_id": 14, + "poly": [ + 457, + 1394, + 1241, + 1394, + 1241, + 1482, + 457, + 1482 + ], + "score": 0.91, + "latex": "\\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } ( \\gamma ) ^ { t ^ { \\prime } - t } \\Delta _ { t ^ { \\prime } } \\quad \\Delta _ { t ^ { \\prime } } = r ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } ) + \\gamma V _ { R } ( s _ { t ^ { \\prime } + 1 } ) - V _ { R } ( s _ { t ^ { \\prime } } )" + }, + { + "category_id": 13, + "poly": [ + 1052, + 424, + 1221, + 424, + 1221, + 463, + 1052, + 463 + ], + "score": 0.91, + "latex": "\\chi - \\hat { V } _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu )" + }, + { + "category_id": 14, + "poly": [ + 310, + 254, + 1406, + 254, + 1406, + 376, + 310, + 376 + ], + "score": 0.91, + "latex": "\\begin{array} { r l } { } & { \\underset { \\pi _ { \\phi } } { \\mathrm { a a x } } \\underset { \\lambda \\geq 0 } { \\mathrm { m i n } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\Big ] - \\lambda \\left( V _ { C } ^ { \\pi _ { \\phi _ { o l d } } } ( \\mu ) + \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\Big [ \\hat { A } _ { C } ( s , a ) \\Big ] - \\chi \\right) } \\\\ { : \\cdot \\hfill } & { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 568, + 1355, + 643, + 1355, + 643, + 1394, + 568, + 1394 + ], + "score": 0.9, + "latex": "A _ { R } ^ { \\pi _ { \\phi _ { o l d } } }" + }, + { + "category_id": 13, + "poly": [ + 804, + 1653, + 881, + 1653, + 881, + 1685, + 804, + 1685 + ], + "score": 0.9, + "latex": "\\tilde { A } ^ { \\pi _ { \\phi _ { o l d } } }" + }, + { + "category_id": 13, + "poly": [ + 359, + 1991, + 413, + 1991, + 413, + 2022, + 359, + 2022 + ], + "score": 0.9, + "latex": "D _ { \\mathrm { K L } }" + }, + { + "category_id": 13, + "poly": [ + 1052, + 1007, + 1302, + 1007, + 1302, + 1042, + 1052, + 1042 + ], + "score": 0.9, + "latex": "\\nabla _ { \\phi } \\pi _ { \\phi } = \\pi _ { \\phi } \\nabla _ { \\phi } \\log \\pi _ { \\phi }" + }, + { + "category_id": 14, + "poly": [ + 468, + 529, + 1201, + 529, + 1201, + 650, + 468, + 650 + ], + "score": 0.9, + "latex": "\\begin{array} { r l } & { \\underset { \\pi _ { \\phi } } { \\operatorname* { m a x } } \\underset { \\lambda \\geq 0 } { \\operatorname* { m i n } } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi } } \\left[ \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\frac { \\lambda } { 1 - \\gamma } \\hat { A } _ { C } ( s , a ) \\right] + \\lambda \\chi ^ { \\prime } } \\\\ & { s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 474, + 741, + 554, + 741, + 554, + 770, + 474, + 770 + ], + "score": 0.9, + "latex": "\\operatorname* { m a x } _ { \\pi _ { \\phi } }" + }, + { + "category_id": 13, + "poly": [ + 421, + 692, + 532, + 692, + 532, + 722, + 421, + 722 + ], + "score": 0.9, + "latex": "a \\sim \\pi _ { \\phi _ { o l d } }" + }, + { + "category_id": 14, + "poly": [ + 317, + 1039, + 1334, + 1039, + 1334, + 1114, + 317, + 1114 + ], + "score": 0.9, + "latex": "\\nabla _ { \\phi } \\bar { A } ( \\phi ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\frac { \\pi _ { \\phi } ( a | s ) } { \\pi _ { \\phi _ { o l d } } ( a | s ) } \\nabla _ { \\phi } \\log \\pi _ { \\phi } ( a | s ) ( A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) ) \\right] \\right]" + }, + { + "category_id": 14, + "poly": [ + 339, + 1563, + 1316, + 1563, + 1316, + 1654, + 339, + 1654 + ], + "score": 0.89, + "latex": "\\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } = \\sum _ { t ^ { \\prime } = t } ^ { \\infty } ( \\gamma ) ^ { t ^ { \\prime } - t } \\Delta _ { t ^ { \\prime } } ~ \\Delta _ { t ^ { \\prime } } = r ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } ) + \\gamma V _ { R } ( s _ { t ^ { \\prime } + 1 } ) - V _ { R } ( s _ { t ^ { \\prime } } ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s _ { t ^ { \\prime } } , a _ { t ^ { \\prime } } )" + }, + { + "category_id": 14, + "poly": [ + 466, + 1688, + 1231, + 1688, + 1231, + 1747, + 466, + 1747 + ], + "score": 0.88, + "latex": "\\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } \\right] \\right]" + }, + { + "category_id": 13, + "poly": [ + 297, + 690, + 382, + 690, + 382, + 722, + 297, + 722 + ], + "score": 0.88, + "latex": "a \\sim \\pi _ { \\phi }" + }, + { + "category_id": 13, + "poly": [ + 298, + 462, + 324, + 462, + 324, + 491, + 298, + 491 + ], + "score": 0.88, + "latex": "\\chi ^ { \\prime }" + }, + { + "category_id": 14, + "poly": [ + 436, + 1280, + 1260, + 1280, + 1260, + 1339, + 436, + 1339 + ], + "score": 0.87, + "latex": "\\nabla _ { \\phi _ { o l d } } a r A ( \\phi _ { o l d } ) = \\overline { { \\nabla _ { \\phi _ { o l d } } J ( \\phi _ { o l d } ) + \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) \\right] \\right] } }" + }, + { + "category_id": 13, + "poly": [ + 1084, + 1113, + 1187, + 1113, + 1187, + 1143, + 1084, + 1143 + ], + "score": 0.87, + "latex": "\\phi = \\phi _ { o l d }" + }, + { + "category_id": 13, + "poly": [ + 1255, + 429, + 1283, + 429, + 1283, + 461, + 1255, + 461 + ], + "score": 0.85, + "latex": "\\chi ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 645, + 461, + 665, + 461, + 665, + 491, + 645, + 491 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 651, + 652, + 679, + 652, + 679, + 680, + 651, + 680 + ], + "score": 0.84, + "latex": "\\lambda ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 1328, + 737, + 1347, + 737, + 1347, + 767, + 1328, + 767 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 891, + 1992, + 914, + 1992, + 914, + 2018, + 891, + 2018 + ], + "score": 0.81, + "latex": "\\mathbf { F }" + }, + { + "category_id": 14, + "poly": [ + 414, + 768, + 1281, + 768, + 1281, + 891, + 414, + 891 + ], + "score": 0.75, + "latex": "\\begin{array} { r l } & { \\phi \\longleftarrow \\arg \\operatorname* { m a x } _ { \\phi } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\frac { \\pi _ { \\phi } ( a | s ) } { \\pi _ { \\phi _ { o l d } } ( a | s ) } ( A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) ) \\right] \\right] } \\\\ & { s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 564, + 1859, + 1010, + 1859, + 1010, + 1916, + 564, + 1916 + ], + "score": 0.71, + "latex": "\\phi \\arg \\operatorname* { m a x } _ { \\phi } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } ( \\phi - \\phi _ { o l d } )" + }, + { + "category_id": 14, + "poly": [ + 324, + 1143, + 1336, + 1143, + 1336, + 1200, + 324, + 1200 + ], + "score": 0.61, + "latex": "\\begin{array} { r } { \\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) ( A _ { R } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) - \\lambda ^ { \\prime } \\hat { A } _ { C } ( s , a ) ) \\right] \\right] } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 570, + 1855, + 1136, + 1855, + 1136, + 1963, + 570, + 1963 + ], + "score": 0.45, + "latex": "\\begin{array} { r l } & { \\phi \\underset { \\phi } { \\mathrm { a r g } \\mathrm { m a x } } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } ( \\phi - \\phi _ { o l d } ) } \\\\ & { \\quad s . t . \\quad \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 564, + 906, + 1011, + 906, + 1011, + 962, + 564, + 962 + ], + "score": 0.39, + "latex": "\\phi \\arg \\operatorname* { m a x } _ { \\phi } \\nabla _ { \\phi _ { o l d } } \\bar { A } ( \\phi _ { o l d } ) ^ { T } ( \\phi - \\phi _ { o l d } )" + }, + { + "category_id": 14, + "poly": [ + 453, + 1206, + 1267, + 1206, + 1267, + 1257, + 453, + 1257 + ], + "score": 0.39, + "latex": "\\begin{array} { r l } { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left. \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left. \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right. \\right. = \\nabla _ { \\phi _ { o l d } } J ( \\phi _ { o l d } ) } & { { } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 605, + 1922, + 1133, + 1922, + 1133, + 1961, + 605, + 1961 + ], + "score": 0.39, + "latex": "\\begin{array} { r } { \\begin{array} { r l } { s . t . } & { { } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array} } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 455, + 1207, + 1268, + 1207, + 1268, + 1256, + 455, + 1256 + ], + "score": 0.33, + "latex": "\\begin{array} { r l } { \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left. \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left. \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( a | s ) \\hat { A } ^ { \\pi _ { \\phi _ { o l d } } } ( s , a ) \\right. \\right. = \\nabla _ { \\phi _ { o l d } } J ( \\phi _ { o l d } ) } & { { } } \\end{array}" + }, + { + "category_id": 14, + "poly": [ + 600, + 964, + 1133, + 964, + 1133, + 1009, + 600, + 1009 + ], + "score": 0.3, + "latex": "\\begin{array} { r } { \\begin{array} { r l } { s . t . } & { { } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } [ D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) ] \\leq \\delta } \\end{array} } \\end{array}" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 108.0, + 296.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 869.0, + 2084.0, + 869.0, + 2124.0, + 830.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 417.0, + 427.0, + 417.0, + 427.0, + 468.0, + 293.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 417.0, + 801.0, + 417.0, + 801.0, + 468.0, + 546.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 417.0, + 1051.0, + 417.0, + 1051.0, + 468.0, + 921.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1222.0, + 417.0, + 1254.0, + 417.0, + 1254.0, + 468.0, + 1222.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 417.0, + 1409.0, + 417.0, + 1409.0, + 468.0, + 1284.0, + 468.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 455.0, + 297.0, + 455.0, + 297.0, + 496.0, + 292.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 455.0, + 644.0, + 455.0, + 644.0, + 496.0, + 325.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 666.0, + 455.0, + 1288.0, + 455.0, + 1288.0, + 496.0, + 666.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 640.0, + 650.0, + 640.0, + 650.0, + 699.0, + 289.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 640.0, + 910.0, + 640.0, + 910.0, + 699.0, + 680.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 961.0, + 640.0, + 1409.0, + 640.0, + 1409.0, + 699.0, + 961.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 686.0, + 296.0, + 686.0, + 296.0, + 722.0, + 293.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 686.0, + 420.0, + 686.0, + 420.0, + 722.0, + 383.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 533.0, + 686.0, + 1207.0, + 686.0, + 1207.0, + 722.0, + 533.0, + 722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1957.0, + 1401.0, + 1957.0, + 1401.0, + 1992.0, + 296.0, + 1992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1988.0, + 358.0, + 1988.0, + 358.0, + 2025.0, + 292.0, + 2025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 414.0, + 1988.0, + 890.0, + 1988.0, + 890.0, + 2025.0, + 414.0, + 2025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1988.0, + 1404.0, + 1988.0, + 1404.0, + 2025.0, + 915.0, + 2025.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 284.0, + 1482.0, + 340.0, + 1482.0, + 340.0, + 1568.0, + 284.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 802.0, + 1482.0, + 1414.0, + 1482.0, + 1414.0, + 1568.0, + 802.0, + 1568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1650.0, + 803.0, + 1650.0, + 803.0, + 1696.0, + 294.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1650.0, + 994.0, + 1650.0, + 994.0, + 1696.0, + 882.0, + 1696.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 732.0, + 473.0, + 732.0, + 473.0, + 771.0, + 294.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 555.0, + 732.0, + 1327.0, + 732.0, + 1327.0, + 771.0, + 555.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1348.0, + 732.0, + 1351.0, + 732.0, + 1351.0, + 771.0, + 1348.0, + 771.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 996.0, + 1051.0, + 996.0, + 1051.0, + 1047.0, + 292.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1303.0, + 996.0, + 1406.0, + 996.0, + 1406.0, + 1047.0, + 1303.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1827.0, + 708.0, + 1827.0, + 708.0, + 1864.0, + 296.0, + 1864.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1269.0, + 1198.0, + 1408.0, + 1198.0, + 1408.0, + 1253.0, + 1269.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1247.0, + 1219.0, + 1247.0, + 1219.0, + 1287.0, + 296.0, + 1287.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.75, + 1198.0, + 564.75, + 1198.0, + 564.75, + 1253.5, + 292.75, + 1253.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 565.25, + 1204.5, + 683.25, + 1204.5, + 683.25, + 1252.5, + 565.25, + 1252.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 283.0, + 1339.0, + 567.0, + 1339.0, + 567.0, + 1414.0, + 283.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 1339.0, + 1278.0, + 1339.0, + 1278.0, + 1414.0, + 644.0, + 1414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 326.0, + 309.0, + 326.0, + 309.0, + 381.0, + 289.0, + 381.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1103.0, + 1083.0, + 1103.0, + 1083.0, + 1153.0, + 294.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1188.0, + 1103.0, + 1193.0, + 1103.0, + 1193.0, + 1153.0, + 1188.0, + 1153.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 21, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1033, + 1405, + 1033, + 1405, + 1284, + 297, + 1284 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 682, + 1402, + 682, + 1402, + 805, + 297, + 805 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 297, + 1498, + 1405, + 1498, + 1405, + 1724, + 297, + 1724 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1737, + 1404, + 1737, + 1404, + 1892, + 298, + 1892 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 297, + 1377, + 1405, + 1377, + 1405, + 1485, + 297, + 1485 + ], + "score": 0.973 + }, + { + "category_id": 1, + "poly": [ + 299, + 879, + 1405, + 879, + 1405, + 941, + 299, + 941 + ], + "score": 0.949 + }, + { + "category_id": 8, + "poly": [ + 539, + 809, + 1128, + 809, + 1128, + 876, + 539, + 876 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 542, + 265, + 1159, + 265, + 1159, + 323, + 542, + 323 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 301, + 529, + 1400, + 529, + 1400, + 590, + 301, + 590 + ], + "score": 0.942 + }, + { + "category_id": 8, + "poly": [ + 433, + 445, + 1263, + 445, + 1263, + 528, + 433, + 528 + ], + "score": 0.928 + }, + { + "category_id": 8, + "poly": [ + 416, + 594, + 1281, + 594, + 1281, + 679, + 416, + 679 + ], + "score": 0.924 + }, + { + "category_id": 1, + "poly": [ + 299, + 228, + 1249, + 228, + 1249, + 262, + 299, + 262 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 300, + 977, + 601, + 977, + 601, + 1009, + 300, + 1009 + ], + "score": 0.92 + }, + { + "category_id": 8, + "poly": [ + 497, + 356, + 1199, + 356, + 1199, + 407, + 497, + 407 + ], + "score": 0.917 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 814, + 74, + 814, + 106, + 298, + 106 + ], + "score": 0.915 + }, + { + "category_id": 1, + "poly": [ + 298, + 322, + 784, + 322, + 784, + 354, + 298, + 354 + ], + "score": 0.91 + }, + { + "category_id": 0, + "poly": [ + 303, + 1321, + 725, + 1321, + 725, + 1352, + 303, + 1352 + ], + "score": 0.903 + }, + { + "category_id": 9, + "poly": [ + 1351, + 825, + 1400, + 825, + 1400, + 856, + 1351, + 856 + ], + "score": 0.885 + }, + { + "category_id": 9, + "poly": [ + 1351, + 277, + 1401, + 277, + 1401, + 307, + 1351, + 307 + ], + "score": 0.882 + }, + { + "category_id": 9, + "poly": [ + 1351, + 469, + 1400, + 469, + 1400, + 500, + 1351, + 500 + ], + "score": 0.881 + }, + { + "category_id": 9, + "poly": [ + 1351, + 620, + 1400, + 620, + 1400, + 650, + 1351, + 650 + ], + "score": 0.88 + }, + { + "category_id": 9, + "poly": [ + 1351, + 366, + 1401, + 366, + 1401, + 397, + 1351, + 397 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 864, + 2088, + 864, + 2112, + 835, + 2112 + ], + "score": 0.858 + }, + { + "category_id": 1, + "poly": [ + 284, + 410, + 1260, + 410, + 1260, + 440, + 284, + 440 + ], + "score": 0.8 + }, + { + "category_id": 13, + "poly": [ + 665, + 1439, + 862, + 1439, + 862, + 1489, + 665, + 1489 + ], + "score": 0.94, + "latex": "\\begin{array} { r } { \\delta \\ge \\frac { ( 1 - \\gamma ) ^ { 4 } ( 2 - \\bar { \\chi } ) ^ { 2 } } { 8 \\gamma ^ { 2 } } } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 684, + 232, + 788, + 232, + 788, + 263, + 684, + 263 + ], + "score": 0.93, + "latex": "\\phi = \\phi _ { o l d }" + }, + { + "category_id": 14, + "poly": [ + 434, + 441, + 1265, + 441, + 1265, + 530, + 434, + 530 + ], + "score": 0.93, + "latex": "\\phi \\phi _ { o l d } + \\beta \\mathbf { F } ^ { - 1 } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) \\quad \\beta = \\sqrt { \\frac { 2 \\delta } { \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } \\mathbf { F } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) } }" + }, + { + "category_id": 14, + "poly": [ + 416, + 589, + 1285, + 589, + 1285, + 681, + 416, + 681 + ], + "score": 0.93, + "latex": "\\phi \\phi _ { o l d } + \\beta { \\pmb { F } } ^ { - 1 } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) \\quad \\beta = \\beta ^ { j } \\sqrt { \\frac { 2 \\delta } { \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) ^ { T } { \\pmb { F } } \\nabla _ { \\phi _ { o l d } } \\tilde { J } ( \\phi _ { o l d } ) } }" + }, + { + "category_id": 14, + "poly": [ + 538, + 804, + 1129, + 804, + 1129, + 879, + 538, + 879 + ], + "score": 0.93, + "latex": "\\lambda \\lambda - ( \\frac { 1 } { 1 - \\gamma } \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } , a \\sim \\pi _ { \\phi _ { o l d } } } [ \\hat { A } _ { C } ( s , a ) ] - \\chi ^ { \\prime } )" + }, + { + "category_id": 13, + "poly": [ + 1187, + 1591, + 1394, + 1591, + 1394, + 1633, + 1187, + 1633 + ], + "score": 0.92, + "latex": "V _ { R } ^ { * } ( \\mu ) \\stackrel { - } { - } V _ { R } ^ { ( T ) } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 297, + 1097, + 388, + 1097, + 388, + 1134, + 297, + 1134 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu )" + }, + { + "category_id": 13, + "poly": [ + 980, + 1064, + 1135, + 1064, + 1135, + 1100, + 980, + 1100 + ], + "score": 0.92, + "latex": "V _ { C } ^ { \\pi _ { \\phi } } ( \\mu ) \\leq \\chi" + }, + { + "category_id": 13, + "poly": [ + 507, + 1448, + 575, + 1448, + 575, + 1481, + 507, + 1481 + ], + "score": 0.92, + "latex": "\\zeta = 0" + }, + { + "category_id": 13, + "poly": [ + 299, + 714, + 419, + 714, + 419, + 746, + 299, + 746 + ], + "score": 0.92, + "latex": "j \\gets j + 1" + }, + { + "category_id": 13, + "poly": [ + 1133, + 680, + 1352, + 680, + 1352, + 716, + 1133, + 716 + ], + "score": 0.91, + "latex": "\\beta ^ { j } = \\beta ^ { j } ( 1 - \\beta ^ { j } ) ^ { j }" + }, + { + "category_id": 13, + "poly": [ + 1205, + 1563, + 1249, + 1563, + 1249, + 1594, + 1205, + 1594 + ], + "score": 0.91, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 850, + 1501, + 978, + 1501, + 978, + 1530, + 850, + 1530 + ], + "score": 0.91, + "latex": "\\delta > 1 \\mathsf { e } - 8" + }, + { + "category_id": 14, + "poly": [ + 498, + 352, + 1198, + 352, + 1198, + 410, + 498, + 410 + ], + "score": 0.91, + "latex": "\\mathbf { F } = \\mathbb { E } _ { s \\sim \\rho _ { \\phi _ { o l d } } } \\left[ \\mathbb { E } _ { a \\sim \\pi _ { \\phi _ { o l d } } } \\left[ \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ( \\nabla _ { \\phi _ { o l d } } \\log \\pi _ { \\phi _ { o l d } } ) ^ { T } \\right] \\right]" + }, + { + "category_id": 13, + "poly": [ + 575, + 1501, + 689, + 1501, + 689, + 1533, + 575, + 1533 + ], + "score": 0.9, + "latex": "\\chi = 0 . 0 5" + }, + { + "category_id": 13, + "poly": [ + 405, + 1502, + 521, + 1502, + 521, + 1533, + 405, + 1533 + ], + "score": 0.9, + "latex": "\\gamma = 0 . 9 9" + }, + { + "category_id": 13, + "poly": [ + 1332, + 1799, + 1375, + 1799, + 1375, + 1830, + 1332, + 1830 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 14, + "poly": [ + 540, + 261, + 1158, + 261, + 1158, + 325, + 540, + 325 + ], + "score": 0.9, + "latex": "D _ { \\mathrm { K L } } ( \\pi _ { \\phi _ { o l d } } ( \\cdot | s ) | | \\pi _ { \\phi } ( \\cdot | s ) ) = \\frac { 1 } { 2 } ( \\phi - \\phi _ { o l d } ) ^ { T } \\mathbf { F } ( \\phi - \\phi _ { o l d } )" + }, + { + "category_id": 13, + "poly": [ + 728, + 746, + 837, + 746, + 837, + 776, + 728, + 776 + ], + "score": 0.9, + "latex": "\\phi \\phi _ { o l d }" + }, + { + "category_id": 13, + "poly": [ + 833, + 879, + 875, + 879, + 875, + 913, + 833, + 913 + ], + "score": 0.88, + "latex": "\\textstyle \\sum _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1001, + 531, + 1054, + 531, + 1054, + 559, + 1001, + 559 + ], + "score": 0.88, + "latex": "D _ { \\mathrm { K L } }" + }, + { + "category_id": 13, + "poly": [ + 731, + 715, + 835, + 715, + 835, + 744, + 731, + 744 + ], + "score": 0.88, + "latex": "\\bar { D } _ { \\mathrm { K L } } \\leq \\delta" + }, + { + "category_id": 13, + "poly": [ + 797, + 1162, + 834, + 1162, + 834, + 1192, + 797, + 1192 + ], + "score": 0.87, + "latex": "V _ { C }" + }, + { + "category_id": 13, + "poly": [ + 669, + 682, + 881, + 682, + 881, + 715, + 669, + 715 + ], + "score": 0.87, + "latex": "\\bar { D } _ { \\mathrm { K L } } ( \\phi | | \\phi _ { o l d } ) \\leq \\delta" + }, + { + "category_id": 13, + "poly": [ + 1266, + 1633, + 1309, + 1633, + 1309, + 1662, + 1266, + 1662 + ], + "score": 0.86, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 473, + 911, + 545, + 911, + 545, + 940, + 473, + 940 + ], + "score": 0.85, + "latex": "a _ { t } , s _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 623, + 745, + 643, + 745, + 643, + 775, + 623, + 775 + ], + "score": 0.84, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 584, + 715, + 607, + 715, + 607, + 741, + 584, + 741 + ], + "score": 0.82, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 877, + 411, + 897, + 411, + 897, + 440, + 877, + 440 + ], + "score": 0.82, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 504, + 1537, + 526, + 1537, + 526, + 1559, + 504, + 1559 + ], + "score": 0.81, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 414, + 776, + 433, + 776, + 433, + 802, + 414, + 802 + ], + "score": 0.81, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 1270, + 717, + 1292, + 717, + 1292, + 741, + 1270, + 741 + ], + "score": 0.8, + "latex": "L" + }, + { + "category_id": 13, + "poly": [ + 1293, + 1532, + 1319, + 1532, + 1319, + 1559, + 1293, + 1559 + ], + "score": 0.8, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 385, + 324, + 408, + 324, + 408, + 350, + 385, + 350 + ], + "score": 0.77, + "latex": "\\mathbf { F }" + }, + { + "category_id": 13, + "poly": [ + 496, + 1600, + 523, + 1600, + 523, + 1628, + 496, + 1628 + ], + "score": 0.76, + "latex": "\\Delta" + }, + { + "category_id": 13, + "poly": [ + 1007, + 1380, + 1023, + 1380, + 1023, + 1407, + 1007, + 1407 + ], + "score": 0.75, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 896, + 231, + 936, + 231, + 936, + 258, + 896, + 258 + ], + "score": 0.28, + "latex": "\\mathrm { K L }" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 971.0, + 605.0, + 971.0, + 605.0, + 1017.0, + 293.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 72.0, + 815.0, + 72.0, + 815.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1321.0, + 729.0, + 1321.0, + 729.0, + 1355.0, + 298.0, + 1355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 828.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 828.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1032.0, + 1406.0, + 1032.0, + 1406.0, + 1071.0, + 294.0, + 1071.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1062.0, + 979.0, + 1062.0, + 979.0, + 1104.0, + 294.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1062.0, + 1407.0, + 1062.0, + 1407.0, + 1104.0, + 1136.0, + 1104.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 389.0, + 1094.0, + 1407.0, + 1094.0, + 1407.0, + 1140.0, + 389.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1130.0, + 1403.0, + 1130.0, + 1403.0, + 1163.0, + 294.0, + 1163.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1160.0, + 796.0, + 1160.0, + 796.0, + 1197.0, + 294.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 1160.0, + 1405.0, + 1160.0, + 1405.0, + 1197.0, + 835.0, + 1197.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1191.0, + 1403.0, + 1191.0, + 1403.0, + 1226.0, + 295.0, + 1226.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1219.0, + 1406.0, + 1219.0, + 1406.0, + 1257.0, + 294.0, + 1257.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1254.0, + 604.0, + 1254.0, + 604.0, + 1284.0, + 297.0, + 1284.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 678.0, + 668.0, + 678.0, + 668.0, + 717.0, + 293.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 678.0, + 1132.0, + 678.0, + 1132.0, + 717.0, + 882.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1353.0, + 678.0, + 1406.0, + 678.0, + 1406.0, + 717.0, + 1353.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 712.0, + 298.0, + 712.0, + 298.0, + 748.0, + 291.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 420.0, + 712.0, + 583.0, + 712.0, + 583.0, + 748.0, + 420.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 712.0, + 730.0, + 712.0, + 730.0, + 748.0, + 608.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 712.0, + 1269.0, + 712.0, + 1269.0, + 748.0, + 836.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 712.0, + 1407.0, + 712.0, + 1407.0, + 748.0, + 1293.0, + 748.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 740.0, + 622.0, + 740.0, + 622.0, + 781.0, + 292.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 644.0, + 740.0, + 727.0, + 740.0, + 727.0, + 781.0, + 644.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 740.0, + 1406.0, + 740.0, + 1406.0, + 781.0, + 838.0, + 781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 774.0, + 413.0, + 774.0, + 413.0, + 810.0, + 295.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 434.0, + 774.0, + 744.0, + 774.0, + 744.0, + 810.0, + 434.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1499.0, + 404.0, + 1499.0, + 404.0, + 1535.0, + 294.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1499.0, + 574.0, + 1499.0, + 574.0, + 1535.0, + 522.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 1499.0, + 849.0, + 1499.0, + 849.0, + 1535.0, + 690.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 979.0, + 1499.0, + 1406.0, + 1499.0, + 1406.0, + 1535.0, + 979.0, + 1535.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1531.0, + 503.0, + 1531.0, + 503.0, + 1565.0, + 294.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 527.0, + 1531.0, + 1292.0, + 1531.0, + 1292.0, + 1565.0, + 527.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1320.0, + 1531.0, + 1405.0, + 1531.0, + 1405.0, + 1565.0, + 1320.0, + 1565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1561.0, + 1204.0, + 1561.0, + 1204.0, + 1597.0, + 295.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1561.0, + 1403.0, + 1561.0, + 1403.0, + 1597.0, + 1250.0, + 1597.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 1580.0, + 495.0, + 1580.0, + 495.0, + 1644.0, + 286.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 524.0, + 1580.0, + 1186.0, + 1580.0, + 1186.0, + 1644.0, + 524.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1580.0, + 1412.0, + 1580.0, + 1412.0, + 1644.0, + 1395.0, + 1644.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1628.0, + 1265.0, + 1628.0, + 1265.0, + 1667.0, + 292.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1310.0, + 1628.0, + 1407.0, + 1628.0, + 1407.0, + 1667.0, + 1310.0, + 1667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1659.0, + 1406.0, + 1659.0, + 1406.0, + 1694.0, + 294.0, + 1694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1691.0, + 1192.0, + 1691.0, + 1192.0, + 1727.0, + 295.0, + 1727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1736.0, + 1404.0, + 1736.0, + 1404.0, + 1773.0, + 294.0, + 1773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1765.0, + 1406.0, + 1765.0, + 1406.0, + 1801.0, + 293.0, + 1801.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1797.0, + 1331.0, + 1797.0, + 1331.0, + 1834.0, + 296.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1797.0, + 1406.0, + 1797.0, + 1406.0, + 1834.0, + 1376.0, + 1834.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1822.0, + 1405.0, + 1822.0, + 1405.0, + 1868.0, + 292.0, + 1868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1858.0, + 697.0, + 1858.0, + 697.0, + 1894.0, + 295.0, + 1894.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1378.0, + 1006.0, + 1378.0, + 1006.0, + 1410.0, + 297.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1024.0, + 1378.0, + 1403.0, + 1378.0, + 1403.0, + 1410.0, + 1024.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1410.0, + 1403.0, + 1410.0, + 1403.0, + 1443.0, + 296.0, + 1443.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1447.0, + 506.0, + 1447.0, + 506.0, + 1480.0, + 293.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 1447.0, + 664.0, + 1447.0, + 664.0, + 1480.0, + 576.0, + 1480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.75, + 1435.0, + 1165.75, + 1435.0, + 1165.75, + 1488.5, + 707.75, + 1488.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 875.0, + 832.0, + 875.0, + 832.0, + 914.0, + 294.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 876.0, + 875.0, + 1405.0, + 875.0, + 1405.0, + 914.0, + 876.0, + 914.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 906.0, + 472.0, + 906.0, + 472.0, + 942.0, + 295.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 546.0, + 906.0, + 960.0, + 906.0, + 960.0, + 942.0, + 546.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 526.0, + 1000.0, + 526.0, + 1000.0, + 563.0, + 292.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 526.0, + 1405.0, + 526.0, + 1405.0, + 563.0, + 1055.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 558.0, + 641.0, + 558.0, + 641.0, + 593.0, + 295.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 224.0, + 683.0, + 224.0, + 683.0, + 271.0, + 292.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 224.0, + 895.0, + 224.0, + 895.0, + 271.0, + 789.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 224.0, + 1254.0, + 224.0, + 1254.0, + 271.0, + 937.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 318.0, + 384.0, + 318.0, + 384.0, + 358.0, + 296.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 318.0, + 785.0, + 318.0, + 785.0, + 358.0, + 409.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 407.0, + 876.0, + 407.0, + 876.0, + 445.0, + 295.0, + 445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 898.0, + 407.0, + 1255.0, + 407.0, + 1255.0, + 445.0, + 898.0, + 445.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 22, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1429, + 1405, + 1429, + 1405, + 1708, + 297, + 1708 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 300, + 286, + 1404, + 286, + 1404, + 409, + 300, + 409 + ], + "score": 0.957 + }, + { + "category_id": 0, + "poly": [ + 302, + 1372, + 717, + 1372, + 717, + 1403, + 302, + 1403 + ], + "score": 0.892 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.887 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2112, + 834, + 2112 + ], + "score": 0.85 + }, + { + "category_id": 0, + "poly": [ + 301, + 230, + 816, + 230, + 816, + 261, + 301, + 261 + ], + "score": 0.719 + }, + { + "category_id": 1, + "poly": [ + 366, + 429, + 1407, + 429, + 1407, + 1351, + 366, + 1351 + ], + "score": 0.658 + }, + { + "category_id": 1, + "poly": [ + 301, + 230, + 816, + 230, + 816, + 261, + 301, + 261 + ], + "score": 0.123 + }, + { + "category_id": 13, + "poly": [ + 1062, + 1581, + 1111, + 1581, + 1111, + 1616, + 1062, + 1616 + ], + "score": 0.92, + "latex": "\\beta ^ { ( 0 ) }" + }, + { + "category_id": 13, + "poly": [ + 855, + 1431, + 898, + 1431, + 898, + 1461, + 855, + 1461 + ], + "score": 0.9, + "latex": "Q _ { C }" + }, + { + "category_id": 13, + "poly": [ + 497, + 1522, + 551, + 1522, + 551, + 1552, + 497, + 1552 + ], + "score": 0.9, + "latex": "D _ { \\mathrm { K L } }" + }, + { + "category_id": 13, + "poly": [ + 655, + 1616, + 741, + 1616, + 741, + 1644, + 655, + 1644 + ], + "score": 0.9, + "latex": "L = 2 0" + }, + { + "category_id": 13, + "poly": [ + 1226, + 988, + 1294, + 988, + 1294, + 1021, + 1226, + 1021 + ], + "score": 0.89, + "latex": "( x , y )" + }, + { + "category_id": 13, + "poly": [ + 1270, + 1492, + 1351, + 1492, + 1351, + 1520, + 1270, + 1520 + ], + "score": 0.87, + "latex": "4 e - 2" + }, + { + "category_id": 13, + "poly": [ + 346, + 1461, + 428, + 1461, + 428, + 1490, + 346, + 1490 + ], + "score": 0.87, + "latex": "2 e - 4" + }, + { + "category_id": 13, + "poly": [ + 963, + 1431, + 1041, + 1431, + 1041, + 1459, + 963, + 1459 + ], + "score": 0.87, + "latex": "2 e - 4" + }, + { + "category_id": 13, + "poly": [ + 1324, + 1430, + 1402, + 1430, + 1402, + 1459, + 1324, + 1459 + ], + "score": 0.86, + "latex": "1 e - 4" + }, + { + "category_id": 13, + "poly": [ + 885, + 1495, + 917, + 1495, + 917, + 1522, + 885, + 1522 + ], + "score": 0.86, + "latex": "\\eta _ { \\lambda }" + }, + { + "category_id": 13, + "poly": [ + 596, + 1435, + 632, + 1435, + 632, + 1463, + 596, + 1463 + ], + "score": 0.85, + "latex": "\\eta _ { Q }" + }, + { + "category_id": 13, + "poly": [ + 1353, + 1465, + 1375, + 1465, + 1375, + 1493, + 1353, + 1493 + ], + "score": 0.83, + "latex": "\\gamma" + }, + { + "category_id": 13, + "poly": [ + 1220, + 1526, + 1241, + 1526, + 1241, + 1549, + 1220, + 1549 + ], + "score": 0.79, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1150, + 1492, + 1170, + 1492, + 1170, + 1518, + 1150, + 1518 + ], + "score": 0.78, + "latex": "\\lambda" + }, + { + "category_id": 13, + "poly": [ + 839, + 1617, + 863, + 1617, + 863, + 1647, + 839, + 1647 + ], + "score": 0.77, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 680, + 1556, + 702, + 1556, + 702, + 1580, + 680, + 1580 + ], + "score": 0.77, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 395, + 1522, + 412, + 1522, + 412, + 1549, + 395, + 1549 + ], + "score": 0.77, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 1116, + 1020, + 1254, + 1020, + 1254, + 1048, + 1116, + 1048 + ], + "score": 0.66, + "latex": "1 5 \\mathrm { c m } \\mathrm { x } 3 5 \\mathrm { c m }" + }, + { + "category_id": 13, + "poly": [ + 674, + 795, + 814, + 795, + 814, + 824, + 674, + 824 + ], + "score": 0.29, + "latex": "2 0 \\mathrm { c m } \\mathrm { x } 4 0 \\mathrm { c m }" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1372.0, + 722.0, + 1372.0, + 722.0, + 1405.0, + 297.0, + 1405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 872.0, + 2084.0, + 872.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 823.0, + 230.0, + 823.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1428.0, + 595.0, + 1428.0, + 595.0, + 1465.0, + 296.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1428.0, + 854.0, + 1428.0, + 854.0, + 1465.0, + 633.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 899.0, + 1428.0, + 962.0, + 1428.0, + 962.0, + 1465.0, + 899.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1428.0, + 1323.0, + 1428.0, + 1323.0, + 1465.0, + 1042.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1403.0, + 1428.0, + 1407.0, + 1428.0, + 1407.0, + 1465.0, + 1403.0, + 1465.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1459.0, + 345.0, + 1459.0, + 345.0, + 1492.0, + 295.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 429.0, + 1459.0, + 1352.0, + 1459.0, + 1352.0, + 1492.0, + 429.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1376.0, + 1459.0, + 1403.0, + 1459.0, + 1403.0, + 1492.0, + 1376.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1490.0, + 884.0, + 1490.0, + 884.0, + 1526.0, + 292.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 918.0, + 1490.0, + 1149.0, + 1490.0, + 1149.0, + 1526.0, + 918.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1171.0, + 1490.0, + 1269.0, + 1490.0, + 1269.0, + 1526.0, + 1171.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1490.0, + 1407.0, + 1490.0, + 1407.0, + 1526.0, + 1352.0, + 1526.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1520.0, + 394.0, + 1520.0, + 394.0, + 1557.0, + 295.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 413.0, + 1520.0, + 496.0, + 1520.0, + 496.0, + 1557.0, + 413.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 1520.0, + 1219.0, + 1520.0, + 1219.0, + 1557.0, + 552.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1242.0, + 1520.0, + 1405.0, + 1520.0, + 1405.0, + 1557.0, + 1242.0, + 1557.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1553.0, + 679.0, + 1553.0, + 679.0, + 1586.0, + 295.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 703.0, + 1553.0, + 1405.0, + 1553.0, + 1405.0, + 1586.0, + 703.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1578.0, + 1061.0, + 1578.0, + 1061.0, + 1622.0, + 291.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 1578.0, + 1407.0, + 1578.0, + 1407.0, + 1622.0, + 1112.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1616.0, + 654.0, + 1616.0, + 654.0, + 1649.0, + 295.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 742.0, + 1616.0, + 838.0, + 1616.0, + 838.0, + 1649.0, + 742.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 1616.0, + 1403.0, + 1616.0, + 1403.0, + 1649.0, + 864.0, + 1649.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1642.0, + 1406.0, + 1642.0, + 1406.0, + 1682.0, + 292.0, + 1682.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1678.0, + 1211.0, + 1678.0, + 1211.0, + 1711.0, + 295.0, + 1711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 286.0, + 1406.0, + 286.0, + 1406.0, + 321.0, + 294.0, + 321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 316.0, + 1405.0, + 316.0, + 1405.0, + 354.0, + 294.0, + 354.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 346.0, + 1404.0, + 346.0, + 1404.0, + 384.0, + 294.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 380.0, + 857.0, + 380.0, + 857.0, + 411.0, + 294.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 437.0, + 1406.0, + 437.0, + 1406.0, + 471.0, + 368.0, + 471.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 467.0, + 1405.0, + 467.0, + 1405.0, + 503.0, + 392.0, + 503.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 498.0, + 1405.0, + 498.0, + 1405.0, + 532.0, + 396.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 529.0, + 1403.0, + 529.0, + 1403.0, + 562.0, + 393.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 559.0, + 1405.0, + 559.0, + 1405.0, + 592.0, + 394.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 591.0, + 1209.0, + 591.0, + 1209.0, + 623.0, + 393.0, + 623.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 630.0, + 1407.0, + 630.0, + 1407.0, + 666.0, + 391.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 662.0, + 1404.0, + 662.0, + 1404.0, + 695.0, + 394.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 692.0, + 1229.0, + 692.0, + 1229.0, + 726.0, + 396.0, + 726.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 731.0, + 1405.0, + 731.0, + 1405.0, + 768.0, + 384.0, + 768.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 765.0, + 1405.0, + 765.0, + 1405.0, + 798.0, + 394.0, + 798.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 795.0, + 673.0, + 795.0, + 673.0, + 827.0, + 392.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 815.0, + 795.0, + 1403.0, + 795.0, + 1403.0, + 827.0, + 815.0, + 827.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 823.0, + 1404.0, + 823.0, + 1404.0, + 859.0, + 394.0, + 859.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 853.0, + 1404.0, + 853.0, + 1404.0, + 889.0, + 393.0, + 889.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 886.0, + 878.0, + 886.0, + 878.0, + 920.0, + 391.0, + 920.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 368.0, + 925.0, + 1405.0, + 925.0, + 1405.0, + 961.0, + 368.0, + 961.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 956.0, + 1405.0, + 956.0, + 1405.0, + 992.0, + 393.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 989.0, + 1225.0, + 989.0, + 1225.0, + 1022.0, + 394.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1295.0, + 989.0, + 1404.0, + 989.0, + 1404.0, + 1022.0, + 1295.0, + 1022.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1019.0, + 1115.0, + 1019.0, + 1115.0, + 1051.0, + 393.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 1019.0, + 1403.0, + 1019.0, + 1403.0, + 1051.0, + 1255.0, + 1051.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1051.0, + 1402.0, + 1051.0, + 1402.0, + 1080.0, + 396.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 396.0, + 1080.0, + 1404.0, + 1080.0, + 1404.0, + 1114.0, + 396.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 1109.0, + 588.0, + 1109.0, + 588.0, + 1142.0, + 395.0, + 1142.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 1149.0, + 1405.0, + 1149.0, + 1405.0, + 1188.0, + 377.0, + 1188.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1181.0, + 1406.0, + 1181.0, + 1406.0, + 1216.0, + 392.0, + 1216.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 1213.0, + 1405.0, + 1213.0, + 1405.0, + 1246.0, + 393.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1244.0, + 1404.0, + 1244.0, + 1404.0, + 1277.0, + 394.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 392.0, + 1273.0, + 1404.0, + 1273.0, + 1404.0, + 1306.0, + 392.0, + 1306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1301.0, + 474.0, + 1301.0, + 474.0, + 1338.0, + 394.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 230.0, + 823.0, + 230.0, + 823.0, + 265.0, + 296.0, + 265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 23, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 296, + 1398, + 296, + 1398, + 1230, + 301, + 1230 + ], + "score": 0.972 + }, + { + "category_id": 4, + "poly": [ + 296, + 1250, + 1405, + 1250, + 1405, + 1418, + 296, + 1418 + ], + "score": 0.943 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.889 + }, + { + "category_id": 2, + "poly": [ + 835, + 2088, + 865, + 2088, + 865, + 2112, + 835, + 2112 + ], + "score": 0.844 + }, + { + "category_id": 2, + "poly": [ + 331, + 234, + 1302, + 234, + 1302, + 260, + 331, + 260 + ], + "score": 0.553 + }, + { + "category_id": 13, + "poly": [ + 1109, + 1337, + 1129, + 1337, + 1129, + 1363, + 1109, + 1363 + ], + "score": 0.83, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 372, + 1365, + 393, + 1365, + 393, + 1391, + 372, + 1391 + ], + "score": 0.82, + "latex": "\\chi" + }, + { + "category_id": 13, + "poly": [ + 779, + 1311, + 800, + 1311, + 800, + 1336, + 779, + 1336 + ], + "score": 0.81, + "latex": "\\chi" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 303.0, + 340.0, + 303.0, + 340.0, + 401.0, + 301.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 307.0, + 396.0, + 307.0, + 396.0, + 335.0, + 353.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 303.0, + 704.0, + 303.0, + 704.0, + 376.0, + 664.0, + 376.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 307.0, + 1012.0, + 307.0, + 1012.0, + 335.0, + 953.0, + 335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 299.0, + 1080.0, + 299.0, + 1080.0, + 404.0, + 1032.0, + 404.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 308.0, + 1133.0, + 308.0, + 1133.0, + 337.0, + 1090.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 328.0, + 394.0, + 328.0, + 394.0, + 345.0, + 356.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 328.0, + 1012.0, + 328.0, + 1012.0, + 345.0, + 969.0, + 345.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 329.0, + 1133.0, + 329.0, + 1133.0, + 347.0, + 1090.0, + 347.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 339.0, + 398.0, + 339.0, + 398.0, + 355.0, + 356.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 970.0, + 339.0, + 1014.0, + 339.0, + 1014.0, + 357.0, + 970.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 340.0, + 1136.0, + 340.0, + 1136.0, + 358.0, + 1090.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 360.0, + 700.0, + 360.0, + 700.0, + 390.0, + 667.0, + 390.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 363.0, + 1071.0, + 363.0, + 1071.0, + 384.0, + 1050.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 363.0, + 1361.0, + 363.0, + 1361.0, + 374.0, + 1281.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1176.0, + 373.0, + 1376.0, + 373.0, + 1376.0, + 388.0, + 1176.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 393.0, + 331.0, + 393.0, + 331.0, + 405.0, + 322.0, + 405.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 411.0, + 351.0, + 411.0, + 351.0, + 421.0, + 342.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 408.0, + 624.0, + 408.0, + 624.0, + 434.0, + 373.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 409.0, + 660.0, + 409.0, + 660.0, + 422.0, + 631.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 394.0, + 697.0, + 394.0, + 697.0, + 409.0, + 679.0, + 409.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 410.0, + 717.0, + 410.0, + 717.0, + 421.0, + 707.0, + 421.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 397.0, + 987.0, + 397.0, + 987.0, + 440.0, + 739.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 409.0, + 1023.0, + 409.0, + 1023.0, + 422.0, + 994.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 398.0, + 1071.0, + 398.0, + 1071.0, + 413.0, + 1056.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 410.0, + 1089.0, + 410.0, + 1089.0, + 422.0, + 1078.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 399.0, + 1358.0, + 399.0, + 1358.0, + 440.0, + 1109.0, + 440.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1361.0, + 406.0, + 1393.0, + 406.0, + 1393.0, + 424.0, + 1361.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 450.0, + 624.0, + 450.0, + 624.0, + 478.0, + 344.0, + 478.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 450.0, + 1009.0, + 450.0, + 1009.0, + 481.0, + 685.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 450.0, + 1382.0, + 450.0, + 1382.0, + 481.0, + 1051.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 491.0, + 341.0, + 491.0, + 341.0, + 509.0, + 315.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 493.0, + 402.0, + 493.0, + 402.0, + 509.0, + 348.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 490.0, + 708.0, + 490.0, + 708.0, + 585.0, + 661.0, + 585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 493.0, + 1012.0, + 493.0, + 1012.0, + 509.0, + 969.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 488.0, + 1134.0, + 488.0, + 1134.0, + 512.0, + 1089.0, + 512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 504.0, + 402.0, + 504.0, + 402.0, + 521.0, + 345.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 957.0, + 503.0, + 1012.0, + 503.0, + 1012.0, + 521.0, + 957.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 501.0, + 1133.0, + 501.0, + 1133.0, + 588.0, + 1028.0, + 588.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 516.0, + 402.0, + 516.0, + 402.0, + 532.0, + 361.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 514.0, + 1012.0, + 514.0, + 1012.0, + 531.0, + 969.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 514.0, + 1133.0, + 514.0, + 1133.0, + 531.0, + 1090.0, + 531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 527.0, + 341.0, + 527.0, + 341.0, + 547.0, + 320.0, + 547.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 361.0, + 526.0, + 406.0, + 526.0, + 406.0, + 542.0, + 361.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 526.0, + 1014.0, + 526.0, + 1014.0, + 542.0, + 969.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 526.0, + 1136.0, + 526.0, + 1136.0, + 542.0, + 1090.0, + 542.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 544.0, + 341.0, + 544.0, + 341.0, + 564.0, + 320.0, + 564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 562.0, + 341.0, + 562.0, + 341.0, + 582.0, + 320.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 560.0, + 1071.0, + 560.0, + 1071.0, + 582.0, + 1049.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 580.0, + 341.0, + 580.0, + 341.0, + 599.0, + 320.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 573.0, + 699.0, + 573.0, + 699.0, + 591.0, + 677.0, + 591.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 584.0, + 1071.0, + 584.0, + 1071.0, + 599.0, + 1056.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 596.0, + 358.0, + 596.0, + 358.0, + 606.0, + 348.0, + 606.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 381.0, + 593.0, + 625.0, + 593.0, + 625.0, + 620.0, + 381.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 593.0, + 661.0, + 593.0, + 661.0, + 610.0, + 631.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 596.0, + 717.0, + 596.0, + 717.0, + 608.0, + 706.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 593.0, + 988.0, + 593.0, + 988.0, + 620.0, + 739.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 593.0, + 1024.0, + 593.0, + 1024.0, + 610.0, + 994.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 595.0, + 1089.0, + 595.0, + 1089.0, + 608.0, + 1078.0, + 608.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 593.0, + 1331.0, + 593.0, + 1331.0, + 620.0, + 1110.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 593.0, + 1393.0, + 593.0, + 1393.0, + 610.0, + 1363.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 636.0, + 620.0, + 636.0, + 620.0, + 666.0, + 348.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 636.0, + 1004.0, + 636.0, + 1004.0, + 667.0, + 690.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 635.0, + 1376.0, + 635.0, + 1376.0, + 666.0, + 1058.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 687.0, + 340.0, + 687.0, + 340.0, + 767.0, + 301.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 678.0, + 411.0, + 678.0, + 411.0, + 716.0, + 353.0, + 716.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 628.0, + 681.0, + 639.0, + 681.0, + 639.0, + 690.0, + 628.0, + 690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 681.0, + 704.0, + 681.0, + 704.0, + 767.0, + 664.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 678.0, + 1012.0, + 678.0, + 1012.0, + 717.0, + 953.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1032.0, + 680.0, + 1076.0, + 680.0, + 1076.0, + 746.0, + 1032.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 678.0, + 1133.0, + 678.0, + 1133.0, + 717.0, + 1090.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1201.0, + 701.0, + 1216.0, + 701.0, + 1216.0, + 711.0, + 1201.0, + 711.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 713.0, + 379.0, + 713.0, + 379.0, + 727.0, + 356.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 712.0, + 1014.0, + 712.0, + 1014.0, + 728.0, + 969.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 712.0, + 1136.0, + 712.0, + 1136.0, + 728.0, + 1091.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 734.0, + 1053.0, + 734.0, + 1053.0, + 762.0, + 1040.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 739.0, + 1069.0, + 739.0, + 1069.0, + 757.0, + 1053.0, + 757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 761.0, + 331.0, + 761.0, + 331.0, + 773.0, + 322.0, + 773.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 754.0, + 366.0, + 754.0, + 366.0, + 769.0, + 346.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 758.0, + 699.0, + 758.0, + 699.0, + 777.0, + 677.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 772.0, + 1069.0, + 772.0, + 1069.0, + 784.0, + 1058.0, + 784.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 407.0, + 782.0, + 434.0, + 782.0, + 434.0, + 794.0, + 407.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 443.0, + 779.0, + 475.0, + 779.0, + 475.0, + 795.0, + 443.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 478.0, + 782.0, + 513.0, + 782.0, + 513.0, + 794.0, + 478.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 782.0, + 549.0, + 782.0, + 549.0, + 794.0, + 517.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 779.0, + 588.0, + 779.0, + 588.0, + 795.0, + 550.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 779.0, + 661.0, + 779.0, + 661.0, + 797.0, + 590.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 782.0, + 717.0, + 782.0, + 717.0, + 794.0, + 706.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 779.0, + 988.0, + 779.0, + 988.0, + 807.0, + 737.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 779.0, + 1024.0, + 779.0, + 1024.0, + 797.0, + 995.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 780.0, + 1090.0, + 780.0, + 1090.0, + 795.0, + 1075.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 779.0, + 1357.0, + 779.0, + 1357.0, + 805.0, + 1108.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 779.0, + 1393.0, + 779.0, + 1393.0, + 797.0, + 1364.0, + 797.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 397.0, + 821.0, + 573.0, + 821.0, + 573.0, + 857.0, + 397.0, + 857.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 823.0, + 955.0, + 823.0, + 955.0, + 854.0, + 739.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 823.0, + 1327.0, + 823.0, + 1327.0, + 854.0, + 1108.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 871.0, + 336.0, + 871.0, + 336.0, + 948.0, + 306.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 554.0, + 884.0, + 568.0, + 884.0, + 568.0, + 893.0, + 554.0, + 893.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 662.0, + 857.0, + 706.0, + 857.0, + 706.0, + 937.0, + 662.0, + 937.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 867.0, + 1012.0, + 867.0, + 1012.0, + 906.0, + 969.0, + 906.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 871.0, + 1080.0, + 871.0, + 1080.0, + 959.0, + 1030.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 865.0, + 1133.0, + 865.0, + 1133.0, + 903.0, + 1090.0, + 903.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1293.0, + 884.0, + 1316.0, + 884.0, + 1316.0, + 892.0, + 1293.0, + 892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 905.0, + 687.0, + 905.0, + 687.0, + 915.0, + 679.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 969.0, + 900.0, + 1014.0, + 900.0, + 1014.0, + 917.0, + 969.0, + 917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 897.0, + 1136.0, + 897.0, + 1136.0, + 915.0, + 1091.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 316.0, + 934.0, + 329.0, + 934.0, + 329.0, + 949.0, + 316.0, + 949.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 608.0, + 917.0, + 649.0, + 917.0, + 649.0, + 953.0, + 608.0, + 953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 920.0, + 701.0, + 920.0, + 701.0, + 951.0, + 667.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 911.0, + 1070.0, + 911.0, + 1070.0, + 951.0, + 1047.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 949.0, + 331.0, + 949.0, + 331.0, + 964.0, + 322.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 969.0, + 351.0, + 969.0, + 351.0, + 979.0, + 342.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 949.0, + 660.0, + 949.0, + 660.0, + 992.0, + 373.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 946.0, + 699.0, + 946.0, + 699.0, + 964.0, + 677.0, + 964.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 968.0, + 717.0, + 968.0, + 717.0, + 979.0, + 706.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 730.0, + 957.0, + 987.0, + 957.0, + 987.0, + 997.0, + 730.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 967.0, + 1023.0, + 967.0, + 1023.0, + 980.0, + 995.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 956.0, + 1071.0, + 956.0, + 1071.0, + 971.0, + 1056.0, + 971.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 968.0, + 1089.0, + 968.0, + 1089.0, + 979.0, + 1078.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 955.0, + 1356.0, + 955.0, + 1356.0, + 998.0, + 1100.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 967.0, + 1392.0, + 967.0, + 1392.0, + 980.0, + 1364.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 337.0, + 1009.0, + 633.0, + 1009.0, + 633.0, + 1040.0, + 337.0, + 1040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 1010.0, + 1020.0, + 1010.0, + 1020.0, + 1038.0, + 676.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1044.0, + 1010.0, + 1389.0, + 1010.0, + 1389.0, + 1038.0, + 1044.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 303.0, + 1048.0, + 341.0, + 1048.0, + 341.0, + 1141.0, + 303.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1051.0, + 396.0, + 1051.0, + 396.0, + 1079.0, + 353.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1044.0, + 704.0, + 1044.0, + 704.0, + 1141.0, + 661.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 949.0, + 1053.0, + 1009.0, + 1053.0, + 1009.0, + 1081.0, + 949.0, + 1081.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1049.0, + 1133.0, + 1049.0, + 1133.0, + 1120.0, + 1030.0, + 1120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1074.0, + 396.0, + 1074.0, + 396.0, + 1090.0, + 354.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1072.0, + 1008.0, + 1072.0, + 1008.0, + 1093.0, + 968.0, + 1093.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1072.0, + 1134.0, + 1072.0, + 1134.0, + 1090.0, + 1090.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 1084.0, + 398.0, + 1084.0, + 398.0, + 1101.0, + 354.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1085.0, + 1013.0, + 1085.0, + 1013.0, + 1102.0, + 968.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1073.0, + 1082.0, + 1138.0, + 1082.0, + 1138.0, + 1102.0, + 1073.0, + 1102.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1094.0, + 1071.0, + 1094.0, + 1071.0, + 1115.0, + 1050.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 1106.0, + 1055.0, + 1106.0, + 1055.0, + 1136.0, + 1038.0, + 1136.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1050.0, + 1108.0, + 1071.0, + 1108.0, + 1071.0, + 1131.0, + 1050.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1133.0, + 331.0, + 1133.0, + 331.0, + 1150.0, + 322.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1131.0, + 699.0, + 1131.0, + 699.0, + 1150.0, + 677.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1125.0, + 1071.0, + 1125.0, + 1071.0, + 1146.0, + 1051.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1055.0, + 1142.0, + 1071.0, + 1142.0, + 1071.0, + 1158.0, + 1055.0, + 1158.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1154.0, + 434.0, + 1154.0, + 434.0, + 1167.0, + 408.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1154.0, + 469.0, + 1154.0, + 469.0, + 1167.0, + 447.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 482.0, + 1154.0, + 512.0, + 1154.0, + 512.0, + 1167.0, + 482.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 1154.0, + 548.0, + 1154.0, + 548.0, + 1167.0, + 517.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 556.0, + 1154.0, + 586.0, + 1154.0, + 586.0, + 1167.0, + 556.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1152.0, + 625.0, + 1152.0, + 625.0, + 1168.0, + 590.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1152.0, + 661.0, + 1152.0, + 661.0, + 1168.0, + 626.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1154.0, + 717.0, + 1154.0, + 717.0, + 1167.0, + 706.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 811.0, + 1154.0, + 835.0, + 1154.0, + 835.0, + 1167.0, + 811.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 846.0, + 1154.0, + 877.0, + 1154.0, + 877.0, + 1167.0, + 846.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1154.0, + 913.0, + 1154.0, + 913.0, + 1167.0, + 881.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 1152.0, + 952.0, + 1152.0, + 952.0, + 1168.0, + 919.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1152.0, + 989.0, + 1152.0, + 989.0, + 1168.0, + 954.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 990.0, + 1152.0, + 1024.0, + 1152.0, + 1024.0, + 1168.0, + 990.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1153.0, + 1090.0, + 1153.0, + 1090.0, + 1168.0, + 1075.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1154.0, + 1170.0, + 1154.0, + 1170.0, + 1167.0, + 1144.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1154.0, + 1207.0, + 1154.0, + 1207.0, + 1167.0, + 1181.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1154.0, + 1247.0, + 1154.0, + 1247.0, + 1167.0, + 1215.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1250.0, + 1154.0, + 1282.0, + 1154.0, + 1282.0, + 1167.0, + 1250.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1283.0, + 1152.0, + 1321.0, + 1152.0, + 1321.0, + 1168.0, + 1283.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 1152.0, + 1393.0, + 1152.0, + 1393.0, + 1169.0, + 1322.0, + 1169.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1194.0, + 598.0, + 1194.0, + 598.0, + 1227.0, + 373.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1193.0, + 980.0, + 1193.0, + 980.0, + 1228.0, + 716.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1193.0, + 1353.0, + 1193.0, + 1353.0, + 1228.0, + 1081.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1329.0, + 593.0, + 1355.0, + 593.0, + 1355.0, + 610.0, + 1329.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 360.0, + 722.0, + 452.0, + 722.0, + 452.0, + 765.0, + 360.0, + 765.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 777.5, + 599.0, + 777.5, + 599.0, + 813.5, + 363.0, + 813.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 570.25, + 875.0, + 596.25, + 875.0, + 596.25, + 892.5, + 570.25, + 892.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 973.0, + 941.0, + 997.0, + 941.0, + 997.0, + 954.5, + 973.0, + 954.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 330.25, + 1106.5, + 447.25, + 1106.5, + 447.25, + 1148.0, + 330.25, + 1148.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 363.0, + 1149.5, + 599.0, + 1149.5, + 599.0, + 1185.5, + 363.0, + 1185.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 727.0, + 1146.5, + 966.0, + 1146.5, + 966.0, + 1186.0, + 727.0, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1098.75, + 1150.0, + 1334.75, + 1150.0, + 1334.75, + 1186.0, + 1098.75, + 1186.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1251.0, + 1405.0, + 1251.0, + 1405.0, + 1281.0, + 295.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1277.0, + 1406.0, + 1277.0, + 1406.0, + 1311.0, + 293.0, + 1311.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1305.0, + 778.0, + 1305.0, + 778.0, + 1338.0, + 294.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 801.0, + 1305.0, + 1403.0, + 1305.0, + 1403.0, + 1338.0, + 801.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1333.0, + 1108.0, + 1333.0, + 1108.0, + 1364.0, + 294.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1130.0, + 1333.0, + 1406.0, + 1333.0, + 1406.0, + 1364.0, + 1130.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1360.0, + 371.0, + 1360.0, + 371.0, + 1395.0, + 293.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 1360.0, + 1406.0, + 1360.0, + 1406.0, + 1395.0, + 394.0, + 1395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1392.0, + 650.0, + 1392.0, + 650.0, + 1419.0, + 295.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 829.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2124.0, + 829.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 230.0, + 1306.0, + 230.0, + 1306.0, + 263.0, + 323.0, + 263.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 24, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 3, + "poly": [ + 301, + 296, + 1398, + 296, + 1398, + 1229, + 301, + 1229 + ], + "score": 0.973 + }, + { + "category_id": 3, + "poly": [ + 303, + 1551, + 1394, + 1551, + 1394, + 1734, + 303, + 1734 + ], + "score": 0.945 + }, + { + "category_id": 4, + "poly": [ + 296, + 1757, + 1404, + 1757, + 1404, + 1870, + 296, + 1870 + ], + "score": 0.941 + }, + { + "category_id": 4, + "poly": [ + 296, + 1250, + 1406, + 1250, + 1406, + 1418, + 296, + 1418 + ], + "score": 0.934 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.89 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 865, + 2088, + 865, + 2113, + 834, + 2113 + ], + "score": 0.856 + }, + { + "category_id": 2, + "poly": [ + 306, + 234, + 1038, + 234, + 1038, + 260, + 306, + 260 + ], + "score": 0.814 + }, + { + "category_id": 4, + "poly": [ + 298, + 1483, + 1142, + 1483, + 1142, + 1512, + 298, + 1512 + ], + "score": 0.407 + }, + { + "category_id": 2, + "poly": [ + 298, + 1483, + 1142, + 1483, + 1142, + 1512, + 298, + 1512 + ], + "score": 0.123 + }, + { + "category_id": 13, + "poly": [ + 1005, + 1362, + 1105, + 1362, + 1105, + 1391, + 1005, + 1391 + ], + "score": 0.9, + "latex": "\\chi = 0 . 0 5" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 302.0, + 338.0, + 302.0, + 338.0, + 396.0, + 301.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 311.0, + 411.0, + 311.0, + 411.0, + 337.0, + 354.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 308.0, + 705.0, + 308.0, + 705.0, + 377.0, + 668.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 946.0, + 307.0, + 1015.0, + 307.0, + 1015.0, + 337.0, + 946.0, + 337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 307.0, + 1149.0, + 307.0, + 1149.0, + 380.0, + 1030.0, + 380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 330.0, + 379.0, + 330.0, + 379.0, + 348.0, + 354.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 330.0, + 983.0, + 330.0, + 983.0, + 344.0, + 959.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 330.0, + 1115.0, + 330.0, + 1115.0, + 344.0, + 1091.0, + 344.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 343.0, + 379.0, + 343.0, + 379.0, + 357.0, + 356.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 342.0, + 984.0, + 342.0, + 984.0, + 355.0, + 959.0, + 355.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 667.0, + 359.0, + 700.0, + 359.0, + 700.0, + 391.0, + 667.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 361.0, + 1074.0, + 361.0, + 1074.0, + 396.0, + 1035.0, + 396.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 389.0, + 699.0, + 389.0, + 699.0, + 407.0, + 677.0, + 407.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 399.0, + 332.0, + 399.0, + 332.0, + 411.0, + 322.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 397.0, + 1071.0, + 397.0, + 1071.0, + 414.0, + 1056.0, + 414.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 410.0, + 352.0, + 410.0, + 352.0, + 422.0, + 342.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 409.0, + 394.0, + 409.0, + 394.0, + 422.0, + 374.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 409.0, + 433.0, + 409.0, + 433.0, + 422.0, + 411.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 409.0, + 469.0, + 409.0, + 469.0, + 422.0, + 449.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 485.0, + 409.0, + 510.0, + 409.0, + 510.0, + 422.0, + 485.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 409.0, + 548.0, + 409.0, + 548.0, + 422.0, + 522.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 558.0, + 409.0, + 584.0, + 409.0, + 584.0, + 422.0, + 558.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 409.0, + 624.0, + 409.0, + 624.0, + 422.0, + 596.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 409.0, + 659.0, + 409.0, + 659.0, + 422.0, + 631.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 737.0, + 407.0, + 763.0, + 407.0, + 763.0, + 425.0, + 737.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 409.0, + 797.0, + 409.0, + 797.0, + 422.0, + 776.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 409.0, + 833.0, + 409.0, + 833.0, + 422.0, + 813.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 848.0, + 409.0, + 873.0, + 409.0, + 873.0, + 422.0, + 848.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 409.0, + 912.0, + 409.0, + 912.0, + 422.0, + 886.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 921.0, + 409.0, + 948.0, + 409.0, + 948.0, + 422.0, + 921.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 409.0, + 987.0, + 409.0, + 987.0, + 422.0, + 960.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 409.0, + 1023.0, + 409.0, + 1023.0, + 422.0, + 995.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 410.0, + 1089.0, + 410.0, + 1089.0, + 422.0, + 1078.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 407.0, + 1134.0, + 407.0, + 1134.0, + 425.0, + 1109.0, + 425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 409.0, + 1167.0, + 409.0, + 1167.0, + 422.0, + 1146.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1185.0, + 409.0, + 1205.0, + 409.0, + 1205.0, + 422.0, + 1185.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1219.0, + 409.0, + 1244.0, + 409.0, + 1244.0, + 422.0, + 1219.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 409.0, + 1282.0, + 409.0, + 1282.0, + 422.0, + 1255.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 409.0, + 1318.0, + 409.0, + 1318.0, + 422.0, + 1290.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 409.0, + 1356.0, + 409.0, + 1356.0, + 422.0, + 1330.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 409.0, + 1392.0, + 409.0, + 1392.0, + 422.0, + 1364.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 416.0, + 595.0, + 416.0, + 595.0, + 432.0, + 391.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 416.0, + 959.0, + 416.0, + 959.0, + 432.0, + 755.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 416.0, + 1331.0, + 416.0, + 1331.0, + 432.0, + 1124.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 344.0, + 451.0, + 624.0, + 451.0, + 624.0, + 481.0, + 344.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 686.0, + 451.0, + 1009.0, + 451.0, + 1009.0, + 482.0, + 686.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 451.0, + 1383.0, + 451.0, + 1383.0, + 481.0, + 1051.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 492.0, + 340.0, + 492.0, + 340.0, + 582.0, + 301.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 494.0, + 412.0, + 494.0, + 412.0, + 523.0, + 343.0, + 523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 498.0, + 705.0, + 498.0, + 705.0, + 582.0, + 661.0, + 582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 493.0, + 1014.0, + 493.0, + 1014.0, + 519.0, + 943.0, + 519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 491.0, + 1149.0, + 491.0, + 1149.0, + 563.0, + 1030.0, + 563.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 517.0, + 378.0, + 517.0, + 378.0, + 530.0, + 356.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 517.0, + 982.0, + 517.0, + 982.0, + 530.0, + 960.0, + 530.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 514.0, + 1116.0, + 514.0, + 1116.0, + 532.0, + 1090.0, + 532.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 528.0, + 379.0, + 528.0, + 379.0, + 541.0, + 356.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 528.0, + 984.0, + 528.0, + 984.0, + 541.0, + 960.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 525.0, + 1119.0, + 525.0, + 1119.0, + 543.0, + 1090.0, + 543.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 545.0, + 1074.0, + 545.0, + 1074.0, + 581.0, + 1035.0, + 581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 575.0, + 697.0, + 575.0, + 697.0, + 590.0, + 679.0, + 590.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 584.0, + 1071.0, + 584.0, + 1071.0, + 599.0, + 1056.0, + 599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 595.0, + 396.0, + 595.0, + 396.0, + 607.0, + 374.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 411.0, + 595.0, + 434.0, + 595.0, + 434.0, + 607.0, + 411.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 449.0, + 595.0, + 472.0, + 595.0, + 472.0, + 609.0, + 449.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 595.0, + 512.0, + 595.0, + 512.0, + 607.0, + 483.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 595.0, + 549.0, + 595.0, + 549.0, + 607.0, + 522.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 559.0, + 595.0, + 585.0, + 595.0, + 585.0, + 607.0, + 559.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 596.0, + 595.0, + 624.0, + 595.0, + 624.0, + 609.0, + 596.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 595.0, + 660.0, + 595.0, + 660.0, + 609.0, + 633.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 596.0, + 717.0, + 596.0, + 717.0, + 607.0, + 707.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 595.0, + 762.0, + 595.0, + 762.0, + 609.0, + 739.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 595.0, + 798.0, + 595.0, + 798.0, + 607.0, + 775.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 595.0, + 836.0, + 595.0, + 836.0, + 609.0, + 813.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 594.0, + 876.0, + 594.0, + 876.0, + 607.0, + 847.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 595.0, + 912.0, + 595.0, + 912.0, + 607.0, + 886.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 923.0, + 595.0, + 948.0, + 595.0, + 948.0, + 607.0, + 923.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 595.0, + 987.0, + 595.0, + 987.0, + 609.0, + 960.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 592.0, + 1024.0, + 592.0, + 1024.0, + 610.0, + 994.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 596.0, + 1089.0, + 596.0, + 1089.0, + 607.0, + 1078.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 595.0, + 1133.0, + 595.0, + 1133.0, + 609.0, + 1110.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 594.0, + 1171.0, + 594.0, + 1171.0, + 607.0, + 1146.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1182.0, + 595.0, + 1206.0, + 595.0, + 1206.0, + 609.0, + 1182.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1214.0, + 594.0, + 1247.0, + 594.0, + 1247.0, + 610.0, + 1214.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 595.0, + 1283.0, + 595.0, + 1283.0, + 609.0, + 1255.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1290.0, + 595.0, + 1320.0, + 595.0, + 1320.0, + 607.0, + 1290.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1330.0, + 595.0, + 1356.0, + 595.0, + 1356.0, + 609.0, + 1330.0, + 609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 592.0, + 1393.0, + 592.0, + 1393.0, + 610.0, + 1363.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 602.0, + 595.0, + 602.0, + 595.0, + 618.0, + 391.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 602.0, + 959.0, + 602.0, + 959.0, + 618.0, + 755.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 602.0, + 1330.0, + 602.0, + 1330.0, + 618.0, + 1125.0, + 618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 348.0, + 636.0, + 620.0, + 636.0, + 620.0, + 666.0, + 348.0, + 666.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 690.0, + 637.0, + 1004.0, + 637.0, + 1004.0, + 668.0, + 690.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 637.0, + 1377.0, + 637.0, + 1377.0, + 668.0, + 1059.0, + 668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 681.0, + 377.0, + 681.0, + 377.0, + 694.0, + 354.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 678.0, + 697.0, + 678.0, + 697.0, + 693.0, + 679.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 947.0, + 683.0, + 957.0, + 683.0, + 957.0, + 692.0, + 947.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 681.0, + 982.0, + 681.0, + 982.0, + 694.0, + 958.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 677.0, + 1071.0, + 677.0, + 1071.0, + 695.0, + 1047.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 681.0, + 1114.0, + 681.0, + 1114.0, + 694.0, + 1091.0, + 694.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 687.0, + 340.0, + 687.0, + 340.0, + 767.0, + 301.0, + 767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 690.0, + 411.0, + 690.0, + 411.0, + 707.0, + 353.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 690.0, + 705.0, + 690.0, + 705.0, + 769.0, + 661.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 943.0, + 690.0, + 1014.0, + 690.0, + 1014.0, + 707.0, + 943.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 688.0, + 1146.0, + 688.0, + 1146.0, + 746.0, + 1030.0, + 746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 700.0, + 379.0, + 700.0, + 379.0, + 718.0, + 354.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 700.0, + 983.0, + 700.0, + 983.0, + 718.0, + 958.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 700.0, + 1116.0, + 700.0, + 1116.0, + 718.0, + 1090.0, + 718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 714.0, + 379.0, + 714.0, + 379.0, + 728.0, + 356.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 712.0, + 1118.0, + 712.0, + 1118.0, + 729.0, + 1089.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 761.0, + 331.0, + 761.0, + 331.0, + 774.0, + 322.0, + 774.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 346.0, + 752.0, + 367.0, + 752.0, + 367.0, + 770.0, + 346.0, + 770.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 761.0, + 697.0, + 761.0, + 697.0, + 776.0, + 679.0, + 776.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1033.0, + 728.0, + 1076.0, + 728.0, + 1076.0, + 769.0, + 1033.0, + 769.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 770.0, + 1071.0, + 770.0, + 1071.0, + 785.0, + 1056.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 345.0, + 785.0, + 351.0, + 785.0, + 351.0, + 791.0, + 345.0, + 791.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 780.0, + 625.0, + 780.0, + 625.0, + 805.0, + 371.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 779.0, + 661.0, + 779.0, + 661.0, + 796.0, + 626.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 782.0, + 717.0, + 782.0, + 717.0, + 794.0, + 706.0, + 794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 781.0, + 959.0, + 781.0, + 959.0, + 805.0, + 739.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 960.0, + 781.0, + 987.0, + 781.0, + 987.0, + 795.0, + 960.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 779.0, + 1024.0, + 779.0, + 1024.0, + 796.0, + 994.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 782.0, + 1088.0, + 782.0, + 1088.0, + 795.0, + 1078.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 781.0, + 1356.0, + 781.0, + 1356.0, + 805.0, + 1109.0, + 805.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 779.0, + 1393.0, + 779.0, + 1393.0, + 796.0, + 1363.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 821.0, + 572.0, + 821.0, + 572.0, + 853.0, + 395.0, + 853.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 822.0, + 955.0, + 822.0, + 955.0, + 854.0, + 739.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1108.0, + 822.0, + 1327.0, + 822.0, + 1327.0, + 854.0, + 1108.0, + 854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 305.0, + 869.0, + 332.0, + 869.0, + 332.0, + 947.0, + 305.0, + 947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 864.0, + 412.0, + 864.0, + 412.0, + 904.0, + 353.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 871.0, + 710.0, + 871.0, + 710.0, + 960.0, + 661.0, + 960.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 722.0, + 866.0, + 777.0, + 866.0, + 777.0, + 904.0, + 722.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 864.0, + 1148.0, + 864.0, + 1148.0, + 935.0, + 1034.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 903.0, + 333.0, + 903.0, + 333.0, + 919.0, + 320.0, + 919.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 898.0, + 381.0, + 898.0, + 381.0, + 915.0, + 353.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 898.0, + 749.0, + 898.0, + 749.0, + 915.0, + 721.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 909.0, + 1070.0, + 909.0, + 1070.0, + 929.0, + 1049.0, + 929.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 898.0, + 1119.0, + 898.0, + 1119.0, + 915.0, + 1089.0, + 915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 920.0, + 331.0, + 920.0, + 331.0, + 934.0, + 322.0, + 934.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1035.0, + 918.0, + 1074.0, + 918.0, + 1074.0, + 954.0, + 1035.0, + 954.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 936.0, + 331.0, + 936.0, + 331.0, + 951.0, + 321.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 954.0, + 331.0, + 954.0, + 331.0, + 966.0, + 322.0, + 966.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 969.0, + 351.0, + 969.0, + 351.0, + 979.0, + 342.0, + 979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 960.0, + 625.0, + 960.0, + 625.0, + 996.0, + 374.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 631.0, + 965.0, + 661.0, + 965.0, + 661.0, + 982.0, + 631.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 957.0, + 700.0, + 957.0, + 700.0, + 970.0, + 689.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 709.0, + 967.0, + 720.0, + 967.0, + 720.0, + 980.0, + 709.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 741.0, + 960.0, + 987.0, + 960.0, + 987.0, + 996.0, + 741.0, + 996.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 993.0, + 965.0, + 1024.0, + 965.0, + 1024.0, + 982.0, + 993.0, + 982.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 957.0, + 1069.0, + 957.0, + 1069.0, + 970.0, + 1058.0, + 970.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 967.0, + 1089.0, + 967.0, + 1089.0, + 980.0, + 1078.0, + 980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 956.0, + 1356.0, + 956.0, + 1356.0, + 999.0, + 1100.0, + 999.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1364.0, + 967.0, + 1392.0, + 967.0, + 1392.0, + 981.0, + 1364.0, + 981.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1008.0, + 632.0, + 1008.0, + 632.0, + 1037.0, + 334.0, + 1037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 675.0, + 1008.0, + 1022.0, + 1008.0, + 1022.0, + 1039.0, + 675.0, + 1039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 1008.0, + 1389.0, + 1008.0, + 1389.0, + 1038.0, + 1042.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1046.0, + 340.0, + 1046.0, + 340.0, + 1141.0, + 301.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 353.0, + 1053.0, + 411.0, + 1053.0, + 411.0, + 1080.0, + 353.0, + 1080.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1059.0, + 704.0, + 1059.0, + 704.0, + 1140.0, + 664.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 953.0, + 1051.0, + 1013.0, + 1051.0, + 1013.0, + 1079.0, + 953.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1050.0, + 1148.0, + 1050.0, + 1148.0, + 1146.0, + 1029.0, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 356.0, + 1075.0, + 378.0, + 1075.0, + 378.0, + 1089.0, + 356.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 955.0, + 1073.0, + 980.0, + 1073.0, + 980.0, + 1090.0, + 955.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1073.0, + 1116.0, + 1073.0, + 1116.0, + 1090.0, + 1090.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 357.0, + 1087.0, + 379.0, + 1087.0, + 379.0, + 1100.0, + 357.0, + 1100.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1084.0, + 983.0, + 1084.0, + 983.0, + 1101.0, + 954.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1089.0, + 1084.0, + 1119.0, + 1084.0, + 1119.0, + 1101.0, + 1089.0, + 1101.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1051.0, + 1116.0, + 1070.0, + 1116.0, + 1070.0, + 1131.0, + 1051.0, + 1131.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1137.0, + 331.0, + 1137.0, + 331.0, + 1150.0, + 322.0, + 1150.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1135.0, + 699.0, + 1135.0, + 699.0, + 1154.0, + 677.0, + 1154.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1056.0, + 1142.0, + 1071.0, + 1142.0, + 1071.0, + 1159.0, + 1056.0, + 1159.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 1154.0, + 398.0, + 1154.0, + 398.0, + 1167.0, + 372.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 408.0, + 1154.0, + 434.0, + 1154.0, + 434.0, + 1166.0, + 408.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 1154.0, + 472.0, + 1154.0, + 472.0, + 1166.0, + 447.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1156.0, + 510.0, + 1156.0, + 510.0, + 1165.0, + 483.0, + 1165.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 518.0, + 1154.0, + 548.0, + 1154.0, + 548.0, + 1166.0, + 518.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 553.0, + 1154.0, + 585.0, + 1154.0, + 585.0, + 1166.0, + 553.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 590.0, + 1152.0, + 625.0, + 1152.0, + 625.0, + 1168.0, + 590.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 626.0, + 1152.0, + 661.0, + 1152.0, + 661.0, + 1168.0, + 626.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1155.0, + 717.0, + 1155.0, + 717.0, + 1167.0, + 706.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 736.0, + 1154.0, + 763.0, + 1154.0, + 763.0, + 1167.0, + 736.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 773.0, + 1154.0, + 798.0, + 1154.0, + 798.0, + 1166.0, + 773.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 810.0, + 1154.0, + 836.0, + 1154.0, + 836.0, + 1166.0, + 810.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 842.0, + 1152.0, + 878.0, + 1152.0, + 878.0, + 1168.0, + 842.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1154.0, + 913.0, + 1154.0, + 913.0, + 1166.0, + 882.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 916.0, + 1154.0, + 950.0, + 1154.0, + 950.0, + 1166.0, + 916.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 954.0, + 1152.0, + 1024.0, + 1152.0, + 1024.0, + 1168.0, + 954.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1075.0, + 1154.0, + 1090.0, + 1154.0, + 1090.0, + 1168.0, + 1075.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1154.0, + 1134.0, + 1154.0, + 1134.0, + 1167.0, + 1106.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1144.0, + 1154.0, + 1170.0, + 1154.0, + 1170.0, + 1167.0, + 1144.0, + 1167.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1154.0, + 1206.0, + 1154.0, + 1206.0, + 1166.0, + 1181.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1215.0, + 1154.0, + 1247.0, + 1154.0, + 1247.0, + 1166.0, + 1215.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1154.0, + 1320.0, + 1154.0, + 1320.0, + 1166.0, + 1251.0, + 1166.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1322.0, + 1152.0, + 1357.0, + 1152.0, + 1357.0, + 1168.0, + 1322.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1359.0, + 1152.0, + 1393.0, + 1152.0, + 1393.0, + 1168.0, + 1359.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 391.0, + 1161.0, + 595.0, + 1161.0, + 595.0, + 1177.0, + 391.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 751.0, + 1160.0, + 958.0, + 1160.0, + 958.0, + 1177.0, + 751.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1124.0, + 1161.0, + 1330.0, + 1161.0, + 1330.0, + 1177.0, + 1124.0, + 1177.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 1195.0, + 596.0, + 1195.0, + 596.0, + 1227.0, + 373.0, + 1227.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 716.0, + 1193.0, + 979.0, + 1193.0, + 979.0, + 1228.0, + 716.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1081.0, + 1193.0, + 1352.0, + 1193.0, + 1352.0, + 1228.0, + 1081.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.75, + 338.0, + 1115.75, + 338.0, + 1115.75, + 357.0, + 1090.75, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 409.0, + 714.0, + 409.0, + 714.0, + 422.0, + 707.0, + 422.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 557.5, + 410.0, + 557.5, + 410.0, + 579.5, + 365.0, + 579.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 711.5, + 982.0, + 711.5, + 982.0, + 729.0, + 959.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.75, + 1097.0, + 485.75, + 1097.0, + 485.75, + 1146.0, + 325.75, + 1146.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 1553.0, + 342.0, + 1553.0, + 342.0, + 1572.0, + 315.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1558.0, + 409.0, + 1558.0, + 409.0, + 1574.0, + 362.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 680.0, + 1559.0, + 696.0, + 1559.0, + 696.0, + 1576.0, + 680.0, + 1576.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 966.0, + 1558.0, + 1016.0, + 1558.0, + 1016.0, + 1574.0, + 966.0, + 1574.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1557.0, + 1141.0, + 1557.0, + 1141.0, + 1577.0, + 1090.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 301.0, + 1568.0, + 350.0, + 1568.0, + 350.0, + 1630.0, + 301.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 362.0, + 1569.0, + 406.0, + 1569.0, + 406.0, + 1586.0, + 362.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 1569.0, + 705.0, + 1569.0, + 705.0, + 1648.0, + 661.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 968.0, + 1569.0, + 1011.0, + 1569.0, + 1011.0, + 1586.0, + 968.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1568.0, + 1078.0, + 1568.0, + 1078.0, + 1654.0, + 1030.0, + 1654.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 1569.0, + 1136.0, + 1569.0, + 1136.0, + 1586.0, + 1090.0, + 1586.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 319.0, + 1597.0, + 340.0, + 1597.0, + 340.0, + 1618.0, + 319.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 306.0, + 1612.0, + 341.0, + 1612.0, + 341.0, + 1640.0, + 306.0, + 1640.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 1607.0, + 1071.0, + 1607.0, + 1071.0, + 1626.0, + 1047.0, + 1626.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1627.0, + 341.0, + 1627.0, + 341.0, + 1648.0, + 320.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1632.0, + 698.0, + 1632.0, + 698.0, + 1653.0, + 677.0, + 1653.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1645.0, + 340.0, + 1645.0, + 340.0, + 1660.0, + 321.0, + 1660.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1650.0, + 698.0, + 1650.0, + 698.0, + 1669.0, + 677.0, + 1669.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1650.0, + 1069.0, + 1650.0, + 1069.0, + 1663.0, + 1058.0, + 1663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 1659.0, + 597.0, + 1659.0, + 597.0, + 1683.0, + 380.0, + 1683.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 598.0, + 1660.0, + 624.0, + 1660.0, + 624.0, + 1674.0, + 598.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 633.0, + 1660.0, + 660.0, + 1660.0, + 660.0, + 1674.0, + 633.0, + 1674.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 707.0, + 1660.0, + 717.0, + 1660.0, + 717.0, + 1673.0, + 707.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1658.0, + 987.0, + 1658.0, + 987.0, + 1686.0, + 739.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 995.0, + 1658.0, + 1025.0, + 1658.0, + 1025.0, + 1675.0, + 995.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1661.0, + 1088.0, + 1661.0, + 1088.0, + 1673.0, + 1077.0, + 1673.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 1659.0, + 1356.0, + 1659.0, + 1356.0, + 1686.0, + 1110.0, + 1686.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1658.0, + 1392.0, + 1658.0, + 1392.0, + 1675.0, + 1363.0, + 1675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 343.0, + 1701.0, + 624.0, + 1701.0, + 624.0, + 1731.0, + 343.0, + 1731.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 685.0, + 1699.0, + 1011.0, + 1699.0, + 1011.0, + 1735.0, + 685.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 1699.0, + 1383.0, + 1699.0, + 1383.0, + 1733.0, + 1049.0, + 1733.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1756.0, + 1402.0, + 1756.0, + 1402.0, + 1790.0, + 295.0, + 1790.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1783.0, + 1406.0, + 1783.0, + 1406.0, + 1819.0, + 294.0, + 1819.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1811.0, + 1405.0, + 1811.0, + 1405.0, + 1843.0, + 294.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1841.0, + 670.0, + 1841.0, + 670.0, + 1871.0, + 296.0, + 1871.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1251.0, + 1404.0, + 1251.0, + 1404.0, + 1281.0, + 296.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1280.0, + 1404.0, + 1280.0, + 1404.0, + 1310.0, + 295.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1308.0, + 1404.0, + 1308.0, + 1404.0, + 1338.0, + 296.0, + 1338.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1334.0, + 1403.0, + 1334.0, + 1403.0, + 1364.0, + 295.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1360.0, + 1004.0, + 1360.0, + 1004.0, + 1393.0, + 291.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1106.0, + 1360.0, + 1406.0, + 1360.0, + 1406.0, + 1393.0, + 1106.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1393.0, + 444.0, + 1393.0, + 444.0, + 1419.0, + 295.0, + 1419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2121.0, + 830.0, + 2121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 230.0, + 1044.0, + 230.0, + 1044.0, + 264.0, + 300.0, + 264.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1144.0, + 1479.0, + 1144.0, + 1516.0, + 295.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1479.0, + 1144.0, + 1479.0, + 1144.0, + 1516.0, + 295.0, + 1516.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 25, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1119, + 1403, + 1119, + 1403, + 1366, + 298, + 1366 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 287, + 1405, + 287, + 1405, + 654, + 297, + 654 + ], + "score": 0.981 + }, + { + "category_id": 4, + "poly": [ + 298, + 895, + 1403, + 895, + 1403, + 1007, + 298, + 1007 + ], + "score": 0.957 + }, + { + "category_id": 3, + "poly": [ + 303, + 689, + 1395, + 689, + 1395, + 871, + 303, + 871 + ], + "score": 0.951 + }, + { + "category_id": 3, + "poly": [ + 304, + 1398, + 1394, + 1398, + 1394, + 1583, + 304, + 1583 + ], + "score": 0.926 + }, + { + "category_id": 2, + "poly": [ + 300, + 76, + 813, + 76, + 813, + 104, + 300, + 104 + ], + "score": 0.882 + }, + { + "category_id": 0, + "poly": [ + 300, + 1065, + 1190, + 1065, + 1190, + 1094, + 300, + 1094 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 834, + 2088, + 864, + 2088, + 864, + 2112, + 834, + 2112 + ], + "score": 0.846 + }, + { + "category_id": 4, + "poly": [ + 403, + 1604, + 1296, + 1604, + 1296, + 1636, + 403, + 1636 + ], + "score": 0.828 + }, + { + "category_id": 0, + "poly": [ + 307, + 233, + 1064, + 233, + 1064, + 260, + 307, + 260 + ], + "score": 0.442 + }, + { + "category_id": 4, + "poly": [ + 1081, + 1551, + 1351, + 1551, + 1351, + 1578, + 1081, + 1578 + ], + "score": 0.102 + }, + { + "category_id": 15, + "poly": [ + 294.0, + 894.0, + 1404.0, + 894.0, + 1404.0, + 927.0, + 294.0, + 927.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 922.0, + 1402.0, + 922.0, + 1402.0, + 951.0, + 297.0, + 951.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 950.0, + 1404.0, + 950.0, + 1404.0, + 983.0, + 296.0, + 983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 976.0, + 653.0, + 976.0, + 653.0, + 1012.0, + 295.0, + 1012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 696.0, + 407.0, + 696.0, + 407.0, + 712.0, + 354.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 698.0, + 633.0, + 698.0, + 633.0, + 707.0, + 609.0, + 707.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 679.0, + 697.0, + 698.0, + 697.0, + 698.0, + 712.0, + 679.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 696.0, + 1012.0, + 696.0, + 1012.0, + 712.0, + 958.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 696.0, + 1145.0, + 696.0, + 1145.0, + 712.0, + 1091.0, + 712.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 702.0, + 340.0, + 702.0, + 340.0, + 783.0, + 300.0, + 783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 354.0, + 707.0, + 409.0, + 707.0, + 409.0, + 732.0, + 354.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 711.0, + 702.0, + 711.0, + 702.0, + 761.0, + 664.0, + 761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 707.0, + 1014.0, + 707.0, + 1014.0, + 732.0, + 958.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 709.0, + 1069.0, + 709.0, + 1069.0, + 725.0, + 1049.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1090.0, + 707.0, + 1147.0, + 707.0, + 1147.0, + 732.0, + 1090.0, + 732.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 727.0, + 380.0, + 727.0, + 380.0, + 745.0, + 352.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 727.0, + 1117.0, + 727.0, + 1117.0, + 745.0, + 1091.0, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 746.0, + 700.0, + 746.0, + 700.0, + 777.0, + 668.0, + 777.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1038.0, + 750.0, + 1069.0, + 750.0, + 1069.0, + 779.0, + 1038.0, + 779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 775.0, + 331.0, + 775.0, + 331.0, + 788.0, + 321.0, + 788.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 678.0, + 776.0, + 699.0, + 776.0, + 699.0, + 795.0, + 678.0, + 795.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 788.0, + 1070.0, + 788.0, + 1070.0, + 799.0, + 1059.0, + 799.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 342.0, + 798.0, + 351.0, + 798.0, + 351.0, + 809.0, + 342.0, + 809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 372.0, + 796.0, + 398.0, + 796.0, + 398.0, + 810.0, + 372.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 409.0, + 798.0, + 436.0, + 798.0, + 436.0, + 810.0, + 409.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 447.0, + 798.0, + 473.0, + 798.0, + 473.0, + 810.0, + 447.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 479.0, + 798.0, + 513.0, + 798.0, + 513.0, + 810.0, + 479.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 517.0, + 798.0, + 549.0, + 798.0, + 549.0, + 810.0, + 517.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 798.0, + 587.0, + 798.0, + 587.0, + 810.0, + 550.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 591.0, + 795.0, + 662.0, + 795.0, + 662.0, + 813.0, + 591.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 798.0, + 717.0, + 798.0, + 717.0, + 810.0, + 706.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 738.0, + 796.0, + 761.0, + 796.0, + 761.0, + 810.0, + 738.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 796.0, + 797.0, + 796.0, + 797.0, + 810.0, + 776.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 813.0, + 796.0, + 834.0, + 796.0, + 834.0, + 811.0, + 813.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 847.0, + 796.0, + 874.0, + 796.0, + 874.0, + 810.0, + 847.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 886.0, + 798.0, + 911.0, + 798.0, + 911.0, + 810.0, + 886.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 919.0, + 795.0, + 951.0, + 795.0, + 951.0, + 811.0, + 919.0, + 811.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 795.0, + 988.0, + 795.0, + 988.0, + 813.0, + 958.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 994.0, + 795.0, + 1024.0, + 795.0, + 1024.0, + 813.0, + 994.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 799.0, + 1087.0, + 799.0, + 1087.0, + 810.0, + 1079.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1110.0, + 796.0, + 1132.0, + 796.0, + 1132.0, + 810.0, + 1110.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 796.0, + 1168.0, + 796.0, + 1168.0, + 810.0, + 1146.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1184.0, + 796.0, + 1205.0, + 796.0, + 1205.0, + 810.0, + 1184.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1218.0, + 796.0, + 1244.0, + 796.0, + 1244.0, + 810.0, + 1218.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1255.0, + 796.0, + 1281.0, + 796.0, + 1281.0, + 810.0, + 1255.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1291.0, + 796.0, + 1318.0, + 796.0, + 1318.0, + 810.0, + 1291.0, + 810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1328.0, + 795.0, + 1358.0, + 795.0, + 1358.0, + 813.0, + 1328.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 795.0, + 1393.0, + 795.0, + 1393.0, + 813.0, + 1363.0, + 813.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 390.0, + 804.0, + 594.0, + 804.0, + 594.0, + 820.0, + 390.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 755.0, + 804.0, + 960.0, + 804.0, + 960.0, + 820.0, + 755.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1125.0, + 804.0, + 1330.0, + 804.0, + 1330.0, + 820.0, + 1125.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 838.0, + 597.0, + 838.0, + 597.0, + 867.0, + 371.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 712.0, + 838.0, + 982.0, + 838.0, + 982.0, + 869.0, + 712.0, + 869.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 836.0, + 1356.0, + 836.0, + 1356.0, + 868.0, + 1077.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.75, + 727.0, + 983.75, + 727.0, + 983.75, + 745.0, + 958.75, + 745.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.75, + 753.5, + 393.75, + 753.5, + 393.75, + 767.5, + 380.75, + 767.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 1418.0, + 340.0, + 1418.0, + 340.0, + 1494.0, + 302.0, + 1494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 664.0, + 1407.0, + 703.0, + 1407.0, + 703.0, + 1473.0, + 664.0, + 1473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 942.0, + 1407.0, + 1015.0, + 1407.0, + 1015.0, + 1445.0, + 942.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1030.0, + 1400.0, + 1146.0, + 1400.0, + 1146.0, + 1474.0, + 1030.0, + 1474.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1438.0, + 1119.0, + 1438.0, + 1119.0, + 1456.0, + 1091.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1470.0, + 331.0, + 1470.0, + 331.0, + 1483.0, + 321.0, + 1483.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 341.0, + 1449.0, + 661.0, + 1449.0, + 661.0, + 1534.0, + 341.0, + 1534.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 668.0, + 1459.0, + 700.0, + 1459.0, + 700.0, + 1508.0, + 668.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 706.0, + 1508.0, + 718.0, + 1508.0, + 718.0, + 1521.0, + 706.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 739.0, + 1450.0, + 1024.0, + 1450.0, + 1024.0, + 1538.0, + 739.0, + 1538.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1034.0, + 1452.0, + 1076.0, + 1452.0, + 1076.0, + 1512.0, + 1034.0, + 1512.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1077.0, + 1508.0, + 1088.0, + 1508.0, + 1088.0, + 1521.0, + 1077.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1091.0, + 1450.0, + 1357.0, + 1450.0, + 1357.0, + 1531.0, + 1091.0, + 1531.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1363.0, + 1506.0, + 1393.0, + 1506.0, + 1393.0, + 1524.0, + 1363.0, + 1524.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 370.0, + 1548.0, + 596.0, + 1548.0, + 596.0, + 1577.0, + 370.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 711.0, + 1546.0, + 983.0, + 1546.0, + 983.0, + 1582.0, + 711.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1078.0, + 1548.0, + 1354.0, + 1548.0, + 1354.0, + 1577.0, + 1078.0, + 1577.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 958.0, + 1438.0, + 984.0, + 1438.0, + 984.0, + 1455.0, + 958.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 72.0, + 816.0, + 72.0, + 816.0, + 110.0, + 294.0, + 110.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1061.0, + 1197.0, + 1061.0, + 1197.0, + 1098.0, + 295.0, + 1098.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 830.0, + 2084.0, + 870.0, + 2084.0, + 870.0, + 2125.0, + 830.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 403.0, + 1601.0, + 1294.0, + 1601.0, + 1294.0, + 1641.0, + 403.0, + 1641.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 302.0, + 230.0, + 1064.0, + 230.0, + 1064.0, + 263.0, + 302.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1079.0, + 1547.0, + 1354.0, + 1547.0, + 1354.0, + 1581.0, + 1079.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1121.0, + 1402.0, + 1121.0, + 1402.0, + 1155.0, + 294.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1153.0, + 1403.0, + 1153.0, + 1403.0, + 1183.0, + 296.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1180.0, + 1405.0, + 1180.0, + 1405.0, + 1217.0, + 294.0, + 1217.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1212.0, + 1404.0, + 1212.0, + 1404.0, + 1246.0, + 293.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1241.0, + 1405.0, + 1241.0, + 1405.0, + 1277.0, + 293.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1270.0, + 1405.0, + 1270.0, + 1405.0, + 1308.0, + 292.0, + 1308.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1303.0, + 1405.0, + 1303.0, + 1405.0, + 1339.0, + 292.0, + 1339.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1334.0, + 874.0, + 1334.0, + 874.0, + 1368.0, + 293.0, + 1368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 286.0, + 1406.0, + 286.0, + 1406.0, + 323.0, + 294.0, + 323.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 318.0, + 1405.0, + 318.0, + 1405.0, + 353.0, + 295.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 348.0, + 1405.0, + 348.0, + 1405.0, + 384.0, + 294.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 378.0, + 1405.0, + 378.0, + 1405.0, + 413.0, + 292.0, + 413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 410.0, + 1405.0, + 410.0, + 1405.0, + 444.0, + 294.0, + 444.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 443.0, + 1403.0, + 443.0, + 1403.0, + 473.0, + 295.0, + 473.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 472.0, + 1403.0, + 472.0, + 1403.0, + 505.0, + 292.0, + 505.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 497.0, + 1405.0, + 497.0, + 1405.0, + 537.0, + 294.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 530.0, + 1406.0, + 530.0, + 1406.0, + 568.0, + 294.0, + 568.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 554.0, + 1406.0, + 554.0, + 1406.0, + 602.0, + 290.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 591.0, + 1405.0, + 591.0, + 1405.0, + 627.0, + 292.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 620.0, + 771.0, + 620.0, + 771.0, + 657.0, + 294.0, + 657.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 26, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/parse/train/rkcya1ZAW/rkcya1ZAW_content_list.json b/parse/train/rkcya1ZAW/rkcya1ZAW_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ec4f0a91cfb33e31aceeb9268963428b7c9b133 --- /dev/null +++ b/parse/train/rkcya1ZAW/rkcya1ZAW_content_list.json @@ -0,0 +1,2033 @@ +[ + { + "type": "text", + "text": "CONTINUOUS-TIME FLOWS FOR EFFICIENT INFERENCE AND DENSITY ESTIMATION ", + "text_level": 1, + "bbox": [ + 176, + 98, + 826, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 170, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two fundamental problems in unsupervised learning are efficient inference for latent-variable models and robust density estimation based on large amounts of unlabeled data. For efficient inference, normalizing flows have been recently developed to approximate a target distribution arbitrarily well. In practice, however, normalizing flows only consist of a finite number of deterministic transformations, and thus they possess no guarantee on the approximation accuracy. For density estimation, the generative adversarial network (GAN) has been advanced as an appealing model, due to its often excellent performance in generating samples. In this paper, we propose the concept of continuous-time flows (CTFs), a family of diffusion-based methods that are able to asymptotically approach a target distribution. Distinct from normalizing flows and GANs, CTFs can be adopted to achieve the above two goals in one framework, with theoretical guarantees. Our framework includes distilling knowledge from a CTF for efficient inference, and learning an explicit energy-based distribution with CTFs for density estimation. Experiments on various tasks demonstrate promising performance of the proposed CTF framework, compared to related techniques. ", + "bbox": [ + 233, + 268, + 766, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 522, + 336, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficient inference and robust density estimation are two important goals in unsupervised learning. In fact, inference and density estimation can be unified from the perspective of learning desired target distributions. In inference problems, one is given an unnormalized distribution (e.g., the posterior distribution), and the goal is to learn a distribution that is close to the unnormalized distribution. In density estimation, one tries to learn an unknown data distribution given samples from it. It is also helpful to make a distinction between two types of representations for learning distributions: explicit and implicit methods (Mohamed & Lakshminarayanan, 2017). Explicit methods provide a prescribed parametric form for the distribution, while implicit methods learn a stochastic procedure to directly generate samples from the unknown distribution. ", + "bbox": [ + 174, + 556, + 825, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing deep generative models can easily be identified from this taxonomy. For example, the standard variational autoencoder (VAE) (Kingma & Welling, 2014; Rezende et al., 2014) is an important example of an explicit inference method. Within the inference arm (encoder) of a VAE, recent research has focused on improving the accuracy of the approximation to the posterior distribution on latent variables (codes) using normalizing flow (Rezende & Mohamed, 2015). Normalizing flow is particularly interesting due to its ability to approximate the posterior distribution arbitrarily well, while maintaining explicit parametric forms. On the other hand, Stein VAE (Feng et al., 2017; Pu et al., 2017b) is an implicit inference method, as it only learns to draw samples to approximate posteriors, without assuming an explicit form for the distribution.. For density estimation on observed data, the generative adversarial network (GAN) can be regarded as an implicit density estimation method (Ranganath et al., 2016; Huszár, 2017; Mohamed & Lakshminarayanan, 2017), in the sense that one may sample from the distribution (regarded as a representation of the unknown distribution), but an explicit form for the distribution is not estimated. GAN has recently been augmented by Flow-GAN (Grover et al., 2017) to incorporate a likelihood term for explicit density estimation. Further, there also are some works trying to perform inference within the implicit density estimation framework, e.g., the real-valued non-volume preserving (real NVP) transformations algorithm (Dinh et al., 2017) was proposed as a tractable yet expressive approach to model high-dimensional data. ", + "bbox": [ + 174, + 688, + 825, + 924 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some aforementioned methods rely on the concept of flows. A flow defines a series of transformations for a random variable (RV), such that the distribution of the RV evolves from a simple distribution to a more complex distribution. When the sequence of transformations are indexed on a discrete-time domain (e.g., indexed with integers) with a finite number of transformations, this method is referred to as a normalizing flow (Rezende & Mohamed, 2015). Various efficient implementations of normalizing flows have been proposed, such as the planar, radial (Rezende & Mohamed, 2015), Householder (Tomczak & Welling, 2016), and inverse autoregressive flows (Kingma et al., 2016). One theoretical limitation of existing normalizing flows is that there is no guarantee on the approximation accuracy due to the finite number of transformations. ", + "bbox": [ + 174, + 103, + 825, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By contrast, little work has explored the applicability of continuous-time flows (CTFs) in deep generative models, where a sequence of transformations are indexed on a continuous-time domain (e.g., indexed with real numbers). There are at least two reasons encouraging research in this direction: $i ,$ ) CTFs are more general than traditional normalizing flows in terms of modeling flexibility, due to the intrinsic infinite number of transformations; $i i$ ) CTFs are more theoretically grounded, in the sense that they are guaranteed to approach a target distribution asymptotically (details provided in Section 2.2). Unfortunately, these advantages also bring challenges for efficient learning, in that: i) it is difficult to optimize over the variational lower bound in the inference framework, due to the extra randomness introduced in CTFs; ii) it is difficult to design algorithms for efficient learning of CTF-based models, due to the induced infinite number of transformations. ", + "bbox": [ + 174, + 236, + 825, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose efficient ways to apply CTFs for the two motivating tasks. Based on the CTF, our framework learns to drawn samples directly from desired distributions (e.g., the unknown posterior and data distributions) for both inference and density estimation tasks. In addition, we are able to learn an explicit form of the unknown data distribution for density estimation∗. This shares a similar flavor as Wang & Liu (2017); Feng et al. (2017). Specifically, $i$ ) for efficient inference, we first show that optimizing the variational lower bound with CTFs can be achieved by decomposing the optimization problem into a sequence of sub-optimization problems, based on a variational formulation of the Fokker-Planck equations from statistical physics (Jordan et al., 1998). Based on this decomposition, we derive bounds on the approximation errors when applying numerical methods to solve a CTF. For computational efficiency, we generalize ideas from Gershman & Goodman (2014) to distill knowledge of a CTF into an efficient inference network; $\\ddot { u }$ ) for density estimation, we propose to use a flexible Gibbsian-style distribution (implemented by a deep neural network) to model an unknown data distribution, whose samples can be drawn by learning a stochastic generator with our CTF framework. The Gibbsian-style data distribution and the stochastic generator are learned alternatively, leading to a learning procedure that is connected to the GAN framework (Goodfellow et al., 2014), but that yields an explicit distribution for the data. We conduct a number of experiments on real datasets, demonstrating excellent performance of the proposed framework, relative to existing representative approaches. ", + "bbox": [ + 174, + 382, + 825, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 PRELIMINARIES ", + "text_level": 1, + "bbox": [ + 176, + 655, + 339, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first review related techniques of performing efficient inference and density estimation in the machine learning literature. We then introduce the general concept of continuous-time flows. ", + "bbox": [ + 174, + 689, + 823, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 EFFICIENT INFERENCE AND DENSITY ESTIMATION ", + "text_level": 1, + "bbox": [ + 174, + 739, + 562, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficient inference with normalizing flows Consider a probabilistic generative model with observation $\\mathbf { x } \\in \\mathbb { R } ^ { D }$ and latent variable $\\textbf { z } \\in \\mathbb { R } ^ { L }$ such that $\\mathbf { x } \\mid \\mathbf { z } \\sim p _ { \\pmb { \\theta } } ( \\mathbf { x } \\mid \\mathbf { z } )$ with $\\mathbf { z } \\sim p ( \\mathbf { z } )$ . For efficient inference of $\\mathbf { z }$ , the VAE (Kingma & Welling, 2014) introduces the concept of an inference network (recognition model or encoder), $q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )$ , as a variational distribution in the VB framework. An inference network is typically a stochastic (nonlinear) mapping from the input $\\mathbf { x }$ to the latent $\\mathbf { z }$ , with associated parameters $\\phi$ . For example, one of the simplest inference networks is defined as $\\begin{array} { r } { q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) = \\mathcal { N } ( \\bar { \\mathbf { z } } ; \\mu _ { \\phi } ( \\mathbf { x } ) , \\mathrm { d i a g } ( \\sigma _ { \\phi } ^ { 2 } ( \\mathbf { x } ) ) ) } \\end{array}$ , where the mean function $\\mu _ { \\phi } ( \\mathbf { x } )$ and the standardderivation function $\\sigma _ { \\phi } ( \\mathbf { x } )$ are specified via deep neural networks parameterized by $\\phi$ . Parameters are learned by minimizing the evidence lower bound (ELBO), i.e., the KL divergence between $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ ", + "bbox": [ + 174, + 766, + 826, + 893 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and $q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )$ $\\mathbf { \\cdot } ) \\colon \\mathrm { K L } \\left( q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) \\triangleq \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) } \\left[ \\log q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right]$ , via stochastic gradient descent (Bottou, 2012). ", + "bbox": [ + 173, + 103, + 825, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One limitation of the VAE framework is that $q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )$ is often restricted to simple distributions for feasibility, e.g., the normal distribution discussed above, and thus the gap between $q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )$ and $p _ { \\pmb { \\theta } } ( \\mathbf { z } | \\mathbf { x } )$ is typically large for complicated posterior distributions. Normalizing flows is a recently proposed VB-based technique designed to mitigate this problem (Rezende & Mohamed, 2015). The idea is to augment $\\mathbf { z }$ via a sequence of deterministic invertible transformations $\\{ \\mathcal { T } _ { k } : \\mathbb { R } ^ { L } \\to \\mathbb { R } ^ { \\dot { L } } \\} _ { k = 1 } ^ { K }$ , such that: $\\mathbf { z } _ { 0 } \\sim q _ { \\phi } ( \\cdot | \\mathbf { x } ) , \\mathbf { z } _ { 1 } = \\mathcal { T } _ { 1 } ( \\mathbf { z } _ { 0 } ) , \\cdot \\cdot \\cdot , \\mathbf { z } _ { K } = \\mathcal { T } _ { K } ( \\mathbf { z } _ { K - 1 } )$ . ", + "bbox": [ + 173, + 138, + 825, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note the transformations $\\{ \\mathcal { T } _ { k } \\}$ are typically endowed with different parameters, and we absorb them into $\\phi$ . Because the transformations are deterministic, the distribution of $\\mathbf { z } _ { K }$ can be written as $\\begin{array} { r } { q ( \\mathbf { z } _ { K } ) = q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) \\prod _ { k = 1 } ^ { K } \\left| \\operatorname* { d e t } \\frac { \\partial { \\mathcal T } _ { k } } { \\partial \\mathbf { z } _ { k } } \\right| ^ { - 1 } } \\end{array}$ via the change of variable formula. As a result, the ELBO for normalizing flows becomes: ", + "bbox": [ + 173, + 228, + 825, + 296 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/b6c35f7eded82e866476d16ef2cdbfb92964dd3c567653511718b2a8f9ebdeb8.jpg", + "text": "$$\n\\begin{array} { l } { \\displaystyle \\mathrm { K L } \\left( q _ { \\phi } ( \\mathbf { z } _ { K } \\mid \\mathbf { x } ) \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) = } \\\\ { \\displaystyle \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\log q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) \\right] - \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } _ { K } ) \\right] - \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\sum _ { k = 1 } ^ { K } \\log \\left| \\mathrm { d e t } \\frac { \\partial T _ { k } } { \\partial \\mathbf { z } _ { k } } \\right| \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 205, + 301, + 790, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, transformations $\\mathcal { T } _ { k }$ of a simple parametric form are employed to make the computations tractable (Rezende & Mohamed, 2015). Our method for inference generalizes these discrete-time transformation to continuous-time transformations, ensuring convergence of the transformations to the target distribution. ", + "bbox": [ + 173, + 372, + 825, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Density estimation overview There exist implicit and explicit density-estimation methods. Implicit density models such as GAN provide a flexible way to draw samples directly from unknown data distributions (via a deep neural network (DNN) called a generator with stochastic inputs) without explicitly modeling their density forms; whereas explicit models such as the pixel RNN/CNN (van den Oord et al., 2016) define and learn explicit forms of the unknown data distributions. This gives the advantage that the likelihood for a test data point can be explicitly evaluated. However, the generation of samples is typically time-consuming due to the sequential generation nature. ", + "bbox": [ + 173, + 444, + 825, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to Wang & Liu (2017), our CTF-based approach in Section 4 provides an alternative way for this problem, by simultaneously learning an explicit Gibbsian-style data distribution (estimated density) and a generator whose generated samples match the learned Gibbsian distribution. This not only gives us the advantage of explicit density modeling but also provides an efficient way to generate samples. ", + "bbox": [ + 174, + 549, + 825, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2 CONTINUOUS-TIME FLOWS ", + "text_level": 1, + "bbox": [ + 176, + 636, + 401, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We notice two potential limitations with traditional normalizing flows: i) given specified transformations $\\{ \\mathcal { T } _ { k } \\}$ , there is no guarantee that the distribution of $\\mathbf { z } _ { K }$ could exactly match $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ ; ii) the randomness is only introduced in $\\mathbf { z } _ { 0 }$ (from the inference network), limiting the representation power. We specify CTFs where the transformations are indexed by real numbers, thus they could be considered as consisting of an infinite number of transformations. Further, we consider stochastic flows where randomness is injected in a continuous-time manner. In fact, the concept of CTFs (such as the Hamiltonian flow) has been introduced in Rezende & Mohamed (2015), without further development on efficient inference. ", + "bbox": [ + 173, + 661, + 825, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider a flow on $\\mathbb { R } ^ { L }$ , defined as the mapping† $\\mathcal { T } : \\mathbb { R } ^ { L } \\times \\mathbb { R } \\to \\mathbb { R } ^ { L }$ such that‡ we have $\\boldsymbol { \\mathcal { T } } ( \\mathbf { Z } , 0 ) = \\mathbf { z }$ and $\\mathcal { T } ( \\mathcal { T } ( \\mathbf { Z } , t ) , s ) = \\mathcal { T } ( \\mathbf { Z } , s + t )$ , for all $\\mathbf { Z } \\in \\mathbb { R } ^ { L }$ and $s , t \\in \\mathbb { R }$ . A typical example of this family is defined as $\\mathcal { T } ( \\mathbf { Z } , t ) = \\mathbf { Z } _ { t }$ , where $\\mathbf { Z } _ { t }$ is driven by a diffusion of the form: ", + "bbox": [ + 174, + 779, + 825, + 823 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/c8ef519cf5f4b039dffc5a4a43dd8e677643835401771d015f79f2b1b33b1d71.jpg", + "text": "$$\n\\mathrm { d } \\mathbf { Z } _ { t } = F ( \\mathbf { Z } _ { t } ) \\mathrm { d } t + V ( \\mathbf { Z } _ { t } ) \\mathrm { d } { \\mathcal { W } } .\n$$", + "text_format": "latex", + "bbox": [ + 393, + 830, + 604, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here ${ \\cal F } : \\mathbb { R } ^ { L } \\mathbb { R } ^ { L }$ , $V ~ : ~ \\mathbb { R } ^ { L \\times L } ~ \\to ~ \\mathbb { R } ^ { L }$ are called the drift term and diffusion term, respectively; $\\mathcal { W }$ is the standard $L$ -dimensional Brownian motion. In the context of inference, we seek to make the stationary distribution of $\\mathbf { Z } _ { t }$ approach $p _ { \\pmb { \\theta } } ( \\mathbf { z } | \\mathbf { x } )$ . One solution for this is to set $\\begin{array} { r } { F ( \\mathbf { Z } _ { t } ) = \\frac { 1 } { 2 } \\nabla _ { \\mathbf { z } } \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } \\stackrel { . } { = } \\mathbf { Z } _ { t } ) } \\end{array}$ and $V ( \\mathbf { Z } _ { t } ) = \\mathbf { I } _ { L }$ with $\\mathbf { I } _ { L }$ the $L \\times L$ identity matrix. The resulting diffusion is called Langevin dynamics Welling & Teh (2011). Denoting the distribution of $\\mathbf { Z } _ { t }$ as $\\rho _ { t }$ , it is well known Risken (1989) that $\\rho _ { t }$ is characterized by the Fokker-Planck (FP) equation: ", + "bbox": [ + 173, + 102, + 825, + 188 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/b92a895db7d555542f06f389aa945d7d08ce8aa1198dc2a6b34e3c7cf0d85a8a.jpg", + "text": "$$\n\\frac { \\partial \\rho _ { t } } { \\partial t } = - \\nabla _ { \\mathbf { z } } \\cdot \\left( \\rho _ { t } F ( \\mathbf { Z } _ { t } ) \\right) + \\nabla _ { \\mathbf { z } } \\nabla _ { \\mathbf { z } } \\colon \\left( \\rho _ { t } V ( \\mathbf { Z } _ { t } ) V ^ { \\top } ( \\mathbf { Z } _ { t } ) \\right) ~ ,\n$$", + "text_format": "latex", + "bbox": [ + 313, + 190, + 681, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf { a } \\cdot \\mathbf { b } \\triangleq \\mathbf { a } ^ { \\intercal } \\mathbf { b }$ for vectors $\\mathbf { a }$ and $\\mathbf { b }$ , $\\mathbf { A } : \\mathbf { B } \\triangleq \\operatorname { t r a c e } ( \\mathbf { A } ^ { \\intercal } \\mathbf { B } )$ for matrices $\\mathbf { A }$ and $\\mathbf { B }$ ", + "bbox": [ + 176, + 224, + 733, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplicity, we consider the flow defined by the Langevin dynamics specified above, though our results generalize to other stochastic flows Dorogovtsev & Nishchenko (2014). In the following, we specify the ELBO under a CTF, which can then be readily solved by a discretized numerical scheme, based on the results from Jordan et al. (1998). An approximation error bound for the scheme is also derived. We defer proofs of our theoretical results to the Supplementary Material (SM) for conciseness. ", + "bbox": [ + 173, + 247, + 825, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3 CONTINUOUS-TIME FLOWS FOR INFERENCE ", + "text_level": 1, + "bbox": [ + 174, + 351, + 576, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first give an overview of our CTF-based method for efficient inference. We adopt the VAE/normalizing-flow framework with an encoder-decoder structure. An important difference is that instead of feeding data to an encoder and sampling a latent representation in the output as in VAE, we concatenate the data with independent noise as input and directly generate output samples§. These output samples are then driven by the CTF to approach the true posterior distribution. In the learning process, the implicit transformations from the CTF are sequentially distilled into the inference network by amortized learning, making the inference network flexible enough to represent the true posterior distribution. In the following subsections, we specify our framework in detail. ", + "bbox": [ + 173, + 381, + 825, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1 THE VARIATIONAL LOWER BOUND AND DISCRETIZED APPROXIMATION ", + "text_level": 1, + "bbox": [ + 174, + 511, + 704, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first incorporate CTF into the normalizing-flow framework by writing out the corresponding ELBO. Note that there are two steps in the inference process. First, an initial $\\mathbf { z } _ { 0 }$ is drawn from the inference network $q _ { \\phi } ( \\cdot | { \\bf x } )$ ; second, $\\mathbf { z } _ { 0 }$ is evolved via a diffusion such as (2) for time $T$ (via the transformation ${ \\bf Z } _ { T } = \\mathcal { T } ( { \\bf z } _ { 0 } , T ) )$ . Consequently, the ELBO for CTF can be written as ", + "bbox": [ + 174, + 535, + 825, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/c1559d6ae3f5d8a3453494946f6abbdb87fe7a5c0fac8cd32ac2af541d15093a.jpg", + "text": "$$\n\\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\mathbb { E } _ { \\rho _ { T } } \\left[ \\log \\rho _ { T } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { Z } _ { T } ) + \\log \\left| \\operatorname* { d e t } \\frac { \\partial \\mathbf { Z } _ { T } } { \\partial \\mathbf { z } _ { 0 } } \\right| \\right] \\triangleq \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\mathcal { F } _ { 1 } ( \\mathbf { x } , \\mathbf { z } _ { 0 } ) \\right] \\ .\n$$", + "text_format": "latex", + "bbox": [ + 181, + 594, + 795, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note the term $\\mathcal { F } _ { 1 } ( \\mathbf { x } , \\mathbf { z } _ { 0 } )$ is intractable to calculate, in that $i$ ) $\\rho _ { T }$ does not have an explicit form; $i i$ ) the Jacobian $\\frac { \\partial \\mathbf { Z } _ { T } } { \\partial \\mathbf { z } _ { 0 } }$ is generally infeasible. In the following, we propose an approximate solution for problem i). Learning by avoiding problem $\\romannumeral 2$ ) is presented in Section 3.2 via amortization. ", + "bbox": [ + 174, + 633, + 825, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For problem $i$ ), a reformulation of the results from Jordan et al. (1998) leads to a nice way to approximate $\\rho _ { t }$ in Lemma 1. Note in practice we adopt an implicit method which uses samples to approximate the solution in Lemma 1 for feasibility, detailed in (6). ", + "bbox": [ + 174, + 686, + 825, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1. Assume that $\\log p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } ) \\leq C _ { 1 }$ is infinitely differentiable, and $\\| \\nabla _ { \\mathbf { z } } \\log p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } ) \\| \\leq$ $C _ { 2 } \\left( 1 + C _ { 1 } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) ( \\forall \\mathbf { x } , \\mathbf { z } )$ for some constants $\\left\\{ C _ { 1 } , C _ { 2 } \\right\\}$ . Let $T = h K$ ( $h$ is the stepsize in discretization and $K$ is the number of transformations), $\\rho _ { 0 } \\triangleq q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ , and $\\{ \\tilde { \\rho } _ { k } \\} _ { k = 1 } ^ { K }$ be the solution of the functional optimization problem: ", + "bbox": [ + 174, + 731, + 825, + 791 + ], + "page_idx": 3 + }, + { + "type": "equation", + "img_path": "images/036ee732096b4b559a01fd9f8be2e56fd397ab0f6997650f124c5929577b9532.jpg", + "text": "$$\n\\tilde { \\rho } _ { k } = \\arg \\operatorname* { m i n } _ { \\rho \\in { \\mathcal { K } } } K L \\left( \\rho \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) + \\frac { 1 } { 2 h } W _ { 2 } ^ { 2 } \\left( \\tilde { \\rho } _ { k - 1 } , \\rho \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 325, + 792, + 671, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\begin{array} { r } { W _ { 2 } ^ { 2 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) \\triangleq \\operatorname* { i n f } _ { p \\in \\mathcal { P } ( \\mu _ { 1 } , \\mu _ { 2 } ) } \\int \\left\\| \\mathbf { x } - \\mathbf { y } \\right\\| _ { 2 } ^ { 2 } p ( \\mathrm { d } \\mathbf { x } , \\mathrm { d } \\mathbf { y } ) , \\mathrm { ~ } } \\end{array}$ $W _ { 2 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right)$ is the 2nd-order Wasserstein distance, with $\\mathcal { P } ( \\mu _ { 1 } , \\mu _ { 2 } )$ being the space of joint distributions on $\\{ \\mu _ { 1 } , \\mu _ { 2 } \\}$ . $\\kappa$ is the space of probability distributions with the finite 2nd-order moment. Then $\\tilde { \\rho } _ { K }$ converges to $\\rho _ { T }$ in the limit of $h 0$ , i.e., $\\mathrm { l i m } _ { h 0 } \\tilde { \\rho } _ { K } = \\rho _ { T }$ , where $\\rho _ { T }$ is the solution of the FP equation (3) at time $T$ . ", + "bbox": [ + 173, + 829, + 826, + 890 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1 reveals an interesting way to compute $\\rho _ { T }$ via a sequence of functional optimization problems. By comparing it with the objective of the traditional normalizing flow, which minimizes the KL-divergence between $\\rho _ { K }$ and $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ , at each sub-optimization-problem in Lemma 1, it minimizes the KL-divergence between $\\tilde { \\rho } _ { k }$ and $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ , plus a regularization term as the Wasserstein distance between $\\tilde { \\rho } _ { k - 1 }$ and $\\tilde { \\rho } _ { k }$ . The extra Wasserstein-distance term arises naturally due to the fact that the Langevin diffusion can be explained as a gradient flow whose geometry is equipped with the Wasserstein distance (Otto, 1998). From another point of view, it is known that the Wasserstein distance is a better metric for probability distributions than the KL-divergence, especially in the case of non-overlapping domains (Arjovsky & Bottou, 2017; Arjovsky et al., 2017). By using the Wasserstein term as a regularizer, the CTF alleviates the issue in non-overlapping domains by introducing the Brownian-motion (noise) term in the evolution (2). This relates to the idea in (Arjovsky & Bottou, 2017), in which noise is added in parameter updates to alleviate the intrinsic drawback of the KL-divergence metric. ", + "bbox": [ + 174, + 103, + 825, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimization problem in Lemma 1 is difficult to deal with directly. In practice, we instead approximate the discretization in an equivalent way by simulation from the CTF. Starting from $\\mathbf { z } _ { 0 }$ $\\mathbf { z } _ { k }$ $( k = 0 , \\cdots , K - 1 )$ is fed into a transformation $\\mathcal { T } _ { k }$ (specified below), resulting in ${ \\mathbf z } _ { k + 1 }$ whose distribution coincides with $\\tilde { \\rho } _ { k + 1 }$ in Lemma 1. The discretization procedure is illustrated in Figure 1. We must specify the transformations $\\mathcal { T } _ { k }$ . For each $k$ , let $t = h k$ ; we can conclude from Lemma 1 that $\\operatorname* { l i m } _ { h \\to 0 } \\tilde { \\rho } _ { k } = \\rho _ { t }$ . From FP theory, $\\rho _ { t }$ is obtained by solving the diffusion (2) with initial condition ${ \\bf Z } _ { 0 } = { \\bf z } _ { 0 }$ . It is thus reasonable to specify the transformation $\\mathcal { T } _ { k }$ as the $k$ -th step of a numerical integrator for (2). Specifically, we specify $\\mathcal { T } _ { k }$ as a stochastic transformation: ", + "bbox": [ + 173, + 290, + 825, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/197d4f4ace878ed60868e893eae9ffb5758c77de999f3c246be7af44f22053e1.jpg", + "text": "$$\n{ \\bf z } _ { k } = { \\mathcal T } _ { k } ( { \\bf z } _ { k - 1 } ) \\triangleq { \\bf z } _ { k - 1 } + { \\cal F } ( { \\bf z } _ { k - 1 } ) h + V ( { \\bf z } _ { k - 1 } ) \\zeta _ { k } ,\n$$", + "text_format": "latex", + "bbox": [ + 325, + 420, + 671, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathsf { \\Sigma } \\xi _ { k } \\sim \\mathcal { N } ( \\mathbf { 0 } , h \\mathbf { I } _ { L } )$ is drawn from an isotropic normal. Note the transformation defined here is stochastic, thus we only get samples from $\\tilde { \\rho } _ { K }$ at the end. A natural way to approximate $\\tilde { \\rho } _ { K }$ is to use the empirical sample distribution, i.e., $\\begin{array} { r } { \\tilde { \\rho } _ { K } \\approx \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\delta _ { { \\bf z } _ { k } } \\triangleq \\bar { \\rho } _ { T } } \\end{array}$ with $\\delta _ { \\mathbf { z } }$ a point mass at $\\mathbf { z }$ Afterwards, $\\tilde { \\rho } _ { K }$ (thus $\\bar { \\rho } _ { T }$ ) will be used to approximate the true $\\rho _ { T }$ from (3). ", + "bbox": [ + 174, + 457, + 825, + 516 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bcb7adee8eb928455458752caa91f82356f295994b68b570ae3c02cde0d35129.jpg", + "image_caption": [ + "Figure 1: Discretized approximation (right) of a continuous-time flow (left). Densities $\\{ \\tilde { \\rho } _ { k } \\}$ of $\\left\\{ { \\bf z } _ { k } \\right\\}$ evolve via transformations $\\{ \\mathcal { T } _ { k } \\}$ , with $\\tilde { \\rho } _ { k } \\to \\rho _ { h k }$ when $h 0$ for each $k$ due to Lemma 1. " + ], + "image_footnote": [], + "bbox": [ + 207, + 530, + 792, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that we use the simple sample averaging for the convenience of analysis, and the approximation for $\\rho _ { T }$ is not necessarily optimal. Better approximation can be obtained by assigning more weights to the more recent samples. However, this leads to more challenges in theoretical analysis, an interesting future direction to pursue. In the following, we study how well $\\bar { \\rho } _ { T }$ approximates $\\rho _ { T }$ . Following literature on numerical approximation for Itô diffusions (Vollmer et al., 2016; Chen et al., 2015), we consider a 1-Lipschitz test function $\\psi : \\mathbb { R } ^ { L } \\mathbb { R }$ , and use the mean square error (MSE) bound to measure the closeness of $\\bar { \\rho } _ { T }$ and $\\rho _ { T }$ , defined as: $\\begin{array} { r } { \\mathbf { M S E } ( \\bar { \\rho } _ { T } , \\rho _ { T } ; \\psi ) \\triangleq \\mathbb { E } \\left( \\int \\psi ( \\mathbf { z } ) ( \\tilde { \\rho } _ { T } - \\rho _ { T } ) ( \\mathbf { z } ) \\mathrm { d } \\mathbf { z } \\right) ^ { 2 } } \\end{array}$ , where the expectation is taken over all the randomness in the construction of $\\tilde { \\rho } _ { T }$ . Note that our goal is related but different from the standard setup as in Vollmer et al. (2016); Chen et al. (2015), which studies the closeness of $\\bar { \\rho } _ { T }$ to $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ . We need to adopt the assumptions from Vollmer et al. (2016); Chen et al. (2015), which are described in the Supplementary Material (SM). The assumptions are somewhat involved but essentially require coefficients of the diffusion (2) to be well-behaved. We derive the following bound for the MSE of the sampled approximation, $\\bar { \\rho } _ { T }$ , and the true distribution. ", + "bbox": [ + 173, + 733, + 826, + 919 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 2. Under Assumption $^ { l }$ in the SM, assume that $\\begin{array} { r } { \\int \\rho _ { T } ( \\mathbf { z } ) p _ { \\pmb \\theta } ^ { - 1 } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } < \\infty } \\end{array}$ and there exists a constant C such tha t dW 22 (ρT ,pθ(x,z))dt ≥ CW 22 (ρT , pθ(x, z)), the MSE is bounded as ", + "bbox": [ + 173, + 102, + 825, + 140 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/ee9e3348c5bc3d9582efb151f896bd1f6af3cf1fc49f57727dad0c5592eecc10.jpg", + "text": "$$\nM S E ( \\bar { \\rho } _ { T } , \\rho _ { T } ; \\psi ) = O \\left( \\frac { 1 } { h K } + h ^ { 2 } + e ^ { - 2 C h K } \\right) .\n$$", + "text_format": "latex", + "bbox": [ + 336, + 146, + 661, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The last assumption in Theorem 2 requires $\\rho _ { T }$ to evolve fast through the FP equation, which is a standard assumption used to establish convergence to equilibrium for FP equations (Bolley et al., 2012). The MSE bound consists of three terms, the first two terms come from numerical approximation of the continuous-time diffusion, whereas the third term comes from the convergence bound of the FP equation in terms of the Wasserstein distance (Bolley et al., 2012). When the time $T = h K$ is large enough, the third term may be ignored due to its exponential-decay rate. Moreover, in the infinite-time limit, the bound endows a bias proportional to $h$ ; this, however, can be removed by adopting a decreasing-step-size scheme in the numerical method, as in standard stochastic gradient MCMC methods (Teh et al., 2016; Chen et al., 2015). ", + "bbox": [ + 173, + 194, + 825, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Remark 3. To examine the optimal bound in Theorem 2, we drop out the term $e ^ { - 2 C h K }$ in the long-time case (when hK is large enough) for simplicity because it is in a much lower order term than the other terms. The optimal MSE bound (over h) decreases at a rate of $O \\left( K ^ { - 2 / 3 } \\right)$ , meaning that $O \\left( \\epsilon ^ { - 3 / 2 } \\right)$ steps of transformations in Figure 1 (right) are needed to reach an $\\epsilon$ -accurate approximation, i.e., $M S E \\leq \\epsilon .$ . This is computationally expensive. An efficient way for inference is thus imperative, developed in the next section. ", + "bbox": [ + 173, + 324, + 825, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2 EFFICIENT INFERENCE VIA AMORTIZATION ", + "text_level": 1, + "bbox": [ + 174, + 430, + 513, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Even though we approximate $\\rho _ { T }$ with $\\bar { \\rho } _ { T }$ , it is still infeasible to directly apply it to the ELBO in (4) as $\\bar { \\rho } _ { T }$ is discrete. To deal with this problem, we adopt the idea of “amortized learning” (Gershman $\\&$ Goodman, 2014) for efficient inference. The main idea is to optimize the two sets of parameters $\\phi$ and $\\pmb { \\theta }$ alternatively, based on different but related objective functions. ", + "bbox": [ + 173, + 455, + 826, + 512 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Updating $\\phi$ To explain the idea, first note that the ELBO can be equivalently written as ", + "bbox": [ + 173, + 529, + 759, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/dcb966fdbebae32ef2ad32f90a218d00774c5e9559f84804ead4f3f9f6f0a9e1.jpg", + "text": "$$\n\\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { \\rho _ { 0 } \\triangleq q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\mathbb { E } _ { \\rho _ { T } } \\left[ \\log \\rho _ { 0 } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { Z } _ { T } ) \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 323, + 551, + 673, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When $\\rho _ { 0 } = \\rho _ { T }$ , it is easy to see that: $\\begin{array} { r } { \\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { \\rho _ { 0 } } \\left[ \\log \\rho _ { 0 } - \\log p _ { \\theta } ( \\mathbf { Z } _ { T } | \\mathbf { x } ) \\right] + \\log p ( \\mathbf { x } ) = \\log p ( \\mathbf { x } ) . } \\end{array}$ which essentially makes the gap between $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ and $p _ { \\pmb { \\theta } } ( \\mathbf { Z } _ { T } \\mid \\mathbf { x } )$ vanished. As a result, our goal is to learn $\\phi$ such that $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ approaches $p _ { \\pmb { \\theta } } ( \\mathbf { Z } _ { T } \\mid \\mathbf { x } )$ . As mentioned previously, we will learn an implicit distribution of $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ (i.e., learn how to draw samples from $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ instead of its explicit form), as it allows us to chose a candidate distribution from a much larger distribution space, compared to explicitly defining $q _ { \\phi } ^ { \\bullet }$ . Consequently, $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ is implemented by a stochastic generator (a DNN parameterized by $\\phi$ ) $Q _ { \\phi } ( \\mathbf { z } _ { 0 } | \\mathbf { x } , \\omega )$ with input as the concatenation of $\\mathbf { x }$ and $\\omega$ , where $\\omega$ is a sample from an isotropic Gaussian distribution $q _ { 0 } ( \\omega )$ . Our goal is now translated to update the parameter $\\phi$ of $Q _ { \\phi } ( \\mathbf { z } _ { 0 } | \\mathbf { x } , \\omega )$ to $\\phi ^ { \\prime }$ such that the distribution of $\\{ \\mathbf { z } _ { 0 } ^ { \\prime } = Q _ { \\phi ^ { \\prime } } ( \\mathbf { z } _ { 0 } ^ { \\prime } \\mid \\mathbf { x } , \\omega ) \\}$ with $\\omega \\sim q _ { 0 } ( \\omega )$ matches that of $\\mathbf { z } _ { 1 }$ in the original generating process with $\\phi$ in Figure 1. In this way, the generating process of $\\mathbf { z } _ { 1 }$ via $\\mathcal { T } _ { 1 }$ is distilled into the parameterized generator $Q _ { \\phi } ( \\cdot )$ , eliminating the need to do a specific transformation via $\\mathcal { T } _ { 1 }$ in testing, and thus is very efficient. Specifically, we update $\\phi ^ { \\prime }$ such that ", + "bbox": [ + 173, + 579, + 826, + 761 + ], + "page_idx": 5 + }, + { + "type": "equation", + "img_path": "images/f1c3ff570f6a04c96384ca8c54209d1171a4e27623f34e24d4ac94b555ec213d.jpg", + "text": "$$\n\\phi ^ { \\prime } = \\arg \\operatorname* { m i n } _ { \\phi } \\mathcal { D } \\left( \\{ \\mathbf { z } _ { 0 } ^ { \\prime ( i ) } \\} , \\{ \\mathbf { z } _ { 1 } ^ { ( i ) } \\} \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 380, + 768, + 614, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\{ \\mathbf { z } _ { 0 } ^ { \\prime ( i ) } \\} _ { i = 1 } ^ { S }$ are a set of samples generated from $q _ { \\phi ^ { \\prime } } ( \\mathbf { z } _ { 0 } ^ { \\prime } \\mid \\mathbf { x } )$ via $Q _ { \\phi } ( \\cdot )$ , and $\\{ \\mathbf { z } _ { 1 } ^ { ( i ) } \\} _ { i = 1 } ^ { S }$ are samples drawn by $\\omega ^ { i } \\sim q _ { 0 } ( \\omega ) , \\tilde { \\mathbf { z } } _ { 0 } ^ { i } = Q _ { \\phi } ( \\cdot | \\mathbf { x } , \\omega ^ { i } ) , \\mathbf { z } _ { 1 } ^ { ( i ) } \\sim \\mathcal { T } _ { 1 } ( \\tilde { \\mathbf { z } } _ { 0 } ^ { i } ) ; \\mathcal { D } ( \\cdot , \\cdot )$ is a metric between samples such as the simple Euclidean distance or the more advanced Wasserstein distance (Arjovsky et al., 2017). The optimization is done by applying standard stochastic gradient descent (SGD). We call this procedure distilling knowledge from $\\mathcal { T } _ { 1 }$ to $Q _ { \\phi } ( \\cdot )$ . ", + "bbox": [ + 173, + 806, + 825, + 886 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After distilling knowledge from $\\mathcal { T } _ { 1 }$ , we apply the same procedure for other transformations $\\mathcal { T } _ { k } ( k > 1 )$ sequentially. The final inference network, represented by $q _ { \\phi } ( \\cdot | { \\bf x } )$ , can then well approximate the continuous-time flows, e.g., the distribution of ${ \\bf z } _ { 0 } \\sim q _ { \\phi } ( { \\bf \\cdot } | { \\bf \\dot { x } } )$ is close to $\\rho _ { T }$ from the CTF. This concept is illustrated in Figure 2. According to Theorem 2, the number of updates for $\\phi$ in training is still bounded by $O ( \\epsilon ^ { - 3 / 2 } )$ for an $\\epsilon$ -accurate MSE, however, inference in testing is significantly boosted since we do not need to simulate a long-time transformations as shown in Figure 1 (right). ", + "bbox": [ + 173, + 103, + 825, + 189 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Updating $\\pmb { \\theta }$ Given $\\phi , \\theta$ can be updated by simply optimizing the ELBO in (7), where $\\rho _ { T }$ is approximated by $\\bar { \\rho } _ { T }$ from the discretized CTF. Specifically, the expectation w.r.t. $\\rho _ { T }$ in (7) is approximated by a sample average from: ", + "bbox": [ + 174, + 203, + 823, + 246 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/ca6eb3f8efef2212242ece3eceda8ffb9a6fbe7b955b4dfea6c99f354fa8ceb7.jpg", + "text": "$$\n\\begin{array} { r } { { \\bf z } _ { 0 } \\sim q _ { \\phi } ( { \\bf z } _ { 0 } \\mid { \\bf x } ) , { \\bf z } _ { 1 } \\sim { \\mathcal T } _ { 1 } ( { \\bf z } _ { 0 } ) , { \\bf z } _ { 2 } \\sim { \\mathcal T } _ { 2 } ( { \\bf z } _ { 1 } ) , \\cdots , { \\bf z } _ { K } \\sim { \\mathcal T } _ { K } ( { \\bf z } _ { K - 1 } ) . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 274, + 248, + 722, + 267 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b4fbce5997de9b3fa8366c2ff82ece19b528c5a3a6783274270ba1e5c8532dc8.jpg", + "image_caption": [ + "Figure 2: Amortized learning of continuous-time flows for VAEs. From left to right: the initial architecture with $K$ -step transformations; For each step $k$ , $q _ { \\phi } ( \\cdot )$ is trained to match the distributin of $\\mathbf { z } _ { k }$ in CTFs; In the end, the CTF is distilled into $q _ { \\phi } ( \\cdot )$ . " + ], + "image_footnote": [], + "bbox": [ + 176, + 281, + 825, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To sum up, there are three main steps in learning a CTF-based VAE: ", + "bbox": [ + 173, + 422, + 620, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "1. Generate a sample path $\\left( { \\bf z } _ { 0 } , \\cdots , { \\bf z } _ { K } \\right)$ according to $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ and the discretized flow with transformations $\\{ \\mathcal { T } _ { k } \\}$ ; \n2. Update $\\phi$ according to (8); \n3. Optimize $\\pmb \\theta$ by minimizing the ELBO (7) with the generated sample path. ", + "bbox": [ + 210, + 446, + 823, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In testing, we use only the finally learned $q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )$ for inference (into which the CTF has been distilled), and hence testing is like the standard VAE. Since the discretized-CTF model is essentially a Markov chain, we call our model Markov-chain-based VAE (MacVAE). ", + "bbox": [ + 176, + 522, + 826, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4 CONTINUOUS TIME FLOWS FOR EXPLICIT DENSITY ESTIMATION ", + "text_level": 1, + "bbox": [ + 174, + 584, + 746, + 601 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We describe how to apply the proposed CTF framework to density estimation of the observed data. We assume that the density of the observation $\\mathbf { x }$ is characterized by a parametric Gibbsian-style probability model $\\begin{array} { r } { p _ { \\pmb \\theta } ( \\mathbf x ) = \\frac { 1 } { \\mathcal Z ( \\pmb \\theta ) } \\tilde { p } _ { \\pmb \\theta } ( \\mathbf x ) \\triangleq \\frac { 1 } { \\mathcal Z ( \\pmb \\theta ) } e ^ { U ( \\mathbf x ; \\pmb \\theta ) } } \\end{array}$ , where $\\tilde { p } _ { \\pmb { \\theta } } ( \\mathbf { x } )$ is an unnormalized version of $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ with parameter $\\pmb \\theta$ , $U ( \\mathbf { x } ; \\pmb { \\theta } ) \\triangleq \\log \\tilde { p } \\pmb { \\theta } ( \\mathbf { x } )$ is called the energy function (Zhao et al., 2017), and $\\begin{array} { r } { \\mathcal { Z } ( \\pmb { \\theta } ) \\triangleq \\int \\tilde { p } _ { \\pmb { \\theta } } ( \\mathbf { x } ) \\mathrm { d } \\mathbf { x } } \\end{array}$ is the normalizer. Note this form of distributions constitutes a very large class of distributions as long as the capacity of the energy function is large enough. This can be easily achieved by adopting a DNN to implement $U ( \\mathbf { x } ; \\pmb { \\theta } )$ , the setting we considered in this paper. Note our model can be placed in between existing implicit and explicit density estimation methods, because we model the data density with an explicit distribution form up to an intractable normalizer. Such distributions have been proved to be useful in real applications, e.g., Haarnoja et al. (2017) used them to model policies in deep reinforcement learning. ", + "bbox": [ + 173, + 614, + 826, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our goal is to learn $\\pmb { \\theta }$ given observations $\\{ { \\mathbf { x } } _ { i } \\} _ { i = 1 } ^ { N }$ , which can be achieved via the standard maximum likelihood estimator (MLE): ", + "bbox": [ + 173, + 784, + 821, + 813 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/298e6b70f6a0c30dffc39c8fdf15e179700b8c8dc6aba4cdb2d5a65499dc4691.jpg", + "text": "$$\n\\pmb \\theta = \\arg \\operatorname* { m a x } _ { \\pmb { \\theta } } \\sum _ { i = 1 } ^ { N } \\log p _ { \\pmb { \\theta } } ( \\mathbf { x } _ { i } ) \\triangleq \\arg \\operatorname* { m a x } _ { \\pmb { \\theta } } \\mathcal { M } ( \\{ \\mathbf { x } _ { i } \\} ; \\pmb { \\theta } )\n$$", + "text_format": "latex", + "bbox": [ + 323, + 815, + 673, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This is usually optimized via SGD, with the following gradient formula: ", + "bbox": [ + 173, + 861, + 647, + 876 + ], + "page_idx": 6 + }, + { + "type": "equation", + "img_path": "images/2a6c979d1f1697faded0e1e4d562e033e41293c21774d0324ea18d82eaf5be8b.jpg", + "text": "$$\n\\frac { \\partial \\mathcal { M } ( \\{ { \\bf x } _ { i } \\} ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\frac { \\partial U ( { \\bf x } _ { i } ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } - \\mathbb { E } _ { p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } ) } \\left[ \\frac { \\partial U ( \\mathbf { x } ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } \\right]\n$$", + "text_format": "latex", + "bbox": [ + 299, + 878, + 686, + 922 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 103, + 745, + 119 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/228270436d54101d23ac214311b62400ab48ffff489d72dfea30e56ca0e75f45.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Igorithm1 CTFs for generative models at the k-th iteration. D(·,:)is the same as (8). Input: parameters from last step 0(k-1),Φ(k-1) Output: updated parameters 0(k),𝜙(k)
1. Generate samples {X1,s}𝑠=1 via a discretized CTF: x0,s ~ q(k-1)(x0), X1,s ~ T1(x0,s); S 2. Update the generator by minimizing ({xo,s}S=1 are generated with the updated parameter (k):
(k) = arg minD({x1,s},{x6,s}） .
? 3. Update the energy-based model 0k by maximum likelihood, with gradient as (9) except replacing Ex~pe(x) with Ex~q(x);
", + "bbox": [ + 183, + 108, + 826, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The gradient formula requires an integration over the model distribution $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ , which can be approximated by Monte Carlo integration with samples. The sampling problem has been well studied for some particular energy-based distributions, for example, via contrastive divergence in restricted Boltzmann machines (Hinton, 2002). However, this does not fit into our setting directly. Here we adopt the idea of CTFs and propose to use a DNN guided by a CTF, which we call a generator, to generate approximate samples from the original model $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ . Specifically, we require that samples from the generator should well approximate the target $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ . This can be done by adopting the CTF idea above, i.e., distilling knowledge of a CTF (which approaches $p _ { \\pmb { \\theta } } ( \\mathbf { x } ) )$ to the generator. In testing, instead of generating samples from $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ via MCMC (which is complicated and time consuming), we generate samples from the generator directly. Furthermore, when evaluating the likelihood for test data, the unknown constant $\\mathcal { Z } ( \\pmb \\theta )$ of $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ can also be approximated by Monte Carlo integration with samples drawn from the generator. ", + "bbox": [ + 174, + 286, + 825, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On the right side of (9), the first term is a model fit to observed data, and the (negative) second term is a model fit to synthetic data drawn from $p _ { \\pmb { \\theta } ( \\mathbf { x } ) }$ ; this is similar to the critic/discriminator in GANs (Arjovsky et al., 2017), but derived directly from the MLE. More connections are discussed below. ", + "bbox": [ + 174, + 460, + 825, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1 LEARNING VIA AMORTIZATION ", + "text_level": 1, + "bbox": [ + 176, + 526, + 433, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our goal is to learn a generator whose generated samples match those from the original model $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ , by adopting the amortization idea with CTF in the inference section above. Similar to inference, the generator is learned implicitly. However, we also learn an explicit density model for the data by SGD, with samples from the implicit generator to estimate gradients in (9). Note that in this case, the CTF is performed directly on the data space, instead of on latentvariable space as in previous sections. Specifically, the sampling procedure from the generator plus a continuoustime-flow transformation are written as: ", + "bbox": [ + 174, + 554, + 553, + 707 + ], + "page_idx": 7 + }, + { + "type": "equation", + "img_path": "images/5b1559c7588790150c2a8adf05d421683bbb3b9ae7a96bf207af30fbda846a55.jpg", + "text": "$$\n\\mathbf { x } _ { 0 } \\sim q _ { \\phi } ( \\mathbf { x } _ { 0 } ) , \\mathbf { x } _ { T } \\sim { \\mathcal { T } } ( \\mathbf { x } _ { 0 } , T ) .\n$$", + "text_format": "latex", + "bbox": [ + 258, + 718, + 467, + 737 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b4afaf4a45229a804fe66cbebef9ba63504a9ce81f293051a8bbd6f04a06cfb.jpg", + "image_caption": [ + "Figure 3: Learning a generator with CTF. The goal is to match the samples $\\mathbf { x } _ { \\mathrm { 0 } }$ from $q _ { \\phi }$ to those after a CTF $\\left( { { \\bf { x } } _ { T } } \\right)$ , or equivalently samples from $p _ { \\pmb { \\theta } }$ . " + ], + "image_footnote": [], + "bbox": [ + 565, + 537, + 803, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here $\\tau ( \\cdot , \\cdot )$ is the continuous-time flow; a sample $\\mathbf { x } _ { \\mathrm { 0 } }$ from $q _ { \\phi } ( \\cdot )$ is implemented by a deep neural network (generator) $G _ { \\phi } ( \\omega )$ with input $\\omega \\sim q _ { 0 } ( \\omega )$ , where $q _ { 0 }$ is a simple distribution for a noise random variable, e.g., the standard isotropic normal distribution. The procedure is illustrated in Figure 3. Note the CTF cannot be replaced by standard normalizing flow (Rezende & Mohamed, 2015) in this model, because there is no objective function to guide the update of parameters in normalizing flows, which is not necessary for CTFs. ", + "bbox": [ + 173, + 747, + 825, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, denote the parameters in the $k$ -th step of our algorithm with subscript “ $( k ) ^ { \\prime }$ . For efficient sample generation, in the $k$ -th step, we again adopt the amortization idea from Section 3.2 to update $\\phi ^ { ( k ^ { - 1 ) } }$ of the generator network $G _ { \\phi } ( \\cdot )$ , such that samples from the updated generator match those from the current generator followed by a one-step transformation $\\mathcal { T } _ { 1 } ( \\cdot )$ . After that, $\\pmb \\theta$ is updated by drawing samples from $q _ { \\phi } ( \\cdot )$ to estimate the expectation in (9). The detailed algorithm is presented in Algorithm 1. ", + "bbox": [ + 173, + 838, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2 CONNECTIONS TO WASSERSTEIN GAN (WGAN) AND MLE ", + "text_level": 1, + "bbox": [ + 174, + 103, + 633, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "There is an interesting relation between our model and the WGAN framework (Arjovsky et al., 2017). To see this, let $p _ { r }$ be the data distribution. Substituting $p _ { \\pmb { \\theta } } ( \\mathbf { x } )$ with $q _ { \\phi } ( \\mathbf { x } )$ for the expectation in the gradient formula (9) and integrating out $\\pmb { \\theta }$ , we have that our objective is ", + "bbox": [ + 174, + 127, + 823, + 167 + ], + "page_idx": 8 + }, + { + "type": "equation", + "img_path": "images/772210da739db4e71d00c69461f108c6e0aa5a93273571509859741e30dde6f7.jpg", + "text": "$$\n\\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\pmb { \\theta } ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ U ( \\mathbf { x } ; \\pmb { \\theta } ) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 367, + 169, + 620, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The objective is an instance of the general integral probability metrics (Arjovsky & Bottou, 2017). When $U$ is chosen to be 1-Lipschitz functions, it recovers WGAN. This connection motivates us to introduce weight clipping (Arjovsky et al., 2017) or alternative regularizers (Gulrajani et al., 2017) when updating $\\pmb \\theta$ for a better theoretical property. For this reason, we call our model Markov-chainbased GAN (MacGAN). ", + "bbox": [ + 174, + 207, + 826, + 277 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Furthermore, it can be shown by Jensen’s inequality that the MLE is bounded by (detailed derivations are provided in Section C of the SM) ", + "bbox": [ + 173, + 284, + 823, + 313 + ], + "page_idx": 8 + }, + { + "type": "equation", + "img_path": "images/df446ddb77704865499a2dd727e53edad8f416e99834eb05e898237bb3a0a4c7.jpg", + "text": "$$\n\\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ \\log q _ { \\phi } \\right] .\n$$", + "text_format": "latex", + "bbox": [ + 199, + 314, + 769, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "By inspecting (10) and (11), it is clear that: $i$ ) when learning the energy-based model parameters $\\pmb \\theta$ , the objective can be interpreted as maximizing an upper bound of the MLE shown in (11); $i i )$ ) when optimizing the parameter $\\phi$ of the inference network, we adopt the amortized learning procedure presented in Algorithm 1, whose objective is $\\operatorname* { m i n } _ { \\phi } \\operatorname { K L } \\left( q _ { \\phi } \\| p _ { \\theta } \\right)$ , coinciding with the last two terms in (11). In other words, both $\\pmb \\theta$ and $\\phi$ are optimized by maximizing the same upper bound of the MLE, guaranteeing convergence of the algorithm. Particularly, we can conclude that ", + "bbox": [ + 174, + 357, + 825, + 440 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Proposition 4. The optimal solution of MacGAN is the maximum likelihood estimator. ", + "bbox": [ + 178, + 443, + 740, + 457 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Note another difference between MacGAN and standard GAN framework is the way of learning the generator $q _ { \\phi }$ . We adopt the amortization idea, which directly guides $q _ { \\phi }$ to approach $p _ { \\pmb { \\theta } }$ ; whereas in GAN, the generator is optimized via a min-max procedure to make it approach the empirical data distribution $p _ { r }$ . By explicitly learning $p _ { \\pmb { \\theta } }$ , MacGAN is able to evaluate likelihood for test data (at least up to a constant). ", + "bbox": [ + 174, + 467, + 825, + 536 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "5 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 555, + 344, + 571 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Our framework extends the idea of normalizing flows (Rezende & Mohamed, 2015) to continuoustime flows, by developing theoretical properties on the convergence behavior. Inference based on CTFs has been studied in Salimans et al. (2015) based on the auxiliary-variable technique. However, Salimans et al. (2015) directly uses discrete approximations for the flow, and the approximation accuracy is unclear. Moreover, the inference network requires simulating a long Markov chain for the auxiliary model, thus is less efficient than ours. Finally, the inference network is implemented as a parametric distribution (e.g., the Gaussian distribution), limiting the representation power, a common setting in existing auxiliary-variable based models (Tran et al., 2016). The idea of amortization (Gershman & Goodman, 2014) has recently been explored in various research topics for Bayesian inference such as in variational inference (Kingma & Welling, 2014; Rezende et al., 2014) and Markov chain Monte Carlo (Wang & Liu, 2017; Li et al., 2017; Pu et al., 2017a). Both Wang & Liu (2017) and $\\mathrm { P u }$ et al. (2017a) extend the idea of Stein variational gradient descent (Liu & Wang, 2016) with amortized inference for a GAN-based and a VAE-based model, respectively, which resemble our proposed MacVAE and MacGAN in concept. Li et al. (2017) applies amortization to distill knowledge from MCMC to learn a student network. The ideas in Li et al. (2017) are similar to ours, but the motivation and underlying theory are different from that developed here. ", + "bbox": [ + 174, + 585, + 825, + 808 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "6 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 174, + 828, + 326, + 843 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We conduct experiments to test our CTF-based framework for efficient inference and density estimation described above, and compared them with related methods. The implementation is based on the excellent code for SteinGANk Wang & Liu (2017), where we adopt their default parameter setting. ", + "bbox": [ + 174, + 858, + 825, + 901 + ], + "page_idx": 8 + }, + { + "type": "image", + "img_path": "images/c77f66c6e7bcd94aa794bbca2d0d82ccb351a3e4c8f0d101e39392c22a5424b2.jpg", + "image_caption": [ + "Figure 4: Knowledge distillation from the CTF (left) and ELBO versus epochs on MNIST (right). VAE with 80-layer NF is not included because it has much more parameters. " + ], + "image_footnote": [], + "bbox": [ + 200, + 88, + 795, + 272 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "The discretization stepsize $h$ is robust as long as it is set in a reasonable range, e.g., we set it the same as the stepsize in SGD. ", + "bbox": [ + 176, + 338, + 823, + 366 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.1 CTFS FOR INFERENCE ", + "text_level": 1, + "bbox": [ + 176, + 383, + 370, + 398 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Synthetic experiment We examine our amortized learning framework with a toy experiment. Following Rezende & Mohamed (2015), we use MacVAE to approximate samples from a two dimensional distribution on $\\mathbf { z } = \\{ \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } \\}$ : $p ( \\mathbf { z } ) \\propto e ^ { - U ( \\mathbf { z } ) }$ with $\\begin{array} { r } { U ( \\mathbf { z } ) \\triangleq \\frac { 1 } { 2 } ( \\frac { \\Vert \\mathbf { z } \\Vert - 2 } { 0 . 4 } ) ^ { 2 } - \\ln ( e ^ { - \\frac { 1 } { 2 } [ \\frac { \\mathbf { z } _ { 1 } - 2 } { 0 . 6 } ] ^ { 2 } } + } \\end{array}$ $e ^ { - \\frac { 1 } { 2 } [ \\frac { { \\bf z } _ { 1 } + 2 } { 0 . 6 } ] ^ { 2 } } )$ . The inference network $q _ { \\phi }$ is defined to be a 2-layer MLP with isotropic normal random variables as input. Figure 4 (top) plots the densities estimated with the samples from transformations $\\{ \\mathcal { T } _ { K = 1 0 0 } \\}$ (before optimizing $\\phi$ ), as well as with samples generated directly from $q _ { \\phi }$ (after optimizing $\\phi )$ ). It is clear that the amortized learning is able to distill knowledge from the CTF to the inference network. ", + "bbox": [ + 173, + 410, + 825, + 530 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "MacVAE on MNIST Following Rezende & Mohamed (2015); Tomczak & Welling (2016), we define the inference network as a deep neural network with two fully connected layers of size 300 with softplus activation functions. We compare MacVAE with the standard VAE and the VAE with normalizing flow, where testing ELBOs are reported (Section D.1 of the SM describes how to calculate the ELBO). We do not compare with other state-of-the-art methods such as the inverse autoregressive flow (Kingma et al., 2016), because they typically endowed more complicated inference networks (with more parameters), unfair for comparison. We use the same inference network architecture for all the models. Figure 4 (bottom) plots the testing ELBO versus training epochs. MacVAE outperforms VAE and normalizing flows with a better ELBO (around -85.62). ", + "bbox": [ + 173, + 545, + 825, + 671 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "6.2 CTFS FOR DENSITY ESTIMATION ", + "text_level": 1, + "bbox": [ + 176, + 689, + 442, + 702 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We test MacGAN on three datasets: MNIST, CIFAR-10 and CelabA. Following GAN-related methods, the model is evaluated by observing its ability to draw samples from the learned data distribution. Inspiring by Wang $\\&$ Liu (2017), we define a parametric form of the energy-based model as $p _ { \\theta } ( \\mathbf { x } ) \\propto \\mathrm { e x p } \\{ - \\| \\mathbf { x } - \\mathrm { D E C } _ { \\theta } \\left( \\mathrm { E N C } _ { \\theta } ( \\mathbf { x } ) \\right) \\| ^ { 2 } \\}$ , where $\\operatorname { E N C } _ { \\theta } ( \\cdot )$ and $\\operatorname { D E C } _ { \\pmb { \\theta } } ( \\cdot )$ are encoder and decoder defined by using deep convolutional neural networks and deconvolutional neural networks, respectively, parameterized by $\\pmb \\theta$ . For simplicity, we adopt the popular DCGAN architecture (Radford et al., 2016) for the encoder and decoder. The generator $G _ { \\phi }$ is defined as a 3-layer convolutional neural network with the ReLU activation function (except for the top layer which uses tanh as the activation function, see SM D for details). Following Wang & Liu (2017), the stepsizes are set to (me−e)×lrm 50 , where e indexes the epoch, me is the total number of epochs, lr = 1e-4 when updating θ, and $l _ { r } = 1 { \\mathrm e } { - } 3$ when updating $\\phi$ . The stepsize in $\\mathcal { L } _ { 1 }$ is set to 1e-3. ", + "bbox": [ + 174, + 713, + 825, + 875 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "We compare MacGAN with DCGAN (Radford et al., 2016), the improved WGAN (WGAN-I) (Gulrajani et al., 2017) and SteinGAN (Wang & Liu, 2017). We plot images generated with MacGAN and its most related method SteinGAN in Figure 5 for CelebA and CIFAR-10 datasets. More results are provided in SM Section D. We observe that visually MacGAN is able to generate clear-looking images. Following Wang & Liu (2017), we also plot the images generated by a random walk in the $\\omega$ space in Figure 5. ", + "bbox": [ + 174, + 881, + 825, + 924 + ], + "page_idx": 9 + }, + { + "type": "image", + "img_path": "images/f32eae48928503928e1e56b310929d8a7099bc5f96ebf9dbd78c5069876c85de.jpg", + "image_caption": [ + "Figure 5: Generated images for CIFAR-10 (top) and CelebA (middle) datasets with MacGAN (left) and SteinGAN (right). The bottom are images generated by a random walk on the $\\omega$ space for the generator of MacGAN, i.e., $\\omega _ { t } = \\omega _ { t - 1 } + 0 . 0 3 \\times \\mathrm { r a n d } ( [ - 1 , 1 ] )$ . " + ], + "image_footnote": [], + "bbox": [ + 173, + 99, + 821, + 619 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 734, + 825, + 777 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Qualitatively evaluating a GAN-like model is challenging. We follow literature and use the inception score (Salimans et al., 2016) to measure the quantity of the generated images. Figure 6 plots inception scores versus training epochs for different models. MacGAN obtains competitive inception scores with the popular DCGAN model. Quantitatively, we get a final inception score of 6.49 for MacGAN, compared to 6.35 for SteinGAN, 6.25 for WGAN-I and 6.58 for DCGAN. ", + "bbox": [ + 174, + 785, + 508, + 922 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/886856f7b0b2b16fcc23aab4c1098b0a0407b016e3f738e952ddc0295d2a0f7c.jpg", + "image_caption": [ + "Figure 6: Inception score versus epochs for different models. " + ], + "image_footnote": [], + "bbox": [ + 557, + 772, + 767, + 890 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "7 CONCLUSION ", + "text_level": 1, + "bbox": [ + 176, + 102, + 318, + 117 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We study the problem of applying CTFs for efficient inference and explicit density estimation in deep generative models, two important tasks in unsupervised machine learning. Compared to discrete-time normalized flows, CTFs are more general and flexible due to the fact that their stationary distributions can be controlled without extra flow parameters. We develop theory on the approximation accuracy when adopting a CTF to approximate a target distribution. We apply CTFs on two classes of deep generative models, a variational autoencoder for efficient inference, and a GAN-like density estimator for explicit density estimation and efficient data generation. Experiments show encouraging results of our framework in both models compared to existing techniques. One interesting direction of future work is to explore more efficient learning algorithms for the proposed CTF-based framework. ", + "bbox": [ + 174, + 132, + 825, + 258 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 279, + 285, + 294 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "M. Arjovsky and L. Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. \nM. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN. Technical Report arXiv:1701.07875, March 2017. \nF. Bolley, I. Gentil, and A. Guillin. Convergence to equilibrium in wasserstein distance for fokker–planck equations. Journal of Functional Analysis, 263(8):2430–2457, 2012. \nL. Bottou. Stochastic gradient descent tricks. Technical report, Microsoft Research, Redmond, WA, 2012. \nC. Chen, N. Ding, and L. Carin. On the convergence of stochastic gradient MCMC algorithms with high-order integrators. In NIPS, 2015. \nLaurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. ICLR, 2017. \nA. A. Dorogovtsev and I. I. Nishchenko. An analysis of stochastic flows. Communications on Stochastic Analysis, 8(3):331–342, 2014. \nY. Feng, D. Wang, and Q. Liu. Learning to draw samples with amortized Stein variational gradient descent. In UAI, 2017. \nS. J. Gershman and N. D. Goodman. Amortized inference in probabilistic reasoning. In Annual Conference of the Cognitive Science Society, 2014. \nC. R. Givens and R. M. Shortt. A class of wasserstein metrics for probability distributions. Michigan Math. J., 31, 1984. \nI. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In NIPS, 2014. \nA. Grover, M. Dhar, and S. Ermon. Flow-GAN: Bridging implicit and prescribed learning in generative models. Technical Report arXiv:1705.08868, 2017. \nI. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of Wasserstein GAN. Technical Report arXiv:1704.00028, March 2017. \nT. Haarnoja, H. Tang, P. Abbeel, and S. Levine. Reinforcement learning with deep energy-based policies. In ICML, 2017. \nG. E. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation, 14(8):1771–1800, 2002. \nF. Huszár. Variational inference using implicit distributions. Technical Report arXiv:1702.08235, 2017. \nR. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker-Planck equation. SIAM J. MATH. ANAL., 29(1):1–17, 1998. \nD. Kingma, T. P. Salimans, and M. Welling. Improving variational inference with inverse autoregressive flow. In NIPS, 2016. \nD. P. Kingma and M. Welling. Auto-encoding variational Bayes. In ICLR, 2014. \nY. Li, R. E. Turner, and Q. Liu. Approximate inference with amortised MCMC. Technical Report arXiv:1702.08343, 2017. \nQ. Liu and D. Wang. Stein variational gradient descent: A general purpose Bayesian inference algorithm. In NIPS, 2016. \nJ. C. Mattingly, A. M. Stuart, and M. V. Tretyakov. Construction of numerical time-average and stationary measures via Poisson equations. SIAM Journal on Numerical Analysis, 48(2):552–577, 2010. \nS. Mohamed and B. Lakshminarayanan. Learning in implicit generative models. Technical Report arXiv:1610.03483, 2017. \nF. Otto. Dynamics of Labyrinthine pattern formation in magnetic fluids: A mean-field theory. Arch. Rational Mech. Anal., pp. 63–103, 1998. \nY. Pu, Z. Gan, R. Henao, C. Li, S. Han, and L. Carin. Stein variational autoencoder. Technical Report arXiv:1704.05155, 2017a. \nY. Pu, Z. Gan, R. Henao, C. Li, S. Han, and L. Carin. Vae learning via stein variational gradient descent. In NIPS, 2017b. \nA. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. Technical Report arXiv:1511.06434, January 2016. \nR. Ranganath, J. Altosaar, D. Tran, and D. M. Blei. Operator variational inference. In NIPS, 2016. \nD. J. Rezende and S. Mohamed. Variational inference with normalizing flows. In ICML, 2015. \nD. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. \nH. Risken. The Fokker-Planck equation. Springer-Verlag, New York, 1989. \nT. Salimans, D. P. Kingma, and M. Welling. Markov chain Monte Carlo and variational inference: Bridging the gap. In ICML, 2015. \nT. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for training GANs. Technical Report arXiv:1606.03498, June 2016. \nY. W. Teh, A. H. Thiery, and S. J. Vollmer. Consistency and fluctuations for stochastic gradient Langevin dynamics. JMLR, 17(1):193–225, 2016. \nJ. M. Tomczak and M. Welling. Improving variational auto-encoders using Householder flow. Technical Report arXiv:1611.09630, November 2016. \nD. Tran, R. Ranganath, and D. M. Blei. The variational gaussian process. In ICLR, 2016. \nA. van den Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks. In ICML, 2016. \nS. J. Vollmer, K. C. Zygalakis, and Y. W. Teh. (exploration of the (Non-)asymptotic bias and variance of stochastic gradient Langevin dynamics. JMLR, 1:1–48, 2016. \nD. Wang and Q. Liu. Learning to draw samples: With application to amortized MLE for generative adversarial learning. In ICLR workshop, 2017. \nM. Welling and Y. W. Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML, 2011. \nJ. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative adversarial networks. In ICLR, 2017. ", + "bbox": [ + 171, + 303, + 826, + 926 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 78, + 828, + 928 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "SUPPLEMENTARY MATERIAL FOR: CONTINUOUS-TIME FLOWS FOR EFFICIENT INFERENCE AND DENSITY ESTIMATION ", + "text_level": 1, + "bbox": [ + 238, + 98, + 759, + 171 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "A ASSUMPTIONS OF THEOREM 2 ", + "text_level": 1, + "bbox": [ + 176, + 232, + 464, + 248 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "First, let us define the infinitesimal generator of the diffusion (2). Formally, the generator $\\mathcal { L }$ of the diffusion (2) is defined for any compactly supported twice differentiable function $\\breve { f } : \\mathbf { R } ^ { L } \\mathbf { R }$ , such that, ", + "bbox": [ + 174, + 263, + 825, + 306 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/836de84f757b84ca0f70a7f98051d6709f159c13ad0e0d33891b3dcfc5b9e778.jpg", + "text": "$$\n\\mathcal { L } f ( \\mathbf { Z } _ { t } ) \\triangleq \\operatorname* { l i m } _ { h 0 ^ { + } } \\frac { \\mathbb { E } [ f ( \\mathbf { Z } _ { t + h } ) ] - f ( \\mathbf { Z } _ { t } ) } { h } = ( F ( \\mathbf { Z } _ { t } ) \\cdot \\nabla + \\frac { 1 } { 2 } ( G ( \\mathbf { Z } _ { t } ) G ( \\mathbf { Z } _ { t } ) ^ { T } ) \\colon \\nabla \\nabla ^ { T } ) f ( \\mathbf { Z } _ { t } ) ,\n$$", + "text_format": "latex", + "bbox": [ + 191, + 310, + 803, + 345 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "where ${ \\mathbf { a } } \\cdot { \\mathbf { b } } \\triangleq { \\mathbf { a } } ^ { T }$ b, $\\mathbf { A } : \\mathbf { B } \\triangleq { \\mathrm { t r } } ( \\mathbf { A } ^ { T } \\mathbf { B } )$ , $h \\to 0 ^ { + }$ means $h$ approaches zero along the positive real axis. ", + "bbox": [ + 169, + 353, + 823, + 383 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Given an ergodic diffusion (2) with an invariant measure $\\rho ( \\mathbf { Z } )$ , the posterior average is defined as: $\\begin{array} { r } { \\bar { \\psi } \\triangleq \\int \\psi ( \\mathbf { Z } ) \\rho ( \\mathbf { Z } ) \\mathrm { d } \\mathbf { Z } } \\end{array}$ for some test function $\\psi ( \\mathbf { Z } )$ of interest. For a given numerical method with generated samples $( \\mathbf { z } _ { k } ) _ { k = 1 } ^ { K }$ , we use the sample average $\\hat { \\psi }$ defined as $\\begin{array} { r } { \\hat { \\psi } _ { K } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) } \\end{array}$ to approximate $\\bar { \\psi }$ . We define a functional $\\tilde { \\psi }$ that solves the following Poisson Equation: ", + "bbox": [ + 173, + 388, + 825, + 455 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/27f14d4df4498051faa9ea9cfbe33c59a6117e2a35022a9cfac5565c4f78cc31.jpg", + "text": "$$\n\\mathcal { L } \\tilde { \\psi } ( \\mathbf { z } _ { k } ) = \\psi ( \\mathbf { z } _ { k } ) - \\bar { \\psi }\n$$", + "text_format": "latex", + "bbox": [ + 424, + 460, + 573, + 481 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We make the following assumptions on $\\tilde { \\psi }$ ", + "bbox": [ + 176, + 496, + 450, + 512 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Assumption 1. $\\tilde { \\psi }$ exists, and its up to 4rd-order derivatives, $\\mathcal { D } ^ { k } \\tilde { \\psi }$ , are bounded by a function $\\nu$ , i.e., $\\lVert \\mathcal { D } ^ { k } \\tilde { \\psi } \\rVert \\leq C _ { k } \\mathcal { V } ^ { p _ { k } }$ for $k = ( 0 , 1 , 2 , 3 , 4 )$ , $C _ { k } , p _ { k } > 0$ . Furthermore, the expectation of $\\nu$ on $\\left\\{ { \\bf z } _ { k } \\right\\}$ is bounded: $\\mathrm { s u p } _ { l } \\mathbb { E } \\mathcal { V } ^ { p } ( { \\mathbf { z } } _ { k } ) < \\infty$ , and $\\nu$ is smooth such that $\\begin{array} { r } { \\operatorname* { s u p } _ { s \\in ( 0 , 1 ) } \\mathcal { V } ^ { p } \\left( s \\mathbf { z } + \\left( 1 - s \\right) \\mathbf { y } \\right) \\leq } \\end{array}$ $C \\left( \\mathcal { V } ^ { p } \\left( \\mathbf { z } \\right) + \\mathcal { V } ^ { p } \\left( \\mathbf { y } \\right) \\right)$ , $\\forall \\mathbf { z } , \\mathbf { y } , p \\leq \\operatorname* { m a x } \\{ 2 p _ { k } \\}$ for some $C > 0$ . ", + "bbox": [ + 173, + 516, + 826, + 579 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "B PROOFS FOR SECTION 3 ", + "text_level": 1, + "bbox": [ + 176, + 597, + 408, + 614 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Sketch Proof of Lemma $^ { l }$ . First note that (5) in Lemma 1 corresponds to eq.13 in Jordan et al. (1998), where $F ( p )$ in Jordan et al. (1998) is in the form of $\\mathrm { K L } ( \\rho \\| p _ { \\pm } ( \\bar { \\bf x } , { \\bf z } ) )$ in our setting. ", + "bbox": [ + 171, + 628, + 825, + 659 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Proposition 4.1 in Jordan et al. (1998) then proves that (5) has a unique solution. Theorem 5.1 in Jordan et al. (1998) then guarantees that the solution of (5) approach the solution of the Fokker-Planck equation in (3), which is $\\rho _ { T }$ in the limit of $h 0$ . ", + "bbox": [ + 174, + 664, + 823, + 707 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Since this is true for each $k$ (thus each $t$ in $\\rho _ { t }$ ), we conclude that $\\tilde { \\rho } _ { k } = \\rho _ { h k }$ in the limit of $h 0$ . ", + "bbox": [ + 171, + 713, + 795, + 729 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "To prove Theorem 2, we first need a convergence result about convergence to equilibrium in Wasserstein distance for Fokker-Planck equations, which is presented in Bolley et al. (2012). Putting in our setting, we can get the following lemma based on Corollary 2.4 in Bolley et al. (2012). ", + "bbox": [ + 174, + 743, + 826, + 786 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "Lemma 5 (Bolley et al. (2012)). Let $\\rho _ { T }$ be the solution of the $F P$ equation (3) at time $T$ , $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ be the joint posterior distribution given x. Assume that $\\begin{array} { r } { \\int \\rho _ { T } ( \\mathbf { z } ) p _ { \\pmb \\theta } ^ { - 1 } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } < \\infty } \\end{array}$ and there exists $a$ constant $C$ such tha t dW 22 (ρT ,pθ(x,z)) ≥ CW 22 (ρT , pθ(x, z)). Then ", + "bbox": [ + 173, + 790, + 825, + 842 + ], + "page_idx": 13 + }, + { + "type": "equation", + "img_path": "images/9940691449e4abfa1d52b38b2247dbdb1018ebd321720f5c3284fedf7997a230.jpg", + "text": "$$\nW _ { 2 } \\left( \\rho _ { T } , p ( { \\bf x } , { \\bf z } ) \\right) \\leq W _ { 2 } \\left( \\rho _ { 0 } , p ( { \\bf x } , { \\bf z } ) \\right) e ^ { - C T } .\n$$", + "text_format": "latex", + "bbox": [ + 352, + 847, + 645, + 866 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "We further need to borrow convergence results from Mattingly et al. (2010); Vollmer et al. (2016); Chen et al. (2015) to characterize error bounds of a numerical integrator for the diffusion (2). Specifically, the goal is to evaluate the posterior average of a test function $\\psi ( \\mathbf { z } )$ , defined as $\\bar { \\psi }$ , $\\begin{array} { r } { \\int \\psi ( \\mathbf { z } ) p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } } \\end{array}$ . When using a numerical integrator to solve (2) to get samples $\\{ \\mathbf { z } _ { k } \\} _ { k = 1 } ^ { K }$ , the sample average $\\begin{array} { r } { \\hat { \\psi } _ { K } \\triangleq \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) } \\end{array}$ k=1 is used to approximate the posterior average. The accuracy is characterized by the mean square error (MSE) defined as: $\\mathbb { E } \\left( \\hat { \\psi } _ { K } - \\bar { \\psi } \\right) ^ { 2 }$ . Lemma 6 derives the bound for the MSE. ", + "bbox": [ + 174, + 880, + 826, + 924 + ], + "page_idx": 13 + }, + { + "type": "text", + "text": "", + "bbox": [ + 173, + 102, + 826, + 176 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Lemma 6 (Vollmer et al. (2016)). Under Assumption $^ { l }$ , and for a 1st-order numerical intergrator, the MSE is bounded, for a constant $C$ independent of $h$ and $K$ , by ", + "bbox": [ + 171, + 179, + 825, + 208 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/f7c58fbb325ea18ca21fccaa787d7af868550507f1deb15800a82eedd4b7babf.jpg", + "text": "$$\n\\mathbb { E } \\left( \\hat { \\psi } _ { K } - \\bar { \\psi } \\right) ^ { 2 } \\leq C \\left( \\frac { 1 } { h K } + h ^ { 2 } \\right) ~ .\n$$", + "text_format": "latex", + "bbox": [ + 380, + 212, + 617, + 247 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Furthermore, except for the 2nd-order Wasserstein distance defined in Lemma 1, we define the 1st-order Wasserstein distance between two probability measures $\\mu _ { 1 }$ and $\\mu _ { 2 }$ as ", + "bbox": [ + 173, + 257, + 823, + 286 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/9524acac499dae0f55bc7ac6ae4eab25ac34cd1f411af9a2f8f32cb20096a344.jpg", + "text": "$$\nW _ { 1 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) \\triangleq \\operatorname* { i n f } _ { p \\in { \\mathscr P } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) } \\int \\| { \\mathbf x } - { \\mathbf y } \\| _ { 2 } p ( \\mathrm { d } { \\mathbf x } , \\mathrm { d } { \\mathbf y } ) .\n$$", + "text_format": "latex", + "bbox": [ + 328, + 290, + 668, + 324 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "According to the Kantorovich-Rubinstein duality Arjovsky et al. (2017), $W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )$ is equivalently represented as ", + "bbox": [ + 174, + 335, + 825, + 364 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/393d61b8fbebc773842fe49e897304ed634efcc6132e9e0add233df1a844c541.jpg", + "text": "$$\nW _ { 1 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) = \\underset { f \\in \\mathcal { L } _ { 1 } } { \\operatorname* { s u p } } \\ \\mathbb { E } _ { \\mathbf { z } \\sim \\mu _ { 1 } } \\left[ f ( \\mathbf { z } ) \\right] - \\mathbb { E } _ { \\mathbf { z } \\sim \\mu _ { 2 } } \\left[ f ( \\mathbf { z } ) \\right] \\ ,\n$$", + "text_format": "latex", + "bbox": [ + 326, + 368, + 668, + 395 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "where $\\mathcal { L } _ { 1 }$ is the space of 1-Lipschitz functions $f : \\mathbb { R } ^ { L } \\mathbb { R }$ . ", + "bbox": [ + 173, + 401, + 568, + 417 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "We have the following relation between $W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )$ and $W _ { 2 } ( \\mu _ { 1 } , \\mu _ { 2 } )$ . ", + "bbox": [ + 173, + 422, + 627, + 439 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Lemma 7 (Givens & Shortt (1984)). We have for any two distributions $\\mu _ { 1 }$ and $\\mu _ { 2 }$ that $W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } ) \\leq$ $W _ { 2 } ( \\mu _ { 1 } , \\mu _ { 2 } )$ . ", + "bbox": [ + 178, + 441, + 825, + 470 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Now it is ready to prove Theorem 2. ", + "bbox": [ + 174, + 479, + 413, + 496 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Proof of Theorem 2. The idea is to simply decompose the MSE into two parts, with one part charactering the MSE of the numerical method, the other part charactering the MSE of $\\rho _ { T }$ and $p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )$ , which consequentially can be bounded using Lemma 5 above. ", + "bbox": [ + 173, + 510, + 826, + 553 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "Specifically, we have ", + "bbox": [ + 173, + 559, + 313, + 574 + ], + "page_idx": 14 + }, + { + "type": "equation", + "img_path": "images/6b824cf02479b7d23d9a9bada5bb8519b6db673f0f18b58997a6977dc083a080.jpg", + "text": "$$\n\\begin{array} { r l } & { \\quad \\mathrm { M S E ~ } \\left( \\left( \\frac { 1 } { N } , \\frac { 1 } { N } , \\rho _ { t } , \\eta _ { t } , \\phi , z ( \\phi ) \\right) + \\mathbb { E } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } } \\\\ & { = \\mathbb { E } \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } \\exp \\left( \\int \\phi ( \\mathrm { d } \\rho z ) \\mathrm { d } z \\mathrm { d } z \\right) \\right) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } - \\int \\psi ( \\mathrm { d } \\rho \\eta ) \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) ^ { - 1 } \\int \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } - \\int \\psi ( \\mathrm { d } \\rho z ) \\mathrm { d } \\rho z \\mathrm { d } z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { 2 } \\right) ^ { - 1 } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { - 1 } \\int \\psi \\mathrm { d } \\eta \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\mathrm { d } \\mathrm { \\rho } \\mathrm { \\rho } \\Bigg ) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\exp \\left( - \\int \\psi ( \\mathrm { d } \\rho z ) \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } + \\mathbb { E } \\left( \\int \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) } \\\\ & \\quad \\times \\mathbb { E } \\left( \\frac { 1 } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 183, + 580, + 813, + 925 + ], + "page_idx": 14 + }, + { + "type": "text", + "text": "where “(1)” follows by the fact that $\\begin{array} { r } { \\mathbb { E } \\left( \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) - \\int \\psi ( \\mathbf { z } ) p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } \\right) = 0 } \\end{array}$ Chen et al. (2015); “(2)” follows by the definition of $W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )$ in (14) and the 1-Lipschitz assumption of the test function $\\psi$ ; “(3)” follows by Lemma 7; “(4)” follows by Lemma 5 and Lemma 6. □ ", + "bbox": [ + 173, + 101, + 826, + 154 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "C CONNECTION TO WGAN ", + "text_level": 1, + "bbox": [ + 176, + 172, + 419, + 190 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We derive the upper bound of the maximum likelihood estimator, which connects MacGAN to WGAN. Let $p _ { r }$ be the data distribution, rewrite our maximum likelihood objective as ", + "bbox": [ + 171, + 204, + 825, + 234 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/522d55a249fc4e1b448ac7a67251345bcdcfe6aeb4209828b88e1b4ac68959ed.jpg", + "text": "$$\n\\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) = \\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\bigg ( U ( \\mathbf { x } _ { i } ; \\theta ) - \\log \\int e ^ { U ( \\mathbf { x } ; \\theta ) } \\mathrm { d } \\mathbf { x } \\bigg ) \\enspace .\n$$", + "text_format": "latex", + "bbox": [ + 259, + 241, + 740, + 285 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The above maximum likelihood estimator can be bounded with Jensen’s inequality as: ", + "bbox": [ + 173, + 290, + 738, + 306 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/ca21dfad937e6480c739cfc6b3141731f22c027f012b367a238d3ec07c8925a1.jpg", + "text": "$$\n\\begin{array} { r l } & { \\displaystyle \\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\log \\int \\frac { e ^ { U ( \\mathbf { x } ; \\theta ) } } { q _ { \\phi } ( \\mathbf { x } ; \\omega ) } q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\mathrm { d } \\mathbf { x } } \\\\ & { \\displaystyle \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log \\frac { e ^ { U ( \\mathbf { x } ; \\theta ) } } { q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\right] } \\\\ & { \\displaystyle = \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 233, + 313, + 766, + 415 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "This results in the same objective form as WGAN except that our model does not restrict $U ( \\mathbf { x } ; \\pmb { \\theta } )$ to be 1-Lipschitz functions and the objective has an extra constant term $\\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\right]$ w.r.t. $\\pmb \\theta$ . ", + "bbox": [ + 174, + 426, + 826, + 469 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Now we prove Proposition 4. ", + "bbox": [ + 174, + 477, + 366, + 491 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Proof of Proposition 4. First it is clear that the equality in (16) is achieved if and only if ", + "bbox": [ + 171, + 507, + 751, + 523 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/a715f4a4f805e84f353ad665f3d014b1be7d8e23fcbccbde1a52bac791302396.jpg", + "text": "$$\nq _ { \\phi } ( \\mathbf { x } ; \\omega ) = p _ { \\theta } ( \\mathbf { x } ) \\propto e ^ { U ( \\mathbf { x } ; \\theta ) } .\n$$", + "text_format": "latex", + "bbox": [ + 398, + 530, + 598, + 550 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "From the description in Section 4 and (16), we know that $\\pmb { \\theta }$ and $\\phi$ share the same objective function, which is an upper bound of the MLE in (16). ", + "bbox": [ + 173, + 564, + 825, + 593 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "Furthermore, based on the property of continuous-time flows (or formally Theorem 2), we know that $q _ { \\phi }$ is learned such that $q _ { \\phi } p _ { \\theta }$ in the limit of $h 0$ (or alternatively, we could achieve this by using a decreasing-step-size sequence in a numerical method, as proved in Chen et al. (2015)). When $q _ { \\phi } = p _ { \\theta }$ , the equality in (16) is achieved, leading to the MLE. □ ", + "bbox": [ + 173, + 599, + 826, + 656 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D ADDITIONAL EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 676, + 444, + 693 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D.1 CALCULATING THE TESTING ELBO FOR MACVAE ", + "text_level": 1, + "bbox": [ + 173, + 707, + 571, + 723 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "We follow the method in $\\mathrm { P u }$ et al. (2017a) for calculating the ELBO for a test data $\\mathbf { x } _ { * }$ . First, after distilling the CTF into the inference network $q _ { \\phi }$ , we have that the ELBO can be represented as ", + "bbox": [ + 173, + 733, + 825, + 763 + ], + "page_idx": 15 + }, + { + "type": "equation", + "img_path": "images/329c76880e24ab3a6b461a553d09af3719f11c498d3ceb4059e4246856c1648a.jpg", + "text": "$$\n\\begin{array} { r } { \\log p ( \\mathbf { x } _ { * } ) \\geq \\mathbb { E } _ { q _ { \\phi } } \\left[ \\log p _ { \\theta } ( \\mathbf { x } _ { * } , \\mathbf { z } _ { * } ) \\right] - \\mathbb { E } _ { q _ { \\phi } } \\left[ \\log q _ { \\phi } \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 334, + 770, + 661, + 789 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The expectation is approximated with samples $\\{ \\mathbf { z } _ { * j } \\} _ { j = 1 } ^ { M }$ with $\\mathbf { z } _ { \\ast j } = f _ { \\phi } ( \\mathbf { x } _ { \\ast } , \\zeta _ { j } )$ , and $\\zeta _ { j } \\sim q _ { 0 } ( \\zeta )$ the standard isotropic normal. Here $f _ { \\phi }$ represents the deep neural network in the inference network. Note $q _ { \\phi } ( \\mathbf { z } _ { * } )$ is not readily obtained. To evaluate it, we use the density transformation formula: $\\begin{array} { r } { q _ { \\phi } \\mathopen { } \\mathclose \\bgroup \\left( \\mathbf { z } _ { * } \\aftergroup \\egroup \\right) \\stackrel { } { = } q _ { 0 } \\mathopen { } \\mathclose \\bgroup \\left( \\zeta \\aftergroup \\egroup \\right) \\mathopen { } \\mathclose \\bgroup \\left| \\mathrm { d e t } \\frac { \\partial f _ { \\phi } \\mathopen { } \\mathclose \\bgroup \\left( \\mathbf { x } _ { * } , \\zeta \\aftergroup \\egroup \\right) } { \\partial \\zeta } \\aftergroup \\egroup \\right| ^ { - 1 } . } \\end{array}$ ", + "bbox": [ + 173, + 796, + 828, + 868 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "D.2 NETWORK ARCHITECTURE ", + "text_level": 1, + "bbox": [ + 174, + 882, + 405, + 897 + ], + "page_idx": 15 + }, + { + "type": "text", + "text": "The architecture of the generator of MacGAN is given in Table 1. ", + "bbox": [ + 173, + 909, + 602, + 924 + ], + "page_idx": 15 + }, + { + "type": "table", + "img_path": "images/0358d199f954b07c4ac42f553c3ad035695782e54e3e210720e39bd698999fc2.jpg", + "table_caption": [ + "Table 1: Architecture of generator in MacGAN " + ], + "table_footnote": [], + "table_body": "
Output SizeArchitecture
100×1100 ×10Linear,BN,ReLU
256×8×8512 × 4 × 4 deconv,256 5 × 5 kernels,ReLU, strike 2, BN
128 ×16 ×16256 × 8 × 8 deconv,128 5 × 5 kernels,ReLU, strike 2,BN
3× 32 × 32128 × 16 × 16 deconv,3 5 × 5 kernels,Tanh,strike 2
", + "bbox": [ + 235, + 127, + 763, + 212 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/baf33b33adb73feefa1f183846c66ce80219a0cb8d10bbd7a1726cb544553764.jpg", + "image_caption": [ + "Figure 7: Generated images for MNIST datasets with MacGAN (top) and SteinGAN (bottom). " + ], + "image_footnote": [], + "bbox": [ + 173, + 224, + 825, + 743 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D.3 ADDITIONAL RESULTS ", + "text_level": 1, + "bbox": [ + 174, + 792, + 375, + 808 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "Additional experimental results are given in Figure 7 – 12. ", + "bbox": [ + 174, + 820, + 555, + 835 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "D.4 ROBUSTNESS OF THE DISCRETIZATION STEPSIZE ", + "text_level": 1, + "bbox": [ + 176, + 854, + 550, + 868 + ], + "page_idx": 16 + }, + { + "type": "text", + "text": "To test the impact of the discretization stepsize $h$ in (6), following SteinGAN Feng et al. (2017), we test MacGAN on the MNIST dataset, where ee use a simple Gaussian-Bernoulli Restricted Boltzmann Machines as the energy-based model. We adopt the annealed importance sampling method to evaluate ", + "bbox": [ + 176, + 881, + 825, + 924 + ], + "page_idx": 16 + }, + { + "type": "image", + "img_path": "images/df70e18f19fb102d117a9f4a20fddc9375221a3993e0003772e47a29cb6ee7c3.jpg", + "image_caption": [ + "Figure 8: Generated images for CelebA datasets with MacGAN. " + ], + "image_footnote": [], + "bbox": [ + 174, + 101, + 823, + 604 + ], + "page_idx": 17 + }, + { + "type": "text", + "text": "log-likelihoods Feng et al. (2017). We vary $h$ in $\\{ 6 e - 4 , 2 . 4 e - 3 , 3 . 6 e - 3 , 6 e - 3 , 1 e - 2 , 1 . 5 e - 2 \\}$ . The trend of log-likelihoods is plotted in Figure 13. We can see that log-likelihoods do not change a lot within the chosen stepsize interval, demonstrating the robustness of $h$ . ", + "bbox": [ + 176, + 656, + 826, + 699 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/510b1ba5679995516cf70d0e8a807c5b979fddd698b18ac7c1a42b5722b649e0.jpg", + "image_caption": [ + "Figure 13: Log-likelihoods vs discretization stepsize for MacGAN on MNIST. " + ], + "image_footnote": [], + "bbox": [ + 271, + 718, + 727, + 895 + ], + "page_idx": 17 + }, + { + "type": "image", + "img_path": "images/2590ba66def78acb17923aea4b076627cdc952e917fb7aa8f72cc66ce1929962.jpg", + "image_caption": [ + "Figure 9: Generated images for CIFAR-10 datasets with MacGAN. " + ], + "image_footnote": [], + "bbox": [ + 174, + 104, + 823, + 356 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/20b4b0a19beab48e103ccf648183466dd297f9eee0a29d1029a2884965d2eb58.jpg", + "image_caption": [ + "Figure 10: Generated images for CelebA datasets with SteinGAN. " + ], + "image_footnote": [], + "bbox": [ + 174, + 391, + 825, + 895 + ], + "page_idx": 18 + }, + { + "type": "image", + "img_path": "images/dafbac1d644b9153e4d3341ddb10e46b6d37a0d08e2929035e46e43910c12d6f.jpg", + "image_caption": [ + "Figure 11: Generated images for CIFAR-10 datasets with SteinGAN. " + ], + "image_footnote": [], + "bbox": [ + 174, + 373, + 825, + 625 + ], + "page_idx": 19 + }, + { + "type": "image", + "img_path": "images/dfc2e36b0e28db379525f7f5e2f26eecca5dd5a6bc6132ffa3a69178104ec0d7.jpg", + "image_caption": [ + "Figure 12: Generated images with a random walk on the $\\omega$ space for CelebA datasets with MacGAN, $\\omega _ { t } \\overline { { = \\omega _ { t - 1 } + 0 . 0 2 \\times \\mathrm { r a n d } ( [ - 1 , 1 ] ) } }$ . " + ], + "image_footnote": [], + "bbox": [ + 238, + 95, + 759, + 902 + ], + "page_idx": 20 + } +] \ No newline at end of file diff --git a/parse/train/rkcya1ZAW/rkcya1ZAW_middle.json b/parse/train/rkcya1ZAW/rkcya1ZAW_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..062449d90086af84f3b08733d2b9ec4c90f08cb1 --- /dev/null +++ b/parse/train/rkcya1ZAW/rkcya1ZAW_middle.json @@ -0,0 +1,55168 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 508, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 508, + 97 + ], + "score": 1.0, + "content": "CONTINUOUS-TIME FLOWS FOR EFFICIENT INFER-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 356, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 356, + 118 + ], + "score": 1.0, + "content": "ENCE AND DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 213, + 469, + 389 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 470, + 226 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 470, + 226 + ], + "score": 1.0, + "content": "Two fundamental problems in unsupervised learning are efficient inference for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 224, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 470, + 236 + ], + "score": 1.0, + "content": "latent-variable models and robust density estimation based on large amounts of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "score": 1.0, + "content": "unlabeled data. For efficient inference, normalizing flows have been recently de-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 246, + 471, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 246, + 471, + 259 + ], + "score": 1.0, + "content": "veloped to approximate a target distribution arbitrarily well. In practice, however,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "score": 1.0, + "content": "normalizing flows only consist of a finite number of deterministic transformations,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "and thus they possess no guarantee on the approximation accuracy. For density", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 279, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "estimation, the generative adversarial network (GAN) has been advanced as an", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 471, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 471, + 302 + ], + "score": 1.0, + "content": "appealing model, due to its often excellent performance in generating samples.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 313 + ], + "score": 1.0, + "content": "In this paper, we propose the concept of continuous-time flows (CTFs), a family", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 471, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 471, + 325 + ], + "score": 1.0, + "content": "of diffusion-based methods that are able to asymptotically approach a target dis-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "tribution. Distinct from normalizing flows and GANs, CTFs can be adopted to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "score": 1.0, + "content": "achieve the above two goals in one framework, with theoretical guarantees. Our", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 345, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 142, + 345, + 469, + 356 + ], + "score": 1.0, + "content": "framework includes distilling knowledge from a CTF for efficient inference, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "score": 1.0, + "content": "learning an explicit energy-based distribution with CTFs for density estimation.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 367, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 379 + ], + "score": 1.0, + "content": "Experiments on various tasks demonstrate promising performance of the proposed", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 376, + 341, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 341, + 390 + ], + "score": 1.0, + "content": "CTF framework, compared to related techniques.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12.5 + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 206, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 208, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 208, + 430 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Efficient inference and robust density estimation are two important goals in unsupervised learning. In", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "fact, inference and density estimation can be unified from the perspective of learning desired target", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "distributions. In inference problems, one is given an unnormalized distribution (e.g., the posterior", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "distribution), and the goal is to learn a distribution that is close to the unnormalized distribution. In", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "density estimation, one tries to learn an unknown data distribution given samples from it. It is also", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "helpful to make a distinction between two types of representations for learning distributions: explicit", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "and implicit methods (Mohamed & Lakshminarayanan, 2017). Explicit methods provide a prescribed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "parametric form for the distribution, while implicit methods learn a stochastic procedure to directly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 529, + 303, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 303, + 540 + ], + "score": 1.0, + "content": "generate samples from the unknown distribution.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 557 + ], + "score": 1.0, + "content": "Existing deep generative models can easily be identified from this taxonomy. For example, the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "standard variational autoencoder (VAE) (Kingma & Welling, 2014; Rezende et al., 2014) is an impor-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "tant example of an explicit inference method. Within the inference arm (encoder) of a VAE, recent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 592 + ], + "score": 1.0, + "content": "research has focused on improving the accuracy of the approximation to the posterior distribution", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "on latent variables (codes) using normalizing flow (Rezende & Mohamed, 2015). Normalizing flow", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "is particularly interesting due to its ability to approximate the posterior distribution arbitrarily well,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "while maintaining explicit parametric forms. On the other hand, Stein VAE (Feng et al., 2017; Pu", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "et al., 2017b) is an implicit inference method, as it only learns to draw samples to approximate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "posteriors, without assuming an explicit form for the distribution.. For density estimation on observed", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "data, the generative adversarial network (GAN) can be regarded as an implicit density estimation", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "method (Ranganath et al., 2016; Huszár, 2017; Mohamed & Lakshminarayanan, 2017), in the sense", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "that one may sample from the distribution (regarded as a representation of the unknown distribution),", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "but an explicit form for the distribution is not estimated. GAN has recently been augmented by", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 507, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 507, + 701 + ], + "score": 1.0, + "content": "Flow-GAN (Grover et al., 2017) to incorporate a likelihood term for explicit density estimation.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Further, there also are some works trying to perform inference within the implicit density estimation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "framework, e.g., the real-valued non-volume preserving (real NVP) transformations algorithm (Dinh", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 497, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 497, + 734 + ], + "score": 1.0, + "content": "et al., 2017) was proposed as a tractable yet expressive approach to model high-dimensional data.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 39 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 308, + 761 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 506, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 508, + 97 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 508, + 97 + ], + "score": 1.0, + "content": "CONTINUOUS-TIME FLOWS FOR EFFICIENT INFER-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 98, + 356, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 98, + 356, + 118 + ], + "score": 1.0, + "content": "ENCE AND DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 135, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 135, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "spans": [ + { + "bbox": [ + 111, + 146, + 245, + 158 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 111, + 135, + 245, + 158 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 213, + 469, + 389 + ], + "lines": [ + { + "bbox": [ + 142, + 213, + 470, + 226 + ], + "spans": [ + { + "bbox": [ + 142, + 213, + 470, + 226 + ], + "score": 1.0, + "content": "Two fundamental problems in unsupervised learning are efficient inference for", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 224, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 141, + 224, + 470, + 236 + ], + "score": 1.0, + "content": "latent-variable models and robust density estimation based on large amounts of", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "spans": [ + { + "bbox": [ + 141, + 234, + 470, + 248 + ], + "score": 1.0, + "content": "unlabeled data. For efficient inference, normalizing flows have been recently de-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 142, + 246, + 471, + 259 + ], + "spans": [ + { + "bbox": [ + 142, + 246, + 471, + 259 + ], + "score": 1.0, + "content": "veloped to approximate a target distribution arbitrarily well. In practice, however,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "spans": [ + { + "bbox": [ + 141, + 257, + 471, + 270 + ], + "score": 1.0, + "content": "normalizing flows only consist of a finite number of deterministic transformations,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "spans": [ + { + "bbox": [ + 141, + 268, + 470, + 281 + ], + "score": 1.0, + "content": "and thus they possess no guarantee on the approximation accuracy. For density", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 142, + 279, + 469, + 291 + ], + "spans": [ + { + "bbox": [ + 142, + 279, + 469, + 291 + ], + "score": 1.0, + "content": "estimation, the generative adversarial network (GAN) has been advanced as an", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 141, + 290, + 471, + 302 + ], + "spans": [ + { + "bbox": [ + 141, + 290, + 471, + 302 + ], + "score": 1.0, + "content": "appealing model, due to its often excellent performance in generating samples.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 470, + 313 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 313 + ], + "score": 1.0, + "content": "In this paper, we propose the concept of continuous-time flows (CTFs), a family", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 471, + 325 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 471, + 325 + ], + "score": 1.0, + "content": "of diffusion-based methods that are able to asymptotically approach a target dis-", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "spans": [ + { + "bbox": [ + 141, + 323, + 470, + 335 + ], + "score": 1.0, + "content": "tribution. Distinct from normalizing flows and GANs, CTFs can be adopted to", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "spans": [ + { + "bbox": [ + 141, + 334, + 470, + 346 + ], + "score": 1.0, + "content": "achieve the above two goals in one framework, with theoretical guarantees. Our", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 142, + 345, + 469, + 356 + ], + "spans": [ + { + "bbox": [ + 142, + 345, + 469, + 356 + ], + "score": 1.0, + "content": "framework includes distilling knowledge from a CTF for efficient inference, and", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "spans": [ + { + "bbox": [ + 141, + 356, + 471, + 369 + ], + "score": 1.0, + "content": "learning an explicit energy-based distribution with CTFs for density estimation.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 141, + 367, + 470, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 470, + 379 + ], + "score": 1.0, + "content": "Experiments on various tasks demonstrate promising performance of the proposed", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 141, + 376, + 341, + 390 + ], + "spans": [ + { + "bbox": [ + 141, + 376, + 341, + 390 + ], + "score": 1.0, + "content": "CTF framework, compared to related techniques.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 12.5, + "bbox_fs": [ + 141, + 213, + 471, + 390 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 414, + 206, + 426 + ], + "lines": [ + { + "bbox": [ + 105, + 413, + 208, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 413, + 208, + 430 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 441, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 440, + 505, + 454 + ], + "score": 1.0, + "content": "Efficient inference and robust density estimation are two important goals in unsupervised learning. In", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 451, + 506, + 465 + ], + "score": 1.0, + "content": "fact, inference and density estimation can be unified from the perspective of learning desired target", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "distributions. In inference problems, one is given an unnormalized distribution (e.g., the posterior", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 505, + 486 + ], + "score": 1.0, + "content": "distribution), and the goal is to learn a distribution that is close to the unnormalized distribution. In", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 506, + 497 + ], + "score": 1.0, + "content": "density estimation, one tries to learn an unknown data distribution given samples from it. It is also", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 505, + 507 + ], + "score": 1.0, + "content": "helpful to make a distinction between two types of representations for learning distributions: explicit", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 506, + 505, + 518 + ], + "score": 1.0, + "content": "and implicit methods (Mohamed & Lakshminarayanan, 2017). Explicit methods provide a prescribed", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 517, + 505, + 531 + ], + "score": 1.0, + "content": "parametric form for the distribution, while implicit methods learn a stochastic procedure to directly", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 529, + 303, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 303, + 540 + ], + "score": 1.0, + "content": "generate samples from the unknown distribution.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 440, + 506, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 545, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 557 + ], + "score": 1.0, + "content": "Existing deep generative models can easily be identified from this taxonomy. For example, the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 506, + 569 + ], + "score": 1.0, + "content": "standard variational autoencoder (VAE) (Kingma & Welling, 2014; Rezende et al., 2014) is an impor-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 580 + ], + "score": 1.0, + "content": "tant example of an explicit inference method. Within the inference arm (encoder) of a VAE, recent", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 505, + 592 + ], + "score": 1.0, + "content": "research has focused on improving the accuracy of the approximation to the posterior distribution", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 589, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 589, + 506, + 602 + ], + "score": 1.0, + "content": "on latent variables (codes) using normalizing flow (Rezende & Mohamed, 2015). Normalizing flow", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 506, + 613 + ], + "score": 1.0, + "content": "is particularly interesting due to its ability to approximate the posterior distribution arbitrarily well,", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 505, + 623 + ], + "score": 1.0, + "content": "while maintaining explicit parametric forms. On the other hand, Stein VAE (Feng et al., 2017; Pu", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "score": 1.0, + "content": "et al., 2017b) is an implicit inference method, as it only learns to draw samples to approximate", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 506, + 645 + ], + "score": 1.0, + "content": "posteriors, without assuming an explicit form for the distribution.. For density estimation on observed", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 506, + 658 + ], + "score": 1.0, + "content": "data, the generative adversarial network (GAN) can be regarded as an implicit density estimation", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 506, + 668 + ], + "score": 1.0, + "content": "method (Ranganath et al., 2016; Huszár, 2017; Mohamed & Lakshminarayanan, 2017), in the sense", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 506, + 678 + ], + "score": 1.0, + "content": "that one may sample from the distribution (regarded as a representation of the unknown distribution),", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "but an explicit form for the distribution is not estimated. GAN has recently been augmented by", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 687, + 507, + 701 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 507, + 701 + ], + "score": 1.0, + "content": "Flow-GAN (Grover et al., 2017) to incorporate a likelihood term for explicit density estimation.", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "Further, there also are some works trying to perform inference within the implicit density estimation", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 723 + ], + "score": 1.0, + "content": "framework, e.g., the real-valued non-volume preserving (real NVP) transformations algorithm (Dinh", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 497, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 497, + 734 + ], + "score": 1.0, + "content": "et al., 2017) was proposed as a tractable yet expressive approach to model high-dimensional data.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 39, + "bbox_fs": [ + 105, + 546, + 507, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Some aforementioned methods rely on the concept of flows. A flow defines a series of transformations", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "for a random variable (RV), such that the distribution of the RV evolves from a simple distribution to", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "a more complex distribution. When the sequence of transformations are indexed on a discrete-time", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "domain (e.g., indexed with integers) with a finite number of transformations, this method is referred to", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "as a normalizing flow (Rezende & Mohamed, 2015). Various efficient implementations of normalizing", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "flows have been proposed, such as the planar, radial (Rezende & Mohamed, 2015), Householder", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "(Tomczak & Welling, 2016), and inverse autoregressive flows (Kingma et al., 2016). One theoretical", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "score": 1.0, + "content": "limitation of existing normalizing flows is that there is no guarantee on the approximation accuracy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 282, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 282, + 183 + ], + "score": 1.0, + "content": "due to the finite number of transformations.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "By contrast, little work has explored the applicability of continuous-time flows (CTFs) in deep", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "generative models, where a sequence of transformations are indexed on a continuous-time domain", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 208, + 507, + 222 + ], + "spans": [ + { + "bbox": [ + 104, + 208, + 507, + 222 + ], + "score": 1.0, + "content": "(e.g., indexed with real numbers). There are at least two reasons encouraging research in this direction:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 112, + 231 + ], + "score": 0.29, + "content": "i ,", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 220, + 505, + 233 + ], + "score": 1.0, + "content": ") CTFs are more general than traditional normalizing flows in terms of modeling flexibility, due", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 310, + 244 + ], + "score": 1.0, + "content": "to the intrinsic infinite number of transformations;", + "type": "text" + }, + { + "bbox": [ + 310, + 232, + 317, + 241 + ], + "score": 0.39, + "content": "i i", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 231, + 505, + 244 + ], + "score": 1.0, + "content": ") CTFs are more theoretically grounded, in the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 254 + ], + "score": 1.0, + "content": "sense that they are guaranteed to approach a target distribution asymptotically (details provided in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "Section 2.2). Unfortunately, these advantages also bring challenges for efficient learning, in that: i)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "it is difficult to optimize over the variational lower bound in the inference framework, due to the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "extra randomness introduced in CTFs; ii) it is difficult to design algorithms for efficient learning of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 405, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 405, + 298 + ], + "score": 1.0, + "content": "CTF-based models, due to the induced infinite number of transformations.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5 + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "In this paper, we propose efficient ways to apply CTFs for the two motivating tasks. Based on the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "CTF, our framework learns to drawn samples directly from desired distributions (e.g., the unknown", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "posterior and data distributions) for both inference and density estimation tasks. In addition, we are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "able to learn an explicit form of the unknown data distribution for density estimation∗. This shares a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 347, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 389, + 359 + ], + "score": 1.0, + "content": "similar flavor as Wang & Liu (2017); Feng et al. (2017). Specifically,", + "type": "text" + }, + { + "bbox": [ + 389, + 347, + 394, + 357 + ], + "score": 0.41, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 347, + 506, + 359 + ], + "score": 1.0, + "content": ") for efficient inference, we", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "first show that optimizing the variational lower bound with CTFs can be achieved by decomposing", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 506, + 381 + ], + "score": 1.0, + "content": "the optimization problem into a sequence of sub-optimization problems, based on a variational", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "formulation of the Fokker-Planck equations from statistical physics (Jordan et al., 1998). Based on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "this decomposition, we derive bounds on the approximation errors when applying numerical methods", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "to solve a CTF. For computational efficiency, we generalize ideas from Gershman & Goodman (2014)", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 381, + 425 + ], + "score": 1.0, + "content": "to distill knowledge of a CTF into an efficient inference network;", + "type": "text" + }, + { + "bbox": [ + 382, + 413, + 389, + 423 + ], + "score": 0.26, + "content": "\\ddot { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 411, + 506, + 425 + ], + "score": 1.0, + "content": ") for density estimation, we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "propose to use a flexible Gibbsian-style distribution (implemented by a deep neural network) to model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "an unknown data distribution, whose samples can be drawn by learning a stochastic generator with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "our CTF framework. The Gibbsian-style data distribution and the stochastic generator are learned", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "alternatively, leading to a learning procedure that is connected to the GAN framework (Goodfellow", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "et al., 2014), but that yields an explicit distribution for the data. We conduct a number of experiments", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 492 + ], + "score": 1.0, + "content": "on real datasets, demonstrating excellent performance of the proposed framework, relative to existing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 490, + 214, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 214, + 501 + ], + "score": 1.0, + "content": "representative approaches.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 27.5 + }, + { + "type": "title", + "bbox": [ + 108, + 519, + 208, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 209, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 209, + 534 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 504, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "We first review related techniques of performing efficient inference and density estimation in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 558, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 479, + 569 + ], + "score": 1.0, + "content": "machine learning literature. We then introduce the general concept of continuous-time flows.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5 + }, + { + "type": "title", + "bbox": [ + 107, + 586, + 344, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 345, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 345, + 598 + ], + "score": 1.0, + "content": "2.1 EFFICIENT INFERENCE AND DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 506, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "Efficient inference with normalizing flows Consider a probabilistic generative model with ob-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 146, + 632 + ], + "score": 1.0, + "content": "servation", + "type": "text" + }, + { + "bbox": [ + 147, + 618, + 184, + 628 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 616, + 266, + 632 + ], + "score": 1.0, + "content": "and latent variable", + "type": "text" + }, + { + "bbox": [ + 267, + 618, + 302, + 628 + ], + "score": 0.9, + "content": "\\textbf { z } \\in \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 616, + 345, + 632 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 346, + 618, + 416, + 630 + ], + "score": 0.92, + "content": "\\mathbf { x } \\mid \\mathbf { z } \\sim p _ { \\pmb { \\theta } } ( \\mathbf { x } \\mid \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 616, + 439, + 632 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 440, + 618, + 481, + 630 + ], + "score": 0.93, + "content": "\\mathbf { z } \\sim p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 616, + 506, + 632 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 192, + 642 + ], + "score": 1.0, + "content": "efficient inference of", + "type": "text" + }, + { + "bbox": [ + 192, + 631, + 199, + 639 + ], + "score": 0.55, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 628, + 506, + 642 + ], + "score": 1.0, + "content": ", the VAE (Kingma & Welling, 2014) introduces the concept of an inference", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 639, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 639, + 277, + 654 + ], + "score": 1.0, + "content": "network (recognition model or encoder),", + "type": "text" + }, + { + "bbox": [ + 277, + 640, + 313, + 653 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 639, + 507, + 654 + ], + "score": 1.0, + "content": ", as a variational distribution in the VB frame-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 470, + 664 + ], + "score": 1.0, + "content": "work. An inference network is typically a stochastic (nonlinear) mapping from the input", + "type": "text" + }, + { + "bbox": [ + 470, + 653, + 479, + 662 + ], + "score": 0.43, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "to the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 661, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 132, + 675 + ], + "score": 1.0, + "content": "latent", + "type": "text" + }, + { + "bbox": [ + 132, + 664, + 139, + 672 + ], + "score": 0.56, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 661, + 258, + 675 + ], + "score": 1.0, + "content": ", with associated parameters", + "type": "text" + }, + { + "bbox": [ + 258, + 663, + 266, + 673 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 661, + 506, + 675 + ], + "score": 1.0, + "content": ". For example, one of the simplest inference networks is", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 148, + 686 + ], + "score": 1.0, + "content": "defined as", + "type": "text" + }, + { + "bbox": [ + 148, + 673, + 305, + 687 + ], + "score": 0.83, + "content": "\\begin{array} { r } { q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) = \\mathcal { N } ( \\bar { \\mathbf { z } } ; \\mu _ { \\phi } ( \\mathbf { x } ) , \\mathrm { d i a g } ( \\sigma _ { \\phi } ^ { 2 } ( \\mathbf { x } ) ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 673, + 408, + 686 + ], + "score": 1.0, + "content": ", where the mean function", + "type": "text" + }, + { + "bbox": [ + 408, + 673, + 435, + 686 + ], + "score": 0.93, + "content": "\\mu _ { \\phi } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 673, + 506, + 686 + ], + "score": 1.0, + "content": "and the standard-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 685, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 183, + 699 + ], + "score": 1.0, + "content": "derivation function", + "type": "text" + }, + { + "bbox": [ + 183, + 686, + 210, + 699 + ], + "score": 0.92, + "content": "\\sigma _ { \\phi } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 685, + 433, + 699 + ], + "score": 1.0, + "content": "are specified via deep neural networks parameterized by", + "type": "text" + }, + { + "bbox": [ + 433, + 686, + 441, + 697 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 685, + 506, + 699 + ], + "score": 1.0, + "content": ". Parameters are", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 695, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 470, + 711 + ], + "score": 1.0, + "content": "learned by minimizing the evidence lower bound (ELBO), i.e., the KL divergence between", + "type": "text" + }, + { + "bbox": [ + 470, + 697, + 505, + 709 + ], + "score": 0.94, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + } + ], + "index": 49 + } + ], + "index": 45 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 114, + 721, + 476, + 732 + ], + "lines": [ + { + "bbox": [ + 118, + 721, + 478, + 733 + ], + "spans": [ + { + "bbox": [ + 118, + 721, + 478, + 733 + ], + "score": 1.0, + "content": "∗Although the density is represented as an energy-based distribution with an intractable normalizer.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 181 + ], + "lines": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 83, + 505, + 94 + ], + "score": 1.0, + "content": "Some aforementioned methods rely on the concept of flows. A flow defines a series of transformations", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 506, + 106 + ], + "score": 1.0, + "content": "for a random variable (RV), such that the distribution of the RV evolves from a simple distribution to", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 117 + ], + "score": 1.0, + "content": "a more complex distribution. When the sequence of transformations are indexed on a discrete-time", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 506, + 128 + ], + "score": 1.0, + "content": "domain (e.g., indexed with integers) with a finite number of transformations, this method is referred to", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 506, + 140 + ], + "score": 1.0, + "content": "as a normalizing flow (Rezende & Mohamed, 2015). Various efficient implementations of normalizing", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 506, + 150 + ], + "score": 1.0, + "content": "flows have been proposed, such as the planar, radial (Rezende & Mohamed, 2015), Householder", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "spans": [ + { + "bbox": [ + 105, + 148, + 505, + 162 + ], + "score": 1.0, + "content": "(Tomczak & Welling, 2016), and inverse autoregressive flows (Kingma et al., 2016). One theoretical", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 505, + 173 + ], + "score": 1.0, + "content": "limitation of existing normalizing flows is that there is no guarantee on the approximation accuracy", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 170, + 282, + 183 + ], + "spans": [ + { + "bbox": [ + 106, + 170, + 282, + 183 + ], + "score": 1.0, + "content": "due to the finite number of transformations.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 4, + "bbox_fs": [ + 105, + 83, + 506, + 183 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 187, + 505, + 297 + ], + "lines": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "spans": [ + { + "bbox": [ + 105, + 186, + 505, + 200 + ], + "score": 1.0, + "content": "By contrast, little work has explored the applicability of continuous-time flows (CTFs) in deep", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "spans": [ + { + "bbox": [ + 105, + 198, + 506, + 211 + ], + "score": 1.0, + "content": "generative models, where a sequence of transformations are indexed on a continuous-time domain", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 104, + 208, + 507, + 222 + ], + "spans": [ + { + "bbox": [ + 104, + 208, + 507, + 222 + ], + "score": 1.0, + "content": "(e.g., indexed with real numbers). There are at least two reasons encouraging research in this direction:", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 220, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 112, + 231 + ], + "score": 0.29, + "content": "i ,", + "type": "inline_equation" + }, + { + "bbox": [ + 112, + 220, + 505, + 233 + ], + "score": 1.0, + "content": ") CTFs are more general than traditional normalizing flows in terms of modeling flexibility, due", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 231, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 231, + 310, + 244 + ], + "score": 1.0, + "content": "to the intrinsic infinite number of transformations;", + "type": "text" + }, + { + "bbox": [ + 310, + 232, + 317, + 241 + ], + "score": 0.39, + "content": "i i", + "type": "inline_equation" + }, + { + "bbox": [ + 318, + 231, + 505, + 244 + ], + "score": 1.0, + "content": ") CTFs are more theoretically grounded, in the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 243, + 505, + 254 + ], + "spans": [ + { + "bbox": [ + 105, + 243, + 505, + 254 + ], + "score": 1.0, + "content": "sense that they are guaranteed to approach a target distribution asymptotically (details provided in", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 105, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "Section 2.2). Unfortunately, these advantages also bring challenges for efficient learning, in that: i)", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 506, + 276 + ], + "score": 1.0, + "content": "it is difficult to optimize over the variational lower bound in the inference framework, due to the", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 275, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 286 + ], + "score": 1.0, + "content": "extra randomness introduced in CTFs; ii) it is difficult to design algorithms for efficient learning of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 286, + 405, + 298 + ], + "spans": [ + { + "bbox": [ + 106, + 286, + 405, + 298 + ], + "score": 1.0, + "content": "CTF-based models, due to the induced infinite number of transformations.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 186, + 507, + 298 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 303, + 505, + 500 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "In this paper, we propose efficient ways to apply CTFs for the two motivating tasks. Based on the", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 506, + 326 + ], + "score": 1.0, + "content": "CTF, our framework learns to drawn samples directly from desired distributions (e.g., the unknown", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 324, + 506, + 338 + ], + "score": 1.0, + "content": "posterior and data distributions) for both inference and density estimation tasks. In addition, we are", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 506, + 348 + ], + "score": 1.0, + "content": "able to learn an explicit form of the unknown data distribution for density estimation∗. This shares a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 347, + 506, + 359 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 389, + 359 + ], + "score": 1.0, + "content": "similar flavor as Wang & Liu (2017); Feng et al. (2017). Specifically,", + "type": "text" + }, + { + "bbox": [ + 389, + 347, + 394, + 357 + ], + "score": 0.41, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 347, + 506, + 359 + ], + "score": 1.0, + "content": ") for efficient inference, we", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "first show that optimizing the variational lower bound with CTFs can be achieved by decomposing", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 368, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 105, + 368, + 506, + 381 + ], + "score": 1.0, + "content": "the optimization problem into a sequence of sub-optimization problems, based on a variational", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "formulation of the Fokker-Planck equations from statistical physics (Jordan et al., 1998). Based on", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 506, + 404 + ], + "score": 1.0, + "content": "this decomposition, we derive bounds on the approximation errors when applying numerical methods", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 506, + 414 + ], + "score": 1.0, + "content": "to solve a CTF. For computational efficiency, we generalize ideas from Gershman & Goodman (2014)", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 411, + 506, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 411, + 381, + 425 + ], + "score": 1.0, + "content": "to distill knowledge of a CTF into an efficient inference network;", + "type": "text" + }, + { + "bbox": [ + 382, + 413, + 389, + 423 + ], + "score": 0.26, + "content": "\\ddot { u }", + "type": "inline_equation" + }, + { + "bbox": [ + 390, + 411, + 506, + 425 + ], + "score": 1.0, + "content": ") for density estimation, we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "spans": [ + { + "bbox": [ + 104, + 423, + 506, + 437 + ], + "score": 1.0, + "content": "propose to use a flexible Gibbsian-style distribution (implemented by a deep neural network) to model", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 105, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "an unknown data distribution, whose samples can be drawn by learning a stochastic generator with", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 446, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 505, + 457 + ], + "score": 1.0, + "content": "our CTF framework. The Gibbsian-style data distribution and the stochastic generator are learned", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 105, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "alternatively, leading to a learning procedure that is connected to the GAN framework (Goodfellow", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 480 + ], + "score": 1.0, + "content": "et al., 2014), but that yields an explicit distribution for the data. We conduct a number of experiments", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 477, + 506, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 506, + 492 + ], + "score": 1.0, + "content": "on real datasets, demonstrating excellent performance of the proposed framework, relative to existing", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 490, + 214, + 501 + ], + "spans": [ + { + "bbox": [ + 105, + 490, + 214, + 501 + ], + "score": 1.0, + "content": "representative approaches.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 303, + 506, + 501 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 519, + 208, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 518, + 209, + 534 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 209, + 534 + ], + "score": 1.0, + "content": "2 PRELIMINARIES", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 546, + 504, + 568 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "We first review related techniques of performing efficient inference and density estimation in the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 558, + 479, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 479, + 569 + ], + "score": 1.0, + "content": "machine learning literature. We then introduce the general concept of continuous-time flows.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38.5, + "bbox_fs": [ + 105, + 546, + 505, + 569 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 586, + 344, + 597 + ], + "lines": [ + { + "bbox": [ + 105, + 585, + 345, + 598 + ], + "spans": [ + { + "bbox": [ + 105, + 585, + 345, + 598 + ], + "score": 1.0, + "content": "2.1 EFFICIENT INFERENCE AND DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "text", + "bbox": [ + 107, + 607, + 506, + 708 + ], + "lines": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "spans": [ + { + "bbox": [ + 105, + 606, + 506, + 620 + ], + "score": 1.0, + "content": "Efficient inference with normalizing flows Consider a probabilistic generative model with ob-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 506, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 146, + 632 + ], + "score": 1.0, + "content": "servation", + "type": "text" + }, + { + "bbox": [ + 147, + 618, + 184, + 628 + ], + "score": 0.91, + "content": "\\mathbf { x } \\in \\mathbb { R } ^ { D }", + "type": "inline_equation" + }, + { + "bbox": [ + 185, + 616, + 266, + 632 + ], + "score": 1.0, + "content": "and latent variable", + "type": "text" + }, + { + "bbox": [ + 267, + 618, + 302, + 628 + ], + "score": 0.9, + "content": "\\textbf { z } \\in \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 616, + 345, + 632 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 346, + 618, + 416, + 630 + ], + "score": 0.92, + "content": "\\mathbf { x } \\mid \\mathbf { z } \\sim p _ { \\pmb { \\theta } } ( \\mathbf { x } \\mid \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 416, + 616, + 439, + 632 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 440, + 618, + 481, + 630 + ], + "score": 0.93, + "content": "\\mathbf { z } \\sim p ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 616, + 506, + 632 + ], + "score": 1.0, + "content": ". For", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 105, + 628, + 192, + 642 + ], + "score": 1.0, + "content": "efficient inference of", + "type": "text" + }, + { + "bbox": [ + 192, + 631, + 199, + 639 + ], + "score": 0.55, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 628, + 506, + 642 + ], + "score": 1.0, + "content": ", the VAE (Kingma & Welling, 2014) introduces the concept of an inference", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 104, + 639, + 507, + 654 + ], + "spans": [ + { + "bbox": [ + 104, + 639, + 277, + 654 + ], + "score": 1.0, + "content": "network (recognition model or encoder),", + "type": "text" + }, + { + "bbox": [ + 277, + 640, + 313, + 653 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 313, + 639, + 507, + 654 + ], + "score": 1.0, + "content": ", as a variational distribution in the VB frame-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 651, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 651, + 470, + 664 + ], + "score": 1.0, + "content": "work. An inference network is typically a stochastic (nonlinear) mapping from the input", + "type": "text" + }, + { + "bbox": [ + 470, + 653, + 479, + 662 + ], + "score": 0.43, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 479, + 651, + 505, + 664 + ], + "score": 1.0, + "content": "to the", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 661, + 506, + 675 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 132, + 675 + ], + "score": 1.0, + "content": "latent", + "type": "text" + }, + { + "bbox": [ + 132, + 664, + 139, + 672 + ], + "score": 0.56, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 139, + 661, + 258, + 675 + ], + "score": 1.0, + "content": ", with associated parameters", + "type": "text" + }, + { + "bbox": [ + 258, + 663, + 266, + 673 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 661, + 506, + 675 + ], + "score": 1.0, + "content": ". For example, one of the simplest inference networks is", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 106, + 673, + 506, + 687 + ], + "spans": [ + { + "bbox": [ + 106, + 673, + 148, + 686 + ], + "score": 1.0, + "content": "defined as", + "type": "text" + }, + { + "bbox": [ + 148, + 673, + 305, + 687 + ], + "score": 0.83, + "content": "\\begin{array} { r } { q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) = \\mathcal { N } ( \\bar { \\mathbf { z } } ; \\mu _ { \\phi } ( \\mathbf { x } ) , \\mathrm { d i a g } ( \\sigma _ { \\phi } ^ { 2 } ( \\mathbf { x } ) ) ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 673, + 408, + 686 + ], + "score": 1.0, + "content": ", where the mean function", + "type": "text" + }, + { + "bbox": [ + 408, + 673, + 435, + 686 + ], + "score": 0.93, + "content": "\\mu _ { \\phi } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 673, + 506, + 686 + ], + "score": 1.0, + "content": "and the standard-", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 685, + 506, + 699 + ], + "spans": [ + { + "bbox": [ + 105, + 685, + 183, + 699 + ], + "score": 1.0, + "content": "derivation function", + "type": "text" + }, + { + "bbox": [ + 183, + 686, + 210, + 699 + ], + "score": 0.92, + "content": "\\sigma _ { \\phi } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 685, + 433, + 699 + ], + "score": 1.0, + "content": "are specified via deep neural networks parameterized by", + "type": "text" + }, + { + "bbox": [ + 433, + 686, + 441, + 697 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 685, + 506, + 699 + ], + "score": 1.0, + "content": ". Parameters are", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 695, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 105, + 695, + 470, + 711 + ], + "score": 1.0, + "content": "learned by minimizing the evidence lower bound (ELBO), i.e., the KL divergence between", + "type": "text" + }, + { + "bbox": [ + 470, + 697, + 505, + 709 + ], + "score": 0.94, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + } + ], + "index": 49 + } + ], + "index": 45, + "bbox_fs": [ + 104, + 606, + 507, + 711 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 123, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 82, + 159, + 95 + ], + "score": 0.79, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 82, + 408, + 96 + ], + "score": 0.76, + "content": "\\mathbf { \\cdot } ) \\colon \\mathrm { K L } \\left( q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) \\triangleq \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) } \\left[ \\log q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ", via stochastic gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 203, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 203, + 106 + ], + "score": 1.0, + "content": "descent (Bottou, 2012).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 288, + 123 + ], + "score": 1.0, + "content": "One limitation of the VAE framework is that", + "type": "text" + }, + { + "bbox": [ + 289, + 110, + 324, + 123 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "is often restricted to simple distributions for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 450, + 134 + ], + "score": 1.0, + "content": "feasibility, e.g., the normal distribution discussed above, and thus the gap between", + "type": "text" + }, + { + "bbox": [ + 450, + 121, + 486, + 134 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 142, + 144 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { z } | \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "is typically large for complicated posterior distributions. Normalizing flows is a recently", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "proposed VB-based technique designed to mitigate this problem (Rezende & Mohamed, 2015). The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 152, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 103, + 152, + 180, + 168 + ], + "score": 1.0, + "content": "idea is to augment", + "type": "text" + }, + { + "bbox": [ + 181, + 156, + 187, + 164 + ], + "score": 0.55, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 152, + 416, + 168 + ], + "score": 1.0, + "content": "via a sequence of deterministic invertible transformations", + "type": "text" + }, + { + "bbox": [ + 416, + 154, + 502, + 167 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } : \\mathbb { R } ^ { L } \\to \\mathbb { R } ^ { \\dot { L } } \\} _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 152, + 506, + 168 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 164, + 354, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 147, + 177 + ], + "score": 1.0, + "content": "such that:", + "type": "text" + }, + { + "bbox": [ + 148, + 165, + 351, + 178 + ], + "score": 0.87, + "content": "\\mathbf { z } _ { 0 } \\sim q _ { \\phi } ( \\cdot | \\mathbf { x } ) , \\mathbf { z } _ { 1 } = \\mathcal { T } _ { 1 } ( \\mathbf { z } _ { 0 } ) , \\cdot \\cdot \\cdot , \\mathbf { z } _ { K } = \\mathcal { T } _ { K } ( \\mathbf { z } _ { K - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 164, + 354, + 177 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5 + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 213, + 195 + ], + "score": 1.0, + "content": "Note the transformations", + "type": "text" + }, + { + "bbox": [ + 213, + 182, + 234, + 194 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "are typically endowed with different parameters, and we absorb", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 146, + 205 + ], + "score": 1.0, + "content": "them into", + "type": "text" + }, + { + "bbox": [ + 147, + 194, + 154, + 204 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 192, + 420, + 205 + ], + "score": 1.0, + "content": ". Because the transformations are deterministic, the distribution of", + "type": "text" + }, + { + "bbox": [ + 420, + 194, + 434, + 204 + ], + "score": 0.87, + "content": "\\mathbf { z } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 204, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 259, + 226 + ], + "score": 0.91, + "content": "\\begin{array} { r } { q ( \\mathbf { z } _ { K } ) = q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) \\prod _ { k = 1 } ^ { K } \\left| \\operatorname* { d e t } \\frac { \\partial { \\mathcal T } _ { k } } { \\partial \\mathbf { z } _ { k } } \\right| ^ { - 1 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "via the change of variable formula. As a result, the ELBO for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 223, + 221, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 221, + 236 + ], + "score": 1.0, + "content": "normalizing flows becomes:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5 + }, + { + "type": "interline_equation", + "bbox": [ + 126, + 239, + 484, + 291 + ], + "lines": [ + { + "bbox": [ + 126, + 239, + 484, + 291 + ], + "spans": [ + { + "bbox": [ + 126, + 239, + 484, + 291 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\mathrm { K L } \\left( q _ { \\phi } ( \\mathbf { z } _ { K } \\mid \\mathbf { x } ) \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) = } \\\\ { \\displaystyle \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\log q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) \\right] - \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } _ { K } ) \\right] - \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\sum _ { k = 1 } ^ { K } \\log \\left| \\mathrm { d e t } \\frac { \\partial T _ { k } } { \\partial \\mathbf { z } _ { k } } \\right| \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "b6c35f7eded82e866476d16ef2cdbfb92964dd3c567653511718b2a8f9ebdeb8.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 126, + 239, + 484, + 256.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 126, + 256.3333333333333, + 484, + 273.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 126, + 273.66666666666663, + 484, + 290.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 214, + 308 + ], + "score": 1.0, + "content": "Typically, transformations", + "type": "text" + }, + { + "bbox": [ + 214, + 296, + 226, + 307 + ], + "score": 0.89, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "of a simple parametric form are employed to make the computations", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 305, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 319 + ], + "score": 1.0, + "content": "tractable (Rezende & Mohamed, 2015). Our method for inference generalizes these discrete-time", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "transformation to continuous-time transformations, ensuring convergence of the transformations to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 329, + 197, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 197, + 341 + ], + "score": 1.0, + "content": "the target distribution.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 352, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "Density estimation overview There exist implicit and explicit density-estimation methods. Implicit", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "density models such as GAN provide a flexible way to draw samples directly from unknown data", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "distributions (via a deep neural network (DNN) called a generator with stochastic inputs) without", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "explicitly modeling their density forms; whereas explicit models such as the pixel RNN/CNN (van den", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "Oord et al., 2016) define and learn explicit forms of the unknown data distributions. This gives the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "advantage that the likelihood for a test data point can be explicitly evaluated. However, the generation", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 419, + 425, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 425, + 431 + ], + "score": 1.0, + "content": "of samples is typically time-consuming due to the sequential generation nature.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 449 + ], + "score": 1.0, + "content": "Similar to Wang & Liu (2017), our CTF-based approach in Section 4 provides an alternative way", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "for this problem, by simultaneously learning an explicit Gibbsian-style data distribution (estimated", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "density) and a generator whose generated samples match the learned Gibbsian distribution. This", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "not only gives us the advantage of explicit density modeling but also provides an efficient way to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 479, + 180, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 180, + 492 + ], + "score": 1.0, + "content": "generate samples.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28 + }, + { + "type": "title", + "bbox": [ + 108, + 504, + 246, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 248, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 248, + 517 + ], + "score": 1.0, + "content": "2.2 CONTINUOUS-TIME FLOWS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "We notice two potential limitations with traditional normalizing flows: i) given specified transfor-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 142, + 548 + ], + "score": 1.0, + "content": "mations", + "type": "text" + }, + { + "bbox": [ + 142, + 536, + 163, + 548 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 536, + 352, + 548 + ], + "score": 1.0, + "content": ", there is no guarantee that the distribution of", + "type": "text" + }, + { + "bbox": [ + 352, + 537, + 367, + 547 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 536, + 454, + 548 + ], + "score": 1.0, + "content": "could exactly match", + "type": "text" + }, + { + "bbox": [ + 454, + 536, + 489, + 548 + ], + "score": 0.75, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "; ii)", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 548, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 259, + 559 + ], + "score": 1.0, + "content": "the randomness is only introduced in", + "type": "text" + }, + { + "bbox": [ + 259, + 549, + 270, + 558 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 548, + 504, + 559 + ], + "score": 1.0, + "content": "(from the inference network), limiting the representation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "power. We specify CTFs where the transformations are indexed by real numbers, thus they could be", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "considered as consisting of an infinite number of transformations. Further, we consider stochastic", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "flows where randomness is injected in a continuous-time manner. In fact, the concept of CTFs", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "(such as the Hamiltonian flow) has been introduced in Rezende & Mohamed (2015), without further", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 602, + 249, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 249, + 613 + ], + "score": 1.0, + "content": "development on efficient inference.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5 + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 205, + 631 + ], + "score": 1.0, + "content": "We consider a flow on", + "type": "text" + }, + { + "bbox": [ + 205, + 618, + 219, + 628 + ], + "score": 0.87, + "content": "\\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 617, + 330, + 631 + ], + "score": 1.0, + "content": ", defined as the mapping†", + "type": "text" + }, + { + "bbox": [ + 330, + 618, + 420, + 629 + ], + "score": 0.91, + "content": "\\mathcal { T } : \\mathbb { R } ^ { L } \\times \\mathbb { R } \\to \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "such that‡ we have", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 107, + 630, + 159, + 642 + ], + "score": 0.93, + "content": "\\boldsymbol { \\mathcal { T } } ( \\mathbf { Z } , 0 ) = \\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 628, + 177, + 642 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 178, + 630, + 295, + 641 + ], + "score": 0.92, + "content": "\\mathcal { T } ( \\mathcal { T } ( \\mathbf { Z } , t ) , s ) = \\mathcal { T } ( \\mathbf { Z } , s + t )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 628, + 326, + 642 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 326, + 629, + 360, + 640 + ], + "score": 0.91, + "content": "\\mathbf { Z } \\in \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 628, + 379, + 642 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 630, + 412, + 641 + ], + "score": 0.9, + "content": "s , t \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 628, + 506, + 642 + ], + "score": 1.0, + "content": ". A typical example of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 447, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 204, + 653 + ], + "score": 1.0, + "content": "this family is defined as", + "type": "text" + }, + { + "bbox": [ + 204, + 641, + 259, + 653 + ], + "score": 0.92, + "content": "\\mathcal { T } ( \\mathbf { Z } , t ) = \\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 639, + 290, + 653 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 290, + 641, + 302, + 651 + ], + "score": 0.89, + "content": "\\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 639, + 447, + 653 + ], + "score": 1.0, + "content": "is driven by a diffusion of the form:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41 + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 658, + 370, + 671 + ], + "lines": [ + { + "bbox": [ + 241, + 658, + 370, + 671 + ], + "spans": [ + { + "bbox": [ + 241, + 658, + 370, + 671 + ], + "score": 0.91, + "content": "\\mathrm { d } \\mathbf { Z } _ { t } = F ( \\mathbf { Z } _ { t } ) \\mathrm { d } t + V ( \\mathbf { Z } _ { t } ) \\mathrm { d } { \\mathcal { W } } .", + "type": "interline_equation", + "image_path": "c8ef519cf5f4b039dffc5a4a43dd8e677643835401771d015f79f2b1b33b1d71.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 241, + 658, + 370, + 671 + ], + "spans": [], + "index": 43 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 680, + 506, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 678, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 118, + 678, + 201, + 693 + ], + "score": 1.0, + "content": "†We reuse the notation", + "type": "text" + }, + { + "bbox": [ + 201, + 681, + 210, + 690 + ], + "score": 0.79, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 678, + 459, + 693 + ], + "score": 1.0, + "content": "as transformations from the discrete case above for simplicity, and use", + "type": "text" + }, + { + "bbox": [ + 460, + 681, + 467, + 690 + ], + "score": 0.7, + "content": "\\mathbf { Z }", + "type": "inline_equation" + }, + { + "bbox": [ + 468, + 678, + 506, + 693 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 689, + 474, + 703 + ], + "spans": [ + { + "bbox": [ + 106, + 692, + 113, + 700 + ], + "score": 0.61, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 113, + 689, + 474, + 703 + ], + "score": 1.0, + "content": "(reserved for the discrete-time setting) to denote the random variable in the continuous-time setting.", + "type": "text" + } + ] + }, + { + "bbox": [ + 117, + 699, + 506, + 714 + ], + "spans": [ + { + "bbox": [ + 117, + 699, + 352, + 714 + ], + "score": 1.0, + "content": "‡Note we define continuous-time flows in terms of latent variable", + "type": "text" + }, + { + "bbox": [ + 353, + 702, + 360, + 711 + ], + "score": 0.74, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 699, + 506, + 714 + ], + "score": 1.0, + "content": "in order to incorporate it into the setting", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 724 + ], + "score": 1.0, + "content": "of inference. However, the same description applies when we define the flow in data space, which is the setting", + "type": "text" + } + ] + }, + { + "bbox": [ + 106, + 722, + 231, + 732 + ], + "spans": [ + { + "bbox": [ + 106, + 722, + 231, + 732 + ], + "score": 1.0, + "content": "of density estimation in Section 4.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 105 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 123, + 96 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 124, + 82, + 159, + 95 + ], + "score": 0.79, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 82, + 408, + 96 + ], + "score": 0.76, + "content": "\\mathbf { \\cdot } ) \\colon \\mathrm { K L } \\left( q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) \\triangleq \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) } \\left[ \\log q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } ) - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 81, + 506, + 96 + ], + "score": 1.0, + "content": ", via stochastic gradient", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 203, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 203, + 106 + ], + "score": 1.0, + "content": "descent (Bottou, 2012).", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 105, + 81, + 506, + 106 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 110, + 505, + 178 + ], + "lines": [ + { + "bbox": [ + 106, + 110, + 506, + 123 + ], + "spans": [ + { + "bbox": [ + 106, + 110, + 288, + 123 + ], + "score": 1.0, + "content": "One limitation of the VAE framework is that", + "type": "text" + }, + { + "bbox": [ + 289, + 110, + 324, + 123 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 110, + 506, + 123 + ], + "score": 1.0, + "content": "is often restricted to simple distributions for", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 121, + 505, + 134 + ], + "spans": [ + { + "bbox": [ + 105, + 121, + 450, + 134 + ], + "score": 1.0, + "content": "feasibility, e.g., the normal distribution discussed above, and thus the gap between", + "type": "text" + }, + { + "bbox": [ + 450, + 121, + 486, + 134 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 121, + 505, + 134 + ], + "score": 1.0, + "content": "and", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 132, + 506, + 145 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 142, + 144 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { z } | \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 132, + 506, + 145 + ], + "score": 1.0, + "content": "is typically large for complicated posterior distributions. Normalizing flows is a recently", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 506, + 156 + ], + "score": 1.0, + "content": "proposed VB-based technique designed to mitigate this problem (Rezende & Mohamed, 2015). The", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 152, + 506, + 168 + ], + "spans": [ + { + "bbox": [ + 103, + 152, + 180, + 168 + ], + "score": 1.0, + "content": "idea is to augment", + "type": "text" + }, + { + "bbox": [ + 181, + 156, + 187, + 164 + ], + "score": 0.55, + "content": "\\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 152, + 416, + 168 + ], + "score": 1.0, + "content": "via a sequence of deterministic invertible transformations", + "type": "text" + }, + { + "bbox": [ + 416, + 154, + 502, + 167 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } : \\mathbb { R } ^ { L } \\to \\mathbb { R } ^ { \\dot { L } } \\} _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 152, + 506, + 168 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 164, + 354, + 178 + ], + "spans": [ + { + "bbox": [ + 106, + 164, + 147, + 177 + ], + "score": 1.0, + "content": "such that:", + "type": "text" + }, + { + "bbox": [ + 148, + 165, + 351, + 178 + ], + "score": 0.87, + "content": "\\mathbf { z } _ { 0 } \\sim q _ { \\phi } ( \\cdot | \\mathbf { x } ) , \\mathbf { z } _ { 1 } = \\mathcal { T } _ { 1 } ( \\mathbf { z } _ { 0 } ) , \\cdot \\cdot \\cdot , \\mathbf { z } _ { K } = \\mathcal { T } _ { K } ( \\mathbf { z } _ { K - 1 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 164, + 354, + 177 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 4.5, + "bbox_fs": [ + 103, + 110, + 506, + 178 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 181, + 505, + 235 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 213, + 195 + ], + "score": 1.0, + "content": "Note the transformations", + "type": "text" + }, + { + "bbox": [ + 213, + 182, + 234, + 194 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 181, + 505, + 195 + ], + "score": 1.0, + "content": "are typically endowed with different parameters, and we absorb", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 192, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 146, + 205 + ], + "score": 1.0, + "content": "them into", + "type": "text" + }, + { + "bbox": [ + 147, + 194, + 154, + 204 + ], + "score": 0.83, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 192, + 420, + 205 + ], + "score": 1.0, + "content": ". Because the transformations are deterministic, the distribution of", + "type": "text" + }, + { + "bbox": [ + 420, + 194, + 434, + 204 + ], + "score": 0.87, + "content": "\\mathbf { z } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 192, + 505, + 205 + ], + "score": 1.0, + "content": "can be written as", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 204, + 506, + 226 + ], + "spans": [ + { + "bbox": [ + 106, + 204, + 259, + 226 + ], + "score": 0.91, + "content": "\\begin{array} { r } { q ( \\mathbf { z } _ { K } ) = q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) \\prod _ { k = 1 } ^ { K } \\left| \\operatorname* { d e t } \\frac { \\partial { \\mathcal T } _ { k } } { \\partial \\mathbf { z } _ { k } } \\right| ^ { - 1 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 209, + 506, + 221 + ], + "score": 1.0, + "content": "via the change of variable formula. As a result, the ELBO for", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 223, + 221, + 236 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 221, + 236 + ], + "score": 1.0, + "content": "normalizing flows becomes:", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 9.5, + "bbox_fs": [ + 106, + 181, + 506, + 236 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 126, + 239, + 484, + 291 + ], + "lines": [ + { + "bbox": [ + 126, + 239, + 484, + 291 + ], + "spans": [ + { + "bbox": [ + 126, + 239, + 484, + 291 + ], + "score": 0.93, + "content": "\\begin{array} { l } { \\displaystyle \\mathrm { K L } \\left( q _ { \\phi } ( \\mathbf { z } _ { K } \\mid \\mathbf { x } ) \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) = } \\\\ { \\displaystyle \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\log q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) \\right] - \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } _ { K } ) \\right] - \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\sum _ { k = 1 } ^ { K } \\log \\left| \\mathrm { d e t } \\frac { \\partial T _ { k } } { \\partial \\mathbf { z } _ { k } } \\right| \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "b6c35f7eded82e866476d16ef2cdbfb92964dd3c567653511718b2a8f9ebdeb8.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 126, + 239, + 484, + 256.3333333333333 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 126, + 256.3333333333333, + 484, + 273.66666666666663 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 126, + 273.66666666666663, + 484, + 290.99999999999994 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 295, + 505, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 295, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 106, + 295, + 214, + 308 + ], + "score": 1.0, + "content": "Typically, transformations", + "type": "text" + }, + { + "bbox": [ + 214, + 296, + 226, + 307 + ], + "score": 0.89, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 295, + 505, + 308 + ], + "score": 1.0, + "content": "of a simple parametric form are employed to make the computations", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 305, + 506, + 319 + ], + "spans": [ + { + "bbox": [ + 105, + 305, + 506, + 319 + ], + "score": 1.0, + "content": "tractable (Rezende & Mohamed, 2015). Our method for inference generalizes these discrete-time", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 506, + 331 + ], + "score": 1.0, + "content": "transformation to continuous-time transformations, ensuring convergence of the transformations to", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 329, + 197, + 341 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 197, + 341 + ], + "score": 1.0, + "content": "the target distribution.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 295, + 506, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 352, + 505, + 430 + ], + "lines": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 506, + 365 + ], + "score": 1.0, + "content": "Density estimation overview There exist implicit and explicit density-estimation methods. Implicit", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 106, + 363, + 505, + 376 + ], + "score": 1.0, + "content": "density models such as GAN provide a flexible way to draw samples directly from unknown data", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 375, + 505, + 387 + ], + "score": 1.0, + "content": "distributions (via a deep neural network (DNN) called a generator with stochastic inputs) without", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 505, + 398 + ], + "score": 1.0, + "content": "explicitly modeling their density forms; whereas explicit models such as the pixel RNN/CNN (van den", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 505, + 409 + ], + "score": 1.0, + "content": "Oord et al., 2016) define and learn explicit forms of the unknown data distributions. This gives the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 407, + 505, + 420 + ], + "score": 1.0, + "content": "advantage that the likelihood for a test data point can be explicitly evaluated. However, the generation", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 419, + 425, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 419, + 425, + 431 + ], + "score": 1.0, + "content": "of samples is typically time-consuming due to the sequential generation nature.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 352, + 506, + 431 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 435, + 505, + 490 + ], + "lines": [ + { + "bbox": [ + 106, + 434, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 449 + ], + "score": 1.0, + "content": "Similar to Wang & Liu (2017), our CTF-based approach in Section 4 provides an alternative way", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 505, + 458 + ], + "score": 1.0, + "content": "for this problem, by simultaneously learning an explicit Gibbsian-style data distribution (estimated", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 505, + 470 + ], + "score": 1.0, + "content": "density) and a generator whose generated samples match the learned Gibbsian distribution. This", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 506, + 482 + ], + "score": 1.0, + "content": "not only gives us the advantage of explicit density modeling but also provides an efficient way to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 479, + 180, + 492 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 180, + 492 + ], + "score": 1.0, + "content": "generate samples.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28, + "bbox_fs": [ + 105, + 434, + 506, + 492 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 504, + 246, + 515 + ], + "lines": [ + { + "bbox": [ + 106, + 504, + 248, + 517 + ], + "spans": [ + { + "bbox": [ + 106, + 504, + 248, + 517 + ], + "score": 1.0, + "content": "2.2 CONTINUOUS-TIME FLOWS", + "type": "text" + } + ], + "index": 31 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 106, + 524, + 505, + 613 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 506, + 537 + ], + "score": 1.0, + "content": "We notice two potential limitations with traditional normalizing flows: i) given specified transfor-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 536, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 142, + 548 + ], + "score": 1.0, + "content": "mations", + "type": "text" + }, + { + "bbox": [ + 142, + 536, + 163, + 548 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 536, + 352, + 548 + ], + "score": 1.0, + "content": ", there is no guarantee that the distribution of", + "type": "text" + }, + { + "bbox": [ + 352, + 537, + 367, + 547 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 536, + 454, + 548 + ], + "score": 1.0, + "content": "could exactly match", + "type": "text" + }, + { + "bbox": [ + 454, + 536, + 489, + 548 + ], + "score": 0.75, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 536, + 505, + 548 + ], + "score": 1.0, + "content": "; ii)", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 548, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 106, + 548, + 259, + 559 + ], + "score": 1.0, + "content": "the randomness is only introduced in", + "type": "text" + }, + { + "bbox": [ + 259, + 549, + 270, + 558 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 548, + 504, + 559 + ], + "score": 1.0, + "content": "(from the inference network), limiting the representation", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 558, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 104, + 558, + 506, + 570 + ], + "score": 1.0, + "content": "power. We specify CTFs where the transformations are indexed by real numbers, thus they could be", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 569, + 505, + 581 + ], + "score": 1.0, + "content": "considered as consisting of an infinite number of transformations. Further, we consider stochastic", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "flows where randomness is injected in a continuous-time manner. In fact, the concept of CTFs", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 506, + 603 + ], + "score": 1.0, + "content": "(such as the Hamiltonian flow) has been introduced in Rezende & Mohamed (2015), without further", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 602, + 249, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 602, + 249, + 613 + ], + "score": 1.0, + "content": "development on efficient inference.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 35.5, + "bbox_fs": [ + 104, + 524, + 506, + 613 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 617, + 505, + 652 + ], + "lines": [ + { + "bbox": [ + 105, + 617, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 617, + 205, + 631 + ], + "score": 1.0, + "content": "We consider a flow on", + "type": "text" + }, + { + "bbox": [ + 205, + 618, + 219, + 628 + ], + "score": 0.87, + "content": "\\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 617, + 330, + 631 + ], + "score": 1.0, + "content": ", defined as the mapping†", + "type": "text" + }, + { + "bbox": [ + 330, + 618, + 420, + 629 + ], + "score": 0.91, + "content": "\\mathcal { T } : \\mathbb { R } ^ { L } \\times \\mathbb { R } \\to \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 617, + 505, + 631 + ], + "score": 1.0, + "content": "such that‡ we have", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 628, + 506, + 642 + ], + "spans": [ + { + "bbox": [ + 107, + 630, + 159, + 642 + ], + "score": 0.93, + "content": "\\boldsymbol { \\mathcal { T } } ( \\mathbf { Z } , 0 ) = \\mathbf { z }", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 628, + 177, + 642 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 178, + 630, + 295, + 641 + ], + "score": 0.92, + "content": "\\mathcal { T } ( \\mathcal { T } ( \\mathbf { Z } , t ) , s ) = \\mathcal { T } ( \\mathbf { Z } , s + t )", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 628, + 326, + 642 + ], + "score": 1.0, + "content": ", for all", + "type": "text" + }, + { + "bbox": [ + 326, + 629, + 360, + 640 + ], + "score": 0.91, + "content": "\\mathbf { Z } \\in \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 628, + 379, + 642 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 379, + 630, + 412, + 641 + ], + "score": 0.9, + "content": "s , t \\in \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 628, + 506, + 642 + ], + "score": 1.0, + "content": ". A typical example of", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 639, + 447, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 639, + 204, + 653 + ], + "score": 1.0, + "content": "this family is defined as", + "type": "text" + }, + { + "bbox": [ + 204, + 641, + 259, + 653 + ], + "score": 0.92, + "content": "\\mathcal { T } ( \\mathbf { Z } , t ) = \\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 639, + 290, + 653 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 290, + 641, + 302, + 651 + ], + "score": 0.89, + "content": "\\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 639, + 447, + 653 + ], + "score": 1.0, + "content": "is driven by a diffusion of the form:", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 41, + "bbox_fs": [ + 105, + 617, + 506, + 653 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 241, + 658, + 370, + 671 + ], + "lines": [ + { + "bbox": [ + 241, + 658, + 370, + 671 + ], + "spans": [ + { + "bbox": [ + 241, + 658, + 370, + 671 + ], + "score": 0.91, + "content": "\\mathrm { d } \\mathbf { Z } _ { t } = F ( \\mathbf { Z } _ { t } ) \\mathrm { d } t + V ( \\mathbf { Z } _ { t } ) \\mathrm { d } { \\mathcal { W } } .", + "type": "interline_equation", + "image_path": "c8ef519cf5f4b039dffc5a4a43dd8e677643835401771d015f79f2b1b33b1d71.jpg" + } + ] + } + ], + "index": 43, + "virtual_lines": [ + { + "bbox": [ + 241, + 658, + 370, + 671 + ], + "spans": [], + "index": 43 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 131, + 96 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 131, + 82, + 204, + 93 + ], + "score": 0.89, + "content": "{ \\cal F } : \\mathbb { R } ^ { L } \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 79, + 210, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 210, + 81, + 296, + 93 + ], + "score": 0.89, + "content": "V ~ : ~ \\mathbb { R } ^ { L \\times L } ~ \\to ~ \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 79, + 507, + 96 + ], + "score": 1.0, + "content": "are called the drift term and diffusion term, re-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 154, + 106 + ], + "score": 1.0, + "content": "spectively;", + "type": "text" + }, + { + "bbox": [ + 154, + 94, + 166, + 104 + ], + "score": 0.8, + "content": "\\mathcal { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 92, + 231, + 106 + ], + "score": 1.0, + "content": "is the standard", + "type": "text" + }, + { + "bbox": [ + 231, + 94, + 239, + 104 + ], + "score": 0.77, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "-dimensional Brownian motion. In the context of inference, we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 284, + 117 + ], + "score": 1.0, + "content": "seek to make the stationary distribution of", + "type": "text" + }, + { + "bbox": [ + 285, + 105, + 296, + 115 + ], + "score": 0.88, + "content": "\\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 104, + 338, + 117 + ], + "score": 1.0, + "content": "approach", + "type": "text" + }, + { + "bbox": [ + 338, + 104, + 374, + 116 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { z } | \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ". One solution for this is to set", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 114, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 107, + 114, + 240, + 128 + ], + "score": 0.9, + "content": "\\begin{array} { r } { F ( \\mathbf { Z } _ { t } ) = \\frac { 1 } { 2 } \\nabla _ { \\mathbf { z } } \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } \\stackrel { . } { = } \\mathbf { Z } _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 114, + 258, + 129 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 258, + 115, + 309, + 127 + ], + "score": 0.92, + "content": "V ( \\mathbf { Z } _ { t } ) = \\mathbf { I } _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 114, + 330, + 129 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 330, + 116, + 342, + 127 + ], + "score": 0.87, + "content": "\\mathbf { I } _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 114, + 357, + 129 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 358, + 116, + 385, + 126 + ], + "score": 0.9, + "content": "L \\times L", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 114, + 506, + 129 + ], + "score": 1.0, + "content": "identity matrix. The resulting", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 460, + 139 + ], + "score": 1.0, + "content": "diffusion is called Langevin dynamics Welling & Teh (2011). Denoting the distribution of", + "type": "text" + }, + { + "bbox": [ + 460, + 127, + 471, + 137 + ], + "score": 0.88, + "content": "\\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 126, + 483, + 139 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 483, + 128, + 493, + 138 + ], + "score": 0.84, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 126, + 506, + 139 + ], + "score": 1.0, + "content": ", it", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 465, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 241, + 150 + ], + "score": 1.0, + "content": "is well known Risken (1989) that", + "type": "text" + }, + { + "bbox": [ + 241, + 139, + 251, + 149 + ], + "score": 0.85, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 136, + 465, + 150 + ], + "score": 1.0, + "content": "is characterized by the Fokker-Planck (FP) equation:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 151, + 417, + 175 + ], + "lines": [ + { + "bbox": [ + 192, + 151, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 192, + 151, + 417, + 175 + ], + "score": 0.91, + "content": "\\frac { \\partial \\rho _ { t } } { \\partial t } = - \\nabla _ { \\mathbf { z } } \\cdot \\left( \\rho _ { t } F ( \\mathbf { Z } _ { t } ) \\right) + \\nabla _ { \\mathbf { z } } \\nabla _ { \\mathbf { z } } \\colon \\left( \\rho _ { t } V ( \\mathbf { Z } _ { t } ) V ^ { \\top } ( \\mathbf { Z } _ { t } ) \\right) ~ ,", + "type": "interline_equation", + "image_path": "b92a895db7d555542f06f389aa945d7d08ce8aa1198dc2a6b34e3c7cf0d85a8a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 192, + 151, + 417, + 175 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 178, + 449, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 176, + 446, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 133, + 193 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 178, + 186, + 190 + ], + "score": 0.88, + "content": "\\mathbf { a } \\cdot \\mathbf { b } \\triangleq \\mathbf { a } ^ { \\intercal } \\mathbf { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 176, + 232, + 193 + ], + "score": 1.0, + "content": "for vectors", + "type": "text" + }, + { + "bbox": [ + 232, + 182, + 240, + 190 + ], + "score": 0.39, + "content": "\\mathbf { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 176, + 257, + 193 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 180, + 265, + 190 + ], + "score": 0.32, + "content": "\\mathbf { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 176, + 269, + 193 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 269, + 178, + 357, + 192 + ], + "score": 0.79, + "content": "\\mathbf { A } : \\mathbf { B } \\triangleq \\operatorname { t r a c e } ( \\mathbf { A } ^ { \\intercal } \\mathbf { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 176, + 408, + 193 + ], + "score": 1.0, + "content": "for matrices", + "type": "text" + }, + { + "bbox": [ + 409, + 180, + 419, + 190 + ], + "score": 0.26, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 176, + 437, + 193 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 180, + 446, + 190 + ], + "score": 0.57, + "content": "\\mathbf { B }", + "type": "inline_equation" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "For simplicity, we consider the flow defined by the Langevin dynamics specified above, though our", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 222 + ], + "score": 1.0, + "content": "results generalize to other stochastic flows Dorogovtsev & Nishchenko (2014). In the following,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "we specify the ELBO under a CTF, which can then be readily solved by a discretized numerical", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "scheme, based on the results from Jordan et al. (1998). An approximation error bound for the scheme", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 241, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 252 + ], + "score": 1.0, + "content": "is also derived. We defer proofs of our theoretical results to the Supplementary Material (SM) for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 253, + 158, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 158, + 263 + ], + "score": 1.0, + "content": "conciseness.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 107, + 278, + 353, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 354, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 354, + 293 + ], + "score": 1.0, + "content": "3 CONTINUOUS-TIME FLOWS FOR INFERENCE", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "We first give an overview of our CTF-based method for efficient inference. We adopt the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "VAE/normalizing-flow framework with an encoder-decoder structure. An important difference", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "is that instead of feeding data to an encoder and sampling a latent representation in the output as in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "score": 1.0, + "content": "VAE, we concatenate the data with independent noise as input and directly generate output samples§.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "These output samples are then driven by the CTF to approach the true posterior distribution. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "the learning process, the implicit transformations from the CTF are sequentially distilled into the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "inference network by amortized learning, making the inference network flexible enough to represent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 380, + 490, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 490, + 392 + ], + "score": 1.0, + "content": "the true posterior distribution. In the following subsections, we specify our framework in detail.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5 + }, + { + "type": "title", + "bbox": [ + 107, + 405, + 431, + 416 + ], + "lines": [ + { + "bbox": [ + 106, + 404, + 432, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 432, + 417 + ], + "score": 1.0, + "content": "3.1 THE VARIATIONAL LOWER BOUND AND DISCRETIZED APPROXIMATION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "We first incorporate CTF into the normalizing-flow framework by writing out the corresponding", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 431, + 448 + ], + "score": 1.0, + "content": "ELBO. Note that there are two steps in the inference process. First, an initial", + "type": "text" + }, + { + "bbox": [ + 431, + 437, + 442, + 447 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "is drawn from", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 197, + 459 + ], + "score": 1.0, + "content": "the inference network", + "type": "text" + }, + { + "bbox": [ + 197, + 447, + 228, + 459 + ], + "score": 0.89, + "content": "q _ { \\phi } ( \\cdot | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 447, + 266, + 459 + ], + "score": 1.0, + "content": "; second,", + "type": "text" + }, + { + "bbox": [ + 266, + 448, + 277, + 458 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 447, + 462, + 459 + ], + "score": 1.0, + "content": "is evolved via a diffusion such as (2) for time", + "type": "text" + }, + { + "bbox": [ + 462, + 447, + 471, + 457 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "(via the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 447, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 168, + 470 + ], + "score": 1.0, + "content": "transformation", + "type": "text" + }, + { + "bbox": [ + 168, + 459, + 235, + 470 + ], + "score": 0.91, + "content": "{ \\bf Z } _ { T } = \\mathcal { T } ( { \\bf z } _ { 0 } , T ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 457, + 447, + 470 + ], + "score": 1.0, + "content": ". Consequently, the ELBO for CTF can be written as", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5 + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 471, + 487, + 500 + ], + "lines": [ + { + "bbox": [ + 111, + 471, + 487, + 500 + ], + "spans": [ + { + "bbox": [ + 111, + 471, + 487, + 500 + ], + "score": 0.92, + "content": "\\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\mathbb { E } _ { \\rho _ { T } } \\left[ \\log \\rho _ { T } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { Z } _ { T } ) + \\log \\left| \\operatorname* { d e t } \\frac { \\partial \\mathbf { Z } _ { T } } { \\partial \\mathbf { z } _ { 0 } } \\right| \\right] \\triangleq \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\mathcal { F } _ { 1 } ( \\mathbf { x } , \\mathbf { z } _ { 0 } ) \\right] \\ .", + "type": "interline_equation", + "image_path": "c1559d6ae3f5d8a3453494946f6abbdb87fe7a5c0fac8cd32ac2af541d15093a.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 111, + 471, + 487, + 480.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 111, + 480.6666666666667, + 487, + 490.33333333333337 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 111, + 490.33333333333337, + 487, + 500.00000000000006 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 165, + 515 + ], + "score": 1.0, + "content": "Note the term", + "type": "text" + }, + { + "bbox": [ + 165, + 502, + 205, + 514 + ], + "score": 0.93, + "content": "\\mathcal { F } _ { 1 } ( \\mathbf { x } , \\mathbf { z } _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 500, + 342, + 515 + ], + "score": 1.0, + "content": "is intractable to calculate, in that", + "type": "text" + }, + { + "bbox": [ + 342, + 504, + 347, + 512 + ], + "score": 0.56, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 500, + 351, + 515 + ], + "score": 1.0, + "content": ")", + "type": "text" + }, + { + "bbox": [ + 352, + 503, + 364, + 514 + ], + "score": 0.8, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 500, + 495, + 515 + ], + "score": 1.0, + "content": "does not have an explicit form;", + "type": "text" + }, + { + "bbox": [ + 495, + 504, + 501, + 512 + ], + "score": 0.63, + "content": "i i", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 500, + 506, + 515 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 507, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 159, + 533 + ], + "score": 1.0, + "content": "the Jacobian", + "type": "text" + }, + { + "bbox": [ + 159, + 514, + 179, + 529 + ], + "score": 0.92, + "content": "\\frac { \\partial \\mathbf { Z } _ { T } } { \\partial \\mathbf { z } _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 509, + 507, + 533 + ], + "score": 1.0, + "content": "is generally infeasible. In the following, we propose an approximate solution for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 467, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 278, + 540 + ], + "score": 1.0, + "content": "problem i). Learning by avoiding problem", + "type": "text" + }, + { + "bbox": [ + 278, + 528, + 285, + 538 + ], + "score": 0.64, + "content": "\\romannumeral 2", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 528, + 467, + 540 + ], + "score": 1.0, + "content": ") is presented in Section 3.2 via amortization.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 161, + 557 + ], + "score": 1.0, + "content": "For problem", + "type": "text" + }, + { + "bbox": [ + 161, + 546, + 166, + 555 + ], + "score": 0.43, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "), a reformulation of the results from Jordan et al. (1998) leads to a nice way to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 160, + 568 + ], + "score": 1.0, + "content": "approximate", + "type": "text" + }, + { + "bbox": [ + 160, + 557, + 169, + 567 + ], + "score": 0.85, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "in Lemma 1. Note in practice we adopt an implicit method which uses samples to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 566, + 378, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 378, + 579 + ], + "score": 1.0, + "content": "approximate the solution in Lemma 1 for feasibility, detailed in (6).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 209, + 593 + ], + "score": 1.0, + "content": "Lemma 1. Assume that", + "type": "text" + }, + { + "bbox": [ + 210, + 580, + 287, + 592 + ], + "score": 0.89, + "content": "\\log p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } ) \\leq C _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 578, + 418, + 593 + ], + "score": 1.0, + "content": "is infinitely differentiable, and", + "type": "text" + }, + { + "bbox": [ + 418, + 580, + 505, + 592 + ], + "score": 0.9, + "content": "\\| \\nabla _ { \\mathbf { z } } \\log p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } ) \\| \\leq", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 590, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 250, + 604 + ], + "score": 0.87, + "content": "C _ { 2 } \\left( 1 + C _ { 1 } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) ( \\forall \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 590, + 332, + 604 + ], + "score": 1.0, + "content": "for some constants", + "type": "text" + }, + { + "bbox": [ + 333, + 591, + 370, + 603 + ], + "score": 0.92, + "content": "\\left\\{ C _ { 1 } , C _ { 2 } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 590, + 393, + 604 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 393, + 592, + 432, + 601 + ], + "score": 0.89, + "content": "T = h K", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 590, + 437, + 604 + ], + "score": 1.0, + "content": "(", + "type": "text" + }, + { + "bbox": [ + 437, + 592, + 444, + 601 + ], + "score": 0.52, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 590, + 505, + 604 + ], + "score": 1.0, + "content": "is the stepsize", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 601, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 601, + 194, + 619 + ], + "score": 1.0, + "content": "in discretization and", + "type": "text" + }, + { + "bbox": [ + 194, + 604, + 204, + 614 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 601, + 351, + 619 + ], + "score": 1.0, + "content": "is the number of transformations),", + "type": "text" + }, + { + "bbox": [ + 351, + 603, + 416, + 616 + ], + "score": 0.93, + "content": "\\rho _ { 0 } \\triangleq q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 601, + 439, + 619 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 440, + 603, + 475, + 616 + ], + "score": 0.93, + "content": "\\{ \\tilde { \\rho } _ { k } \\} _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 601, + 506, + 619 + ], + "score": 1.0, + "content": "be the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 615, + 299, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 299, + 627 + ], + "score": 1.0, + "content": "solution of the functional optimization problem:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 628, + 411, + 653 + ], + "lines": [ + { + "bbox": [ + 199, + 628, + 411, + 653 + ], + "spans": [ + { + "bbox": [ + 199, + 628, + 411, + 653 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k } = \\arg \\operatorname* { m i n } _ { \\rho \\in { \\mathcal { K } } } K L \\left( \\rho \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) + \\frac { 1 } { 2 h } W _ { 2 } ^ { 2 } \\left( \\tilde { \\rho } _ { k - 1 } , \\rho \\right) ,", + "type": "interline_equation", + "image_path": "036ee732096b4b559a01fd9f8be2e56fd397ab0f6997650f124c5929577b9532.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 199, + 628, + 411, + 653 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 657, + 506, + 705 + ], + "lines": [ + { + "bbox": [ + 104, + 656, + 508, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 133, + 674 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 657, + 354, + 672 + ], + "score": 0.88, + "content": "\\begin{array} { r } { W _ { 2 } ^ { 2 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) \\triangleq \\operatorname* { i n f } _ { p \\in \\mathcal { P } ( \\mu _ { 1 } , \\mu _ { 2 } ) } \\int \\left\\| \\mathbf { x } - \\mathbf { y } \\right\\| _ { 2 } ^ { 2 } p ( \\mathrm { d } \\mathbf { x } , \\mathrm { d } \\mathbf { y } ) , \\mathrm { ~ } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 658, + 401, + 671 + ], + "score": 0.52, + "content": "W _ { 2 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 656, + 508, + 674 + ], + "score": 1.0, + "content": "is the 2nd-order Wasser-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 187, + 684 + ], + "score": 1.0, + "content": "stein distance, with", + "type": "text" + }, + { + "bbox": [ + 187, + 671, + 228, + 683 + ], + "score": 0.9, + "content": "\\mathcal { P } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 670, + 394, + 684 + ], + "score": 1.0, + "content": "being the space of joint distributions on", + "type": "text" + }, + { + "bbox": [ + 394, + 671, + 430, + 683 + ], + "score": 0.91, + "content": "\\{ \\mu _ { 1 } , \\mu _ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 670, + 434, + 684 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 434, + 671, + 443, + 681 + ], + "score": 0.71, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 670, + 507, + 684 + ], + "score": 1.0, + "content": "is the space of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 367, + 695 + ], + "score": 1.0, + "content": "probability distributions with the finite 2nd-order moment. Then", + "type": "text" + }, + { + "bbox": [ + 367, + 683, + 380, + 694 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 681, + 434, + 695 + ], + "score": 1.0, + "content": "converges to", + "type": "text" + }, + { + "bbox": [ + 435, + 684, + 447, + 694 + ], + "score": 0.64, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 691, + 462, + 705 + ], + "spans": [ + { + "bbox": [ + 107, + 693, + 134, + 703 + ], + "score": 0.87, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 691, + 155, + 705 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 155, + 693, + 226, + 704 + ], + "score": 0.91, + "content": "\\mathrm { l i m } _ { h 0 } \\tilde { \\rho } _ { K } = \\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 691, + 256, + 705 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 257, + 694, + 269, + 704 + ], + "score": 0.79, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 691, + 450, + 705 + ], + "score": 1.0, + "content": "is the solution of the FP equation (3) at time", + "type": "text" + }, + { + "bbox": [ + 450, + 694, + 457, + 703 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 691, + 462, + 705 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 711, + 507, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 709, + 506, + 724 + ], + "spans": [ + { + "bbox": [ + 117, + 709, + 506, + 724 + ], + "score": 1.0, + "content": "§Such structure can represent much more complex distributions than a parametric form, useful for following", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 506, + 733 + ], + "score": 1.0, + "content": "procedures. We argue an explicit distribution form is not as important in inference as that in density estimation.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 12, + "width": 9 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 507, + 96 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 131, + 96 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 131, + 82, + 204, + 93 + ], + "score": 0.89, + "content": "{ \\cal F } : \\mathbb { R } ^ { L } \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 79, + 210, + 96 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 210, + 81, + 296, + 93 + ], + "score": 0.89, + "content": "V ~ : ~ \\mathbb { R } ^ { L \\times L } ~ \\to ~ \\mathbb { R } ^ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 79, + 507, + 96 + ], + "score": 1.0, + "content": "are called the drift term and diffusion term, re-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 92, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 105, + 92, + 154, + 106 + ], + "score": 1.0, + "content": "spectively;", + "type": "text" + }, + { + "bbox": [ + 154, + 94, + 166, + 104 + ], + "score": 0.8, + "content": "\\mathcal { W }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 92, + 231, + 106 + ], + "score": 1.0, + "content": "is the standard", + "type": "text" + }, + { + "bbox": [ + 231, + 94, + 239, + 104 + ], + "score": 0.77, + "content": "L", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 92, + 506, + 106 + ], + "score": 1.0, + "content": "-dimensional Brownian motion. In the context of inference, we", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 117 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 284, + 117 + ], + "score": 1.0, + "content": "seek to make the stationary distribution of", + "type": "text" + }, + { + "bbox": [ + 285, + 105, + 296, + 115 + ], + "score": 0.88, + "content": "\\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 104, + 338, + 117 + ], + "score": 1.0, + "content": "approach", + "type": "text" + }, + { + "bbox": [ + 338, + 104, + 374, + 116 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { z } | \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 104, + 506, + 117 + ], + "score": 1.0, + "content": ". One solution for this is to set", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 114, + 506, + 129 + ], + "spans": [ + { + "bbox": [ + 107, + 114, + 240, + 128 + ], + "score": 0.9, + "content": "\\begin{array} { r } { F ( \\mathbf { Z } _ { t } ) = \\frac { 1 } { 2 } \\nabla _ { \\mathbf { z } } \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } \\stackrel { . } { = } \\mathbf { Z } _ { t } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 114, + 258, + 129 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 258, + 115, + 309, + 127 + ], + "score": 0.92, + "content": "V ( \\mathbf { Z } _ { t } ) = \\mathbf { I } _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 309, + 114, + 330, + 129 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 330, + 116, + 342, + 127 + ], + "score": 0.87, + "content": "\\mathbf { I } _ { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 114, + 357, + 129 + ], + "score": 1.0, + "content": "the", + "type": "text" + }, + { + "bbox": [ + 358, + 116, + 385, + 126 + ], + "score": 0.9, + "content": "L \\times L", + "type": "inline_equation" + }, + { + "bbox": [ + 385, + 114, + 506, + 129 + ], + "score": 1.0, + "content": "identity matrix. The resulting", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 460, + 139 + ], + "score": 1.0, + "content": "diffusion is called Langevin dynamics Welling & Teh (2011). Denoting the distribution of", + "type": "text" + }, + { + "bbox": [ + 460, + 127, + 471, + 137 + ], + "score": 0.88, + "content": "\\mathbf { Z } _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 472, + 126, + 483, + 139 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 483, + 128, + 493, + 138 + ], + "score": 0.84, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 126, + 506, + 139 + ], + "score": 1.0, + "content": ", it", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 465, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 241, + 150 + ], + "score": 1.0, + "content": "is well known Risken (1989) that", + "type": "text" + }, + { + "bbox": [ + 241, + 139, + 251, + 149 + ], + "score": 0.85, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 136, + 465, + 150 + ], + "score": 1.0, + "content": "is characterized by the Fokker-Planck (FP) equation:", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 79, + 507, + 150 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 192, + 151, + 417, + 175 + ], + "lines": [ + { + "bbox": [ + 192, + 151, + 417, + 175 + ], + "spans": [ + { + "bbox": [ + 192, + 151, + 417, + 175 + ], + "score": 0.91, + "content": "\\frac { \\partial \\rho _ { t } } { \\partial t } = - \\nabla _ { \\mathbf { z } } \\cdot \\left( \\rho _ { t } F ( \\mathbf { Z } _ { t } ) \\right) + \\nabla _ { \\mathbf { z } } \\nabla _ { \\mathbf { z } } \\colon \\left( \\rho _ { t } V ( \\mathbf { Z } _ { t } ) V ^ { \\top } ( \\mathbf { Z } _ { t } ) \\right) ~ ,", + "type": "interline_equation", + "image_path": "b92a895db7d555542f06f389aa945d7d08ce8aa1198dc2a6b34e3c7cf0d85a8a.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 192, + 151, + 417, + 175 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 178, + 449, + 191 + ], + "lines": [ + { + "bbox": [ + 105, + 176, + 446, + 193 + ], + "spans": [ + { + "bbox": [ + 105, + 176, + 133, + 193 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 178, + 186, + 190 + ], + "score": 0.88, + "content": "\\mathbf { a } \\cdot \\mathbf { b } \\triangleq \\mathbf { a } ^ { \\intercal } \\mathbf { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 176, + 232, + 193 + ], + "score": 1.0, + "content": "for vectors", + "type": "text" + }, + { + "bbox": [ + 232, + 182, + 240, + 190 + ], + "score": 0.39, + "content": "\\mathbf { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 176, + 257, + 193 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 257, + 180, + 265, + 190 + ], + "score": 0.32, + "content": "\\mathbf { b }", + "type": "inline_equation" + }, + { + "bbox": [ + 265, + 176, + 269, + 193 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 269, + 178, + 357, + 192 + ], + "score": 0.79, + "content": "\\mathbf { A } : \\mathbf { B } \\triangleq \\operatorname { t r a c e } ( \\mathbf { A } ^ { \\intercal } \\mathbf { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 176, + 408, + 193 + ], + "score": 1.0, + "content": "for matrices", + "type": "text" + }, + { + "bbox": [ + 409, + 180, + 419, + 190 + ], + "score": 0.26, + "content": "\\mathbf { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 419, + 176, + 437, + 193 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 437, + 180, + 446, + 190 + ], + "score": 0.57, + "content": "\\mathbf { B }", + "type": "inline_equation" + } + ], + "index": 7 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 176, + 446, + 193 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 196, + 505, + 263 + ], + "lines": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 505, + 209 + ], + "score": 1.0, + "content": "For simplicity, we consider the flow defined by the Langevin dynamics specified above, though our", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 206, + 506, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 206, + 506, + 222 + ], + "score": 1.0, + "content": "results generalize to other stochastic flows Dorogovtsev & Nishchenko (2014). In the following,", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "spans": [ + { + "bbox": [ + 105, + 218, + 506, + 231 + ], + "score": 1.0, + "content": "we specify the ELBO under a CTF, which can then be readily solved by a discretized numerical", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 230, + 505, + 243 + ], + "score": 1.0, + "content": "scheme, based on the results from Jordan et al. (1998). An approximation error bound for the scheme", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 241, + 505, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 505, + 252 + ], + "score": 1.0, + "content": "is also derived. We defer proofs of our theoretical results to the Supplementary Material (SM) for", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 253, + 158, + 263 + ], + "spans": [ + { + "bbox": [ + 106, + 253, + 158, + 263 + ], + "score": 1.0, + "content": "conciseness.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 196, + 506, + 263 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 278, + 353, + 291 + ], + "lines": [ + { + "bbox": [ + 105, + 277, + 354, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 354, + 293 + ], + "score": 1.0, + "content": "3 CONTINUOUS-TIME FLOWS FOR INFERENCE", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 106, + 302, + 505, + 392 + ], + "lines": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 505, + 315 + ], + "score": 1.0, + "content": "We first give an overview of our CTF-based method for efficient inference. We adopt the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 505, + 326 + ], + "score": 1.0, + "content": "VAE/normalizing-flow framework with an encoder-decoder structure. An important difference", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 505, + 338 + ], + "score": 1.0, + "content": "is that instead of feeding data to an encoder and sampling a latent representation in the output as in", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 335, + 507, + 348 + ], + "score": 1.0, + "content": "VAE, we concatenate the data with independent noise as input and directly generate output samples§.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "spans": [ + { + "bbox": [ + 105, + 347, + 506, + 360 + ], + "score": 1.0, + "content": "These output samples are then driven by the CTF to approach the true posterior distribution. In", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 358, + 505, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 358, + 505, + 370 + ], + "score": 1.0, + "content": "the learning process, the implicit transformations from the CTF are sequentially distilled into the", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 369, + 506, + 381 + ], + "score": 1.0, + "content": "inference network by amortized learning, making the inference network flexible enough to represent", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 380, + 490, + 392 + ], + "spans": [ + { + "bbox": [ + 106, + 380, + 490, + 392 + ], + "score": 1.0, + "content": "the true posterior distribution. In the following subsections, we specify our framework in detail.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 303, + 507, + 392 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 405, + 431, + 416 + ], + "lines": [ + { + "bbox": [ + 106, + 404, + 432, + 417 + ], + "spans": [ + { + "bbox": [ + 106, + 404, + 432, + 417 + ], + "score": 1.0, + "content": "3.1 THE VARIATIONAL LOWER BOUND AND DISCRETIZED APPROXIMATION", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "text", + "bbox": [ + 107, + 424, + 505, + 469 + ], + "lines": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "We first incorporate CTF into the normalizing-flow framework by writing out the corresponding", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 435, + 506, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 431, + 448 + ], + "score": 1.0, + "content": "ELBO. Note that there are two steps in the inference process. First, an initial", + "type": "text" + }, + { + "bbox": [ + 431, + 437, + 442, + 447 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 443, + 435, + 506, + 448 + ], + "score": 1.0, + "content": "is drawn from", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 447, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 197, + 459 + ], + "score": 1.0, + "content": "the inference network", + "type": "text" + }, + { + "bbox": [ + 197, + 447, + 228, + 459 + ], + "score": 0.89, + "content": "q _ { \\phi } ( \\cdot | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 447, + 266, + 459 + ], + "score": 1.0, + "content": "; second,", + "type": "text" + }, + { + "bbox": [ + 266, + 448, + 277, + 458 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 447, + 462, + 459 + ], + "score": 1.0, + "content": "is evolved via a diffusion such as (2) for time", + "type": "text" + }, + { + "bbox": [ + 462, + 447, + 471, + 457 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 447, + 505, + 459 + ], + "score": 1.0, + "content": "(via the", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 457, + 447, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 168, + 470 + ], + "score": 1.0, + "content": "transformation", + "type": "text" + }, + { + "bbox": [ + 168, + 459, + 235, + 470 + ], + "score": 0.91, + "content": "{ \\bf Z } _ { T } = \\mathcal { T } ( { \\bf z } _ { 0 } , T ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 457, + 447, + 470 + ], + "score": 1.0, + "content": ". Consequently, the ELBO for CTF can be written as", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 25.5, + "bbox_fs": [ + 105, + 425, + 506, + 470 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 111, + 471, + 487, + 500 + ], + "lines": [ + { + "bbox": [ + 111, + 471, + 487, + 500 + ], + "spans": [ + { + "bbox": [ + 111, + 471, + 487, + 500 + ], + "score": 0.92, + "content": "\\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\mathbb { E } _ { \\rho _ { T } } \\left[ \\log \\rho _ { T } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { Z } _ { T } ) + \\log \\left| \\operatorname* { d e t } \\frac { \\partial \\mathbf { Z } _ { T } } { \\partial \\mathbf { z } _ { 0 } } \\right| \\right] \\triangleq \\mathbb { E } _ { q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\left[ \\mathcal { F } _ { 1 } ( \\mathbf { x } , \\mathbf { z } _ { 0 } ) \\right] \\ .", + "type": "interline_equation", + "image_path": "c1559d6ae3f5d8a3453494946f6abbdb87fe7a5c0fac8cd32ac2af541d15093a.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 111, + 471, + 487, + 480.6666666666667 + ], + "spans": [], + "index": 28 + }, + { + "bbox": [ + 111, + 480.6666666666667, + 487, + 490.33333333333337 + ], + "spans": [], + "index": 29 + }, + { + "bbox": [ + 111, + 490.33333333333337, + 487, + 500.00000000000006 + ], + "spans": [], + "index": 30 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 502, + 505, + 540 + ], + "lines": [ + { + "bbox": [ + 105, + 500, + 506, + 515 + ], + "spans": [ + { + "bbox": [ + 105, + 500, + 165, + 515 + ], + "score": 1.0, + "content": "Note the term", + "type": "text" + }, + { + "bbox": [ + 165, + 502, + 205, + 514 + ], + "score": 0.93, + "content": "\\mathcal { F } _ { 1 } ( \\mathbf { x } , \\mathbf { z } _ { 0 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 206, + 500, + 342, + 515 + ], + "score": 1.0, + "content": "is intractable to calculate, in that", + "type": "text" + }, + { + "bbox": [ + 342, + 504, + 347, + 512 + ], + "score": 0.56, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 500, + 351, + 515 + ], + "score": 1.0, + "content": ")", + "type": "text" + }, + { + "bbox": [ + 352, + 503, + 364, + 514 + ], + "score": 0.8, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 500, + 495, + 515 + ], + "score": 1.0, + "content": "does not have an explicit form;", + "type": "text" + }, + { + "bbox": [ + 495, + 504, + 501, + 512 + ], + "score": 0.63, + "content": "i i", + "type": "inline_equation" + }, + { + "bbox": [ + 501, + 500, + 506, + 515 + ], + "score": 1.0, + "content": ")", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 507, + 533 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 159, + 533 + ], + "score": 1.0, + "content": "the Jacobian", + "type": "text" + }, + { + "bbox": [ + 159, + 514, + 179, + 529 + ], + "score": 0.92, + "content": "\\frac { \\partial \\mathbf { Z } _ { T } } { \\partial \\mathbf { z } _ { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 180, + 509, + 507, + 533 + ], + "score": 1.0, + "content": "is generally infeasible. In the following, we propose an approximate solution for", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 528, + 467, + 540 + ], + "spans": [ + { + "bbox": [ + 105, + 528, + 278, + 540 + ], + "score": 1.0, + "content": "problem i). Learning by avoiding problem", + "type": "text" + }, + { + "bbox": [ + 278, + 528, + 285, + 538 + ], + "score": 0.64, + "content": "\\romannumeral 2", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 528, + 467, + 540 + ], + "score": 1.0, + "content": ") is presented in Section 3.2 via amortization.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 105, + 500, + 507, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 544, + 505, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 543, + 505, + 557 + ], + "spans": [ + { + "bbox": [ + 105, + 543, + 161, + 557 + ], + "score": 1.0, + "content": "For problem", + "type": "text" + }, + { + "bbox": [ + 161, + 546, + 166, + 555 + ], + "score": 0.43, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 166, + 543, + 505, + 557 + ], + "score": 1.0, + "content": "), a reformulation of the results from Jordan et al. (1998) leads to a nice way to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 556, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 556, + 160, + 568 + ], + "score": 1.0, + "content": "approximate", + "type": "text" + }, + { + "bbox": [ + 160, + 557, + 169, + 567 + ], + "score": 0.85, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 556, + 505, + 568 + ], + "score": 1.0, + "content": "in Lemma 1. Note in practice we adopt an implicit method which uses samples to", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 566, + 378, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 566, + 378, + 579 + ], + "score": 1.0, + "content": "approximate the solution in Lemma 1 for feasibility, detailed in (6).", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 543, + 505, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 579, + 505, + 627 + ], + "lines": [ + { + "bbox": [ + 105, + 578, + 505, + 593 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 209, + 593 + ], + "score": 1.0, + "content": "Lemma 1. Assume that", + "type": "text" + }, + { + "bbox": [ + 210, + 580, + 287, + 592 + ], + "score": 0.89, + "content": "\\log p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } ) \\leq C _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 288, + 578, + 418, + 593 + ], + "score": 1.0, + "content": "is infinitely differentiable, and", + "type": "text" + }, + { + "bbox": [ + 418, + 580, + 505, + 592 + ], + "score": 0.9, + "content": "\\| \\nabla _ { \\mathbf { z } } \\log p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } ) \\| \\leq", + "type": "inline_equation" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 590, + 505, + 604 + ], + "spans": [ + { + "bbox": [ + 106, + 591, + 250, + 604 + ], + "score": 0.87, + "content": "C _ { 2 } \\left( 1 + C _ { 1 } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) ( \\forall \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 251, + 590, + 332, + 604 + ], + "score": 1.0, + "content": "for some constants", + "type": "text" + }, + { + "bbox": [ + 333, + 591, + 370, + 603 + ], + "score": 0.92, + "content": "\\left\\{ C _ { 1 } , C _ { 2 } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 590, + 393, + 604 + ], + "score": 1.0, + "content": ". Let", + "type": "text" + }, + { + "bbox": [ + 393, + 592, + 432, + 601 + ], + "score": 0.89, + "content": "T = h K", + "type": "inline_equation" + }, + { + "bbox": [ + 432, + 590, + 437, + 604 + ], + "score": 1.0, + "content": "(", + "type": "text" + }, + { + "bbox": [ + 437, + 592, + 444, + 601 + ], + "score": 0.52, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 590, + 505, + 604 + ], + "score": 1.0, + "content": "is the stepsize", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 104, + 601, + 506, + 619 + ], + "spans": [ + { + "bbox": [ + 104, + 601, + 194, + 619 + ], + "score": 1.0, + "content": "in discretization and", + "type": "text" + }, + { + "bbox": [ + 194, + 604, + 204, + 614 + ], + "score": 0.76, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 205, + 601, + 351, + 619 + ], + "score": 1.0, + "content": "is the number of transformations),", + "type": "text" + }, + { + "bbox": [ + 351, + 603, + 416, + 616 + ], + "score": 0.93, + "content": "\\rho _ { 0 } \\triangleq q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 417, + 601, + 439, + 619 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 440, + 603, + 475, + 616 + ], + "score": 0.93, + "content": "\\{ \\tilde { \\rho } _ { k } \\} _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 601, + 506, + 619 + ], + "score": 1.0, + "content": "be the", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 615, + 299, + 627 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 299, + 627 + ], + "score": 1.0, + "content": "solution of the functional optimization problem:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38.5, + "bbox_fs": [ + 104, + 578, + 506, + 627 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 628, + 411, + 653 + ], + "lines": [ + { + "bbox": [ + 199, + 628, + 411, + 653 + ], + "spans": [ + { + "bbox": [ + 199, + 628, + 411, + 653 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k } = \\arg \\operatorname* { m i n } _ { \\rho \\in { \\mathcal { K } } } K L \\left( \\rho \\| p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\right) + \\frac { 1 } { 2 h } W _ { 2 } ^ { 2 } \\left( \\tilde { \\rho } _ { k - 1 } , \\rho \\right) ,", + "type": "interline_equation", + "image_path": "036ee732096b4b559a01fd9f8be2e56fd397ab0f6997650f124c5929577b9532.jpg" + } + ] + } + ], + "index": 41, + "virtual_lines": [ + { + "bbox": [ + 199, + 628, + 411, + 653 + ], + "spans": [], + "index": 41 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 657, + 506, + 705 + ], + "lines": [ + { + "bbox": [ + 104, + 656, + 508, + 674 + ], + "spans": [ + { + "bbox": [ + 104, + 656, + 133, + 674 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 657, + 354, + 672 + ], + "score": 0.88, + "content": "\\begin{array} { r } { W _ { 2 } ^ { 2 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) \\triangleq \\operatorname* { i n f } _ { p \\in \\mathcal { P } ( \\mu _ { 1 } , \\mu _ { 2 } ) } \\int \\left\\| \\mathbf { x } - \\mathbf { y } \\right\\| _ { 2 } ^ { 2 } p ( \\mathrm { d } \\mathbf { x } , \\mathrm { d } \\mathbf { y } ) , \\mathrm { ~ } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 658, + 401, + 671 + ], + "score": 0.52, + "content": "W _ { 2 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 402, + 656, + 508, + 674 + ], + "score": 1.0, + "content": "is the 2nd-order Wasser-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 670, + 507, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 670, + 187, + 684 + ], + "score": 1.0, + "content": "stein distance, with", + "type": "text" + }, + { + "bbox": [ + 187, + 671, + 228, + 683 + ], + "score": 0.9, + "content": "\\mathcal { P } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 670, + 394, + 684 + ], + "score": 1.0, + "content": "being the space of joint distributions on", + "type": "text" + }, + { + "bbox": [ + 394, + 671, + 430, + 683 + ], + "score": 0.91, + "content": "\\{ \\mu _ { 1 } , \\mu _ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 670, + 434, + 684 + ], + "score": 1.0, + "content": ".", + "type": "text" + }, + { + "bbox": [ + 434, + 671, + 443, + 681 + ], + "score": 0.71, + "content": "\\kappa", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 670, + 507, + 684 + ], + "score": 1.0, + "content": "is the space of", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 681, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 367, + 695 + ], + "score": 1.0, + "content": "probability distributions with the finite 2nd-order moment. Then", + "type": "text" + }, + { + "bbox": [ + 367, + 683, + 380, + 694 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 681, + 434, + 695 + ], + "score": 1.0, + "content": "converges to", + "type": "text" + }, + { + "bbox": [ + 435, + 684, + 447, + 694 + ], + "score": 0.64, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 447, + 681, + 506, + 695 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 107, + 691, + 462, + 705 + ], + "spans": [ + { + "bbox": [ + 107, + 693, + 134, + 703 + ], + "score": 0.87, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 134, + 691, + 155, + 705 + ], + "score": 1.0, + "content": ", i.e.,", + "type": "text" + }, + { + "bbox": [ + 155, + 693, + 226, + 704 + ], + "score": 0.91, + "content": "\\mathrm { l i m } _ { h 0 } \\tilde { \\rho } _ { K } = \\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 226, + 691, + 256, + 705 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 257, + 694, + 269, + 704 + ], + "score": 0.79, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 269, + 691, + 450, + 705 + ], + "score": 1.0, + "content": "is the solution of the FP equation (3) at time", + "type": "text" + }, + { + "bbox": [ + 450, + 694, + 457, + 703 + ], + "score": 0.82, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 691, + 462, + 705 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 656, + 508, + 705 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 313, + 95 + ], + "score": 1.0, + "content": "Lemma 1 reveals an interesting way to compute", + "type": "text" + }, + { + "bbox": [ + 313, + 85, + 326, + 94 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "via a sequence of functional optimization", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "problems. By comparing it with the objective of the traditional normalizing flow, which minimizes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 223, + 118 + ], + "score": 1.0, + "content": "the KL-divergence between", + "type": "text" + }, + { + "bbox": [ + 224, + 106, + 238, + 116 + ], + "score": 0.86, + "content": "\\rho _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 104, + 258, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 258, + 105, + 292, + 116 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 104, + 506, + 118 + ], + "score": 1.0, + "content": ", at each sub-optimization-problem in Lemma 1, it", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 260, + 130 + ], + "score": 1.0, + "content": "minimizes the KL-divergence between", + "type": "text" + }, + { + "bbox": [ + 261, + 117, + 272, + 127 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 114, + 289, + 130 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 115, + 324, + 127 + ], + "score": 0.93, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 114, + 506, + 130 + ], + "score": 1.0, + "content": ", plus a regularization term as the Wasserstein", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 178, + 139 + ], + "score": 1.0, + "content": "distance between", + "type": "text" + }, + { + "bbox": [ + 178, + 127, + 199, + 138 + ], + "score": 0.91, + "content": "\\tilde { \\rho } _ { k - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 126, + 217, + 139 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 218, + 127, + 229, + 138 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 126, + 506, + 139 + ], + "score": 1.0, + "content": ". The extra Wasserstein-distance term arises naturally due to the fact", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "that the Langevin diffusion can be explained as a gradient flow whose geometry is equipped with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "score": 1.0, + "content": "the Wasserstein distance (Otto, 1998). From another point of view, it is known that the Wasserstein", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "distance is a better metric for probability distributions than the KL-divergence, especially in the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "score": 1.0, + "content": "case of non-overlapping domains (Arjovsky & Bottou, 2017; Arjovsky et al., 2017). By using", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "the Wasserstein term as a regularizer, the CTF alleviates the issue in non-overlapping domains", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "by introducing the Brownian-motion (noise) term in the evolution (2). This relates to the idea in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "(Arjovsky & Bottou, 2017), in which noise is added in parameter updates to alleviate the intrinsic", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 213, + 265, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 265, + 227 + ], + "score": 1.0, + "content": "drawback of the KL-divergence metric.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "The optimization problem in Lemma 1 is difficult to deal with directly. In practice, we instead", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 502, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 491, + 256 + ], + "score": 1.0, + "content": "approximate the discretization in an equivalent way by simulation from the CTF. Starting from", + "type": "text" + }, + { + "bbox": [ + 492, + 244, + 502, + 253 + ], + "score": 0.82, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 118, + 264 + ], + "score": 0.73, + "content": "\\mathbf { z } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 252, + 121, + 266 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 121, + 253, + 202, + 264 + ], + "score": 0.82, + "content": "( k = 0 , \\cdots , K - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 252, + 317, + 266 + ], + "score": 1.0, + "content": "is fed into a transformation", + "type": "text" + }, + { + "bbox": [ + 317, + 253, + 329, + 264 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 252, + 453, + 266 + ], + "score": 1.0, + "content": "(specified below), resulting in", + "type": "text" + }, + { + "bbox": [ + 454, + 254, + 475, + 264 + ], + "score": 0.9, + "content": "{ \\mathbf z } _ { k + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "whose", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 507, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 214, + 277 + ], + "score": 1.0, + "content": "distribution coincides with", + "type": "text" + }, + { + "bbox": [ + 215, + 264, + 236, + 276 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 263, + 507, + 277 + ], + "score": 1.0, + "content": "in Lemma 1. The discretization procedure is illustrated in Figure 1.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 251, + 287 + ], + "score": 1.0, + "content": "We must specify the transformations", + "type": "text" + }, + { + "bbox": [ + 252, + 275, + 263, + 286 + ], + "score": 0.87, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 275, + 303, + 287 + ], + "score": 1.0, + "content": ". For each", + "type": "text" + }, + { + "bbox": [ + 304, + 275, + 310, + 285 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 275, + 326, + 287 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 326, + 275, + 355, + 285 + ], + "score": 0.91, + "content": "t = h k", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "; we can conclude from Lemma 1 that", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 107, + 286, + 172, + 298 + ], + "score": 0.9, + "content": "\\operatorname* { l i m } _ { h \\to 0 } \\tilde { \\rho } _ { k } = \\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 286, + 245, + 299 + ], + "score": 1.0, + "content": ". From FP theory,", + "type": "text" + }, + { + "bbox": [ + 246, + 288, + 256, + 298 + ], + "score": 0.85, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "is obtained by solving the diffusion (2) with initial condition", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 145, + 308 + ], + "score": 0.88, + "content": "{ \\bf Z } _ { 0 } = { \\bf z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 297, + 361, + 309 + ], + "score": 1.0, + "content": ". It is thus reasonable to specify the transformation", + "type": "text" + }, + { + "bbox": [ + 361, + 297, + 373, + 308 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 297, + 402, + 309 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 402, + 297, + 409, + 307 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "-th step of a numerical", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 308, + 412, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 276, + 321 + ], + "score": 1.0, + "content": "integrator for (2). Specifically, we specify", + "type": "text" + }, + { + "bbox": [ + 276, + 308, + 287, + 319 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 308, + 412, + 321 + ], + "score": 1.0, + "content": "as a stochastic transformation:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5 + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 333, + 411, + 348 + ], + "lines": [ + { + "bbox": [ + 199, + 333, + 411, + 348 + ], + "spans": [ + { + "bbox": [ + 199, + 333, + 411, + 348 + ], + "score": 0.88, + "content": "{ \\bf z } _ { k } = { \\mathcal T } _ { k } ( { \\bf z } _ { k - 1 } ) \\triangleq { \\bf z } _ { k - 1 } + { \\cal F } ( { \\bf z } _ { k - 1 } ) h + V ( { \\bf z } _ { k - 1 } ) \\zeta _ { k } ,", + "type": "interline_equation", + "image_path": "197d4f4ace878ed60868e893eae9ffb5758c77de999f3c246be7af44f22053e1.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 199, + 333, + 411, + 348 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 133, + 374 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 362, + 203, + 375 + ], + "score": 0.92, + "content": "\\mathsf { \\Sigma } \\xi _ { k } \\sim \\mathcal { N } ( \\mathbf { 0 } , h \\mathbf { I } _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "is drawn from an isotropic normal. Note the transformation defined here", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 291, + 386 + ], + "score": 1.0, + "content": "is stochastic, thus we only get samples from", + "type": "text" + }, + { + "bbox": [ + 291, + 374, + 305, + 385 + ], + "score": 0.89, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 374, + 480, + 386 + ], + "score": 1.0, + "content": "at the end. A natural way to approximate", + "type": "text" + }, + { + "bbox": [ + 480, + 374, + 494, + 385 + ], + "score": 0.89, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 384, + 503, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 291, + 401 + ], + "score": 1.0, + "content": "to use the empirical sample distribution, i.e.,", + "type": "text" + }, + { + "bbox": [ + 292, + 384, + 398, + 399 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\tilde { \\rho } _ { K } \\approx \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\delta _ { { \\bf z } _ { k } } \\triangleq \\bar { \\rho } _ { T } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 384, + 420, + 401 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 420, + 387, + 431, + 398 + ], + "score": 0.87, + "content": "\\delta _ { \\mathbf { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 384, + 496, + 401 + ], + "score": 1.0, + "content": "a point mass at", + "type": "text" + }, + { + "bbox": [ + 497, + 389, + 503, + 396 + ], + "score": 0.62, + "content": "\\mathbf { z }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 397, + 409, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 156, + 410 + ], + "score": 1.0, + "content": "Afterwards,", + "type": "text" + }, + { + "bbox": [ + 156, + 398, + 170, + 409 + ], + "score": 0.89, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 397, + 193, + 410 + ], + "score": 1.0, + "content": "(thus", + "type": "text" + }, + { + "bbox": [ + 194, + 398, + 206, + 409 + ], + "score": 0.84, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 397, + 356, + 410 + ], + "score": 1.0, + "content": ") will be used to approximate the true", + "type": "text" + }, + { + "bbox": [ + 357, + 400, + 369, + 409 + ], + "score": 0.8, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 397, + 409, + 410 + ], + "score": 1.0, + "content": "from (3).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5 + }, + { + "type": "image", + "bbox": [ + 127, + 420, + 485, + 530 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 420, + 485, + 530 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 420, + 485, + 530 + ], + "spans": [ + { + "bbox": [ + 127, + 420, + 485, + 530 + ], + "score": 0.975, + "type": "image", + "image_path": "bcb7adee8eb928455458752caa91f82356f295994b68b570ae3c02cde0d35129.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 127, + 420, + 485, + 456.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 127, + 456.6666666666667, + 485, + 493.33333333333337 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 127, + 493.33333333333337, + 485, + 530.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 538, + 504, + 562 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 538, + 504, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 449, + 552 + ], + "score": 1.0, + "content": "Figure 1: Discretized approximation (right) of a continuous-time flow (left). Densities", + "type": "text" + }, + { + "bbox": [ + 450, + 539, + 470, + 551 + ], + "score": 0.92, + "content": "\\{ \\tilde { \\rho } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 538, + 483, + 552 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 483, + 539, + 504, + 551 + ], + "score": 0.9, + "content": "\\left\\{ { \\bf z } _ { k } \\right\\}", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 549, + 469, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 215, + 562 + ], + "score": 1.0, + "content": "evolve via transformations", + "type": "text" + }, + { + "bbox": [ + 215, + 550, + 236, + 562 + ], + "score": 0.91, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 549, + 260, + 562 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 260, + 550, + 302, + 561 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k } \\to \\rho _ { h k }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 549, + 327, + 562 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 327, + 550, + 354, + 560 + ], + "score": 0.91, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 549, + 390, + 562 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 390, + 550, + 397, + 560 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 549, + 469, + 562 + ], + "score": 1.0, + "content": "due to Lemma 1.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + } + ], + "index": 28.25 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 506, + 728 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "Note that we use the simple sample averaging for the convenience of analysis, and the approximation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 120, + 605 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 120, + 594, + 133, + 604 + ], + "score": 0.87, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "is not necessarily optimal. Better approximation can be obtained by assigning more weights to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "the more recent samples. However, this leads to more challenges in theoretical analysis, an interesting", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 370, + 628 + ], + "score": 1.0, + "content": "future direction to pursue. In the following, we study how well", + "type": "text" + }, + { + "bbox": [ + 370, + 615, + 383, + 626 + ], + "score": 0.87, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 614, + 442, + 628 + ], + "score": 1.0, + "content": "approximates", + "type": "text" + }, + { + "bbox": [ + 442, + 616, + 455, + 626 + ], + "score": 0.84, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 614, + 506, + 628 + ], + "score": 1.0, + "content": ". Following", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "score": 1.0, + "content": "literature on numerical approximation for Itô diffusions (Vollmer et al., 2016; Chen et al., 2015), we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 253, + 649 + ], + "score": 1.0, + "content": "consider a 1-Lipschitz test function", + "type": "text" + }, + { + "bbox": [ + 254, + 636, + 307, + 648 + ], + "score": 0.93, + "content": "\\psi : \\mathbb { R } ^ { L } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 636, + 506, + 649 + ], + "score": 1.0, + "content": ", and use the mean square error (MSE) bound to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 103, + 648, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 103, + 649, + 209, + 664 + ], + "score": 1.0, + "content": "measure the closeness of", + "type": "text" + }, + { + "bbox": [ + 209, + 651, + 222, + 662 + ], + "score": 0.88, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 649, + 240, + 664 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 651, + 253, + 662 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 649, + 303, + 664 + ], + "score": 1.0, + "content": ", defined as:", + "type": "text" + }, + { + "bbox": [ + 303, + 648, + 502, + 663 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbf { M S E } ( \\bar { \\rho } _ { T } , \\rho _ { T } ; \\psi ) \\triangleq \\mathbb { E } \\left( \\int \\psi ( \\mathbf { z } ) ( \\tilde { \\rho } _ { T } - \\rho _ { T } ) ( \\mathbf { z } ) \\mathrm { d } \\mathbf { z } \\right) ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 649, + 506, + 664 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 413, + 674 + ], + "score": 1.0, + "content": "where the expectation is taken over all the randomness in the construction of", + "type": "text" + }, + { + "bbox": [ + 414, + 663, + 426, + 673 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 661, + 506, + 674 + ], + "score": 1.0, + "content": ". Note that our goal", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 672, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 684 + ], + "score": 1.0, + "content": "is related but different from the standard setup as in Vollmer et al. (2016); Chen et al. (2015), which", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 683, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 683, + 200, + 695 + ], + "score": 1.0, + "content": "studies the closeness of", + "type": "text" + }, + { + "bbox": [ + 200, + 684, + 213, + 695 + ], + "score": 0.88, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 683, + 224, + 695 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 224, + 683, + 259, + 695 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 683, + 506, + 695 + ], + "score": 1.0, + "content": ". We need to adopt the assumptions from Vollmer et al. (2016);", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 693, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 506, + 707 + ], + "score": 1.0, + "content": "Chen et al. (2015), which are described in the Supplementary Material (SM). The assumptions are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 705, + 506, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 506, + 717 + ], + "score": 1.0, + "content": "somewhat involved but essentially require coefficients of the diffusion (2) to be well-behaved. We", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 716, + 506, + 729 + ], + "spans": [ + { + "bbox": [ + 105, + 716, + 390, + 729 + ], + "score": 1.0, + "content": "derive the following bound for the MSE of the sampled approximation,", + "type": "text" + }, + { + "bbox": [ + 391, + 717, + 403, + 728 + ], + "score": 0.86, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 716, + 506, + 729 + ], + "score": 1.0, + "content": ", and the true distribution.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 225 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 313, + 95 + ], + "score": 1.0, + "content": "Lemma 1 reveals an interesting way to compute", + "type": "text" + }, + { + "bbox": [ + 313, + 85, + 326, + 94 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "via a sequence of functional optimization", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 506, + 107 + ], + "score": 1.0, + "content": "problems. By comparing it with the objective of the traditional normalizing flow, which minimizes", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 104, + 223, + 118 + ], + "score": 1.0, + "content": "the KL-divergence between", + "type": "text" + }, + { + "bbox": [ + 224, + 106, + 238, + 116 + ], + "score": 0.86, + "content": "\\rho _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 238, + 104, + 258, + 118 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 258, + 105, + 292, + 116 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 104, + 506, + 118 + ], + "score": 1.0, + "content": ", at each sub-optimization-problem in Lemma 1, it", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 114, + 506, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 114, + 260, + 130 + ], + "score": 1.0, + "content": "minimizes the KL-divergence between", + "type": "text" + }, + { + "bbox": [ + 261, + 117, + 272, + 127 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 114, + 289, + 130 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 290, + 115, + 324, + 127 + ], + "score": 0.93, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 114, + 506, + 130 + ], + "score": 1.0, + "content": ", plus a regularization term as the Wasserstein", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 106, + 126, + 178, + 139 + ], + "score": 1.0, + "content": "distance between", + "type": "text" + }, + { + "bbox": [ + 178, + 127, + 199, + 138 + ], + "score": 0.91, + "content": "\\tilde { \\rho } _ { k - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 126, + 217, + 139 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 218, + 127, + 229, + 138 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 126, + 506, + 139 + ], + "score": 1.0, + "content": ". The extra Wasserstein-distance term arises naturally due to the fact", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 137, + 505, + 150 + ], + "score": 1.0, + "content": "that the Langevin diffusion can be explained as a gradient flow whose geometry is equipped with", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "spans": [ + { + "bbox": [ + 105, + 147, + 506, + 161 + ], + "score": 1.0, + "content": "the Wasserstein distance (Otto, 1998). From another point of view, it is known that the Wasserstein", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "spans": [ + { + "bbox": [ + 105, + 158, + 505, + 172 + ], + "score": 1.0, + "content": "distance is a better metric for probability distributions than the KL-divergence, especially in the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "spans": [ + { + "bbox": [ + 104, + 169, + 506, + 185 + ], + "score": 1.0, + "content": "case of non-overlapping domains (Arjovsky & Bottou, 2017; Arjovsky et al., 2017). By using", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 180, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 180, + 505, + 194 + ], + "score": 1.0, + "content": "the Wasserstein term as a regularizer, the CTF alleviates the issue in non-overlapping domains", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "spans": [ + { + "bbox": [ + 106, + 192, + 505, + 204 + ], + "score": 1.0, + "content": "by introducing the Brownian-motion (noise) term in the evolution (2). This relates to the idea in", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "(Arjovsky & Bottou, 2017), in which noise is added in parameter updates to alleviate the intrinsic", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 213, + 265, + 227 + ], + "spans": [ + { + "bbox": [ + 106, + 213, + 265, + 227 + ], + "score": 1.0, + "content": "drawback of the KL-divergence metric.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 6, + "bbox_fs": [ + 104, + 82, + 506, + 227 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 505, + 320 + ], + "lines": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "spans": [ + { + "bbox": [ + 106, + 231, + 506, + 243 + ], + "score": 1.0, + "content": "The optimization problem in Lemma 1 is difficult to deal with directly. In practice, we instead", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 241, + 502, + 256 + ], + "spans": [ + { + "bbox": [ + 105, + 241, + 491, + 256 + ], + "score": 1.0, + "content": "approximate the discretization in an equivalent way by simulation from the CTF. Starting from", + "type": "text" + }, + { + "bbox": [ + 492, + 244, + 502, + 253 + ], + "score": 0.82, + "content": "\\mathbf { z } _ { 0 }", + "type": "inline_equation" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 254, + 118, + 264 + ], + "score": 0.73, + "content": "\\mathbf { z } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 252, + 121, + 266 + ], + "score": 0.0, + "content": "", + "type": "text" + }, + { + "bbox": [ + 121, + 253, + 202, + 264 + ], + "score": 0.82, + "content": "( k = 0 , \\cdots , K - 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 252, + 317, + 266 + ], + "score": 1.0, + "content": "is fed into a transformation", + "type": "text" + }, + { + "bbox": [ + 317, + 253, + 329, + 264 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 252, + 453, + 266 + ], + "score": 1.0, + "content": "(specified below), resulting in", + "type": "text" + }, + { + "bbox": [ + 454, + 254, + 475, + 264 + ], + "score": 0.9, + "content": "{ \\mathbf z } _ { k + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 475, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "whose", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 263, + 507, + 277 + ], + "spans": [ + { + "bbox": [ + 105, + 263, + 214, + 277 + ], + "score": 1.0, + "content": "distribution coincides with", + "type": "text" + }, + { + "bbox": [ + 215, + 264, + 236, + 276 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 263, + 507, + 277 + ], + "score": 1.0, + "content": "in Lemma 1. The discretization procedure is illustrated in Figure 1.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 275, + 506, + 287 + ], + "spans": [ + { + "bbox": [ + 106, + 275, + 251, + 287 + ], + "score": 1.0, + "content": "We must specify the transformations", + "type": "text" + }, + { + "bbox": [ + 252, + 275, + 263, + 286 + ], + "score": 0.87, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 275, + 303, + 287 + ], + "score": 1.0, + "content": ". For each", + "type": "text" + }, + { + "bbox": [ + 304, + 275, + 310, + 285 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 275, + 326, + 287 + ], + "score": 1.0, + "content": ", let", + "type": "text" + }, + { + "bbox": [ + 326, + 275, + 355, + 285 + ], + "score": 0.91, + "content": "t = h k", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 275, + 506, + 287 + ], + "score": 1.0, + "content": "; we can conclude from Lemma 1 that", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 286, + 505, + 299 + ], + "spans": [ + { + "bbox": [ + 107, + 286, + 172, + 298 + ], + "score": 0.9, + "content": "\\operatorname* { l i m } _ { h \\to 0 } \\tilde { \\rho } _ { k } = \\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 286, + 245, + 299 + ], + "score": 1.0, + "content": ". From FP theory,", + "type": "text" + }, + { + "bbox": [ + 246, + 288, + 256, + 298 + ], + "score": 0.85, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 256, + 286, + 505, + 299 + ], + "score": 1.0, + "content": "is obtained by solving the diffusion (2) with initial condition", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 297, + 505, + 309 + ], + "spans": [ + { + "bbox": [ + 106, + 298, + 145, + 308 + ], + "score": 0.88, + "content": "{ \\bf Z } _ { 0 } = { \\bf z } _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 145, + 297, + 361, + 309 + ], + "score": 1.0, + "content": ". It is thus reasonable to specify the transformation", + "type": "text" + }, + { + "bbox": [ + 361, + 297, + 373, + 308 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 373, + 297, + 402, + 309 + ], + "score": 1.0, + "content": "as the", + "type": "text" + }, + { + "bbox": [ + 402, + 297, + 409, + 307 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 297, + 505, + 309 + ], + "score": 1.0, + "content": "-th step of a numerical", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 308, + 412, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 276, + 321 + ], + "score": 1.0, + "content": "integrator for (2). Specifically, we specify", + "type": "text" + }, + { + "bbox": [ + 276, + 308, + 287, + 319 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 308, + 412, + 321 + ], + "score": 1.0, + "content": "as a stochastic transformation:", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 16.5, + "bbox_fs": [ + 105, + 231, + 507, + 321 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 199, + 333, + 411, + 348 + ], + "lines": [ + { + "bbox": [ + 199, + 333, + 411, + 348 + ], + "spans": [ + { + "bbox": [ + 199, + 333, + 411, + 348 + ], + "score": 0.88, + "content": "{ \\bf z } _ { k } = { \\mathcal T } _ { k } ( { \\bf z } _ { k - 1 } ) \\triangleq { \\bf z } _ { k - 1 } + { \\cal F } ( { \\bf z } _ { k - 1 } ) h + V ( { \\bf z } _ { k - 1 } ) \\zeta _ { k } ,", + "type": "interline_equation", + "image_path": "197d4f4ace878ed60868e893eae9ffb5758c77de999f3c246be7af44f22053e1.jpg" + } + ] + } + ], + "index": 21, + "virtual_lines": [ + { + "bbox": [ + 199, + 333, + 411, + 348 + ], + "spans": [], + "index": 21 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 362, + 505, + 409 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 505, + 375 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 133, + 374 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 362, + 203, + 375 + ], + "score": 0.92, + "content": "\\mathsf { \\Sigma } \\xi _ { k } \\sim \\mathcal { N } ( \\mathbf { 0 } , h \\mathbf { I } _ { L } )", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 362, + 505, + 374 + ], + "score": 1.0, + "content": "is drawn from an isotropic normal. Note the transformation defined here", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 374, + 505, + 386 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 291, + 386 + ], + "score": 1.0, + "content": "is stochastic, thus we only get samples from", + "type": "text" + }, + { + "bbox": [ + 291, + 374, + 305, + 385 + ], + "score": 0.89, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 374, + 480, + 386 + ], + "score": 1.0, + "content": "at the end. A natural way to approximate", + "type": "text" + }, + { + "bbox": [ + 480, + 374, + 494, + 385 + ], + "score": 0.89, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 495, + 374, + 505, + 386 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 384, + 503, + 401 + ], + "spans": [ + { + "bbox": [ + 105, + 384, + 291, + 401 + ], + "score": 1.0, + "content": "to use the empirical sample distribution, i.e.,", + "type": "text" + }, + { + "bbox": [ + 292, + 384, + 398, + 399 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\tilde { \\rho } _ { K } \\approx \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\delta _ { { \\bf z } _ { k } } \\triangleq \\bar { \\rho } _ { T } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 384, + 420, + 401 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 420, + 387, + 431, + 398 + ], + "score": 0.87, + "content": "\\delta _ { \\mathbf { z } }", + "type": "inline_equation" + }, + { + "bbox": [ + 431, + 384, + 496, + 401 + ], + "score": 1.0, + "content": "a point mass at", + "type": "text" + }, + { + "bbox": [ + 497, + 389, + 503, + 396 + ], + "score": 0.62, + "content": "\\mathbf { z }", + "type": "inline_equation" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 397, + 409, + 410 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 156, + 410 + ], + "score": 1.0, + "content": "Afterwards,", + "type": "text" + }, + { + "bbox": [ + 156, + 398, + 170, + 409 + ], + "score": 0.89, + "content": "\\tilde { \\rho } _ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 397, + 193, + 410 + ], + "score": 1.0, + "content": "(thus", + "type": "text" + }, + { + "bbox": [ + 194, + 398, + 206, + 409 + ], + "score": 0.84, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 207, + 397, + 356, + 410 + ], + "score": 1.0, + "content": ") will be used to approximate the true", + "type": "text" + }, + { + "bbox": [ + 357, + 400, + 369, + 409 + ], + "score": 0.8, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 397, + 409, + 410 + ], + "score": 1.0, + "content": "from (3).", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 23.5, + "bbox_fs": [ + 105, + 362, + 505, + 410 + ] + }, + { + "type": "image", + "bbox": [ + 127, + 420, + 485, + 530 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 127, + 420, + 485, + 530 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 127, + 420, + 485, + 530 + ], + "spans": [ + { + "bbox": [ + 127, + 420, + 485, + 530 + ], + "score": 0.975, + "type": "image", + "image_path": "bcb7adee8eb928455458752caa91f82356f295994b68b570ae3c02cde0d35129.jpg" + } + ] + } + ], + "index": 27, + "virtual_lines": [ + { + "bbox": [ + 127, + 420, + 485, + 456.6666666666667 + ], + "spans": [], + "index": 26 + }, + { + "bbox": [ + 127, + 456.6666666666667, + 485, + 493.33333333333337 + ], + "spans": [], + "index": 27 + }, + { + "bbox": [ + 127, + 493.33333333333337, + 485, + 530.0 + ], + "spans": [], + "index": 28 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 105, + 538, + 504, + 562 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 538, + 504, + 552 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 449, + 552 + ], + "score": 1.0, + "content": "Figure 1: Discretized approximation (right) of a continuous-time flow (left). Densities", + "type": "text" + }, + { + "bbox": [ + 450, + 539, + 470, + 551 + ], + "score": 0.92, + "content": "\\{ \\tilde { \\rho } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 471, + 538, + 483, + 552 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 483, + 539, + 504, + 551 + ], + "score": 0.9, + "content": "\\left\\{ { \\bf z } _ { k } \\right\\}", + "type": "inline_equation" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 549, + 469, + 562 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 215, + 562 + ], + "score": 1.0, + "content": "evolve via transformations", + "type": "text" + }, + { + "bbox": [ + 215, + 550, + 236, + 562 + ], + "score": 0.91, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 549, + 260, + 562 + ], + "score": 1.0, + "content": ", with", + "type": "text" + }, + { + "bbox": [ + 260, + 550, + 302, + 561 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k } \\to \\rho _ { h k }", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 549, + 327, + 562 + ], + "score": 1.0, + "content": "when", + "type": "text" + }, + { + "bbox": [ + 327, + 550, + 354, + 560 + ], + "score": 0.91, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 355, + 549, + 390, + 562 + ], + "score": 1.0, + "content": "for each", + "type": "text" + }, + { + "bbox": [ + 390, + 550, + 397, + 560 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 549, + 469, + 562 + ], + "score": 1.0, + "content": "due to Lemma 1.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 29.5 + } + ], + "index": 28.25 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 506, + 728 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 505, + 595 + ], + "score": 1.0, + "content": "Note that we use the simple sample averaging for the convenience of analysis, and the approximation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 592, + 506, + 605 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 120, + 605 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 120, + 594, + 133, + 604 + ], + "score": 0.87, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 133, + 592, + 506, + 605 + ], + "score": 1.0, + "content": "is not necessarily optimal. Better approximation can be obtained by assigning more weights to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 505, + 617 + ], + "score": 1.0, + "content": "the more recent samples. However, this leads to more challenges in theoretical analysis, an interesting", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 614, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 614, + 370, + 628 + ], + "score": 1.0, + "content": "future direction to pursue. In the following, we study how well", + "type": "text" + }, + { + "bbox": [ + 370, + 615, + 383, + 626 + ], + "score": 0.87, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 383, + 614, + 442, + 628 + ], + "score": 1.0, + "content": "approximates", + "type": "text" + }, + { + "bbox": [ + 442, + 616, + 455, + 626 + ], + "score": 0.84, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 614, + 506, + 628 + ], + "score": 1.0, + "content": ". Following", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 625, + 506, + 639 + ], + "score": 1.0, + "content": "literature on numerical approximation for Itô diffusions (Vollmer et al., 2016; Chen et al., 2015), we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 636, + 506, + 649 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 253, + 649 + ], + "score": 1.0, + "content": "consider a 1-Lipschitz test function", + "type": "text" + }, + { + "bbox": [ + 254, + 636, + 307, + 648 + ], + "score": 0.93, + "content": "\\psi : \\mathbb { R } ^ { L } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 307, + 636, + 506, + 649 + ], + "score": 1.0, + "content": ", and use the mean square error (MSE) bound to", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 103, + 648, + 506, + 664 + ], + "spans": [ + { + "bbox": [ + 103, + 649, + 209, + 664 + ], + "score": 1.0, + "content": "measure the closeness of", + "type": "text" + }, + { + "bbox": [ + 209, + 651, + 222, + 662 + ], + "score": 0.88, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 649, + 240, + 664 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 240, + 651, + 253, + 662 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 649, + 303, + 664 + ], + "score": 1.0, + "content": ", defined as:", + "type": "text" + }, + { + "bbox": [ + 303, + 648, + 502, + 663 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbf { M S E } ( \\bar { \\rho } _ { T } , \\rho _ { T } ; \\psi ) \\triangleq \\mathbb { E } \\left( \\int \\psi ( \\mathbf { z } ) ( \\tilde { \\rho } _ { T } - \\rho _ { T } ) ( \\mathbf { z } ) \\mathrm { d } \\mathbf { z } \\right) ^ { 2 } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 649, + 506, + 664 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 661, + 506, + 674 + ], + "spans": [ + { + "bbox": [ + 105, + 661, + 413, + 674 + ], + "score": 1.0, + "content": "where the expectation is taken over all the randomness in the construction of", + "type": "text" + }, + { + "bbox": [ + 414, + 663, + 426, + 673 + ], + "score": 0.88, + "content": "\\tilde { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 426, + 661, + 506, + 674 + ], + "score": 1.0, + "content": ". Note that our goal", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 672, + 506, + 684 + ], + "spans": [ + { + "bbox": [ + 105, + 672, + 506, + 684 + ], + "score": 1.0, + "content": "is related but different from the standard setup as in Vollmer et al. (2016); Chen et al. (2015), which", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 683, + 506, + 695 + ], + "spans": [ + { + "bbox": [ + 106, + 683, + 200, + 695 + ], + "score": 1.0, + "content": "studies the closeness of", + "type": "text" + }, + { + "bbox": [ + 200, + 684, + 213, + 695 + ], + "score": 0.88, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 213, + 683, + 224, + 695 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 224, + 683, + 259, + 695 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 683, + 506, + 695 + ], + "score": 1.0, + "content": ". We need to adopt the assumptions from Vollmer et al. (2016);", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 693, + 506, + 707 + ], + "spans": [ + { + "bbox": [ + 106, + 693, + 506, + 707 + ], + "score": 1.0, + "content": "Chen et al. (2015), which are described in the Supplementary Material (SM). The assumptions are", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 705, + 506, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 705, + 506, + 717 + ], + "score": 1.0, + "content": "somewhat involved but essentially require coefficients of the diffusion (2) to be well-behaved. We", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 716, + 506, + 729 + ], + "spans": [ + { + "bbox": [ + 105, + 716, + 390, + 729 + ], + "score": 1.0, + "content": "derive the following bound for the MSE of the sampled approximation,", + "type": "text" + }, + { + "bbox": [ + 391, + 717, + 403, + 728 + ], + "score": 0.86, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 403, + 716, + 506, + 729 + ], + "score": 1.0, + "content": ", and the true distribution.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 37, + "bbox_fs": [ + 103, + 581, + 506, + 729 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 111 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 235, + 95 + ], + "score": 1.0, + "content": "Theorem 2. Under Assumption", + "type": "text" + }, + { + "bbox": [ + 235, + 83, + 240, + 92 + ], + "score": 0.41, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 81, + 333, + 95 + ], + "score": 1.0, + "content": "in the SM, assume that", + "type": "text" + }, + { + "bbox": [ + 333, + 81, + 441, + 95 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\int \\rho _ { T } ( \\mathbf { z } ) p _ { \\pmb \\theta } ^ { - 1 } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } < \\infty } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "and there exists", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 94, + 459, + 112 + ], + "spans": [ + { + "bbox": [ + 104, + 96, + 206, + 111 + ], + "score": 1.0, + "content": "a constant C such tha", + "type": "text" + }, + { + "bbox": [ + 195, + 94, + 459, + 112 + ], + "score": 1.0, + "content": "t dW 22 (ρT ,pθ(x,z))dt ≥ CW 22 (ρT , pθ(x, z)), the MSE is bounded as", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 116, + 405, + 144 + ], + "lines": [ + { + "bbox": [ + 206, + 116, + 405, + 144 + ], + "spans": [ + { + "bbox": [ + 206, + 116, + 405, + 144 + ], + "score": 0.91, + "content": "M S E ( \\bar { \\rho } _ { T } , \\rho _ { T } ; \\psi ) = O \\left( \\frac { 1 } { h K } + h ^ { 2 } + e ^ { - 2 C h K } \\right) .", + "type": "interline_equation", + "image_path": "ee9e3348c5bc3d9582efb151f896bd1f6af3cf1fc49f57727dad0c5592eecc10.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 206, + 116, + 405, + 144 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 292, + 168 + ], + "score": 1.0, + "content": "The last assumption in Theorem 2 requires", + "type": "text" + }, + { + "bbox": [ + 292, + 157, + 305, + 167 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "to evolve fast through the FP equation, which", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 165, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 505, + 179 + ], + "score": 1.0, + "content": "is a standard assumption used to establish convergence to equilibrium for FP equations (Bolley", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "et al., 2012). The MSE bound consists of three terms, the first two terms come from numerical", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 187, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 506, + 203 + ], + "score": 1.0, + "content": "approximation of the continuous-time diffusion, whereas the third term comes from the convergence", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 199, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 506, + 212 + ], + "score": 1.0, + "content": "bound of the FP equation in terms of the Wasserstein distance (Bolley et al., 2012). When the time", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 209, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 144, + 220 + ], + "score": 0.9, + "content": "T = h K", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 209, + 506, + 223 + ], + "score": 1.0, + "content": "is large enough, the third term may be ignored due to its exponential-decay rate. Moreover,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 362, + 234 + ], + "score": 1.0, + "content": "in the infinite-time limit, the bound endows a bias proportional to", + "type": "text" + }, + { + "bbox": [ + 363, + 222, + 369, + 231 + ], + "score": 0.78, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "; this, however, can be removed by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "adopting a decreasing-step-size scheme in the numerical method, as in standard stochastic gradient", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 243, + 322, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 322, + 255 + ], + "score": 1.0, + "content": "MCMC methods (Teh et al., 2016; Chen et al., 2015).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 257, + 505, + 327 + ], + "lines": [ + { + "bbox": [ + 104, + 256, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 104, + 256, + 441, + 271 + ], + "score": 1.0, + "content": "Remark 3. To examine the optimal bound in Theorem 2, we drop out the term", + "type": "text" + }, + { + "bbox": [ + 442, + 257, + 477, + 268 + ], + "score": 0.9, + "content": "e ^ { - 2 C h K }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 256, + 506, + 271 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 269, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 504, + 281 + ], + "score": 1.0, + "content": "long-time case (when hK is large enough) for simplicity because it is in a much lower order term", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 275, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 104, + 275, + 416, + 296 + ], + "score": 1.0, + "content": "than the other terms. The optimal MSE bound (over h) decreases at a rate of", + "type": "text" + }, + { + "bbox": [ + 416, + 279, + 463, + 293 + ], + "score": 0.91, + "content": "O \\left( K ^ { - 2 / 3 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 275, + 507, + 296 + ], + "score": 1.0, + "content": ", meaning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 126, + 308 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 126, + 292, + 168, + 307 + ], + "score": 0.92, + "content": "O \\left( \\epsilon ^ { - 3 / 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 291, + 460, + 308 + ], + "score": 1.0, + "content": "steps of transformations in Figure 1 (right) are needed to reach an", + "type": "text" + }, + { + "bbox": [ + 460, + 296, + 465, + 303 + ], + "score": 0.42, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 291, + 506, + 308 + ], + "score": 1.0, + "content": "-accurate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 188, + 318 + ], + "score": 1.0, + "content": "approximation, i.e.,", + "type": "text" + }, + { + "bbox": [ + 188, + 305, + 227, + 316 + ], + "score": 0.87, + "content": "M S E \\leq \\epsilon .", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 304, + 506, + 318 + ], + "score": 1.0, + "content": ". This is computationally expensive. An efficient way for inference is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 316, + 292, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 292, + 327 + ], + "score": 1.0, + "content": "thus imperative, developed in the next section.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 107, + 341, + 314, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 315, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 315, + 354 + ], + "score": 1.0, + "content": "3.2 EFFICIENT INFERENCE VIA AMORTIZATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 506, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 226, + 374 + ], + "score": 1.0, + "content": "Even though we approximate", + "type": "text" + }, + { + "bbox": [ + 226, + 364, + 239, + 374 + ], + "score": 0.86, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 362, + 260, + 374 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 260, + 363, + 272, + 374 + ], + "score": 0.87, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 362, + 506, + 374 + ], + "score": 1.0, + "content": ", it is still infeasible to directly apply it to the ELBO in (4)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 373, + 504, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 117, + 385 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 117, + 374, + 129, + 385 + ], + "score": 0.88, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 373, + 496, + 385 + ], + "score": 1.0, + "content": "is discrete. To deal with this problem, we adopt the idea of “amortized learning” (Gershman", + "type": "text" + }, + { + "bbox": [ + 496, + 374, + 504, + 383 + ], + "score": 0.39, + "content": "\\&", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 384, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 496, + 397 + ], + "score": 1.0, + "content": "Goodman, 2014) for efficient inference. The main idea is to optimize the two sets of parameters", + "type": "text" + }, + { + "bbox": [ + 496, + 385, + 504, + 396 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 395, + 385, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 123, + 407 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 396, + 131, + 405 + ], + "score": 0.78, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 395, + 385, + 407 + ], + "score": 1.0, + "content": "alternatively, based on different but related objective functions.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 465, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 465, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 148, + 434 + ], + "score": 1.0, + "content": "Updating", + "type": "text" + }, + { + "bbox": [ + 149, + 420, + 156, + 431 + ], + "score": 0.77, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 418, + 465, + 434 + ], + "score": 1.0, + "content": "To explain the idea, first note that the ELBO can be equivalently written as", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 437, + 412, + 454 + ], + "lines": [ + { + "bbox": [ + 198, + 437, + 412, + 454 + ], + "spans": [ + { + "bbox": [ + 198, + 437, + 412, + 454 + ], + "score": 0.87, + "content": "\\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { \\rho _ { 0 } \\triangleq q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\mathbb { E } _ { \\rho _ { T } } \\left[ \\log \\rho _ { 0 } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { Z } _ { T } ) \\right] .", + "type": "interline_equation", + "image_path": "dcb966fdbebae32ef2ad32f90a218d00774c5e9559f84804ead4f3f9f6f0a9e1.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 198, + 437, + 412, + 454 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 459, + 506, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 503, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 133, + 474 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 133, + 462, + 168, + 472 + ], + "score": 0.91, + "content": "\\rho _ { 0 } = \\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 459, + 256, + 474 + ], + "score": 1.0, + "content": ", it is easy to see that:", + "type": "text" + }, + { + "bbox": [ + 257, + 460, + 503, + 472 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { \\rho _ { 0 } } \\left[ \\log \\rho _ { 0 } - \\log p _ { \\theta } ( \\mathbf { Z } _ { T } | \\mathbf { x } ) \\right] + \\log p ( \\mathbf { x } ) = \\log p ( \\mathbf { x } ) . } \\end{array}", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 469, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 469, + 275, + 485 + ], + "score": 1.0, + "content": "which essentially makes the gap between", + "type": "text" + }, + { + "bbox": [ + 275, + 472, + 315, + 483 + ], + "score": 0.9, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 469, + 334, + 485 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 334, + 471, + 378, + 483 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { Z } _ { T } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 469, + 506, + 485 + ], + "score": 1.0, + "content": "vanished. As a result, our goal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 150, + 495 + ], + "score": 1.0, + "content": "is to learn", + "type": "text" + }, + { + "bbox": [ + 150, + 482, + 158, + 493 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 482, + 199, + 495 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 199, + 482, + 239, + 494 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 482, + 290, + 495 + ], + "score": 1.0, + "content": "approaches", + "type": "text" + }, + { + "bbox": [ + 290, + 482, + 334, + 494 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { Z } _ { T } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 482, + 506, + 495 + ], + "score": 1.0, + "content": ". As mentioned previously, we will learn", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 216, + 506 + ], + "score": 1.0, + "content": "an implicit distribution of", + "type": "text" + }, + { + "bbox": [ + 217, + 493, + 257, + 505 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 492, + 419, + 506 + ], + "score": 1.0, + "content": "(i.e., learn how to draw samples from", + "type": "text" + }, + { + "bbox": [ + 419, + 493, + 460, + 505 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "score": 1.0, + "content": "its explicit form), as it allows us to chose a candidate distribution from a much larger distribution", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 259, + 528 + ], + "score": 1.0, + "content": "space, compared to explicitly defining", + "type": "text" + }, + { + "bbox": [ + 260, + 514, + 276, + 527 + ], + "score": 0.88, + "content": "q _ { \\phi } ^ { \\bullet }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 514, + 339, + 528 + ], + "score": 1.0, + "content": ". Consequently,", + "type": "text" + }, + { + "bbox": [ + 339, + 515, + 380, + 527 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "is implemented by a stochastic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 525, + 507, + 539 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 257, + 539 + ], + "score": 1.0, + "content": "generator (a DNN parameterized by", + "type": "text" + }, + { + "bbox": [ + 257, + 527, + 264, + 537 + ], + "score": 0.76, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 525, + 269, + 539 + ], + "score": 1.0, + "content": ")", + "type": "text" + }, + { + "bbox": [ + 270, + 526, + 324, + 538 + ], + "score": 0.92, + "content": "Q _ { \\phi } ( \\mathbf { z } _ { 0 } | \\mathbf { x } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 525, + 468, + 539 + ], + "score": 1.0, + "content": "with input as the concatenation of", + "type": "text" + }, + { + "bbox": [ + 468, + 527, + 476, + 536 + ], + "score": 0.53, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 525, + 494, + 539 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 495, + 528, + 502, + 536 + ], + "score": 0.74, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 525, + 507, + 539 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 133, + 550 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 538, + 142, + 547 + ], + "score": 0.77, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 537, + 354, + 550 + ], + "score": 1.0, + "content": "is a sample from an isotropic Gaussian distribution", + "type": "text" + }, + { + "bbox": [ + 354, + 537, + 378, + 549 + ], + "score": 0.92, + "content": "q _ { 0 } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 537, + 506, + 550 + ], + "score": 1.0, + "content": ". Our goal is now translated to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 547, + 504, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 195, + 561 + ], + "score": 1.0, + "content": "update the parameter", + "type": "text" + }, + { + "bbox": [ + 195, + 548, + 203, + 559 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 547, + 216, + 561 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 216, + 548, + 271, + 560 + ], + "score": 0.92, + "content": "Q _ { \\phi } ( \\mathbf { z } _ { 0 } | \\mathbf { x } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 547, + 283, + 561 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 283, + 548, + 294, + 559 + ], + "score": 0.88, + "content": "\\phi ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 547, + 412, + 561 + ], + "score": 1.0, + "content": "such that the distribution of", + "type": "text" + }, + { + "bbox": [ + 412, + 547, + 504, + 560 + ], + "score": 0.91, + "content": "\\{ \\mathbf { z } _ { 0 } ^ { \\prime } = Q _ { \\phi ^ { \\prime } } ( \\mathbf { z } _ { 0 } ^ { \\prime } \\mid \\mathbf { x } , \\omega ) \\}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 557, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 126, + 573 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 126, + 559, + 170, + 571 + ], + "score": 0.93, + "content": "\\omega \\sim q _ { 0 } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 557, + 234, + 573 + ], + "score": 1.0, + "content": "matches that of", + "type": "text" + }, + { + "bbox": [ + 235, + 561, + 245, + 570 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 557, + 400, + 573 + ], + "score": 1.0, + "content": "in the original generating process with", + "type": "text" + }, + { + "bbox": [ + 401, + 559, + 408, + 570 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 557, + 507, + 573 + ], + "score": 1.0, + "content": "in Figure 1. In this way,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 568, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 210, + 584 + ], + "score": 1.0, + "content": "the generating process of", + "type": "text" + }, + { + "bbox": [ + 210, + 571, + 221, + 581 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 568, + 237, + 584 + ], + "score": 1.0, + "content": "via", + "type": "text" + }, + { + "bbox": [ + 237, + 570, + 248, + 581 + ], + "score": 0.87, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 568, + 428, + 584 + ], + "score": 1.0, + "content": "is distilled into the parameterized generator", + "type": "text" + }, + { + "bbox": [ + 428, + 570, + 452, + 582 + ], + "score": 0.91, + "content": "Q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 568, + 506, + 584 + ], + "score": 1.0, + "content": ", eliminating", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 579, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 281, + 594 + ], + "score": 1.0, + "content": "the need to do a specific transformation via", + "type": "text" + }, + { + "bbox": [ + 281, + 581, + 293, + 592 + ], + "score": 0.89, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 579, + 506, + 594 + ], + "score": 1.0, + "content": "in testing, and thus is very efficient. Specifically, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 591, + 186, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 135, + 604 + ], + "score": 1.0, + "content": "update", + "type": "text" + }, + { + "bbox": [ + 135, + 592, + 145, + 603 + ], + "score": 0.88, + "content": "\\phi ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 591, + 186, + 604 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 609, + 376, + 632 + ], + "lines": [ + { + "bbox": [ + 233, + 609, + 376, + 632 + ], + "spans": [ + { + "bbox": [ + 233, + 609, + 376, + 632 + ], + "score": 0.94, + "content": "\\phi ^ { \\prime } = \\arg \\operatorname* { m i n } _ { \\phi } \\mathcal { D } \\left( \\{ \\mathbf { z } _ { 0 } ^ { \\prime ( i ) } \\} , \\{ \\mathbf { z } _ { 1 } ^ { ( i ) } \\} \\right) ,", + "type": "interline_equation", + "image_path": "f1c3ff570f6a04c96384ca8c54209d1171a4e27623f34e24d4ac94b555ec213d.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 233, + 609, + 376, + 632 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 639, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 102, + 636, + 508, + 659 + ], + "spans": [ + { + "bbox": [ + 102, + 636, + 132, + 659 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 639, + 173, + 654 + ], + "score": 0.94, + "content": "\\{ \\mathbf { z } _ { 0 } ^ { \\prime ( i ) } \\} _ { i = 1 } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 636, + 314, + 659 + ], + "score": 1.0, + "content": "are a set of samples generated from", + "type": "text" + }, + { + "bbox": [ + 314, + 641, + 357, + 655 + ], + "score": 0.92, + "content": "q _ { \\phi ^ { \\prime } } ( \\mathbf { z } _ { 0 } ^ { \\prime } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 636, + 372, + 659 + ], + "score": 1.0, + "content": "via", + "type": "text" + }, + { + "bbox": [ + 372, + 641, + 397, + 654 + ], + "score": 0.91, + "content": "Q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 636, + 417, + 659 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 417, + 639, + 456, + 654 + ], + "score": 0.94, + "content": "\\{ \\mathbf { z } _ { 1 } ^ { ( i ) } \\} _ { i = 1 } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 636, + 508, + 659 + ], + "score": 1.0, + "content": "are samples", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 653, + 507, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 146, + 670 + ], + "score": 1.0, + "content": "drawn by", + "type": "text" + }, + { + "bbox": [ + 146, + 654, + 360, + 669 + ], + "score": 0.87, + "content": "\\omega ^ { i } \\sim q _ { 0 } ( \\omega ) , \\tilde { \\mathbf { z } } _ { 0 } ^ { i } = Q _ { \\phi } ( \\cdot | \\mathbf { x } , \\omega ^ { i } ) , \\mathbf { z } _ { 1 } ^ { ( i ) } \\sim \\mathcal { T } _ { 1 } ( \\tilde { \\mathbf { z } } _ { 0 } ^ { i } ) ; \\mathcal { D } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 653, + 507, + 670 + ], + "score": 1.0, + "content": "is a metric between samples such as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 667, + 504, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 504, + 680 + ], + "score": 1.0, + "content": "the simple Euclidean distance or the more advanced Wasserstein distance (Arjovsky et al., 2017). The", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "optimization is done by applying standard stochastic gradient descent (SGD). We call this procedure", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 264, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 212, + 703 + ], + "score": 1.0, + "content": "distilling knowledge from", + "type": "text" + }, + { + "bbox": [ + 213, + 690, + 223, + 700 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 688, + 235, + 703 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 235, + 689, + 259, + 702 + ], + "score": 0.91, + "content": "Q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 688, + 264, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 105, + 711, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 117, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 117, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "¶This is distinct from our density-estimation framework described in the next section, where an explicit form", + "type": "text" + } + ] + }, + { + "bbox": [ + 105, + 721, + 278, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 278, + 732 + ], + "score": 1.0, + "content": "is assumed at the beginning for practical needs.", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 310, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 505, + 111 + ], + "lines": [ + { + "bbox": [ + 106, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 106, + 81, + 235, + 95 + ], + "score": 1.0, + "content": "Theorem 2. Under Assumption", + "type": "text" + }, + { + "bbox": [ + 235, + 83, + 240, + 92 + ], + "score": 0.41, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 81, + 333, + 95 + ], + "score": 1.0, + "content": "in the SM, assume that", + "type": "text" + }, + { + "bbox": [ + 333, + 81, + 441, + 95 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\int \\rho _ { T } ( \\mathbf { z } ) p _ { \\pmb \\theta } ^ { - 1 } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } < \\infty } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "and there exists", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 94, + 459, + 112 + ], + "spans": [ + { + "bbox": [ + 104, + 96, + 206, + 111 + ], + "score": 1.0, + "content": "a constant C such tha", + "type": "text" + }, + { + "bbox": [ + 195, + 94, + 459, + 112 + ], + "score": 1.0, + "content": "t dW 22 (ρT ,pθ(x,z))dt ≥ CW 22 (ρT , pθ(x, z)), the MSE is bounded as", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5, + "bbox_fs": [ + 104, + 81, + 505, + 112 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 206, + 116, + 405, + 144 + ], + "lines": [ + { + "bbox": [ + 206, + 116, + 405, + 144 + ], + "spans": [ + { + "bbox": [ + 206, + 116, + 405, + 144 + ], + "score": 0.91, + "content": "M S E ( \\bar { \\rho } _ { T } , \\rho _ { T } ; \\psi ) = O \\left( \\frac { 1 } { h K } + h ^ { 2 } + e ^ { - 2 C h K } \\right) .", + "type": "interline_equation", + "image_path": "ee9e3348c5bc3d9582efb151f896bd1f6af3cf1fc49f57727dad0c5592eecc10.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 206, + 116, + 405, + 144 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 154, + 505, + 254 + ], + "lines": [ + { + "bbox": [ + 105, + 155, + 505, + 168 + ], + "spans": [ + { + "bbox": [ + 105, + 155, + 292, + 168 + ], + "score": 1.0, + "content": "The last assumption in Theorem 2 requires", + "type": "text" + }, + { + "bbox": [ + 292, + 157, + 305, + 167 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 155, + 505, + 168 + ], + "score": 1.0, + "content": "to evolve fast through the FP equation, which", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 165, + 505, + 179 + ], + "spans": [ + { + "bbox": [ + 104, + 165, + 505, + 179 + ], + "score": 1.0, + "content": "is a standard assumption used to establish convergence to equilibrium for FP equations (Bolley", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 105, + 177, + 506, + 190 + ], + "score": 1.0, + "content": "et al., 2012). The MSE bound consists of three terms, the first two terms come from numerical", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 187, + 506, + 203 + ], + "spans": [ + { + "bbox": [ + 104, + 187, + 506, + 203 + ], + "score": 1.0, + "content": "approximation of the continuous-time diffusion, whereas the third term comes from the convergence", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 199, + 506, + 212 + ], + "spans": [ + { + "bbox": [ + 105, + 199, + 506, + 212 + ], + "score": 1.0, + "content": "bound of the FP equation in terms of the Wasserstein distance (Bolley et al., 2012). When the time", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 209, + 506, + 223 + ], + "spans": [ + { + "bbox": [ + 106, + 210, + 144, + 220 + ], + "score": 0.9, + "content": "T = h K", + "type": "inline_equation" + }, + { + "bbox": [ + 144, + 209, + 506, + 223 + ], + "score": 1.0, + "content": "is large enough, the third term may be ignored due to its exponential-decay rate. Moreover,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 220, + 505, + 234 + ], + "spans": [ + { + "bbox": [ + 105, + 220, + 362, + 234 + ], + "score": 1.0, + "content": "in the infinite-time limit, the bound endows a bias proportional to", + "type": "text" + }, + { + "bbox": [ + 363, + 222, + 369, + 231 + ], + "score": 0.78, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 220, + 505, + 234 + ], + "score": 1.0, + "content": "; this, however, can be removed by", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "spans": [ + { + "bbox": [ + 106, + 232, + 505, + 244 + ], + "score": 1.0, + "content": "adopting a decreasing-step-size scheme in the numerical method, as in standard stochastic gradient", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 243, + 322, + 255 + ], + "spans": [ + { + "bbox": [ + 106, + 243, + 322, + 255 + ], + "score": 1.0, + "content": "MCMC methods (Teh et al., 2016; Chen et al., 2015).", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 7, + "bbox_fs": [ + 104, + 155, + 506, + 255 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 257, + 505, + 327 + ], + "lines": [ + { + "bbox": [ + 104, + 256, + 506, + 271 + ], + "spans": [ + { + "bbox": [ + 104, + 256, + 441, + 271 + ], + "score": 1.0, + "content": "Remark 3. To examine the optimal bound in Theorem 2, we drop out the term", + "type": "text" + }, + { + "bbox": [ + 442, + 257, + 477, + 268 + ], + "score": 0.9, + "content": "e ^ { - 2 C h K }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 256, + 506, + 271 + ], + "score": 1.0, + "content": "in the", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 269, + 504, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 269, + 504, + 281 + ], + "score": 1.0, + "content": "long-time case (when hK is large enough) for simplicity because it is in a much lower order term", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 275, + 507, + 296 + ], + "spans": [ + { + "bbox": [ + 104, + 275, + 416, + 296 + ], + "score": 1.0, + "content": "than the other terms. The optimal MSE bound (over h) decreases at a rate of", + "type": "text" + }, + { + "bbox": [ + 416, + 279, + 463, + 293 + ], + "score": 0.91, + "content": "O \\left( K ^ { - 2 / 3 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 464, + 275, + 507, + 296 + ], + "score": 1.0, + "content": ", meaning", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 291, + 506, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 126, + 308 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 126, + 292, + 168, + 307 + ], + "score": 0.92, + "content": "O \\left( \\epsilon ^ { - 3 / 2 } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 291, + 460, + 308 + ], + "score": 1.0, + "content": "steps of transformations in Figure 1 (right) are needed to reach an", + "type": "text" + }, + { + "bbox": [ + 460, + 296, + 465, + 303 + ], + "score": 0.42, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 465, + 291, + 506, + 308 + ], + "score": 1.0, + "content": "-accurate", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 304, + 506, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 304, + 188, + 318 + ], + "score": 1.0, + "content": "approximation, i.e.,", + "type": "text" + }, + { + "bbox": [ + 188, + 305, + 227, + 316 + ], + "score": 0.87, + "content": "M S E \\leq \\epsilon .", + "type": "inline_equation" + }, + { + "bbox": [ + 227, + 304, + 506, + 318 + ], + "score": 1.0, + "content": ". This is computationally expensive. An efficient way for inference is", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 316, + 292, + 327 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 292, + 327 + ], + "score": 1.0, + "content": "thus imperative, developed in the next section.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 14.5, + "bbox_fs": [ + 104, + 256, + 507, + 327 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 341, + 314, + 353 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 315, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 315, + 354 + ], + "score": 1.0, + "content": "3.2 EFFICIENT INFERENCE VIA AMORTIZATION", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + }, + { + "type": "text", + "bbox": [ + 106, + 361, + 506, + 406 + ], + "lines": [ + { + "bbox": [ + 106, + 362, + 506, + 374 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 226, + 374 + ], + "score": 1.0, + "content": "Even though we approximate", + "type": "text" + }, + { + "bbox": [ + 226, + 364, + 239, + 374 + ], + "score": 0.86, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 239, + 362, + 260, + 374 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 260, + 363, + 272, + 374 + ], + "score": 0.87, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 362, + 506, + 374 + ], + "score": 1.0, + "content": ", it is still infeasible to directly apply it to the ELBO in (4)", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 373, + 504, + 385 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 117, + 385 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 117, + 374, + 129, + 385 + ], + "score": 0.88, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 130, + 373, + 496, + 385 + ], + "score": 1.0, + "content": "is discrete. To deal with this problem, we adopt the idea of “amortized learning” (Gershman", + "type": "text" + }, + { + "bbox": [ + 496, + 374, + 504, + 383 + ], + "score": 0.39, + "content": "\\&", + "type": "inline_equation" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 384, + 504, + 397 + ], + "spans": [ + { + "bbox": [ + 106, + 384, + 496, + 397 + ], + "score": 1.0, + "content": "Goodman, 2014) for efficient inference. The main idea is to optimize the two sets of parameters", + "type": "text" + }, + { + "bbox": [ + 496, + 385, + 504, + 396 + ], + "score": 0.82, + "content": "\\phi", + "type": "inline_equation" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 395, + 385, + 407 + ], + "spans": [ + { + "bbox": [ + 105, + 395, + 123, + 407 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 396, + 131, + 405 + ], + "score": 0.78, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 395, + 385, + 407 + ], + "score": 1.0, + "content": "alternatively, based on different but related objective functions.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 362, + 506, + 407 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 419, + 465, + 431 + ], + "lines": [ + { + "bbox": [ + 106, + 418, + 465, + 434 + ], + "spans": [ + { + "bbox": [ + 106, + 418, + 148, + 434 + ], + "score": 1.0, + "content": "Updating", + "type": "text" + }, + { + "bbox": [ + 149, + 420, + 156, + 431 + ], + "score": 0.77, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 157, + 418, + 465, + 434 + ], + "score": 1.0, + "content": "To explain the idea, first note that the ELBO can be equivalently written as", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 23, + "bbox_fs": [ + 106, + 418, + 465, + 434 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 437, + 412, + 454 + ], + "lines": [ + { + "bbox": [ + 198, + 437, + 412, + 454 + ], + "spans": [ + { + "bbox": [ + 198, + 437, + 412, + 454 + ], + "score": 0.87, + "content": "\\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { \\rho _ { 0 } \\triangleq q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } ) } \\mathbb { E } _ { \\rho _ { T } } \\left[ \\log \\rho _ { 0 } - \\log p _ { \\theta } ( \\mathbf { x } , \\mathbf { Z } _ { T } ) \\right] .", + "type": "interline_equation", + "image_path": "dcb966fdbebae32ef2ad32f90a218d00774c5e9559f84804ead4f3f9f6f0a9e1.jpg" + } + ] + } + ], + "index": 24, + "virtual_lines": [ + { + "bbox": [ + 198, + 437, + 412, + 454 + ], + "spans": [], + "index": 24 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 459, + 506, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 459, + 503, + 474 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 133, + 474 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 133, + 462, + 168, + 472 + ], + "score": 0.91, + "content": "\\rho _ { 0 } = \\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 168, + 459, + 256, + 474 + ], + "score": 1.0, + "content": ", it is easy to see that:", + "type": "text" + }, + { + "bbox": [ + 257, + 460, + 503, + 472 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\mathcal { F } ( \\mathbf { x } ) = \\mathbb { E } _ { \\rho _ { 0 } } \\left[ \\log \\rho _ { 0 } - \\log p _ { \\theta } ( \\mathbf { Z } _ { T } | \\mathbf { x } ) \\right] + \\log p ( \\mathbf { x } ) = \\log p ( \\mathbf { x } ) . } \\end{array}", + "type": "inline_equation" + } + ], + "index": 25 + }, + { + "bbox": [ + 104, + 469, + 506, + 485 + ], + "spans": [ + { + "bbox": [ + 104, + 469, + 275, + 485 + ], + "score": 1.0, + "content": "which essentially makes the gap between", + "type": "text" + }, + { + "bbox": [ + 275, + 472, + 315, + 483 + ], + "score": 0.9, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 469, + 334, + 485 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 334, + 471, + 378, + 483 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { Z } _ { T } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 469, + 506, + 485 + ], + "score": 1.0, + "content": "vanished. As a result, our goal", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 482, + 506, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 482, + 150, + 495 + ], + "score": 1.0, + "content": "is to learn", + "type": "text" + }, + { + "bbox": [ + 150, + 482, + 158, + 493 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 159, + 482, + 199, + 495 + ], + "score": 1.0, + "content": "such that", + "type": "text" + }, + { + "bbox": [ + 199, + 482, + 239, + 494 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 482, + 290, + 495 + ], + "score": 1.0, + "content": "approaches", + "type": "text" + }, + { + "bbox": [ + 290, + 482, + 334, + 494 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { Z } _ { T } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 482, + 506, + 495 + ], + "score": 1.0, + "content": ". As mentioned previously, we will learn", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 104, + 492, + 506, + 506 + ], + "spans": [ + { + "bbox": [ + 104, + 492, + 216, + 506 + ], + "score": 1.0, + "content": "an implicit distribution of", + "type": "text" + }, + { + "bbox": [ + 217, + 493, + 257, + 505 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 257, + 492, + 419, + 506 + ], + "score": 1.0, + "content": "(i.e., learn how to draw samples from", + "type": "text" + }, + { + "bbox": [ + 419, + 493, + 460, + 505 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 492, + 506, + 506 + ], + "score": 1.0, + "content": "instead of", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 504, + 506, + 517 + ], + "score": 1.0, + "content": "its explicit form), as it allows us to chose a candidate distribution from a much larger distribution", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 514, + 506, + 528 + ], + "spans": [ + { + "bbox": [ + 104, + 514, + 259, + 528 + ], + "score": 1.0, + "content": "space, compared to explicitly defining", + "type": "text" + }, + { + "bbox": [ + 260, + 514, + 276, + 527 + ], + "score": 0.88, + "content": "q _ { \\phi } ^ { \\bullet }", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 514, + 339, + 528 + ], + "score": 1.0, + "content": ". Consequently,", + "type": "text" + }, + { + "bbox": [ + 339, + 515, + 380, + 527 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 514, + 506, + 528 + ], + "score": 1.0, + "content": "is implemented by a stochastic", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 525, + 507, + 539 + ], + "spans": [ + { + "bbox": [ + 104, + 525, + 257, + 539 + ], + "score": 1.0, + "content": "generator (a DNN parameterized by", + "type": "text" + }, + { + "bbox": [ + 257, + 527, + 264, + 537 + ], + "score": 0.76, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 525, + 269, + 539 + ], + "score": 1.0, + "content": ")", + "type": "text" + }, + { + "bbox": [ + 270, + 526, + 324, + 538 + ], + "score": 0.92, + "content": "Q _ { \\phi } ( \\mathbf { z } _ { 0 } | \\mathbf { x } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 525, + 468, + 539 + ], + "score": 1.0, + "content": "with input as the concatenation of", + "type": "text" + }, + { + "bbox": [ + 468, + 527, + 476, + 536 + ], + "score": 0.53, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 476, + 525, + 494, + 539 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 495, + 528, + 502, + 536 + ], + "score": 0.74, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 525, + 507, + 539 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 537, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 537, + 133, + 550 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 538, + 142, + 547 + ], + "score": 0.77, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 537, + 354, + 550 + ], + "score": 1.0, + "content": "is a sample from an isotropic Gaussian distribution", + "type": "text" + }, + { + "bbox": [ + 354, + 537, + 378, + 549 + ], + "score": 0.92, + "content": "q _ { 0 } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 537, + 506, + 550 + ], + "score": 1.0, + "content": ". Our goal is now translated to", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 547, + 504, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 547, + 195, + 561 + ], + "score": 1.0, + "content": "update the parameter", + "type": "text" + }, + { + "bbox": [ + 195, + 548, + 203, + 559 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 204, + 547, + 216, + 561 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 216, + 548, + 271, + 560 + ], + "score": 0.92, + "content": "Q _ { \\phi } ( \\mathbf { z } _ { 0 } | \\mathbf { x } , \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 547, + 283, + 561 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 283, + 548, + 294, + 559 + ], + "score": 0.88, + "content": "\\phi ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 294, + 547, + 412, + 561 + ], + "score": 1.0, + "content": "such that the distribution of", + "type": "text" + }, + { + "bbox": [ + 412, + 547, + 504, + 560 + ], + "score": 0.91, + "content": "\\{ \\mathbf { z } _ { 0 } ^ { \\prime } = Q _ { \\phi ^ { \\prime } } ( \\mathbf { z } _ { 0 } ^ { \\prime } \\mid \\mathbf { x } , \\omega ) \\}", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 104, + 557, + 507, + 573 + ], + "spans": [ + { + "bbox": [ + 104, + 557, + 126, + 573 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 126, + 559, + 170, + 571 + ], + "score": 0.93, + "content": "\\omega \\sim q _ { 0 } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 171, + 557, + 234, + 573 + ], + "score": 1.0, + "content": "matches that of", + "type": "text" + }, + { + "bbox": [ + 235, + 561, + 245, + 570 + ], + "score": 0.85, + "content": "\\mathbf { z } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 557, + 400, + 573 + ], + "score": 1.0, + "content": "in the original generating process with", + "type": "text" + }, + { + "bbox": [ + 401, + 559, + 408, + 570 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 409, + 557, + 507, + 573 + ], + "score": 1.0, + "content": "in Figure 1. In this way,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 568, + 506, + 584 + ], + "spans": [ + { + "bbox": [ + 104, + 568, + 210, + 584 + ], + "score": 1.0, + "content": "the generating process of", + "type": "text" + }, + { + "bbox": [ + 210, + 571, + 221, + 581 + ], + "score": 0.86, + "content": "\\mathbf { z } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 221, + 568, + 237, + 584 + ], + "score": 1.0, + "content": "via", + "type": "text" + }, + { + "bbox": [ + 237, + 570, + 248, + 581 + ], + "score": 0.87, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 568, + 428, + 584 + ], + "score": 1.0, + "content": "is distilled into the parameterized generator", + "type": "text" + }, + { + "bbox": [ + 428, + 570, + 452, + 582 + ], + "score": 0.91, + "content": "Q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 452, + 568, + 506, + 584 + ], + "score": 1.0, + "content": ", eliminating", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 579, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 281, + 594 + ], + "score": 1.0, + "content": "the need to do a specific transformation via", + "type": "text" + }, + { + "bbox": [ + 281, + 581, + 293, + 592 + ], + "score": 0.89, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 579, + 506, + 594 + ], + "score": 1.0, + "content": "in testing, and thus is very efficient. Specifically, we", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 591, + 186, + 604 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 135, + 604 + ], + "score": 1.0, + "content": "update", + "type": "text" + }, + { + "bbox": [ + 135, + 592, + 145, + 603 + ], + "score": 0.88, + "content": "\\phi ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 591, + 186, + 604 + ], + "score": 1.0, + "content": "such that", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 459, + 507, + 604 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 609, + 376, + 632 + ], + "lines": [ + { + "bbox": [ + 233, + 609, + 376, + 632 + ], + "spans": [ + { + "bbox": [ + 233, + 609, + 376, + 632 + ], + "score": 0.94, + "content": "\\phi ^ { \\prime } = \\arg \\operatorname* { m i n } _ { \\phi } \\mathcal { D } \\left( \\{ \\mathbf { z } _ { 0 } ^ { \\prime ( i ) } \\} , \\{ \\mathbf { z } _ { 1 } ^ { ( i ) } \\} \\right) ,", + "type": "interline_equation", + "image_path": "f1c3ff570f6a04c96384ca8c54209d1171a4e27623f34e24d4ac94b555ec213d.jpg" + } + ] + } + ], + "index": 38, + "virtual_lines": [ + { + "bbox": [ + 233, + 609, + 376, + 632 + ], + "spans": [], + "index": 38 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 639, + 505, + 702 + ], + "lines": [ + { + "bbox": [ + 102, + 636, + 508, + 659 + ], + "spans": [ + { + "bbox": [ + 102, + 636, + 132, + 659 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 639, + 173, + 654 + ], + "score": 0.94, + "content": "\\{ \\mathbf { z } _ { 0 } ^ { \\prime ( i ) } \\} _ { i = 1 } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 174, + 636, + 314, + 659 + ], + "score": 1.0, + "content": "are a set of samples generated from", + "type": "text" + }, + { + "bbox": [ + 314, + 641, + 357, + 655 + ], + "score": 0.92, + "content": "q _ { \\phi ^ { \\prime } } ( \\mathbf { z } _ { 0 } ^ { \\prime } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 636, + 372, + 659 + ], + "score": 1.0, + "content": "via", + "type": "text" + }, + { + "bbox": [ + 372, + 641, + 397, + 654 + ], + "score": 0.91, + "content": "Q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 636, + 417, + 659 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 417, + 639, + 456, + 654 + ], + "score": 0.94, + "content": "\\{ \\mathbf { z } _ { 1 } ^ { ( i ) } \\} _ { i = 1 } ^ { S }", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 636, + 508, + 659 + ], + "score": 1.0, + "content": "are samples", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 653, + 507, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 146, + 670 + ], + "score": 1.0, + "content": "drawn by", + "type": "text" + }, + { + "bbox": [ + 146, + 654, + 360, + 669 + ], + "score": 0.87, + "content": "\\omega ^ { i } \\sim q _ { 0 } ( \\omega ) , \\tilde { \\mathbf { z } } _ { 0 } ^ { i } = Q _ { \\phi } ( \\cdot | \\mathbf { x } , \\omega ^ { i } ) , \\mathbf { z } _ { 1 } ^ { ( i ) } \\sim \\mathcal { T } _ { 1 } ( \\tilde { \\mathbf { z } } _ { 0 } ^ { i } ) ; \\mathcal { D } ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 653, + 507, + 670 + ], + "score": 1.0, + "content": "is a metric between samples such as", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 667, + 504, + 680 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 504, + 680 + ], + "score": 1.0, + "content": "the simple Euclidean distance or the more advanced Wasserstein distance (Arjovsky et al., 2017). The", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 505, + 691 + ], + "score": 1.0, + "content": "optimization is done by applying standard stochastic gradient descent (SGD). We call this procedure", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 264, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 212, + 703 + ], + "score": 1.0, + "content": "distilling knowledge from", + "type": "text" + }, + { + "bbox": [ + 213, + 690, + 223, + 700 + ], + "score": 0.88, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 223, + 688, + 235, + 703 + ], + "score": 1.0, + "content": "to", + "type": "text" + }, + { + "bbox": [ + 235, + 689, + 259, + 702 + ], + "score": 0.91, + "content": "Q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 260, + 688, + 264, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 41, + "bbox_fs": [ + 102, + 636, + 508, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 232, + 96 + ], + "score": 1.0, + "content": "After distilling knowledge from", + "type": "text" + }, + { + "bbox": [ + 232, + 83, + 243, + 93 + ], + "score": 0.86, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 81, + 462, + 96 + ], + "score": 1.0, + "content": ", we apply the same procedure for other transformations", + "type": "text" + }, + { + "bbox": [ + 462, + 82, + 505, + 95 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { k } ( k > 1 )", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 343, + 108 + ], + "score": 1.0, + "content": "sequentially. The final inference network, represented by", + "type": "text" + }, + { + "bbox": [ + 343, + 93, + 375, + 106 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\cdot | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 93, + 506, + 108 + ], + "score": 1.0, + "content": ", can then well approximate the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 301, + 118 + ], + "score": 1.0, + "content": "continuous-time flows, e.g., the distribution of", + "type": "text" + }, + { + "bbox": [ + 302, + 105, + 359, + 117 + ], + "score": 0.92, + "content": "{ \\bf z } _ { 0 } \\sim q _ { \\phi } ( { \\bf \\cdot } | { \\bf \\dot { x } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 104, + 407, + 118 + ], + "score": 1.0, + "content": "is close to", + "type": "text" + }, + { + "bbox": [ + 407, + 106, + 420, + 116 + ], + "score": 0.83, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "from the CTF. This", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 452, + 129 + ], + "score": 1.0, + "content": "concept is illustrated in Figure 2. According to Theorem 2, the number of updates for", + "type": "text" + }, + { + "bbox": [ + 452, + 116, + 460, + 127 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "in training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 125, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 104, + 125, + 185, + 141 + ], + "score": 1.0, + "content": "is still bounded by", + "type": "text" + }, + { + "bbox": [ + 185, + 126, + 224, + 139 + ], + "score": 0.93, + "content": "O ( \\epsilon ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 125, + 252, + 141 + ], + "score": 1.0, + "content": "for an", + "type": "text" + }, + { + "bbox": [ + 253, + 130, + 258, + 137 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 125, + 506, + 141 + ], + "score": 1.0, + "content": "-accurate MSE, however, inference in testing is significantly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 139, + 502, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 502, + 151 + ], + "score": 1.0, + "content": "boosted since we do not need to simulate a long-time transformations as shown in Figure 1 (right).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 161, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 150, + 174 + ], + "score": 1.0, + "content": "Updating", + "type": "text" + }, + { + "bbox": [ + 150, + 162, + 158, + 172 + ], + "score": 0.58, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 161, + 193, + 174 + ], + "score": 1.0, + "content": "Given", + "type": "text" + }, + { + "bbox": [ + 193, + 162, + 214, + 173 + ], + "score": 0.33, + "content": "\\phi , \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 161, + 480, + 174 + ], + "score": 1.0, + "content": "can be updated by simply optimizing the ELBO in (7), where", + "type": "text" + }, + { + "bbox": [ + 480, + 163, + 493, + 173 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 161, + 506, + 174 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 180, + 186 + ], + "score": 1.0, + "content": "approximated by", + "type": "text" + }, + { + "bbox": [ + 180, + 174, + 193, + 184 + ], + "score": 0.87, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 172, + 453, + 186 + ], + "score": 1.0, + "content": "from the discretized CTF. Specifically, the expectation w.r.t.", + "type": "text" + }, + { + "bbox": [ + 453, + 174, + 466, + 184 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 172, + 506, + 186 + ], + "score": 1.0, + "content": "in (7) is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 272, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 272, + 196 + ], + "score": 1.0, + "content": "approximated by a sample average from:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 197, + 442, + 212 + ], + "lines": [ + { + "bbox": [ + 168, + 197, + 442, + 212 + ], + "spans": [ + { + "bbox": [ + 168, + 197, + 442, + 212 + ], + "score": 0.86, + "content": "\\begin{array} { r } { { \\bf z } _ { 0 } \\sim q _ { \\phi } ( { \\bf z } _ { 0 } \\mid { \\bf x } ) , { \\bf z } _ { 1 } \\sim { \\mathcal T } _ { 1 } ( { \\bf z } _ { 0 } ) , { \\bf z } _ { 2 } \\sim { \\mathcal T } _ { 2 } ( { \\bf z } _ { 1 } ) , \\cdots , { \\bf z } _ { K } \\sim { \\mathcal T } _ { K } ( { \\bf z } _ { K - 1 } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ca6eb3f8efef2212242ece3eceda8ffb9a6fbe7b955b4dfea6c99f354fa8ceb7.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 168, + 197, + 442, + 212 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image", + "bbox": [ + 108, + 223, + 505, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 223, + 505, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 223, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 108, + 223, + 505, + 285 + ], + "score": 0.958, + "type": "image", + "image_path": "b4fbce5997de9b3fa8366c2ff82ece19b528c5a3a6783274270ba1e5c8532dc8.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 223, + 505, + 243.66666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 243.66666666666666, + 505, + 264.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 264.3333333333333, + 505, + 285.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 292, + 506, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 292, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 506, + 304 + ], + "score": 1.0, + "content": "Figure 2: Amortized learning of continuous-time flows for VAEs. From left to right: the initial", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 175, + 316 + ], + "score": 1.0, + "content": "architecture with", + "type": "text" + }, + { + "bbox": [ + 176, + 304, + 186, + 313 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 303, + 328, + 316 + ], + "score": 1.0, + "content": "-step transformations; For each step", + "type": "text" + }, + { + "bbox": [ + 328, + 304, + 335, + 313 + ], + "score": 0.52, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 303, + 338, + 316 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 339, + 303, + 360, + 316 + ], + "score": 0.76, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "is trained to match the distributin of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 312, + 325, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 118, + 325 + ], + "score": 0.84, + "content": "\\mathbf { z } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 312, + 298, + 328 + ], + "score": 1.0, + "content": "in CTFs; In the end, the CTF is distilled into", + "type": "text" + }, + { + "bbox": [ + 299, + 314, + 320, + 327 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 312, + 325, + 328 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 380, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 380, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 380, + 348 + ], + "score": 1.0, + "content": "To sum up, there are three main steps in learning a CTF-based VAE:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 129, + 354, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 129, + 354, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 129, + 354, + 238, + 369 + ], + "score": 1.0, + "content": "1. Generate a sample path", + "type": "text" + }, + { + "bbox": [ + 239, + 356, + 293, + 367 + ], + "score": 0.9, + "content": "\\left( { \\bf z } _ { 0 } , \\cdots , { \\bf z } _ { K } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 354, + 346, + 369 + ], + "score": 1.0, + "content": "according to", + "type": "text" + }, + { + "bbox": [ + 346, + 355, + 386, + 368 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 354, + 505, + 369 + ], + "score": 1.0, + "content": "and the discretized flow with", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 141, + 366, + 234, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 207, + 379 + ], + "score": 1.0, + "content": "transformations", + "type": "text" + }, + { + "bbox": [ + 208, + 366, + 228, + 378 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 366, + 234, + 379 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 129, + 380, + 251, + 393 + ], + "spans": [ + { + "bbox": [ + 129, + 380, + 173, + 393 + ], + "score": 1.0, + "content": "2. Update", + "type": "text" + }, + { + "bbox": [ + 173, + 381, + 181, + 392 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 380, + 251, + 393 + ], + "score": 1.0, + "content": "according to (8);", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 129, + 394, + 437, + 408 + ], + "spans": [ + { + "bbox": [ + 129, + 394, + 182, + 408 + ], + "score": 1.0, + "content": "3. Optimize", + "type": "text" + }, + { + "bbox": [ + 182, + 395, + 189, + 405 + ], + "score": 0.7, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 394, + 437, + 408 + ], + "score": 1.0, + "content": "by minimizing the ELBO (7) with the generated sample path.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 108, + 414, + 506, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 281, + 427 + ], + "score": 1.0, + "content": "In testing, we use only the finally learned", + "type": "text" + }, + { + "bbox": [ + 281, + 415, + 321, + 427 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "for inference (into which the CTF has been", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "distilled), and hence testing is like the standard VAE. Since the discretized-CTF model is essentially", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 437, + 402, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 402, + 449 + ], + "score": 1.0, + "content": "a Markov chain, we call our model Markov-chain-based VAE (MacVAE).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22 + }, + { + "type": "title", + "bbox": [ + 107, + 463, + 457, + 476 + ], + "lines": [ + { + "bbox": [ + 104, + 462, + 460, + 478 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 460, + 478 + ], + "score": 1.0, + "content": "4 CONTINUOUS TIME FLOWS FOR EXPLICIT DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 506, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 507, + 501 + ], + "score": 1.0, + "content": "We describe how to apply the proposed CTF framework to density estimation of the observed data.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 498, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 299, + 512 + ], + "score": 1.0, + "content": "We assume that the density of the observation", + "type": "text" + }, + { + "bbox": [ + 299, + 501, + 307, + 509 + ], + "score": 0.49, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 498, + 506, + 512 + ], + "score": 1.0, + "content": "is characterized by a parametric Gibbsian-style", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 102, + 508, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 102, + 509, + 181, + 529 + ], + "score": 1.0, + "content": "probability model", + "type": "text" + }, + { + "bbox": [ + 181, + 510, + 325, + 526 + ], + "score": 0.93, + "content": "\\begin{array} { r } { p _ { \\pmb \\theta } ( \\mathbf x ) = \\frac { 1 } { \\mathcal Z ( \\pmb \\theta ) } \\tilde { p } _ { \\pmb \\theta } ( \\mathbf x ) \\triangleq \\frac { 1 } { \\mathcal Z ( \\pmb \\theta ) } e ^ { U ( \\mathbf x ; \\pmb \\theta ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 508, + 356, + 527 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 356, + 511, + 381, + 523 + ], + "score": 0.91, + "content": "\\tilde { p } _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 508, + 506, + 527 + ], + "score": 1.0, + "content": "is an unnormalized version of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 131, + 538 + ], + "score": 0.89, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 525, + 196, + 540 + ], + "score": 1.0, + "content": "with parameter", + "type": "text" + }, + { + "bbox": [ + 196, + 527, + 203, + 537 + ], + "score": 0.53, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 525, + 207, + 540 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 207, + 526, + 291, + 539 + ], + "score": 0.9, + "content": "U ( \\mathbf { x } ; \\pmb { \\theta } ) \\triangleq \\log \\tilde { p } \\pmb { \\theta } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "is called the energy function (Zhao et al., 2017), and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 107, + 538, + 188, + 552 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathcal { Z } ( \\pmb { \\theta } ) \\triangleq \\int \\tilde { p } _ { \\pmb { \\theta } } ( \\mathbf { x } ) \\mathrm { d } \\mathbf { x } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "is the normalizer. Note this form of distributions constitutes a very large class", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 549, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 104, + 549, + 506, + 564 + ], + "score": 1.0, + "content": "of distributions as long as the capacity of the energy function is large enough. This can be easily", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 277, + 574 + ], + "score": 1.0, + "content": "achieved by adopting a DNN to implement", + "type": "text" + }, + { + "bbox": [ + 277, + 561, + 309, + 573 + ], + "score": 0.92, + "content": "U ( \\mathbf { x } ; \\pmb { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 561, + 506, + 574 + ], + "score": 1.0, + "content": ", the setting we considered in this paper. Note our", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "model can be placed in between existing implicit and explicit density estimation methods, because", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 584, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 596 + ], + "score": 1.0, + "content": "we model the data density with an explicit distribution form up to an intractable normalizer. Such", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "distributions have been proved to be useful in real applications, e.g., Haarnoja et al. (2017) used them", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 305, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 305, + 619 + ], + "score": 1.0, + "content": "to model policies in deep reinforcement learning.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 503, + 644 + ], + "lines": [ + { + "bbox": [ + 104, + 619, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 185, + 638 + ], + "score": 1.0, + "content": "Our goal is to learn", + "type": "text" + }, + { + "bbox": [ + 185, + 623, + 192, + 632 + ], + "score": 0.74, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 619, + 269, + 638 + ], + "score": 1.0, + "content": "given observations", + "type": "text" + }, + { + "bbox": [ + 270, + 621, + 303, + 634 + ], + "score": 0.93, + "content": "\\{ { \\mathbf { x } } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 619, + 506, + 638 + ], + "score": 1.0, + "content": ", which can be achieved via the standard maximum", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 632, + 221, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 221, + 645 + ], + "score": 1.0, + "content": "likelihood estimator (MLE):", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5 + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 646, + 412, + 681 + ], + "lines": [ + { + "bbox": [ + 198, + 646, + 412, + 681 + ], + "spans": [ + { + "bbox": [ + 198, + 646, + 412, + 681 + ], + "score": 0.93, + "content": "\\pmb \\theta = \\arg \\operatorname* { m a x } _ { \\pmb { \\theta } } \\sum _ { i = 1 } ^ { N } \\log p _ { \\pmb { \\theta } } ( \\mathbf { x } _ { i } ) \\triangleq \\arg \\operatorname* { m a x } _ { \\pmb { \\theta } } \\mathcal { M } ( \\{ \\mathbf { x } _ { i } \\} ; \\pmb { \\theta } )", + "type": "interline_equation", + "image_path": "298e6b70f6a0c30dffc39c8fdf15e179700b8c8dc6aba4cdb2d5a65499dc4691.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 198, + 646, + 412, + 663.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 198, + 663.5, + 412, + 681.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 682, + 396, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 396, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 396, + 695 + ], + "score": 1.0, + "content": "This is usually optimized via SGD, with the following gradient formula:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40 + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 696, + 420, + 731 + ], + "lines": [ + { + "bbox": [ + 183, + 696, + 420, + 731 + ], + "spans": [ + { + "bbox": [ + 183, + 696, + 420, + 731 + ], + "score": 0.93, + "content": "\\frac { \\partial \\mathcal { M } ( \\{ { \\bf x } _ { i } \\} ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\frac { \\partial U ( { \\bf x } _ { i } ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } - \\mathbb { E } _ { p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } ) } \\left[ \\frac { \\partial U ( \\mathbf { x } ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } \\right]", + "type": "interline_equation", + "image_path": "2a6c979d1f1697faded0e1e4d562e033e41293c21774d0324ea18d82eaf5be8b.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 183, + 696, + 420, + 713.5 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 183, + 713.5, + 420, + 731.0 + ], + "spans": [], + "index": 42 + } + ] + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 505, + 150 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 232, + 96 + ], + "score": 1.0, + "content": "After distilling knowledge from", + "type": "text" + }, + { + "bbox": [ + 232, + 83, + 243, + 93 + ], + "score": 0.86, + "content": "\\mathcal { T } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 81, + 462, + 96 + ], + "score": 1.0, + "content": ", we apply the same procedure for other transformations", + "type": "text" + }, + { + "bbox": [ + 462, + 82, + 505, + 95 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { k } ( k > 1 )", + "type": "inline_equation" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 93, + 506, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 93, + 343, + 108 + ], + "score": 1.0, + "content": "sequentially. The final inference network, represented by", + "type": "text" + }, + { + "bbox": [ + 343, + 93, + 375, + 106 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\cdot | { \\bf x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 375, + 93, + 506, + 108 + ], + "score": 1.0, + "content": ", can then well approximate the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 506, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 301, + 118 + ], + "score": 1.0, + "content": "continuous-time flows, e.g., the distribution of", + "type": "text" + }, + { + "bbox": [ + 302, + 105, + 359, + 117 + ], + "score": 0.92, + "content": "{ \\bf z } _ { 0 } \\sim q _ { \\phi } ( { \\bf \\cdot } | { \\bf \\dot { x } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 104, + 407, + 118 + ], + "score": 1.0, + "content": "is close to", + "type": "text" + }, + { + "bbox": [ + 407, + 106, + 420, + 116 + ], + "score": 0.83, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 420, + 104, + 506, + 118 + ], + "score": 1.0, + "content": "from the CTF. This", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 115, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 105, + 115, + 452, + 129 + ], + "score": 1.0, + "content": "concept is illustrated in Figure 2. According to Theorem 2, the number of updates for", + "type": "text" + }, + { + "bbox": [ + 452, + 116, + 460, + 127 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 115, + 505, + 129 + ], + "score": 1.0, + "content": "in training", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 125, + 506, + 141 + ], + "spans": [ + { + "bbox": [ + 104, + 125, + 185, + 141 + ], + "score": 1.0, + "content": "is still bounded by", + "type": "text" + }, + { + "bbox": [ + 185, + 126, + 224, + 139 + ], + "score": 0.93, + "content": "O ( \\epsilon ^ { - 3 / 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 125, + 252, + 141 + ], + "score": 1.0, + "content": "for an", + "type": "text" + }, + { + "bbox": [ + 253, + 130, + 258, + 137 + ], + "score": 0.67, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 125, + 506, + 141 + ], + "score": 1.0, + "content": "-accurate MSE, however, inference in testing is significantly", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 139, + 502, + 151 + ], + "spans": [ + { + "bbox": [ + 106, + 139, + 502, + 151 + ], + "score": 1.0, + "content": "boosted since we do not need to simulate a long-time transformations as shown in Figure 1 (right).", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 81, + 506, + 151 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 161, + 504, + 195 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 150, + 174 + ], + "score": 1.0, + "content": "Updating", + "type": "text" + }, + { + "bbox": [ + 150, + 162, + 158, + 172 + ], + "score": 0.58, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 161, + 193, + 174 + ], + "score": 1.0, + "content": "Given", + "type": "text" + }, + { + "bbox": [ + 193, + 162, + 214, + 173 + ], + "score": 0.33, + "content": "\\phi , \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 214, + 161, + 480, + 174 + ], + "score": 1.0, + "content": "can be updated by simply optimizing the ELBO in (7), where", + "type": "text" + }, + { + "bbox": [ + 480, + 163, + 493, + 173 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 161, + 506, + 174 + ], + "score": 1.0, + "content": "is", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 180, + 186 + ], + "score": 1.0, + "content": "approximated by", + "type": "text" + }, + { + "bbox": [ + 180, + 174, + 193, + 184 + ], + "score": 0.87, + "content": "\\bar { \\rho } _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 193, + 172, + 453, + 186 + ], + "score": 1.0, + "content": "from the discretized CTF. Specifically, the expectation w.r.t.", + "type": "text" + }, + { + "bbox": [ + 453, + 174, + 466, + 184 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 466, + 172, + 506, + 186 + ], + "score": 1.0, + "content": "in (7) is", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 184, + 272, + 196 + ], + "spans": [ + { + "bbox": [ + 106, + 184, + 272, + 196 + ], + "score": 1.0, + "content": "approximated by a sample average from:", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 161, + 506, + 196 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 168, + 197, + 442, + 212 + ], + "lines": [ + { + "bbox": [ + 168, + 197, + 442, + 212 + ], + "spans": [ + { + "bbox": [ + 168, + 197, + 442, + 212 + ], + "score": 0.86, + "content": "\\begin{array} { r } { { \\bf z } _ { 0 } \\sim q _ { \\phi } ( { \\bf z } _ { 0 } \\mid { \\bf x } ) , { \\bf z } _ { 1 } \\sim { \\mathcal T } _ { 1 } ( { \\bf z } _ { 0 } ) , { \\bf z } _ { 2 } \\sim { \\mathcal T } _ { 2 } ( { \\bf z } _ { 1 } ) , \\cdots , { \\bf z } _ { K } \\sim { \\mathcal T } _ { K } ( { \\bf z } _ { K - 1 } ) . } \\end{array}", + "type": "interline_equation", + "image_path": "ca6eb3f8efef2212242ece3eceda8ffb9a6fbe7b955b4dfea6c99f354fa8ceb7.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 168, + 197, + 442, + 212 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image", + "bbox": [ + 108, + 223, + 505, + 285 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 108, + 223, + 505, + 285 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 108, + 223, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 108, + 223, + 505, + 285 + ], + "score": 0.958, + "type": "image", + "image_path": "b4fbce5997de9b3fa8366c2ff82ece19b528c5a3a6783274270ba1e5c8532dc8.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 108, + 223, + 505, + 243.66666666666666 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 108, + 243.66666666666666, + 505, + 264.3333333333333 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 108, + 264.3333333333333, + 505, + 285.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 292, + 506, + 326 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 292, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 106, + 292, + 506, + 304 + ], + "score": 1.0, + "content": "Figure 2: Amortized learning of continuous-time flows for VAEs. From left to right: the initial", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 303, + 506, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 175, + 316 + ], + "score": 1.0, + "content": "architecture with", + "type": "text" + }, + { + "bbox": [ + 176, + 304, + 186, + 313 + ], + "score": 0.83, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 186, + 303, + 328, + 316 + ], + "score": 1.0, + "content": "-step transformations; For each step", + "type": "text" + }, + { + "bbox": [ + 328, + 304, + 335, + 313 + ], + "score": 0.52, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 335, + 303, + 338, + 316 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 339, + 303, + 360, + 316 + ], + "score": 0.76, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 360, + 303, + 506, + 316 + ], + "score": 1.0, + "content": "is trained to match the distributin of", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 312, + 325, + 328 + ], + "spans": [ + { + "bbox": [ + 106, + 315, + 118, + 325 + ], + "score": 0.84, + "content": "\\mathbf { z } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 312, + 298, + 328 + ], + "score": 1.0, + "content": "in CTFs; In the end, the CTF is distilled into", + "type": "text" + }, + { + "bbox": [ + 299, + 314, + 320, + 327 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 312, + 325, + 328 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 380, + 347 + ], + "lines": [ + { + "bbox": [ + 105, + 334, + 380, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 334, + 380, + 348 + ], + "score": 1.0, + "content": "To sum up, there are three main steps in learning a CTF-based VAE:", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 334, + 380, + 348 + ] + }, + { + "type": "list", + "bbox": [ + 129, + 354, + 504, + 407 + ], + "lines": [ + { + "bbox": [ + 129, + 354, + 505, + 369 + ], + "spans": [ + { + "bbox": [ + 129, + 354, + 238, + 369 + ], + "score": 1.0, + "content": "1. Generate a sample path", + "type": "text" + }, + { + "bbox": [ + 239, + 356, + 293, + 367 + ], + "score": 0.9, + "content": "\\left( { \\bf z } _ { 0 } , \\cdots , { \\bf z } _ { K } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 293, + 354, + 346, + 369 + ], + "score": 1.0, + "content": "according to", + "type": "text" + }, + { + "bbox": [ + 346, + 355, + 386, + 368 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 387, + 354, + 505, + 369 + ], + "score": 1.0, + "content": "and the discretized flow with", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 141, + 366, + 234, + 379 + ], + "spans": [ + { + "bbox": [ + 141, + 366, + 207, + 379 + ], + "score": 1.0, + "content": "transformations", + "type": "text" + }, + { + "bbox": [ + 208, + 366, + 228, + 378 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 229, + 366, + 234, + 379 + ], + "score": 1.0, + "content": ";", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 380, + 251, + 393 + ], + "spans": [ + { + "bbox": [ + 129, + 380, + 173, + 393 + ], + "score": 1.0, + "content": "2. Update", + "type": "text" + }, + { + "bbox": [ + 173, + 381, + 181, + 392 + ], + "score": 0.84, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 380, + 251, + 393 + ], + "score": 1.0, + "content": "according to (8);", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 129, + 394, + 437, + 408 + ], + "spans": [ + { + "bbox": [ + 129, + 394, + 182, + 408 + ], + "score": 1.0, + "content": "3. Optimize", + "type": "text" + }, + { + "bbox": [ + 182, + 395, + 189, + 405 + ], + "score": 0.7, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 190, + 394, + 437, + 408 + ], + "score": 1.0, + "content": "by minimizing the ELBO (7) with the generated sample path.", + "type": "text" + } + ], + "index": 20, + "is_list_start_line": true, + "is_list_end_line": true + } + ], + "index": 18.5, + "bbox_fs": [ + 129, + 354, + 505, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 414, + 506, + 448 + ], + "lines": [ + { + "bbox": [ + 105, + 414, + 506, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 281, + 427 + ], + "score": 1.0, + "content": "In testing, we use only the finally learned", + "type": "text" + }, + { + "bbox": [ + 281, + 415, + 321, + 427 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\mathbf { z } _ { 0 } \\mid \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 414, + 506, + 427 + ], + "score": 1.0, + "content": "for inference (into which the CTF has been", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 505, + 438 + ], + "score": 1.0, + "content": "distilled), and hence testing is like the standard VAE. Since the discretized-CTF model is essentially", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 437, + 402, + 449 + ], + "spans": [ + { + "bbox": [ + 106, + 437, + 402, + 449 + ], + "score": 1.0, + "content": "a Markov chain, we call our model Markov-chain-based VAE (MacVAE).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 414, + 506, + 449 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 463, + 457, + 476 + ], + "lines": [ + { + "bbox": [ + 104, + 462, + 460, + 478 + ], + "spans": [ + { + "bbox": [ + 104, + 462, + 460, + 478 + ], + "score": 1.0, + "content": "4 CONTINUOUS TIME FLOWS FOR EXPLICIT DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 106, + 487, + 506, + 617 + ], + "lines": [ + { + "bbox": [ + 106, + 488, + 507, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 507, + 501 + ], + "score": 1.0, + "content": "We describe how to apply the proposed CTF framework to density estimation of the observed data.", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 498, + 506, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 299, + 512 + ], + "score": 1.0, + "content": "We assume that the density of the observation", + "type": "text" + }, + { + "bbox": [ + 299, + 501, + 307, + 509 + ], + "score": 0.49, + "content": "\\mathbf { x }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 498, + 506, + 512 + ], + "score": 1.0, + "content": "is characterized by a parametric Gibbsian-style", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 102, + 508, + 506, + 529 + ], + "spans": [ + { + "bbox": [ + 102, + 509, + 181, + 529 + ], + "score": 1.0, + "content": "probability model", + "type": "text" + }, + { + "bbox": [ + 181, + 510, + 325, + 526 + ], + "score": 0.93, + "content": "\\begin{array} { r } { p _ { \\pmb \\theta } ( \\mathbf x ) = \\frac { 1 } { \\mathcal Z ( \\pmb \\theta ) } \\tilde { p } _ { \\pmb \\theta } ( \\mathbf x ) \\triangleq \\frac { 1 } { \\mathcal Z ( \\pmb \\theta ) } e ^ { U ( \\mathbf x ; \\pmb \\theta ) } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 325, + 508, + 356, + 527 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 356, + 511, + 381, + 523 + ], + "score": 0.91, + "content": "\\tilde { p } _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 508, + 506, + 527 + ], + "score": 1.0, + "content": "is an unnormalized version of", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 525, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 526, + 131, + 538 + ], + "score": 0.89, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 132, + 525, + 196, + 540 + ], + "score": 1.0, + "content": "with parameter", + "type": "text" + }, + { + "bbox": [ + 196, + 527, + 203, + 537 + ], + "score": 0.53, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 203, + 525, + 207, + 540 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 207, + 526, + 291, + 539 + ], + "score": 0.9, + "content": "U ( \\mathbf { x } ; \\pmb { \\theta } ) \\triangleq \\log \\tilde { p } \\pmb { \\theta } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 525, + 506, + 540 + ], + "score": 1.0, + "content": "is called the energy function (Zhao et al., 2017), and", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 107, + 538, + 506, + 552 + ], + "spans": [ + { + "bbox": [ + 107, + 538, + 188, + 552 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\mathcal { Z } ( \\pmb { \\theta } ) \\triangleq \\int \\tilde { p } _ { \\pmb { \\theta } } ( \\mathbf { x } ) \\mathrm { d } \\mathbf { x } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 538, + 506, + 552 + ], + "score": 1.0, + "content": "is the normalizer. Note this form of distributions constitutes a very large class", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 104, + 549, + 506, + 564 + ], + "spans": [ + { + "bbox": [ + 104, + 549, + 506, + 564 + ], + "score": 1.0, + "content": "of distributions as long as the capacity of the energy function is large enough. This can be easily", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 561, + 506, + 574 + ], + "spans": [ + { + "bbox": [ + 105, + 561, + 277, + 574 + ], + "score": 1.0, + "content": "achieved by adopting a DNN to implement", + "type": "text" + }, + { + "bbox": [ + 277, + 561, + 309, + 573 + ], + "score": 0.92, + "content": "U ( \\mathbf { x } ; \\pmb { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 561, + 506, + 574 + ], + "score": 1.0, + "content": ", the setting we considered in this paper. Note our", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "spans": [ + { + "bbox": [ + 105, + 572, + 506, + 585 + ], + "score": 1.0, + "content": "model can be placed in between existing implicit and explicit density estimation methods, because", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 584, + 505, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 505, + 596 + ], + "score": 1.0, + "content": "we model the data density with an explicit distribution form up to an intractable normalizer. Such", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 606 + ], + "score": 1.0, + "content": "distributions have been proved to be useful in real applications, e.g., Haarnoja et al. (2017) used them", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 604, + 305, + 619 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 305, + 619 + ], + "score": 1.0, + "content": "to model policies in deep reinforcement learning.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 30, + "bbox_fs": [ + 102, + 488, + 507, + 619 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 621, + 503, + 644 + ], + "lines": [ + { + "bbox": [ + 104, + 619, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 104, + 619, + 185, + 638 + ], + "score": 1.0, + "content": "Our goal is to learn", + "type": "text" + }, + { + "bbox": [ + 185, + 623, + 192, + 632 + ], + "score": 0.74, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 192, + 619, + 269, + 638 + ], + "score": 1.0, + "content": "given observations", + "type": "text" + }, + { + "bbox": [ + 270, + 621, + 303, + 634 + ], + "score": 0.93, + "content": "\\{ { \\mathbf { x } } _ { i } \\} _ { i = 1 } ^ { N }", + "type": "inline_equation" + }, + { + "bbox": [ + 303, + 619, + 506, + 638 + ], + "score": 1.0, + "content": ", which can be achieved via the standard maximum", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 632, + 221, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 632, + 221, + 645 + ], + "score": 1.0, + "content": "likelihood estimator (MLE):", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36.5, + "bbox_fs": [ + 104, + 619, + 506, + 645 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 198, + 646, + 412, + 681 + ], + "lines": [ + { + "bbox": [ + 198, + 646, + 412, + 681 + ], + "spans": [ + { + "bbox": [ + 198, + 646, + 412, + 681 + ], + "score": 0.93, + "content": "\\pmb \\theta = \\arg \\operatorname* { m a x } _ { \\pmb { \\theta } } \\sum _ { i = 1 } ^ { N } \\log p _ { \\pmb { \\theta } } ( \\mathbf { x } _ { i } ) \\triangleq \\arg \\operatorname* { m a x } _ { \\pmb { \\theta } } \\mathcal { M } ( \\{ \\mathbf { x } _ { i } \\} ; \\pmb { \\theta } )", + "type": "interline_equation", + "image_path": "298e6b70f6a0c30dffc39c8fdf15e179700b8c8dc6aba4cdb2d5a65499dc4691.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 198, + 646, + 412, + 663.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 198, + 663.5, + 412, + 681.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 682, + 396, + 694 + ], + "lines": [ + { + "bbox": [ + 105, + 681, + 396, + 695 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 396, + 695 + ], + "score": 1.0, + "content": "This is usually optimized via SGD, with the following gradient formula:", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 40, + "bbox_fs": [ + 105, + 681, + 396, + 695 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 183, + 696, + 420, + 731 + ], + "lines": [ + { + "bbox": [ + 183, + 696, + 420, + 731 + ], + "spans": [ + { + "bbox": [ + 183, + 696, + 420, + 731 + ], + "score": 0.93, + "content": "\\frac { \\partial \\mathcal { M } ( \\{ { \\bf x } _ { i } \\} ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } = \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\frac { \\partial U ( { \\bf x } _ { i } ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } - \\mathbb { E } _ { p _ { \\boldsymbol { \\theta } } ( \\mathbf { x } ) } \\left[ \\frac { \\partial U ( \\mathbf { x } ; \\mathbf { \\boldsymbol { \\theta } } ) } { \\partial \\mathbf { \\boldsymbol { \\theta } } } \\right]", + "type": "interline_equation", + "image_path": "2a6c979d1f1697faded0e1e4d562e033e41293c21774d0324ea18d82eaf5be8b.jpg" + } + ] + } + ], + "index": 41.5, + "virtual_lines": [ + { + "bbox": [ + 183, + 696, + 420, + 713.5 + ], + "spans": [], + "index": 41 + }, + { + "bbox": [ + 183, + 713.5, + 420, + 731.0 + ], + "spans": [], + "index": 42 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 456, + 95 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 456, + 95 + ], + "spans": [], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 112, + 86, + 506, + 200 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 112, + 86, + 506, + 200 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 86, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 112, + 86, + 506, + 200 + ], + "score": 0.456, + "html": "
Igorithm1 CTFs for generative models at the k-th iteration. D(·,:)is the same as (8). Input: parameters from last step 0(k-1),Φ(k-1) Output: updated parameters 0(k),𝜙(k)
1. Generate samples {X1,s}𝑠=1 via a discretized CTF: x0,s ~ q(k-1)(x0), X1,s ~ T1(x0,s); S 2. Update the generator by minimizing ({xo,s}S=1 are generated with the updated parameter (k):
(k) = arg minD({x1,s},{x6,s}） .
? 3. Update the energy-based model 0k by maximum likelihood, with gradient as (9) except replacing Ex~pe(x) with Ex~q(x);
", + "type": "table", + "image_path": "228270436d54101d23ac214311b62400ab48ffff489d72dfea30e56ca0e75f45.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 112, + 86, + 506, + 124.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 124.0, + 506, + 162.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 162.0, + 506, + 200.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 415, + 241 + ], + "score": 1.0, + "content": "The gradient formula requires an integration over the model distribution", + "type": "text" + }, + { + "bbox": [ + 415, + 228, + 440, + 240 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 227, + 505, + 241 + ], + "score": 1.0, + "content": ", which can be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "approximated by Monte Carlo integration with samples. The sampling problem has been well studied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "for some particular energy-based distributions, for example, via contrastive divergence in restricted", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 259, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 274 + ], + "score": 1.0, + "content": "Boltzmann machines (Hinton, 2002). However, this does not fit into our setting directly. Here we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "adopt the idea of CTFs and propose to use a DNN guided by a CTF, which we call a generator, to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 327, + 296 + ], + "score": 1.0, + "content": "generate approximate samples from the original model", + "type": "text" + }, + { + "bbox": [ + 327, + 283, + 352, + 294 + ], + "score": 0.9, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 282, + 506, + 296 + ], + "score": 1.0, + "content": ". Specifically, we require that samples", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 293, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 321, + 307 + ], + "score": 1.0, + "content": "from the generator should well approximate the target", + "type": "text" + }, + { + "bbox": [ + 322, + 294, + 347, + 305 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 293, + 506, + 307 + ], + "score": 1.0, + "content": ". This can be done by adopting the CTF", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 303, + 507, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 367, + 318 + ], + "score": 1.0, + "content": "idea above, i.e., distilling knowledge of a CTF (which approaches", + "type": "text" + }, + { + "bbox": [ + 368, + 305, + 395, + 316 + ], + "score": 0.89, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 303, + 507, + 318 + ], + "score": 1.0, + "content": "to the generator. In testing,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 315, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 250, + 329 + ], + "score": 1.0, + "content": "instead of generating samples from", + "type": "text" + }, + { + "bbox": [ + 250, + 316, + 275, + 327 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 315, + 506, + 329 + ], + "score": 1.0, + "content": "via MCMC (which is complicated and time consuming),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "we generate samples from the generator directly. Furthermore, when evaluating the likelihood for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 232, + 352 + ], + "score": 1.0, + "content": "test data, the unknown constant", + "type": "text" + }, + { + "bbox": [ + 232, + 337, + 255, + 349 + ], + "score": 0.92, + "content": "\\mathcal { Z } ( \\pmb \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 337, + 267, + 352 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 267, + 337, + 292, + 349 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "can also be approximated by Monte Carlo integration", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 348, + 267, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 267, + 361 + ], + "score": 1.0, + "content": "with samples drawn from the generator.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5 + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "On the right side of (9), the first term is a model fit to observed data, and the (negative) second term", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 279, + 389 + ], + "score": 1.0, + "content": "is a model fit to synthetic data drawn from", + "type": "text" + }, + { + "bbox": [ + 279, + 378, + 301, + 389 + ], + "score": 0.89, + "content": "p _ { \\pmb { \\theta } ( \\mathbf { x } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "; this is similar to the critic/discriminator in GANs", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 502, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 502, + 399 + ], + "score": 1.0, + "content": "(Arjovsky et al., 2017), but derived directly from the MLE. More connections are discussed below.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17 + }, + { + "type": "title", + "bbox": [ + 108, + 417, + 265, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 267, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 267, + 429 + ], + "score": 1.0, + "content": "4.1 LEARNING VIA AMORTIZATION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 439, + 339, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 339, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 339, + 451 + ], + "score": 1.0, + "content": "Our goal is to learn a generator whose generated samples", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 449, + 339, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 248, + 462 + ], + "score": 1.0, + "content": "match those from the original model", + "type": "text" + }, + { + "bbox": [ + 248, + 450, + 273, + 462 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 449, + 339, + 462 + ], + "score": 1.0, + "content": ", by adopting the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 461, + 340, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 340, + 473 + ], + "score": 1.0, + "content": "amortization idea with CTF in the inference section above.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 471, + 341, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 341, + 484 + ], + "score": 1.0, + "content": "Similar to inference, the generator is learned implicitly.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 483, + 339, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 339, + 495 + ], + "score": 1.0, + "content": "However, we also learn an explicit density model for the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 494, + 339, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 339, + 506 + ], + "score": 1.0, + "content": "data by SGD, with samples from the implicit generator to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 506, + 339, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 339, + 516 + ], + "score": 1.0, + "content": "estimate gradients in (9). Note that in this case, the CTF is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 516, + 340, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 340, + 528 + ], + "score": 1.0, + "content": "performed directly on the data space, instead of on latent-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 527, + 339, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 339, + 539 + ], + "score": 1.0, + "content": "variable space as in previous sections. Specifically, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 538, + 341, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 341, + 550 + ], + "score": 1.0, + "content": "sampling procedure from the generator plus a continuous-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 549, + 268, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 268, + 560 + ], + "score": 1.0, + "content": "time-flow transformation are written as:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25 + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 569, + 286, + 584 + ], + "lines": [ + { + "bbox": [ + 158, + 569, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 158, + 569, + 286, + 584 + ], + "score": 0.89, + "content": "\\mathbf { x } _ { 0 } \\sim q _ { \\phi } ( \\mathbf { x } _ { 0 } ) , \\mathbf { x } _ { T } \\sim { \\mathcal { T } } ( \\mathbf { x } _ { 0 } , T ) .", + "type": "interline_equation", + "image_path": "5b1559c7588790150c2a8adf05d421683bbb3b9ae7a96bf207af30fbda846a55.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 158, + 569, + 286, + 584 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image", + "bbox": [ + 346, + 426, + 492, + 525 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 346, + 426, + 492, + 525 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 346, + 426, + 492, + 525 + ], + "spans": [ + { + "bbox": [ + 346, + 426, + 492, + 525 + ], + "score": 0.967, + "type": "image", + "image_path": "6b4afaf4a45229a804fe66cbebef9ba63504a9ce81f293051a8bbd6f04a06cfb.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 346, + 426, + 492, + 475.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 346, + 475.5, + 492, + 525.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 346, + 537, + 505, + 581 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 346, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 346, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Figure 3: Learning a generator with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 345, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 345, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "CTF. The goal is to match the samples", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 346, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 346, + 560, + 358, + 570 + ], + "score": 0.83, + "content": "\\mathbf { x } _ { \\mathrm { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 559, + 382, + 572 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 383, + 561, + 394, + 571 + ], + "score": 0.86, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 559, + 482, + 572 + ], + "score": 1.0, + "content": "to those after a CTF", + "type": "text" + }, + { + "bbox": [ + 483, + 560, + 502, + 570 + ], + "score": 0.8, + "content": "\\left( { { \\bf { x } } _ { T } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 559, + 506, + 572 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 345, + 569, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 345, + 569, + 465, + 584 + ], + "score": 1.0, + "content": "or equivalently samples from", + "type": "text" + }, + { + "bbox": [ + 465, + 572, + 476, + 582 + ], + "score": 0.69, + "content": "p _ { \\pmb { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 569, + 481, + 584 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "index": 34.0 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 505, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 128, + 606 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 129, + 593, + 155, + 605 + ], + "score": 0.92, + "content": "\\tau ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 592, + 311, + 606 + ], + "score": 1.0, + "content": "is the continuous-time flow; a sample", + "type": "text" + }, + { + "bbox": [ + 312, + 595, + 324, + 604 + ], + "score": 0.87, + "content": "\\mathbf { x } _ { \\mathrm { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 592, + 347, + 606 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 348, + 593, + 369, + 605 + ], + "score": 0.91, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "is implemented by a deep neural", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 191, + 617 + ], + "score": 1.0, + "content": "network (generator)", + "type": "text" + }, + { + "bbox": [ + 191, + 604, + 220, + 616 + ], + "score": 0.93, + "content": "G _ { \\phi } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 603, + 267, + 617 + ], + "score": 1.0, + "content": "with input", + "type": "text" + }, + { + "bbox": [ + 268, + 604, + 315, + 616 + ], + "score": 0.93, + "content": "\\omega \\sim q _ { 0 } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 603, + 347, + 617 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 348, + 606, + 358, + 615 + ], + "score": 0.86, + "content": "q _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "is a simple distribution for a noise", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "random variable, e.g., the standard isotropic normal distribution. The procedure is illustrated in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "Figure 3. Note the CTF cannot be replaced by standard normalizing flow (Rezende & Mohamed,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "2015) in this model, because there is no objective function to guide the update of parameters in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 317, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 317, + 660 + ], + "score": 1.0, + "content": "normalizing flows, which is not necessary for CTFs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 106, + 664, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 268, + 677 + ], + "score": 1.0, + "content": "Specifically, denote the parameters in the", + "type": "text" + }, + { + "bbox": [ + 268, + 665, + 275, + 675 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 664, + 432, + 677 + ], + "score": 1.0, + "content": "-th step of our algorithm with subscript “", + "type": "text" + }, + { + "bbox": [ + 433, + 664, + 449, + 676 + ], + "score": 0.73, + "content": "( k ) ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 664, + 506, + 677 + ], + "score": 1.0, + "content": ". For efficient", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 208, + 689 + ], + "score": 1.0, + "content": "sample generation, in the", + "type": "text" + }, + { + "bbox": [ + 209, + 676, + 215, + 686 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "-th step, we again adopt the amortization idea from Section 3.2 to update", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 685, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 135, + 699 + ], + "score": 0.91, + "content": "\\phi ^ { ( k ^ { - 1 ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 685, + 239, + 702 + ], + "score": 1.0, + "content": "of the generator network", + "type": "text" + }, + { + "bbox": [ + 239, + 687, + 263, + 700 + ], + "score": 0.91, + "content": "G _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 685, + 506, + 702 + ], + "score": 1.0, + "content": ", such that samples from the updated generator match those", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 697, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 697, + 371, + 713 + ], + "score": 1.0, + "content": "from the current generator followed by a one-step transformation", + "type": "text" + }, + { + "bbox": [ + 371, + 699, + 392, + 711 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { 1 } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 697, + 441, + 713 + ], + "score": 1.0, + "content": ". After that,", + "type": "text" + }, + { + "bbox": [ + 441, + 699, + 448, + 709 + ], + "score": 0.72, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 697, + 505, + 713 + ], + "score": 1.0, + "content": "is updated by", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 196, + 723 + ], + "score": 1.0, + "content": "drawing samples from", + "type": "text" + }, + { + "bbox": [ + 197, + 710, + 218, + 722 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "to estimate the expectation in (9). The detailed algorithm is presented in", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 161, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 161, + 734 + ], + "score": 1.0, + "content": "Algorithm 1.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 456, + 95 + ], + "lines": [ + { + "bbox": [ + 107, + 82, + 456, + 95 + ], + "spans": [], + "index": 0 + } + ], + "index": 0, + "bbox_fs": [ + 107, + 82, + 456, + 95 + ] + }, + { + "type": "table", + "bbox": [ + 112, + 86, + 506, + 200 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 112, + 86, + 506, + 200 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 112, + 86, + 506, + 200 + ], + "spans": [ + { + "bbox": [ + 112, + 86, + 506, + 200 + ], + "score": 0.456, + "html": "
Igorithm1 CTFs for generative models at the k-th iteration. D(·,:)is the same as (8). Input: parameters from last step 0(k-1),Φ(k-1) Output: updated parameters 0(k),𝜙(k)
1. Generate samples {X1,s}𝑠=1 via a discretized CTF: x0,s ~ q(k-1)(x0), X1,s ~ T1(x0,s); S 2. Update the generator by minimizing ({xo,s}S=1 are generated with the updated parameter (k):
(k) = arg minD({x1,s},{x6,s}） .
? 3. Update the energy-based model 0k by maximum likelihood, with gradient as (9) except replacing Ex~pe(x) with Ex~q(x);
", + "type": "table", + "image_path": "228270436d54101d23ac214311b62400ab48ffff489d72dfea30e56ca0e75f45.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 112, + 86, + 506, + 124.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 112, + 124.0, + 506, + 162.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 112, + 162.0, + 506, + 200.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 360 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 415, + 241 + ], + "score": 1.0, + "content": "The gradient formula requires an integration over the model distribution", + "type": "text" + }, + { + "bbox": [ + 415, + 228, + 440, + 240 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 227, + 505, + 241 + ], + "score": 1.0, + "content": ", which can be", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 506, + 252 + ], + "score": 1.0, + "content": "approximated by Monte Carlo integration with samples. The sampling problem has been well studied", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "spans": [ + { + "bbox": [ + 105, + 249, + 506, + 263 + ], + "score": 1.0, + "content": "for some particular energy-based distributions, for example, via contrastive divergence in restricted", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 259, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 259, + 506, + 274 + ], + "score": 1.0, + "content": "Boltzmann machines (Hinton, 2002). However, this does not fit into our setting directly. Here we", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "spans": [ + { + "bbox": [ + 106, + 271, + 505, + 284 + ], + "score": 1.0, + "content": "adopt the idea of CTFs and propose to use a DNN guided by a CTF, which we call a generator, to", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 282, + 506, + 296 + ], + "spans": [ + { + "bbox": [ + 105, + 282, + 327, + 296 + ], + "score": 1.0, + "content": "generate approximate samples from the original model", + "type": "text" + }, + { + "bbox": [ + 327, + 283, + 352, + 294 + ], + "score": 0.9, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 282, + 506, + 296 + ], + "score": 1.0, + "content": ". Specifically, we require that samples", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 293, + 506, + 307 + ], + "spans": [ + { + "bbox": [ + 105, + 293, + 321, + 307 + ], + "score": 1.0, + "content": "from the generator should well approximate the target", + "type": "text" + }, + { + "bbox": [ + 322, + 294, + 347, + 305 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 347, + 293, + 506, + 307 + ], + "score": 1.0, + "content": ". This can be done by adopting the CTF", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 303, + 507, + 318 + ], + "spans": [ + { + "bbox": [ + 105, + 303, + 367, + 318 + ], + "score": 1.0, + "content": "idea above, i.e., distilling knowledge of a CTF (which approaches", + "type": "text" + }, + { + "bbox": [ + 368, + 305, + 395, + 316 + ], + "score": 0.89, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 395, + 303, + 507, + 318 + ], + "score": 1.0, + "content": "to the generator. In testing,", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 315, + 506, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 315, + 250, + 329 + ], + "score": 1.0, + "content": "instead of generating samples from", + "type": "text" + }, + { + "bbox": [ + 250, + 316, + 275, + 327 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 315, + 506, + 329 + ], + "score": 1.0, + "content": "via MCMC (which is complicated and time consuming),", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 326, + 506, + 339 + ], + "score": 1.0, + "content": "we generate samples from the generator directly. Furthermore, when evaluating the likelihood for", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 337, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 337, + 232, + 352 + ], + "score": 1.0, + "content": "test data, the unknown constant", + "type": "text" + }, + { + "bbox": [ + 232, + 337, + 255, + 349 + ], + "score": 0.92, + "content": "\\mathcal { Z } ( \\pmb \\theta )", + "type": "inline_equation" + }, + { + "bbox": [ + 255, + 337, + 267, + 352 + ], + "score": 1.0, + "content": "of", + "type": "text" + }, + { + "bbox": [ + 267, + 337, + 292, + 349 + ], + "score": 0.91, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 292, + 337, + 506, + 352 + ], + "score": 1.0, + "content": "can also be approximated by Monte Carlo integration", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 348, + 267, + 361 + ], + "spans": [ + { + "bbox": [ + 105, + 348, + 267, + 361 + ], + "score": 1.0, + "content": "with samples drawn from the generator.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 9.5, + "bbox_fs": [ + 105, + 227, + 507, + 361 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 365, + 505, + 399 + ], + "lines": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 105, + 364, + 505, + 378 + ], + "score": 1.0, + "content": "On the right side of (9), the first term is a model fit to observed data, and the (negative) second term", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 376, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 104, + 376, + 279, + 389 + ], + "score": 1.0, + "content": "is a model fit to synthetic data drawn from", + "type": "text" + }, + { + "bbox": [ + 279, + 378, + 301, + 389 + ], + "score": 0.89, + "content": "p _ { \\pmb { \\theta } ( \\mathbf { x } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 376, + 505, + 389 + ], + "score": 1.0, + "content": "; this is similar to the critic/discriminator in GANs", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 387, + 502, + 399 + ], + "spans": [ + { + "bbox": [ + 105, + 387, + 502, + 399 + ], + "score": 1.0, + "content": "(Arjovsky et al., 2017), but derived directly from the MLE. More connections are discussed below.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 17, + "bbox_fs": [ + 104, + 364, + 505, + 399 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 417, + 265, + 428 + ], + "lines": [ + { + "bbox": [ + 105, + 416, + 267, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 416, + 267, + 429 + ], + "score": 1.0, + "content": "4.1 LEARNING VIA AMORTIZATION", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 439, + 339, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 439, + 339, + 451 + ], + "spans": [ + { + "bbox": [ + 106, + 439, + 339, + 451 + ], + "score": 1.0, + "content": "Our goal is to learn a generator whose generated samples", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 449, + 339, + 462 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 248, + 462 + ], + "score": 1.0, + "content": "match those from the original model", + "type": "text" + }, + { + "bbox": [ + 248, + 450, + 273, + 462 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 273, + 449, + 339, + 462 + ], + "score": 1.0, + "content": ", by adopting the", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 461, + 340, + 473 + ], + "spans": [ + { + "bbox": [ + 106, + 461, + 340, + 473 + ], + "score": 1.0, + "content": "amortization idea with CTF in the inference section above.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 471, + 341, + 484 + ], + "spans": [ + { + "bbox": [ + 106, + 471, + 341, + 484 + ], + "score": 1.0, + "content": "Similar to inference, the generator is learned implicitly.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 483, + 339, + 495 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 339, + 495 + ], + "score": 1.0, + "content": "However, we also learn an explicit density model for the", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 494, + 339, + 506 + ], + "spans": [ + { + "bbox": [ + 105, + 494, + 339, + 506 + ], + "score": 1.0, + "content": "data by SGD, with samples from the implicit generator to", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 506, + 339, + 516 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 339, + 516 + ], + "score": 1.0, + "content": "estimate gradients in (9). Note that in this case, the CTF is", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 516, + 340, + 528 + ], + "spans": [ + { + "bbox": [ + 105, + 516, + 340, + 528 + ], + "score": 1.0, + "content": "performed directly on the data space, instead of on latent-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 527, + 339, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 527, + 339, + 539 + ], + "score": 1.0, + "content": "variable space as in previous sections. Specifically, the", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 538, + 341, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 538, + 341, + 550 + ], + "score": 1.0, + "content": "sampling procedure from the generator plus a continuous-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 549, + 268, + 560 + ], + "spans": [ + { + "bbox": [ + 106, + 549, + 268, + 560 + ], + "score": 1.0, + "content": "time-flow transformation are written as:", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 25, + "bbox_fs": [ + 105, + 439, + 341, + 560 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 158, + 569, + 286, + 584 + ], + "lines": [ + { + "bbox": [ + 158, + 569, + 286, + 584 + ], + "spans": [ + { + "bbox": [ + 158, + 569, + 286, + 584 + ], + "score": 0.89, + "content": "\\mathbf { x } _ { 0 } \\sim q _ { \\phi } ( \\mathbf { x } _ { 0 } ) , \\mathbf { x } _ { T } \\sim { \\mathcal { T } } ( \\mathbf { x } _ { 0 } , T ) .", + "type": "interline_equation", + "image_path": "5b1559c7588790150c2a8adf05d421683bbb3b9ae7a96bf207af30fbda846a55.jpg" + } + ] + } + ], + "index": 31, + "virtual_lines": [ + { + "bbox": [ + 158, + 569, + 286, + 584 + ], + "spans": [], + "index": 31 + } + ] + }, + { + "type": "image", + "bbox": [ + 346, + 426, + 492, + 525 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 346, + 426, + 492, + 525 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 346, + 426, + 492, + 525 + ], + "spans": [ + { + "bbox": [ + 346, + 426, + 492, + 525 + ], + "score": 0.967, + "type": "image", + "image_path": "6b4afaf4a45229a804fe66cbebef9ba63504a9ce81f293051a8bbd6f04a06cfb.jpg" + } + ] + } + ], + "index": 32.5, + "virtual_lines": [ + { + "bbox": [ + 346, + 426, + 492, + 475.5 + ], + "spans": [], + "index": 32 + }, + { + "bbox": [ + 346, + 475.5, + 492, + 525.0 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 346, + 537, + 505, + 581 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 346, + 537, + 505, + 549 + ], + "spans": [ + { + "bbox": [ + 346, + 537, + 505, + 549 + ], + "score": 1.0, + "content": "Figure 3: Learning a generator with", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 345, + 547, + 505, + 560 + ], + "spans": [ + { + "bbox": [ + 345, + 547, + 505, + 560 + ], + "score": 1.0, + "content": "CTF. The goal is to match the samples", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 346, + 559, + 506, + 572 + ], + "spans": [ + { + "bbox": [ + 346, + 560, + 358, + 570 + ], + "score": 0.83, + "content": "\\mathbf { x } _ { \\mathrm { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 559, + 382, + 572 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 383, + 561, + 394, + 571 + ], + "score": 0.86, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 394, + 559, + 482, + 572 + ], + "score": 1.0, + "content": "to those after a CTF", + "type": "text" + }, + { + "bbox": [ + 483, + 560, + 502, + 570 + ], + "score": 0.8, + "content": "\\left( { { \\bf { x } } _ { T } } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 559, + 506, + 572 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 345, + 569, + 481, + 584 + ], + "spans": [ + { + "bbox": [ + 345, + 569, + 465, + 584 + ], + "score": 1.0, + "content": "or equivalently samples from", + "type": "text" + }, + { + "bbox": [ + 465, + 572, + 476, + 582 + ], + "score": 0.69, + "content": "p _ { \\pmb { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 477, + 569, + 481, + 584 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 35.5 + } + ], + "index": 34.0 + }, + { + "type": "text", + "bbox": [ + 106, + 592, + 505, + 659 + ], + "lines": [ + { + "bbox": [ + 105, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 105, + 592, + 128, + 606 + ], + "score": 1.0, + "content": "Here", + "type": "text" + }, + { + "bbox": [ + 129, + 593, + 155, + 605 + ], + "score": 0.92, + "content": "\\tau ( \\cdot , \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 592, + 311, + 606 + ], + "score": 1.0, + "content": "is the continuous-time flow; a sample", + "type": "text" + }, + { + "bbox": [ + 312, + 595, + 324, + 604 + ], + "score": 0.87, + "content": "\\mathbf { x } _ { \\mathrm { 0 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 324, + 592, + 347, + 606 + ], + "score": 1.0, + "content": "from", + "type": "text" + }, + { + "bbox": [ + 348, + 593, + 369, + 605 + ], + "score": 0.91, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "is implemented by a deep neural", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 603, + 506, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 603, + 191, + 617 + ], + "score": 1.0, + "content": "network (generator)", + "type": "text" + }, + { + "bbox": [ + 191, + 604, + 220, + 616 + ], + "score": 0.93, + "content": "G _ { \\phi } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 603, + 267, + 617 + ], + "score": 1.0, + "content": "with input", + "type": "text" + }, + { + "bbox": [ + 268, + 604, + 315, + 616 + ], + "score": 0.93, + "content": "\\omega \\sim q _ { 0 } ( \\omega )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 603, + 347, + 617 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 348, + 606, + 358, + 615 + ], + "score": 0.86, + "content": "q _ { 0 }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 603, + 506, + 617 + ], + "score": 1.0, + "content": "is a simple distribution for a noise", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "spans": [ + { + "bbox": [ + 106, + 615, + 506, + 628 + ], + "score": 1.0, + "content": "random variable, e.g., the standard isotropic normal distribution. The procedure is illustrated in", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "spans": [ + { + "bbox": [ + 106, + 625, + 506, + 638 + ], + "score": 1.0, + "content": "Figure 3. Note the CTF cannot be replaced by standard normalizing flow (Rezende & Mohamed,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 105, + 636, + 506, + 650 + ], + "score": 1.0, + "content": "2015) in this model, because there is no objective function to guide the update of parameters in", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 648, + 317, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 648, + 317, + 660 + ], + "score": 1.0, + "content": "normalizing flows, which is not necessary for CTFs.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 40.5, + "bbox_fs": [ + 105, + 592, + 506, + 660 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 664, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 664, + 506, + 677 + ], + "spans": [ + { + "bbox": [ + 106, + 664, + 268, + 677 + ], + "score": 1.0, + "content": "Specifically, denote the parameters in the", + "type": "text" + }, + { + "bbox": [ + 268, + 665, + 275, + 675 + ], + "score": 0.83, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 664, + 432, + 677 + ], + "score": 1.0, + "content": "-th step of our algorithm with subscript “", + "type": "text" + }, + { + "bbox": [ + 433, + 664, + 449, + 676 + ], + "score": 0.73, + "content": "( k ) ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 664, + 506, + 677 + ], + "score": 1.0, + "content": ". For efficient", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 676, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 208, + 689 + ], + "score": 1.0, + "content": "sample generation, in the", + "type": "text" + }, + { + "bbox": [ + 209, + 676, + 215, + 686 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 216, + 676, + 505, + 689 + ], + "score": 1.0, + "content": "-th step, we again adopt the amortization idea from Section 3.2 to update", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 106, + 685, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 106, + 686, + 135, + 699 + ], + "score": 0.91, + "content": "\\phi ^ { ( k ^ { - 1 ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 135, + 685, + 239, + 702 + ], + "score": 1.0, + "content": "of the generator network", + "type": "text" + }, + { + "bbox": [ + 239, + 687, + 263, + 700 + ], + "score": 0.91, + "content": "G _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 685, + 506, + 702 + ], + "score": 1.0, + "content": ", such that samples from the updated generator match those", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 104, + 697, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 104, + 697, + 371, + 713 + ], + "score": 1.0, + "content": "from the current generator followed by a one-step transformation", + "type": "text" + }, + { + "bbox": [ + 371, + 699, + 392, + 711 + ], + "score": 0.91, + "content": "\\mathcal { T } _ { 1 } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 697, + 441, + 713 + ], + "score": 1.0, + "content": ". After that,", + "type": "text" + }, + { + "bbox": [ + 441, + 699, + 448, + 709 + ], + "score": 0.72, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 448, + 697, + 505, + 713 + ], + "score": 1.0, + "content": "is updated by", + "type": "text" + } + ], + "index": 47 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 196, + 723 + ], + "score": 1.0, + "content": "drawing samples from", + "type": "text" + }, + { + "bbox": [ + 197, + 710, + 218, + 722 + ], + "score": 0.92, + "content": "q _ { \\phi } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 218, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "to estimate the expectation in (9). The detailed algorithm is presented in", + "type": "text" + } + ], + "index": 48 + }, + { + "bbox": [ + 105, + 720, + 161, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 161, + 734 + ], + "score": 1.0, + "content": "Algorithm 1.", + "type": "text" + } + ], + "index": 49 + } + ], + "index": 46.5, + "bbox_fs": [ + 104, + 664, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 388, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 389, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 389, + 96 + ], + "score": 1.0, + "content": "4.2 CONNECTIONS TO WASSERSTEIN GAN (WGAN) AND MLE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "score": 1.0, + "content": "There is an interesting relation between our model and the WGAN framework (Arjovsky et al., 2017).", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 166, + 125 + ], + "score": 1.0, + "content": "To see this, let", + "type": "text" + }, + { + "bbox": [ + 167, + 113, + 177, + 123 + ], + "score": 0.86, + "content": "p _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 111, + 328, + 125 + ], + "score": 1.0, + "content": "be the data distribution. Substituting", + "type": "text" + }, + { + "bbox": [ + 329, + 112, + 354, + 124 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 111, + 375, + 125 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 375, + 112, + 400, + 124 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 111, + 506, + 125 + ], + "score": 1.0, + "content": "for the expectation in the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 396, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 267, + 135 + ], + "score": 1.0, + "content": "gradient formula (9) and integrating out", + "type": "text" + }, + { + "bbox": [ + 268, + 123, + 274, + 132 + ], + "score": 0.77, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 122, + 396, + 135 + ], + "score": 1.0, + "content": ", we have that our objective is", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 134, + 380, + 147 + ], + "lines": [ + { + "bbox": [ + 225, + 134, + 380, + 147 + ], + "spans": [ + { + "bbox": [ + 225, + 134, + 380, + 147 + ], + "score": 0.87, + "content": "\\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\pmb { \\theta } ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ U ( \\mathbf { x } ; \\pmb { \\theta } ) \\right]", + "type": "interline_equation", + "image_path": "772210da739db4e71d00c69461f108c6e0aa5a93273571509859741e30dde6f7.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 225, + 134, + 380, + 147 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 164, + 506, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 507, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 507, + 177 + ], + "score": 1.0, + "content": "The objective is an instance of the general integral probability metrics (Arjovsky & Bottou, 2017).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 132, + 189 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 132, + 176, + 141, + 185 + ], + "score": 0.8, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 174, + 506, + 189 + ], + "score": 1.0, + "content": "is chosen to be 1-Lipschitz functions, it recovers WGAN. This connection motivates us to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 187, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 199 + ], + "score": 1.0, + "content": "introduce weight clipping (Arjovsky et al., 2017) or alternative regularizers (Gulrajani et al., 2017)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 196, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 168, + 210 + ], + "score": 1.0, + "content": "when updating", + "type": "text" + }, + { + "bbox": [ + 168, + 198, + 175, + 207 + ], + "score": 0.72, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 196, + 507, + 210 + ], + "score": 1.0, + "content": "for a better theoretical property. For this reason, we call our model Markov-chain-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 208, + 206, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 206, + 222 + ], + "score": 1.0, + "content": "based GAN (MacGAN).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 504, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "score": 1.0, + "content": "Furthermore, it can be shown by Jensen’s inequality that the MLE is bounded by (detailed derivations", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 256, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 256, + 249 + ], + "score": 1.0, + "content": "are provided in Section C of the SM)", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 122, + 249, + 471, + 282 + ], + "lines": [ + { + "bbox": [ + 122, + 249, + 471, + 282 + ], + "spans": [ + { + "bbox": [ + 122, + 249, + 471, + 282 + ], + "score": 0.9, + "content": "\\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ \\log q _ { \\phi } \\right] .", + "type": "interline_equation", + "image_path": "df446ddb77704865499a2dd727e53edad8f416e99834eb05e898237bb3a0a4c7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 122, + 249, + 471, + 260.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 122, + 260.0, + 471, + 271.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 122, + 271.0, + 471, + 282.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 283, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 282, + 295 + ], + "score": 1.0, + "content": "By inspecting (10) and (11), it is clear that:", + "type": "text" + }, + { + "bbox": [ + 282, + 284, + 287, + 293 + ], + "score": 0.51, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 282, + 495, + 295 + ], + "score": 1.0, + "content": ") when learning the energy-based model parameters", + "type": "text" + }, + { + "bbox": [ + 496, + 284, + 502, + 293 + ], + "score": 0.78, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 282, + 506, + 295 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 469, + 307 + ], + "score": 1.0, + "content": "the objective can be interpreted as maximizing an upper bound of the MLE shown in (11);", + "type": "text" + }, + { + "bbox": [ + 469, + 295, + 477, + 304 + ], + "score": 0.57, + "content": "i i )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 294, + 505, + 307 + ], + "score": 1.0, + "content": ") when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 212, + 318 + ], + "score": 1.0, + "content": "optimizing the parameter", + "type": "text" + }, + { + "bbox": [ + 212, + 306, + 220, + 316 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "of the inference network, we adopt the amortized learning procedure", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 316, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 284, + 330 + ], + "score": 1.0, + "content": "presented in Algorithm 1, whose objective is", + "type": "text" + }, + { + "bbox": [ + 284, + 316, + 356, + 329 + ], + "score": 0.86, + "content": "\\operatorname* { m i n } _ { \\phi } \\operatorname { K L } \\left( q _ { \\phi } \\| p _ { \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 316, + 506, + 330 + ], + "score": 1.0, + "content": ", coinciding with the last two terms in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 326, + 507, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 326, + 211, + 340 + ], + "score": 1.0, + "content": "(11). In other words, both", + "type": "text" + }, + { + "bbox": [ + 212, + 327, + 219, + 337 + ], + "score": 0.75, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 326, + 236, + 340 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 327, + 244, + 338 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 326, + 507, + 340 + ], + "score": 1.0, + "content": "are optimized by maximizing the same upper bound of the MLE,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 420, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 420, + 351 + ], + "score": 1.0, + "content": "guaranteeing convergence of the algorithm. Particularly, we can conclude that", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5 + }, + { + "type": "text", + "bbox": [ + 109, + 351, + 453, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 455, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 455, + 364 + ], + "score": 1.0, + "content": "Proposition 4. The optimal solution of MacGAN is the maximum likelihood estimator.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "Note another difference between MacGAN and standard GAN framework is the way of learning the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 147, + 394 + ], + "score": 1.0, + "content": "generator", + "type": "text" + }, + { + "bbox": [ + 147, + 382, + 158, + 393 + ], + "score": 0.85, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 381, + 381, + 394 + ], + "score": 1.0, + "content": ". We adopt the amortization idea, which directly guides", + "type": "text" + }, + { + "bbox": [ + 382, + 383, + 393, + 393 + ], + "score": 0.85, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 381, + 443, + 394 + ], + "score": 1.0, + "content": "to approach", + "type": "text" + }, + { + "bbox": [ + 444, + 382, + 455, + 393 + ], + "score": 0.84, + "content": "p _ { \\pmb { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "; whereas in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "GAN, the generator is optimized via a min-max procedure to make it approach the empirical data", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 155, + 416 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 156, + 405, + 166, + 415 + ], + "score": 0.84, + "content": "p _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 403, + 262, + 416 + ], + "score": 1.0, + "content": ". By explicitly learning", + "type": "text" + }, + { + "bbox": [ + 263, + 405, + 274, + 415 + ], + "score": 0.81, + "content": "p _ { \\pmb { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 403, + 506, + 416 + ], + "score": 1.0, + "content": ", MacGAN is able to evaluate likelihood for test data (at", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 414, + 198, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 198, + 427 + ], + "score": 1.0, + "content": "least up to a constant).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24 + }, + { + "type": "title", + "bbox": [ + 108, + 440, + 211, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 213, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 213, + 455 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "Our framework extends the idea of normalizing flows (Rezende & Mohamed, 2015) to continuous-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "time flows, by developing theoretical properties on the convergence behavior. Inference based on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "CTFs has been studied in Salimans et al. (2015) based on the auxiliary-variable technique. However,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "Salimans et al. (2015) directly uses discrete approximations for the flow, and the approximation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "accuracy is unclear. Moreover, the inference network requires simulating a long Markov chain for the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "score": 1.0, + "content": "auxiliary model, thus is less efficient than ours. Finally, the inference network is implemented as a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 530, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 544 + ], + "score": 1.0, + "content": "parametric distribution (e.g., the Gaussian distribution), limiting the representation power, a common", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "setting in existing auxiliary-variable based models (Tran et al., 2016). The idea of amortization", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "(Gershman & Goodman, 2014) has recently been explored in various research topics for Bayesian", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "inference such as in variational inference (Kingma & Welling, 2014; Rezende et al., 2014) and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Markov chain Monte Carlo (Wang & Liu, 2017; Li et al., 2017; Pu et al., 2017a). Both Wang & Liu", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 586, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 151, + 598 + ], + "score": 1.0, + "content": "(2017) and", + "type": "text" + }, + { + "bbox": [ + 151, + 586, + 163, + 596 + ], + "score": 0.29, + "content": "\\mathrm { P u }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 586, + 506, + 598 + ], + "score": 1.0, + "content": "et al. (2017a) extend the idea of Stein variational gradient descent (Liu & Wang, 2016)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "with amortized inference for a GAN-based and a VAE-based model, respectively, which resemble", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "our proposed MacVAE and MacGAN in concept. Li et al. (2017) applies amortization to distill", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 507, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 507, + 632 + ], + "score": 1.0, + "content": "knowledge from MCMC to learn a student network. The ideas in Li et al. (2017) are similar to ours,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 629, + 427, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 427, + 641 + ], + "score": 1.0, + "content": "but the motivation and underlying theory are different from that developed here.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 35.5 + }, + { + "type": "title", + "bbox": [ + 107, + 656, + 200, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 201, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 201, + 670 + ], + "score": 1.0, + "content": "6 EXPERIMENTS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 506, + 693 + ], + "score": 1.0, + "content": "We conduct experiments to test our CTF-based framework for efficient inference and density estima-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "tion described above, and compared them with related methods. The implementation is based on the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 702, + 507, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 507, + 717 + ], + "score": 1.0, + "content": "excellent code for SteinGANk Wang & Liu (2017), where we adopt their default parameter setting.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 721, + 261, + 731 + ], + "lines": [ + { + "bbox": [ + 118, + 720, + 263, + 734 + ], + "spans": [ + { + "bbox": [ + 118, + 720, + 263, + 734 + ], + "score": 1.0, + "content": "khttps://github.com/DartML/SteinGAN", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 82, + 388, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 389, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 389, + 96 + ], + "score": 1.0, + "content": "4.2 CONNECTIONS TO WASSERSTEIN GAN (WGAN) AND MLE", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 101, + 504, + 133 + ], + "lines": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "spans": [ + { + "bbox": [ + 105, + 101, + 505, + 114 + ], + "score": 1.0, + "content": "There is an interesting relation between our model and the WGAN framework (Arjovsky et al., 2017).", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 111, + 506, + 125 + ], + "spans": [ + { + "bbox": [ + 105, + 111, + 166, + 125 + ], + "score": 1.0, + "content": "To see this, let", + "type": "text" + }, + { + "bbox": [ + 167, + 113, + 177, + 123 + ], + "score": 0.86, + "content": "p _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 111, + 328, + 125 + ], + "score": 1.0, + "content": "be the data distribution. Substituting", + "type": "text" + }, + { + "bbox": [ + 329, + 112, + 354, + 124 + ], + "score": 0.92, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 354, + 111, + 375, + 125 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 375, + 112, + 400, + 124 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { x } )", + "type": "inline_equation" + }, + { + "bbox": [ + 400, + 111, + 506, + 125 + ], + "score": 1.0, + "content": "for the expectation in the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 122, + 396, + 135 + ], + "spans": [ + { + "bbox": [ + 105, + 122, + 267, + 135 + ], + "score": 1.0, + "content": "gradient formula (9) and integrating out", + "type": "text" + }, + { + "bbox": [ + 268, + 123, + 274, + 132 + ], + "score": 0.77, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 122, + 396, + 135 + ], + "score": 1.0, + "content": ", we have that our objective is", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 101, + 506, + 135 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 225, + 134, + 380, + 147 + ], + "lines": [ + { + "bbox": [ + 225, + 134, + 380, + 147 + ], + "spans": [ + { + "bbox": [ + 225, + 134, + 380, + 147 + ], + "score": 0.87, + "content": "\\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\pmb { \\theta } ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ U ( \\mathbf { x } ; \\pmb { \\theta } ) \\right]", + "type": "interline_equation", + "image_path": "772210da739db4e71d00c69461f108c6e0aa5a93273571509859741e30dde6f7.jpg" + } + ] + } + ], + "index": 4, + "virtual_lines": [ + { + "bbox": [ + 225, + 134, + 380, + 147 + ], + "spans": [], + "index": 4 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 164, + 506, + 220 + ], + "lines": [ + { + "bbox": [ + 105, + 163, + 507, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 163, + 507, + 177 + ], + "score": 1.0, + "content": "The objective is an instance of the general integral probability metrics (Arjovsky & Bottou, 2017).", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 174, + 506, + 189 + ], + "spans": [ + { + "bbox": [ + 105, + 174, + 132, + 189 + ], + "score": 1.0, + "content": "When", + "type": "text" + }, + { + "bbox": [ + 132, + 176, + 141, + 185 + ], + "score": 0.8, + "content": "U", + "type": "inline_equation" + }, + { + "bbox": [ + 142, + 174, + 506, + 189 + ], + "score": 1.0, + "content": "is chosen to be 1-Lipschitz functions, it recovers WGAN. This connection motivates us to", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 187, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 199 + ], + "score": 1.0, + "content": "introduce weight clipping (Arjovsky et al., 2017) or alternative regularizers (Gulrajani et al., 2017)", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 196, + 507, + 210 + ], + "spans": [ + { + "bbox": [ + 106, + 196, + 168, + 210 + ], + "score": 1.0, + "content": "when updating", + "type": "text" + }, + { + "bbox": [ + 168, + 198, + 175, + 207 + ], + "score": 0.72, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 196, + 507, + 210 + ], + "score": 1.0, + "content": "for a better theoretical property. For this reason, we call our model Markov-chain-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 208, + 206, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 208, + 206, + 222 + ], + "score": 1.0, + "content": "based GAN (MacGAN).", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 163, + 507, + 222 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 225, + 504, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "spans": [ + { + "bbox": [ + 105, + 224, + 506, + 239 + ], + "score": 1.0, + "content": "Furthermore, it can be shown by Jensen’s inequality that the MLE is bounded by (detailed derivations", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 236, + 256, + 249 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 256, + 249 + ], + "score": 1.0, + "content": "are provided in Section C of the SM)", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 224, + 506, + 249 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 122, + 249, + 471, + 282 + ], + "lines": [ + { + "bbox": [ + 122, + 249, + 471, + 282 + ], + "spans": [ + { + "bbox": [ + 122, + 249, + 471, + 282 + ], + "score": 0.9, + "content": "\\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } } \\left[ \\log q _ { \\phi } \\right] .", + "type": "interline_equation", + "image_path": "df446ddb77704865499a2dd727e53edad8f416e99834eb05e898237bb3a0a4c7.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 122, + 249, + 471, + 260.0 + ], + "spans": [], + "index": 12 + }, + { + "bbox": [ + 122, + 260.0, + 471, + 271.0 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 122, + 271.0, + 471, + 282.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 283, + 505, + 349 + ], + "lines": [ + { + "bbox": [ + 106, + 282, + 506, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 282, + 282, + 295 + ], + "score": 1.0, + "content": "By inspecting (10) and (11), it is clear that:", + "type": "text" + }, + { + "bbox": [ + 282, + 284, + 287, + 293 + ], + "score": 0.51, + "content": "i", + "type": "inline_equation" + }, + { + "bbox": [ + 287, + 282, + 495, + 295 + ], + "score": 1.0, + "content": ") when learning the energy-based model parameters", + "type": "text" + }, + { + "bbox": [ + 496, + 284, + 502, + 293 + ], + "score": 0.78, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 282, + 506, + 295 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 294, + 505, + 307 + ], + "spans": [ + { + "bbox": [ + 106, + 294, + 469, + 307 + ], + "score": 1.0, + "content": "the objective can be interpreted as maximizing an upper bound of the MLE shown in (11);", + "type": "text" + }, + { + "bbox": [ + 469, + 295, + 477, + 304 + ], + "score": 0.57, + "content": "i i )", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 294, + 505, + 307 + ], + "score": 1.0, + "content": ") when", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 305, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 212, + 318 + ], + "score": 1.0, + "content": "optimizing the parameter", + "type": "text" + }, + { + "bbox": [ + 212, + 306, + 220, + 316 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 220, + 305, + 505, + 318 + ], + "score": 1.0, + "content": "of the inference network, we adopt the amortized learning procedure", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 316, + 506, + 330 + ], + "spans": [ + { + "bbox": [ + 105, + 316, + 284, + 330 + ], + "score": 1.0, + "content": "presented in Algorithm 1, whose objective is", + "type": "text" + }, + { + "bbox": [ + 284, + 316, + 356, + 329 + ], + "score": 0.86, + "content": "\\operatorname* { m i n } _ { \\phi } \\operatorname { K L } \\left( q _ { \\phi } \\| p _ { \\theta } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 316, + 506, + 330 + ], + "score": 1.0, + "content": ", coinciding with the last two terms in", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 104, + 326, + 507, + 340 + ], + "spans": [ + { + "bbox": [ + 104, + 326, + 211, + 340 + ], + "score": 1.0, + "content": "(11). In other words, both", + "type": "text" + }, + { + "bbox": [ + 212, + 327, + 219, + 337 + ], + "score": 0.75, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 326, + 236, + 340 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 236, + 327, + 244, + 338 + ], + "score": 0.85, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 245, + 326, + 507, + 340 + ], + "score": 1.0, + "content": "are optimized by maximizing the same upper bound of the MLE,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 338, + 420, + 351 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 420, + 351 + ], + "score": 1.0, + "content": "guaranteeing convergence of the algorithm. Particularly, we can conclude that", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 17.5, + "bbox_fs": [ + 104, + 282, + 507, + 351 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 351, + 453, + 362 + ], + "lines": [ + { + "bbox": [ + 106, + 350, + 455, + 364 + ], + "spans": [ + { + "bbox": [ + 106, + 350, + 455, + 364 + ], + "score": 1.0, + "content": "Proposition 4. The optimal solution of MacGAN is the maximum likelihood estimator.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 350, + 455, + 364 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 425 + ], + "lines": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "spans": [ + { + "bbox": [ + 106, + 370, + 505, + 382 + ], + "score": 1.0, + "content": "Note another difference between MacGAN and standard GAN framework is the way of learning the", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 147, + 394 + ], + "score": 1.0, + "content": "generator", + "type": "text" + }, + { + "bbox": [ + 147, + 382, + 158, + 393 + ], + "score": 0.85, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 381, + 381, + 394 + ], + "score": 1.0, + "content": ". We adopt the amortization idea, which directly guides", + "type": "text" + }, + { + "bbox": [ + 382, + 383, + 393, + 393 + ], + "score": 0.85, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 393, + 381, + 443, + 394 + ], + "score": 1.0, + "content": "to approach", + "type": "text" + }, + { + "bbox": [ + 444, + 382, + 455, + 393 + ], + "score": 0.84, + "content": "p _ { \\pmb { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "; whereas in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 391, + 505, + 405 + ], + "score": 1.0, + "content": "GAN, the generator is optimized via a min-max procedure to make it approach the empirical data", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 403, + 506, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 155, + 416 + ], + "score": 1.0, + "content": "distribution", + "type": "text" + }, + { + "bbox": [ + 156, + 405, + 166, + 415 + ], + "score": 0.84, + "content": "p _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 403, + 262, + 416 + ], + "score": 1.0, + "content": ". By explicitly learning", + "type": "text" + }, + { + "bbox": [ + 263, + 405, + 274, + 415 + ], + "score": 0.81, + "content": "p _ { \\pmb { \\theta } }", + "type": "inline_equation" + }, + { + "bbox": [ + 275, + 403, + 506, + 416 + ], + "score": 1.0, + "content": ", MacGAN is able to evaluate likelihood for test data (at", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 414, + 198, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 198, + 427 + ], + "score": 1.0, + "content": "least up to a constant).", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 370, + 506, + 427 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 440, + 211, + 453 + ], + "lines": [ + { + "bbox": [ + 105, + 439, + 213, + 455 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 213, + 455 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 464, + 505, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 465, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 106, + 465, + 506, + 477 + ], + "score": 1.0, + "content": "Our framework extends the idea of normalizing flows (Rezende & Mohamed, 2015) to continuous-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 476, + 505, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 505, + 488 + ], + "score": 1.0, + "content": "time flows, by developing theoretical properties on the convergence behavior. Inference based on", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "CTFs has been studied in Salimans et al. (2015) based on the auxiliary-variable technique. However,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 505, + 510 + ], + "score": 1.0, + "content": "Salimans et al. (2015) directly uses discrete approximations for the flow, and the approximation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 506, + 522 + ], + "score": 1.0, + "content": "accuracy is unclear. Moreover, the inference network requires simulating a long Markov chain for the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 506, + 532 + ], + "score": 1.0, + "content": "auxiliary model, thus is less efficient than ours. Finally, the inference network is implemented as a", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 530, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 105, + 530, + 506, + 544 + ], + "score": 1.0, + "content": "parametric distribution (e.g., the Gaussian distribution), limiting the representation power, a common", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 106, + 542, + 505, + 554 + ], + "score": 1.0, + "content": "setting in existing auxiliary-variable based models (Tran et al., 2016). The idea of amortization", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 553, + 505, + 564 + ], + "score": 1.0, + "content": "(Gershman & Goodman, 2014) has recently been explored in various research topics for Bayesian", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 506, + 577 + ], + "score": 1.0, + "content": "inference such as in variational inference (Kingma & Welling, 2014; Rezende et al., 2014) and", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "Markov chain Monte Carlo (Wang & Liu, 2017; Li et al., 2017; Pu et al., 2017a). Both Wang & Liu", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 586, + 506, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 151, + 598 + ], + "score": 1.0, + "content": "(2017) and", + "type": "text" + }, + { + "bbox": [ + 151, + 586, + 163, + 596 + ], + "score": 0.29, + "content": "\\mathrm { P u }", + "type": "inline_equation" + }, + { + "bbox": [ + 164, + 586, + 506, + 598 + ], + "score": 1.0, + "content": "et al. (2017a) extend the idea of Stein variational gradient descent (Liu & Wang, 2016)", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 505, + 609 + ], + "score": 1.0, + "content": "with amortized inference for a GAN-based and a VAE-based model, respectively, which resemble", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "spans": [ + { + "bbox": [ + 106, + 608, + 505, + 620 + ], + "score": 1.0, + "content": "our proposed MacVAE and MacGAN in concept. Li et al. (2017) applies amortization to distill", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 616, + 507, + 632 + ], + "spans": [ + { + "bbox": [ + 105, + 616, + 507, + 632 + ], + "score": 1.0, + "content": "knowledge from MCMC to learn a student network. The ideas in Li et al. (2017) are similar to ours,", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 629, + 427, + 641 + ], + "spans": [ + { + "bbox": [ + 105, + 629, + 427, + 641 + ], + "score": 1.0, + "content": "but the motivation and underlying theory are different from that developed here.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 35.5, + "bbox_fs": [ + 105, + 465, + 507, + 641 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 656, + 200, + 668 + ], + "lines": [ + { + "bbox": [ + 105, + 654, + 201, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 654, + 201, + 670 + ], + "score": 1.0, + "content": "6 EXPERIMENTS", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 44 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 505, + 714 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 506, + 693 + ], + "score": 1.0, + "content": "We conduct experiments to test our CTF-based framework for efficient inference and density estima-", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "spans": [ + { + "bbox": [ + 105, + 691, + 505, + 703 + ], + "score": 1.0, + "content": "tion described above, and compared them with related methods. The implementation is based on the", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 702, + 507, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 702, + 507, + 717 + ], + "score": 1.0, + "content": "excellent code for SteinGANk Wang & Liu (2017), where we adopt their default parameter setting.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 46, + "bbox_fs": [ + 105, + 680, + 507, + 717 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 70, + 487, + 216 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 70, + 487, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 70, + 487, + 216 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 487, + 216 + ], + "score": 0.96, + "type": "image", + "image_path": "c77f66c6e7bcd94aa794bbca2d0d82ccb351a3e4c8f0d101e39392c22a5424b2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 123, + 70, + 487, + 118.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 123, + 118.66666666666666, + 487, + 167.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 123, + 167.33333333333331, + 487, + 215.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 225, + 505, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "Figure 4: Knowledge distillation from the CTF (left) and ELBO versus epochs on MNIST (right).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 235, + 414, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 414, + 249 + ], + "score": 1.0, + "content": "VAE with 80-layer NF is not included because it has much more parameters.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 212, + 281 + ], + "score": 1.0, + "content": "The discretization stepsize", + "type": "text" + }, + { + "bbox": [ + 212, + 269, + 219, + 279 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "is robust as long as it is set in a reasonable range, e.g., we set it the same", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 279, + 201, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 201, + 291 + ], + "score": 1.0, + "content": "as the stepsize in SGD.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5 + }, + { + "type": "title", + "bbox": [ + 108, + 304, + 227, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 228, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 228, + 316 + ], + "score": 1.0, + "content": "6.1 CTFS FOR INFERENCE", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "Synthetic experiment We examine our amortized learning framework with a toy experiment.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 336, + 507, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 507, + 350 + ], + "score": 1.0, + "content": "Following Rezende & Mohamed (2015), we use MacVAE to approximate samples from a two di-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 101, + 346, + 507, + 367 + ], + "spans": [ + { + "bbox": [ + 101, + 346, + 208, + 367 + ], + "score": 1.0, + "content": "mensional distribution on", + "type": "text" + }, + { + "bbox": [ + 208, + 349, + 261, + 362 + ], + "score": 0.8, + "content": "\\mathbf { z } = \\{ \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 346, + 265, + 367 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 266, + 348, + 325, + 362 + ], + "score": 0.81, + "content": "p ( \\mathbf { z } ) \\propto e ^ { - U ( \\mathbf { z } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 346, + 346, + 367 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 347, + 347, + 507, + 363 + ], + "score": 0.9, + "content": "\\begin{array} { r } { U ( \\mathbf { z } ) \\triangleq \\frac { 1 } { 2 } ( \\frac { \\Vert \\mathbf { z } \\Vert - 2 } { 0 . 4 } ) ^ { 2 } - \\ln ( e ^ { - \\frac { 1 } { 2 } [ \\frac { \\mathbf { z } _ { 1 } - 2 } { 0 . 6 } ] ^ { 2 } } + } \\end{array}", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 361, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 155, + 377 + ], + "score": 0.88, + "content": "e ^ { - \\frac { 1 } { 2 } [ \\frac { { \\bf z } _ { 1 } + 2 } { 0 . 6 } ] ^ { 2 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 361, + 252, + 380 + ], + "score": 1.0, + "content": ". The inference network", + "type": "text" + }, + { + "bbox": [ + 252, + 366, + 263, + 378 + ], + "score": 0.87, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 361, + 507, + 380 + ], + "score": 1.0, + "content": "is defined to be a 2-layer MLP with isotropic normal random", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "score": 1.0, + "content": "variables as input. Figure 4 (top) plots the densities estimated with the samples from transformations", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 107, + 387, + 149, + 399 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { K = 1 0 0 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 386, + 224, + 401 + ], + "score": 1.0, + "content": "(before optimizing", + "type": "text" + }, + { + "bbox": [ + 225, + 388, + 232, + 399 + ], + "score": 0.79, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 386, + 425, + 401 + ], + "score": 1.0, + "content": "), as well as with samples generated directly from", + "type": "text" + }, + { + "bbox": [ + 425, + 388, + 436, + 399 + ], + "score": 0.85, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "(after optimizing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 115, + 410 + ], + "score": 0.77, + "content": "\\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "). It is clear that the amortized learning is able to distill knowledge from the CTF to the inference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 409, + 144, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 409, + 144, + 420 + ], + "score": 1.0, + "content": "network.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "MacVAE on MNIST Following Rezende & Mohamed (2015); Tomczak & Welling (2016), we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "define the inference network as a deep neural network with two fully connected layers of size 300", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "with softplus activation functions. We compare MacVAE with the standard VAE and the VAE with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "normalizing flow, where testing ELBOs are reported (Section D.1 of the SM describes how to calculate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "the ELBO). We do not compare with other state-of-the-art methods such as the inverse autoregressive", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "flow (Kingma et al., 2016), because they typically endowed more complicated inference networks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "(with more parameters), unfair for comparison. We use the same inference network architecture for all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "the models. Figure 4 (bottom) plots the testing ELBO versus training epochs. MacVAE outperforms", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 520, + 367, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 367, + 533 + ], + "score": 1.0, + "content": "VAE and normalizing flows with a better ELBO (around -85.62).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20 + }, + { + "type": "title", + "bbox": [ + 108, + 546, + 271, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 272, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 272, + 558 + ], + "score": 1.0, + "content": "6.2 CTFS FOR DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "We test MacGAN on three datasets: MNIST, CIFAR-10 and CelabA. Following GAN-related", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "methods, the model is evaluated by observing its ability to draw samples from the learned data", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 240, + 601 + ], + "score": 1.0, + "content": "distribution. Inspiring by Wang", + "type": "text" + }, + { + "bbox": [ + 240, + 589, + 249, + 599 + ], + "score": 0.34, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Liu (2017), we define a parametric form of the energy-based", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 599, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 144, + 615 + ], + "score": 1.0, + "content": "model as", + "type": "text" + }, + { + "bbox": [ + 144, + 600, + 320, + 613 + ], + "score": 0.79, + "content": "p _ { \\theta } ( \\mathbf { x } ) \\propto \\mathrm { e x p } \\{ - \\| \\mathbf { x } - \\mathrm { D E C } _ { \\theta } \\left( \\mathrm { E N C } _ { \\theta } ( \\mathbf { x } ) \\right) \\| ^ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 599, + 349, + 615 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 349, + 600, + 386, + 613 + ], + "score": 0.93, + "content": "\\operatorname { E N C } _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 599, + 403, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 403, + 600, + 440, + 613 + ], + "score": 0.91, + "content": "\\operatorname { D E C } _ { \\pmb { \\theta } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 599, + 506, + 615 + ], + "score": 1.0, + "content": "are encoder and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "decoder defined by using deep convolutional neural networks and deconvolutional neural networks,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 227, + 635 + ], + "score": 1.0, + "content": "respectively, parameterized by", + "type": "text" + }, + { + "bbox": [ + 227, + 623, + 234, + 633 + ], + "score": 0.56, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 622, + 506, + 635 + ], + "score": 1.0, + "content": ". For simplicity, we adopt the popular DCGAN architecture (Radford", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 339, + 648 + ], + "score": 1.0, + "content": "et al., 2016) for the encoder and decoder. The generator", + "type": "text" + }, + { + "bbox": [ + 339, + 634, + 353, + 646 + ], + "score": 0.89, + "content": "G _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "is defined as a 3-layer convolutional", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "neural network with the ReLU activation function (except for the top layer which uses tanh as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "activation function, see SM D for details). Following Wang & Liu (2017), the stepsizes are set to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 665, + 508, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 508, + 684 + ], + "score": 1.0, + "content": "(me−e)×lrm 50 , where e indexes the epoch, me is the total number of epochs, lr = 1e-4 when updating θ,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 369, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 123, + 694 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 682, + 163, + 693 + ], + "score": 0.86, + "content": "l _ { r } = 1 { \\mathrm e } { - } 3", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 681, + 225, + 694 + ], + "score": 1.0, + "content": "when updating", + "type": "text" + }, + { + "bbox": [ + 225, + 682, + 233, + 693 + ], + "score": 0.81, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 681, + 299, + 694 + ], + "score": 1.0, + "content": ". The stepsize in", + "type": "text" + }, + { + "bbox": [ + 299, + 682, + 312, + 693 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 681, + 369, + 694 + ], + "score": 1.0, + "content": "is set to 1e-3.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "We compare MacGAN with DCGAN (Radford et al., 2016), the improved WGAN (WGAN-I)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "(Gulrajani et al., 2017) and SteinGAN (Wang & Liu, 2017). We plot images generated with MacGAN", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "and its most related method SteinGAN in Figure 5 for CelebA and CIFAR-10 datasets. More results", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 70, + 487, + 216 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 123, + 70, + 487, + 216 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 123, + 70, + 487, + 216 + ], + "spans": [ + { + "bbox": [ + 123, + 70, + 487, + 216 + ], + "score": 0.96, + "type": "image", + "image_path": "c77f66c6e7bcd94aa794bbca2d0d82ccb351a3e4c8f0d101e39392c22a5424b2.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 123, + 70, + 487, + 118.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 123, + 118.66666666666666, + 487, + 167.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 123, + 167.33333333333331, + 487, + 215.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 225, + 505, + 248 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 506, + 238 + ], + "score": 1.0, + "content": "Figure 4: Knowledge distillation from the CTF (left) and ELBO versus epochs on MNIST (right).", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 235, + 414, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 235, + 414, + 249 + ], + "score": 1.0, + "content": "VAE with 80-layer NF is not included because it has much more parameters.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 108, + 268, + 504, + 290 + ], + "lines": [ + { + "bbox": [ + 106, + 268, + 505, + 281 + ], + "spans": [ + { + "bbox": [ + 106, + 268, + 212, + 281 + ], + "score": 1.0, + "content": "The discretization stepsize", + "type": "text" + }, + { + "bbox": [ + 212, + 269, + 219, + 279 + ], + "score": 0.79, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 268, + 505, + 281 + ], + "score": 1.0, + "content": "is robust as long as it is set in a reasonable range, e.g., we set it the same", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 279, + 201, + 291 + ], + "spans": [ + { + "bbox": [ + 106, + 279, + 201, + 291 + ], + "score": 1.0, + "content": "as the stepsize in SGD.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5.5, + "bbox_fs": [ + 106, + 268, + 505, + 291 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 304, + 227, + 316 + ], + "lines": [ + { + "bbox": [ + 106, + 304, + 228, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 228, + 316 + ], + "score": 1.0, + "content": "6.1 CTFS FOR INFERENCE", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 106, + 325, + 505, + 420 + ], + "lines": [ + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 506, + 338 + ], + "score": 1.0, + "content": "Synthetic experiment We examine our amortized learning framework with a toy experiment.", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 336, + 507, + 350 + ], + "spans": [ + { + "bbox": [ + 105, + 336, + 507, + 350 + ], + "score": 1.0, + "content": "Following Rezende & Mohamed (2015), we use MacVAE to approximate samples from a two di-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 101, + 346, + 507, + 367 + ], + "spans": [ + { + "bbox": [ + 101, + 346, + 208, + 367 + ], + "score": 1.0, + "content": "mensional distribution on", + "type": "text" + }, + { + "bbox": [ + 208, + 349, + 261, + 362 + ], + "score": 0.8, + "content": "\\mathbf { z } = \\{ \\mathbf { z } _ { 1 } , \\mathbf { z } _ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 261, + 346, + 265, + 367 + ], + "score": 1.0, + "content": ":", + "type": "text" + }, + { + "bbox": [ + 266, + 348, + 325, + 362 + ], + "score": 0.81, + "content": "p ( \\mathbf { z } ) \\propto e ^ { - U ( \\mathbf { z } ) }", + "type": "inline_equation" + }, + { + "bbox": [ + 326, + 346, + 346, + 367 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 347, + 347, + 507, + 363 + ], + "score": 0.9, + "content": "\\begin{array} { r } { U ( \\mathbf { z } ) \\triangleq \\frac { 1 } { 2 } ( \\frac { \\Vert \\mathbf { z } \\Vert - 2 } { 0 . 4 } ) ^ { 2 } - \\ln ( e ^ { - \\frac { 1 } { 2 } [ \\frac { \\mathbf { z } _ { 1 } - 2 } { 0 . 6 } ] ^ { 2 } } + } \\end{array}", + "type": "inline_equation" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 361, + 507, + 380 + ], + "spans": [ + { + "bbox": [ + 106, + 362, + 155, + 377 + ], + "score": 0.88, + "content": "e ^ { - \\frac { 1 } { 2 } [ \\frac { { \\bf z } _ { 1 } + 2 } { 0 . 6 } ] ^ { 2 } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 156, + 361, + 252, + 380 + ], + "score": 1.0, + "content": ". The inference network", + "type": "text" + }, + { + "bbox": [ + 252, + 366, + 263, + 378 + ], + "score": 0.87, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 361, + 507, + 380 + ], + "score": 1.0, + "content": "is defined to be a 2-layer MLP with isotropic normal random", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "spans": [ + { + "bbox": [ + 106, + 377, + 505, + 388 + ], + "score": 1.0, + "content": "variables as input. Figure 4 (top) plots the densities estimated with the samples from transformations", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 107, + 386, + 506, + 401 + ], + "spans": [ + { + "bbox": [ + 107, + 387, + 149, + 399 + ], + "score": 0.92, + "content": "\\{ \\mathcal { T } _ { K = 1 0 0 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 149, + 386, + 224, + 401 + ], + "score": 1.0, + "content": "(before optimizing", + "type": "text" + }, + { + "bbox": [ + 225, + 388, + 232, + 399 + ], + "score": 0.79, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 232, + 386, + 425, + 401 + ], + "score": 1.0, + "content": "), as well as with samples generated directly from", + "type": "text" + }, + { + "bbox": [ + 425, + 388, + 436, + 399 + ], + "score": 0.85, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 436, + 386, + 506, + 401 + ], + "score": 1.0, + "content": "(after optimizing", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 397, + 505, + 411 + ], + "spans": [ + { + "bbox": [ + 106, + 399, + 115, + 410 + ], + "score": 0.77, + "content": "\\phi )", + "type": "inline_equation" + }, + { + "bbox": [ + 115, + 397, + 505, + 411 + ], + "score": 1.0, + "content": "). It is clear that the amortized learning is able to distill knowledge from the CTF to the inference", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 409, + 144, + 420 + ], + "spans": [ + { + "bbox": [ + 104, + 409, + 144, + 420 + ], + "score": 1.0, + "content": "network.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 11.5, + "bbox_fs": [ + 101, + 325, + 507, + 420 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 432, + 505, + 532 + ], + "lines": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "spans": [ + { + "bbox": [ + 105, + 432, + 505, + 446 + ], + "score": 1.0, + "content": "MacVAE on MNIST Following Rezende & Mohamed (2015); Tomczak & Welling (2016), we", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 505, + 457 + ], + "score": 1.0, + "content": "define the inference network as a deep neural network with two fully connected layers of size 300", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "spans": [ + { + "bbox": [ + 105, + 454, + 506, + 467 + ], + "score": 1.0, + "content": "with softplus activation functions. We compare MacVAE with the standard VAE and the VAE with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 466, + 505, + 478 + ], + "score": 1.0, + "content": "normalizing flow, where testing ELBOs are reported (Section D.1 of the SM describes how to calculate", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "spans": [ + { + "bbox": [ + 105, + 477, + 505, + 489 + ], + "score": 1.0, + "content": "the ELBO). We do not compare with other state-of-the-art methods such as the inverse autoregressive", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "flow (Kingma et al., 2016), because they typically endowed more complicated inference networks", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "spans": [ + { + "bbox": [ + 105, + 498, + 505, + 511 + ], + "score": 1.0, + "content": "(with more parameters), unfair for comparison. We use the same inference network architecture for all", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 510, + 505, + 522 + ], + "score": 1.0, + "content": "the models. Figure 4 (bottom) plots the testing ELBO versus training epochs. MacVAE outperforms", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 520, + 367, + 533 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 367, + 533 + ], + "score": 1.0, + "content": "VAE and normalizing flows with a better ELBO (around -85.62).", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 432, + 506, + 533 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 546, + 271, + 556 + ], + "lines": [ + { + "bbox": [ + 106, + 546, + 272, + 558 + ], + "spans": [ + { + "bbox": [ + 106, + 546, + 272, + 558 + ], + "score": 1.0, + "content": "6.2 CTFS FOR DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "text", + "bbox": [ + 107, + 565, + 505, + 693 + ], + "lines": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "spans": [ + { + "bbox": [ + 106, + 567, + 505, + 578 + ], + "score": 1.0, + "content": "We test MacGAN on three datasets: MNIST, CIFAR-10 and CelabA. Following GAN-related", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 577, + 506, + 590 + ], + "spans": [ + { + "bbox": [ + 106, + 577, + 506, + 590 + ], + "score": 1.0, + "content": "methods, the model is evaluated by observing its ability to draw samples from the learned data", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 588, + 505, + 601 + ], + "spans": [ + { + "bbox": [ + 106, + 588, + 240, + 601 + ], + "score": 1.0, + "content": "distribution. Inspiring by Wang", + "type": "text" + }, + { + "bbox": [ + 240, + 589, + 249, + 599 + ], + "score": 0.34, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 250, + 588, + 505, + 601 + ], + "score": 1.0, + "content": "Liu (2017), we define a parametric form of the energy-based", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 599, + 506, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 599, + 144, + 615 + ], + "score": 1.0, + "content": "model as", + "type": "text" + }, + { + "bbox": [ + 144, + 600, + 320, + 613 + ], + "score": 0.79, + "content": "p _ { \\theta } ( \\mathbf { x } ) \\propto \\mathrm { e x p } \\{ - \\| \\mathbf { x } - \\mathrm { D E C } _ { \\theta } \\left( \\mathrm { E N C } _ { \\theta } ( \\mathbf { x } ) \\right) \\| ^ { 2 } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 320, + 599, + 349, + 615 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 349, + 600, + 386, + 613 + ], + "score": 0.93, + "content": "\\operatorname { E N C } _ { \\theta } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 386, + 599, + 403, + 615 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 403, + 600, + 440, + 613 + ], + "score": 0.91, + "content": "\\operatorname { D E C } _ { \\pmb { \\theta } } ( \\cdot )", + "type": "inline_equation" + }, + { + "bbox": [ + 441, + 599, + 506, + 615 + ], + "score": 1.0, + "content": "are encoder and", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 506, + 625 + ], + "score": 1.0, + "content": "decoder defined by using deep convolutional neural networks and deconvolutional neural networks,", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 622, + 506, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 227, + 635 + ], + "score": 1.0, + "content": "respectively, parameterized by", + "type": "text" + }, + { + "bbox": [ + 227, + 623, + 234, + 633 + ], + "score": 0.56, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 235, + 622, + 506, + 635 + ], + "score": 1.0, + "content": ". For simplicity, we adopt the popular DCGAN architecture (Radford", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 633, + 506, + 648 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 339, + 648 + ], + "score": 1.0, + "content": "et al., 2016) for the encoder and decoder. The generator", + "type": "text" + }, + { + "bbox": [ + 339, + 634, + 353, + 646 + ], + "score": 0.89, + "content": "G _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 353, + 633, + 506, + 648 + ], + "score": 1.0, + "content": "is defined as a 3-layer convolutional", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 505, + 657 + ], + "score": 1.0, + "content": "neural network with the ReLU activation function (except for the top layer which uses tanh as the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "spans": [ + { + "bbox": [ + 106, + 656, + 505, + 668 + ], + "score": 1.0, + "content": "activation function, see SM D for details). Following Wang & Liu (2017), the stepsizes are set to", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 665, + 508, + 684 + ], + "spans": [ + { + "bbox": [ + 104, + 665, + 508, + 684 + ], + "score": 1.0, + "content": "(me−e)×lrm 50 , where e indexes the epoch, me is the total number of epochs, lr = 1e-4 when updating θ,", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 681, + 369, + 694 + ], + "spans": [ + { + "bbox": [ + 105, + 681, + 123, + 694 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 682, + 163, + 693 + ], + "score": 0.86, + "content": "l _ { r } = 1 { \\mathrm e } { - } 3", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 681, + 225, + 694 + ], + "score": 1.0, + "content": "when updating", + "type": "text" + }, + { + "bbox": [ + 225, + 682, + 233, + 693 + ], + "score": 0.81, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 233, + 681, + 299, + 694 + ], + "score": 1.0, + "content": ". The stepsize in", + "type": "text" + }, + { + "bbox": [ + 299, + 682, + 312, + 693 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 312, + 681, + 369, + 694 + ], + "score": 1.0, + "content": "is set to 1e-3.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 567, + 508, + 694 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "We compare MacGAN with DCGAN (Radford et al., 2016), the improved WGAN (WGAN-I)", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "(Gulrajani et al., 2017) and SteinGAN (Wang & Liu, 2017). We plot images generated with MacGAN", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 732 + ], + "score": 1.0, + "content": "and its most related method SteinGAN in Figure 5 for CelebA and CIFAR-10 datasets. More results", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 104, + 581, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 597 + ], + "score": 1.0, + "content": "are provided in SM Section D. We observe that visually MacGAN is able to generate clear-looking", + "type": "text", + "cross_page": true + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 594, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 496, + 606 + ], + "score": 1.0, + "content": "images. Following Wang & Liu (2017), we also plot the images generated by a random walk in the", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 496, + 596, + 504, + 604 + ], + "score": 0.72, + "content": "\\omega", + "type": "inline_equation", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 605, + 181, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 181, + 617 + ], + "score": 1.0, + "content": "space in Figure 5.", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 698, + 506, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 79, + 503, + 491 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 79, + 503, + 491 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 79, + 503, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 503, + 491 + ], + "score": 0.978, + "type": "image", + "image_path": "f32eae48928503928e1e56b310929d8a7099bc5f96ebf9dbd78c5069876c85de.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 79, + 503, + 216.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 216.33333333333334, + 503, + 353.6666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 353.6666666666667, + 503, + 491.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 496, + 506, + 531 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "Figure 5: Generated images for CIFAR-10 (top) and CelebA (middle) datasets with MacGAN (left)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 441, + 520 + ], + "score": 1.0, + "content": "and SteinGAN (right). The bottom are images generated by a random walk on the", + "type": "text" + }, + { + "bbox": [ + 441, + 509, + 450, + 517 + ], + "score": 0.7, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "space for the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 518, + 362, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 218, + 532 + ], + "score": 1.0, + "content": "generator of MacGAN, i.e.,", + "type": "text" + }, + { + "bbox": [ + 219, + 518, + 358, + 531 + ], + "score": 0.88, + "content": "\\omega _ { t } = \\omega _ { t - 1 } + 0 . 0 3 \\times \\mathrm { r a n d } ( [ - 1 , 1 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 518, + 362, + 532 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 616 + ], + "lines": [ + { + "bbox": [ + 104, + 581, + 506, + 597 + ], + "spans": [ + { + "bbox": [ + 104, + 581, + 506, + 597 + ], + "score": 1.0, + "content": "are provided in SM Section D. We observe that visually MacGAN is able to generate clear-looking", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 594, + 504, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 594, + 496, + 606 + ], + "score": 1.0, + "content": "images. Following Wang & Liu (2017), we also plot the images generated by a random walk in the", + "type": "text" + }, + { + "bbox": [ + 496, + 596, + 504, + 604 + ], + "score": 0.72, + "content": "\\omega", + "type": "inline_equation" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 605, + 181, + 617 + ], + "spans": [ + { + "bbox": [ + 105, + 605, + 181, + 617 + ], + "score": 1.0, + "content": "space in Figure 5.", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 311, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 312, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 312, + 633 + ], + "score": 1.0, + "content": "Qualitatively evaluating a GAN-like model is chal-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 633, + 311, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 311, + 645 + ], + "score": 1.0, + "content": "lenging. We follow literature and use the inception", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 643, + 311, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 311, + 656 + ], + "score": 1.0, + "content": "score (Salimans et al., 2016) to measure the quantity", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 655, + 311, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 311, + 667 + ], + "score": 1.0, + "content": "of the generated images. Figure 6 plots inception", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 666, + 312, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 312, + 677 + ], + "score": 1.0, + "content": "scores versus training epochs for different models.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 676, + 311, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 311, + 689 + ], + "score": 1.0, + "content": "MacGAN obtains competitive inception scores with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 687, + 311, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 311, + 700 + ], + "score": 1.0, + "content": "the popular DCGAN model. Quantitatively, we get", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 699, + 312, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 312, + 711 + ], + "score": 1.0, + "content": "a final inception score of 6.49 for MacGAN, com-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 708, + 312, + 721 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 312, + 721 + ], + "score": 1.0, + "content": "pared to 6.35 for SteinGAN, 6.25 for WGAN-I and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 720, + 181, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 181, + 731 + ], + "score": 1.0, + "content": "6.58 for DCGAN.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5 + }, + { + "type": "image", + "bbox": [ + 341, + 612, + 470, + 705 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 341, + 612, + 470, + 705 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 341, + 612, + 470, + 705 + ], + "spans": [ + { + "bbox": [ + 341, + 612, + 470, + 705 + ], + "score": 0.961, + "type": "image", + "image_path": "886856f7b0b2b16fcc23aab4c1098b0a0407b016e3f738e952ddc0295d2a0f7c.jpg" + } + ] + } + ], + "index": 18.0, + "virtual_lines": [ + { + "bbox": [ + 341, + 612, + 470, + 658.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 341, + 658.5, + 470, + 705.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 318, + 711, + 505, + 732 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 317, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 317, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Figure 6: Inception score versus epochs for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 721, + 388, + 733 + ], + "spans": [ + { + "bbox": [ + 318, + 721, + 388, + 733 + ], + "score": 1.0, + "content": "different models.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 19.75 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 79, + 503, + 491 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 79, + 503, + 491 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 79, + 503, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 503, + 491 + ], + "score": 0.978, + "type": "image", + "image_path": "f32eae48928503928e1e56b310929d8a7099bc5f96ebf9dbd78c5069876c85de.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 106, + 79, + 503, + 216.33333333333334 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 106, + 216.33333333333334, + 503, + 353.6666666666667 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 106, + 353.6666666666667, + 503, + 491.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 496, + 506, + 531 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "Figure 5: Generated images for CIFAR-10 (top) and CelebA (middle) datasets with MacGAN (left)", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 507, + 506, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 507, + 441, + 520 + ], + "score": 1.0, + "content": "and SteinGAN (right). The bottom are images generated by a random walk on the", + "type": "text" + }, + { + "bbox": [ + 441, + 509, + 450, + 517 + ], + "score": 0.7, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 507, + 506, + 520 + ], + "score": 1.0, + "content": "space for the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 518, + 362, + 532 + ], + "spans": [ + { + "bbox": [ + 105, + 518, + 218, + 532 + ], + "score": 1.0, + "content": "generator of MacGAN, i.e.,", + "type": "text" + }, + { + "bbox": [ + 219, + 518, + 358, + 531 + ], + "score": 0.88, + "content": "\\omega _ { t } = \\omega _ { t - 1 } + 0 . 0 3 \\times \\mathrm { r a n d } ( [ - 1 , 1 ] )", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 518, + 362, + 532 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 106, + 582, + 505, + 616 + ], + "lines": [], + "index": 7, + "bbox_fs": [ + 104, + 581, + 506, + 617 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 622, + 311, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 312, + 633 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 312, + 633 + ], + "score": 1.0, + "content": "Qualitatively evaluating a GAN-like model is chal-", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 633, + 311, + 645 + ], + "spans": [ + { + "bbox": [ + 106, + 633, + 311, + 645 + ], + "score": 1.0, + "content": "lenging. We follow literature and use the inception", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 643, + 311, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 643, + 311, + 656 + ], + "score": 1.0, + "content": "score (Salimans et al., 2016) to measure the quantity", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 655, + 311, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 311, + 667 + ], + "score": 1.0, + "content": "of the generated images. Figure 6 plots inception", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 666, + 312, + 677 + ], + "spans": [ + { + "bbox": [ + 105, + 666, + 312, + 677 + ], + "score": 1.0, + "content": "scores versus training epochs for different models.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 676, + 311, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 676, + 311, + 689 + ], + "score": 1.0, + "content": "MacGAN obtains competitive inception scores with", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 687, + 311, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 311, + 700 + ], + "score": 1.0, + "content": "the popular DCGAN model. Quantitatively, we get", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 699, + 312, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 312, + 711 + ], + "score": 1.0, + "content": "a final inception score of 6.49 for MacGAN, com-", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 104, + 708, + 312, + 721 + ], + "spans": [ + { + "bbox": [ + 104, + 708, + 312, + 721 + ], + "score": 1.0, + "content": "pared to 6.35 for SteinGAN, 6.25 for WGAN-I and", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 720, + 181, + 731 + ], + "spans": [ + { + "bbox": [ + 106, + 720, + 181, + 731 + ], + "score": 1.0, + "content": "6.58 for DCGAN.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 13.5, + "bbox_fs": [ + 104, + 622, + 312, + 731 + ] + }, + { + "type": "image", + "bbox": [ + 341, + 612, + 470, + 705 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 341, + 612, + 470, + 705 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 341, + 612, + 470, + 705 + ], + "spans": [ + { + "bbox": [ + 341, + 612, + 470, + 705 + ], + "score": 0.961, + "type": "image", + "image_path": "886856f7b0b2b16fcc23aab4c1098b0a0407b016e3f738e952ddc0295d2a0f7c.jpg" + } + ] + } + ], + "index": 18.0, + "virtual_lines": [ + { + "bbox": [ + 341, + 612, + 470, + 658.5 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 341, + 658.5, + 470, + 705.0 + ], + "spans": [], + "index": 20 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 318, + 711, + 505, + 732 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 317, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 317, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "Figure 6: Inception score versus epochs for", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 721, + 388, + 733 + ], + "spans": [ + { + "bbox": [ + 318, + 721, + 388, + 733 + ], + "score": 1.0, + "content": "different models.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + } + ], + "index": 19.75 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 195, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "score": 1.0, + "content": "7 CONCLUSION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "We study the problem of applying CTFs for efficient inference and explicit density estimation in deep", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "score": 1.0, + "content": "generative models, two important tasks in unsupervised machine learning. Compared to discrete-time", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "normalized flows, CTFs are more general and flexible due to the fact that their stationary distributions", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 153 + ], + "score": 1.0, + "content": "can be controlled without extra flow parameters. We develop theory on the approximation accuracy", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 164 + ], + "score": 1.0, + "content": "when adopting a CTF to approximate a target distribution. We apply CTFs on two classes of deep", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "score": 1.0, + "content": "generative models, a variational autoencoder for efficient inference, and a GAN-like density estimator", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "for explicit density estimation and efficient data generation. Experiments show encouraging results", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "of our framework in both models compared to existing techniques. One interesting direction of future", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 194, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 480, + 206 + ], + "score": 1.0, + "content": "work is to explore more efficient learning algorithms for the proposed CTF-based framework.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5 + }, + { + "type": "title", + "bbox": [ + 107, + 221, + 175, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 176, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 176, + 234 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 105, + 240, + 506, + 734 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 507, + 253 + ], + "score": 1.0, + "content": "M. Arjovsky and L. Bottou. Towards principled methods for training generative adversarial networks.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 115, + 249, + 179, + 263 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 179, + 263 + ], + "score": 1.0, + "content": "In ICLR, 2017.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN. Technical Report arXiv:1701.07875,", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 115, + 280, + 170, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 170, + 292 + ], + "score": 1.0, + "content": "March 2017.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 104, + 297, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 506, + 312 + ], + "score": 1.0, + "content": "F. Bolley, I. Gentil, and A. Guillin. Convergence to equilibrium in wasserstein distance for", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 115, + 310, + 452, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 310, + 452, + 322 + ], + "score": 1.0, + "content": "fokker–planck equations. Journal of Functional Analysis, 263(8):2430–2457, 2012.", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 104, + 326, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 104, + 326, + 507, + 341 + ], + "score": 1.0, + "content": "L. Bottou. Stochastic gradient descent tricks. Technical report, Microsoft Research, Redmond, WA,", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 115, + 339, + 142, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 142, + 352 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "C. Chen, N. Ding, and L. Carin. On the convergence of stochastic gradient MCMC algorithms with", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 115, + 369, + 271, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 369, + 271, + 381 + ], + "score": 1.0, + "content": "high-order integrators. In NIPS, 2015.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 385, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 400 + ], + "score": 1.0, + "content": "Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. ICLR,", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 397, + 142, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 142, + 410 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "A. A. Dorogovtsev and I. I. Nishchenko. An analysis of stochastic flows. Communications on", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 115, + 428, + 282, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 282, + 440 + ], + "score": 1.0, + "content": "Stochastic Analysis, 8(3):331–342, 2014.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "score": 1.0, + "content": "Y. Feng, D. Wang, and Q. Liu. Learning to draw samples with amortized Stein variational gradient", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 456, + 210, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 456, + 210, + 468 + ], + "score": 1.0, + "content": "descent. In UAI, 2017.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "S. J. Gershman and N. D. Goodman. Amortized inference in probabilistic reasoning. In Annual", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 486, + 322, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 486, + 322, + 499 + ], + "score": 1.0, + "content": "Conference of the Cognitive Science Society, 2014.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 502, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 518 + ], + "score": 1.0, + "content": "C. R. Givens and R. M. Shortt. A class of wasserstein metrics for probability distributions. Michigan", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 516, + 193, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 516, + 193, + 526 + ], + "score": 1.0, + "content": "Math. J., 31, 1984.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 545, + 337, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 337, + 557 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial nets. In NIPS, 2014.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 562, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 577 + ], + "score": 1.0, + "content": "A. Grover, M. Dhar, and S. Ermon. Flow-GAN: Bridging implicit and prescribed learning in", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 114, + 573, + 369, + 588 + ], + "spans": [ + { + "bbox": [ + 114, + 573, + 369, + 588 + ], + "score": 1.0, + "content": "generative models. Technical Report arXiv:1705.08868, 2017.", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 104, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of Wasser-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 114, + 603, + 368, + 616 + ], + "spans": [ + { + "bbox": [ + 114, + 603, + 368, + 616 + ], + "score": 1.0, + "content": "stein GAN. Technical Report arXiv:1704.00028, March 2017.", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "T. Haarnoja, H. Tang, P. Abbeel, and S. Levine. Reinforcement learning with deep energy-based", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 114, + 633, + 219, + 645 + ], + "spans": [ + { + "bbox": [ + 114, + 633, + 219, + 645 + ], + "score": 1.0, + "content": "policies. In ICML, 2017.", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 650, + 507, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 507, + 665 + ], + "score": 1.0, + "content": "G. E. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation,", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 115, + 662, + 213, + 674 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 213, + 674 + ], + "score": 1.0, + "content": "14(8):1771–1800, 2002.", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "score": 1.0, + "content": "F. Huszár. Variational inference using implicit distributions. Technical Report arXiv:1702.08235,", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 115, + 690, + 143, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 143, + 704 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker-Planck equation.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 115, + 721, + 288, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 288, + 733 + ], + "score": 1.0, + "content": "SIAM J. MATH. ANAL., 29(1):1–17, 1998.", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 27.5 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 195, + 93 + ], + "lines": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "spans": [ + { + "bbox": [ + 104, + 79, + 197, + 97 + ], + "score": 1.0, + "content": "7 CONCLUSION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 107, + 105, + 505, + 205 + ], + "lines": [ + { + "bbox": [ + 106, + 105, + 505, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 105, + 505, + 119 + ], + "score": 1.0, + "content": "We study the problem of applying CTFs for efficient inference and explicit density estimation in deep", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "spans": [ + { + "bbox": [ + 105, + 117, + 505, + 130 + ], + "score": 1.0, + "content": "generative models, two important tasks in unsupervised machine learning. Compared to discrete-time", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "normalized flows, CTFs are more general and flexible due to the fact that their stationary distributions", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 136, + 506, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 506, + 153 + ], + "score": 1.0, + "content": "can be controlled without extra flow parameters. We develop theory on the approximation accuracy", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 149, + 505, + 164 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 505, + 164 + ], + "score": 1.0, + "content": "when adopting a CTF to approximate a target distribution. We apply CTFs on two classes of deep", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 506, + 174 + ], + "score": 1.0, + "content": "generative models, a variational autoencoder for efficient inference, and a GAN-like density estimator", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "spans": [ + { + "bbox": [ + 106, + 172, + 505, + 185 + ], + "score": 1.0, + "content": "for explicit density estimation and efficient data generation. Experiments show encouraging results", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "spans": [ + { + "bbox": [ + 106, + 183, + 505, + 195 + ], + "score": 1.0, + "content": "of our framework in both models compared to existing techniques. One interesting direction of future", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 194, + 480, + 206 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 480, + 206 + ], + "score": 1.0, + "content": "work is to explore more efficient learning algorithms for the proposed CTF-based framework.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 105, + 506, + 206 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 221, + 175, + 233 + ], + "lines": [ + { + "bbox": [ + 106, + 221, + 176, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 176, + 234 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "list", + "bbox": [ + 105, + 240, + 506, + 734 + ], + "lines": [ + { + "bbox": [ + 105, + 239, + 507, + 253 + ], + "spans": [ + { + "bbox": [ + 105, + 239, + 507, + 253 + ], + "score": 1.0, + "content": "M. Arjovsky and L. Bottou. Towards principled methods for training generative adversarial networks.", + "type": "text" + } + ], + "index": 11, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 249, + 179, + 263 + ], + "spans": [ + { + "bbox": [ + 115, + 249, + 179, + 263 + ], + "score": 1.0, + "content": "In ICLR, 2017.", + "type": "text" + } + ], + "index": 12, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "spans": [ + { + "bbox": [ + 105, + 268, + 506, + 282 + ], + "score": 1.0, + "content": "M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein GAN. Technical Report arXiv:1701.07875,", + "type": "text" + } + ], + "index": 13, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 280, + 170, + 292 + ], + "spans": [ + { + "bbox": [ + 115, + 280, + 170, + 292 + ], + "score": 1.0, + "content": "March 2017.", + "type": "text" + } + ], + "index": 14, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 297, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 297, + 506, + 312 + ], + "score": 1.0, + "content": "F. Bolley, I. Gentil, and A. Guillin. Convergence to equilibrium in wasserstein distance for", + "type": "text" + } + ], + "index": 15, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 310, + 452, + 322 + ], + "spans": [ + { + "bbox": [ + 115, + 310, + 452, + 322 + ], + "score": 1.0, + "content": "fokker–planck equations. Journal of Functional Analysis, 263(8):2430–2457, 2012.", + "type": "text" + } + ], + "index": 16, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 326, + 507, + 341 + ], + "spans": [ + { + "bbox": [ + 104, + 326, + 507, + 341 + ], + "score": 1.0, + "content": "L. Bottou. Stochastic gradient descent tricks. Technical report, Microsoft Research, Redmond, WA,", + "type": "text" + } + ], + "index": 17, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 339, + 142, + 352 + ], + "spans": [ + { + "bbox": [ + 115, + 339, + 142, + 352 + ], + "score": 1.0, + "content": "2012.", + "type": "text" + } + ], + "index": 18, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 356, + 506, + 372 + ], + "score": 1.0, + "content": "C. Chen, N. Ding, and L. Carin. On the convergence of stochastic gradient MCMC algorithms with", + "type": "text" + } + ], + "index": 19, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 369, + 271, + 381 + ], + "spans": [ + { + "bbox": [ + 115, + 369, + 271, + 381 + ], + "score": 1.0, + "content": "high-order integrators. In NIPS, 2015.", + "type": "text" + } + ], + "index": 20, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 385, + 506, + 400 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 400 + ], + "score": 1.0, + "content": "Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. ICLR,", + "type": "text" + } + ], + "index": 21, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 397, + 142, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 397, + 142, + 410 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 22, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 505, + 429 + ], + "score": 1.0, + "content": "A. A. Dorogovtsev and I. I. Nishchenko. An analysis of stochastic flows. Communications on", + "type": "text" + } + ], + "index": 23, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 428, + 282, + 440 + ], + "spans": [ + { + "bbox": [ + 115, + 428, + 282, + 440 + ], + "score": 1.0, + "content": "Stochastic Analysis, 8(3):331–342, 2014.", + "type": "text" + } + ], + "index": 24, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "spans": [ + { + "bbox": [ + 104, + 443, + 506, + 460 + ], + "score": 1.0, + "content": "Y. Feng, D. Wang, and Q. Liu. Learning to draw samples with amortized Stein variational gradient", + "type": "text" + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 456, + 210, + 468 + ], + "spans": [ + { + "bbox": [ + 115, + 456, + 210, + 468 + ], + "score": 1.0, + "content": "descent. In UAI, 2017.", + "type": "text" + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 474, + 506, + 488 + ], + "score": 1.0, + "content": "S. J. Gershman and N. D. Goodman. Amortized inference in probabilistic reasoning. In Annual", + "type": "text" + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 486, + 322, + 499 + ], + "spans": [ + { + "bbox": [ + 115, + 486, + 322, + 499 + ], + "score": 1.0, + "content": "Conference of the Cognitive Science Society, 2014.", + "type": "text" + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 502, + 505, + 518 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 505, + 518 + ], + "score": 1.0, + "content": "C. R. Givens and R. M. Shortt. A class of wasserstein metrics for probability distributions. Michigan", + "type": "text" + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 516, + 193, + 526 + ], + "spans": [ + { + "bbox": [ + 115, + 516, + 193, + 526 + ], + "score": 1.0, + "content": "Math. J., 31, 1984.", + "type": "text" + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 546 + ], + "score": 1.0, + "content": "I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and", + "type": "text" + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 545, + 337, + 557 + ], + "spans": [ + { + "bbox": [ + 115, + 545, + 337, + 557 + ], + "score": 1.0, + "content": "Y. Bengio. Generative adversarial nets. In NIPS, 2014.", + "type": "text" + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 562, + 506, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 562, + 506, + 577 + ], + "score": 1.0, + "content": "A. Grover, M. Dhar, and S. Ermon. Flow-GAN: Bridging implicit and prescribed learning in", + "type": "text" + } + ], + "index": 33, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 573, + 369, + 588 + ], + "spans": [ + { + "bbox": [ + 114, + 573, + 369, + 588 + ], + "score": 1.0, + "content": "generative models. Technical Report arXiv:1705.08868, 2017.", + "type": "text" + } + ], + "index": 34, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 592, + 506, + 606 + ], + "spans": [ + { + "bbox": [ + 104, + 592, + 506, + 606 + ], + "score": 1.0, + "content": "I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. Courville. Improved training of Wasser-", + "type": "text" + } + ], + "index": 35, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 603, + 368, + 616 + ], + "spans": [ + { + "bbox": [ + 114, + 603, + 368, + 616 + ], + "score": 1.0, + "content": "stein GAN. Technical Report arXiv:1704.00028, March 2017.", + "type": "text" + } + ], + "index": 36, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 505, + 634 + ], + "score": 1.0, + "content": "T. Haarnoja, H. Tang, P. Abbeel, and S. Levine. Reinforcement learning with deep energy-based", + "type": "text" + } + ], + "index": 37, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 633, + 219, + 645 + ], + "spans": [ + { + "bbox": [ + 114, + 633, + 219, + 645 + ], + "score": 1.0, + "content": "policies. In ICML, 2017.", + "type": "text" + } + ], + "index": 38, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 650, + 507, + 665 + ], + "spans": [ + { + "bbox": [ + 105, + 650, + 507, + 665 + ], + "score": 1.0, + "content": "G. E. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation,", + "type": "text" + } + ], + "index": 39, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 662, + 213, + 674 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 213, + 674 + ], + "score": 1.0, + "content": "14(8):1771–1800, 2002.", + "type": "text" + } + ], + "index": 40, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 105, + 680, + 506, + 693 + ], + "score": 1.0, + "content": "F. Huszár. Variational inference using implicit distributions. Technical Report arXiv:1702.08235,", + "type": "text" + } + ], + "index": 41, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 690, + 143, + 704 + ], + "spans": [ + { + "bbox": [ + 115, + 690, + 143, + 704 + ], + "score": 1.0, + "content": "2017.", + "type": "text" + } + ], + "index": 42, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "R. Jordan, D. Kinderlehrer, and F. Otto. The variational formulation of the Fokker-Planck equation.", + "type": "text" + } + ], + "index": 43, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 721, + 288, + 733 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 288, + 733 + ], + "score": 1.0, + "content": "SIAM J. MATH. ANAL., 29(1):1–17, 1998.", + "type": "text" + } + ], + "index": 44, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "D. Kingma, T. P. Salimans, and M. Welling. Improving variational inference with inverse autoregres-", + "type": "text", + "cross_page": true + } + ], + "index": 0, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 93, + 219, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 219, + 107 + ], + "score": 1.0, + "content": "sive flow. In NIPS, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 1, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 113, + 429, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 429, + 124 + ], + "score": 1.0, + "content": "D. P. Kingma and M. Welling. Auto-encoding variational Bayes. In ICLR, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 2, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "Y. Li, R. E. Turner, and Q. Liu. Approximate inference with amortised MCMC. Technical Report", + "type": "text", + "cross_page": true + } + ], + "index": 3, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 143, + 219, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 219, + 155 + ], + "score": 1.0, + "content": "arXiv:1702.08343, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 4, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "Q. Liu and D. Wang. Stein variational gradient descent: A general purpose Bayesian inference", + "type": "text", + "cross_page": true + } + ], + "index": 5, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 172, + 224, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 224, + 186 + ], + "score": 1.0, + "content": "algorithm. In NIPS, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 6, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "J. C. Mattingly, A. M. Stuart, and M. V. Tretyakov. Construction of numerical time-average and", + "type": "text", + "cross_page": true + } + ], + "index": 7, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 201, + 507, + 215 + ], + "spans": [ + { + "bbox": [ + 114, + 201, + 507, + 215 + ], + "score": 1.0, + "content": "stationary measures via Poisson equations. SIAM Journal on Numerical Analysis, 48(2):552–577,", + "type": "text", + "cross_page": true + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 212, + 143, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 143, + 226 + ], + "score": 1.0, + "content": "2010.", + "type": "text", + "cross_page": true + } + ], + "index": 9, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 230, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 245 + ], + "score": 1.0, + "content": "S. Mohamed and B. Lakshminarayanan. Learning in implicit generative models. Technical Report", + "type": "text", + "cross_page": true + } + ], + "index": 10, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 243, + 219, + 254 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 219, + 254 + ], + "score": 1.0, + "content": "arXiv:1610.03483, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 11, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 260, + 507, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 507, + 274 + ], + "score": 1.0, + "content": "F. Otto. Dynamics of Labyrinthine pattern formation in magnetic fluids: A mean-field theory. Arch.", + "type": "text", + "cross_page": true + } + ], + "index": 12, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 271, + 282, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 282, + 285 + ], + "score": 1.0, + "content": "Rational Mech. Anal., pp. 63–103, 1998.", + "type": "text", + "cross_page": true + } + ], + "index": 13, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "score": 1.0, + "content": "Y. Pu, Z. Gan, R. Henao, C. Li, S. Han, and L. Carin. Stein variational autoencoder. Technical Report", + "type": "text", + "cross_page": true + } + ], + "index": 14, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 301, + 223, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 301, + 223, + 313 + ], + "score": 1.0, + "content": "arXiv:1704.05155, 2017a.", + "type": "text", + "cross_page": true + } + ], + "index": 15, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 318, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 333 + ], + "score": 1.0, + "content": "Y. Pu, Z. Gan, R. Henao, C. Li, S. Han, and L. Carin. Vae learning via stein variational gradient", + "type": "text", + "cross_page": true + } + ], + "index": 16, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 330, + 220, + 344 + ], + "spans": [ + { + "bbox": [ + 114, + 330, + 220, + 344 + ], + "score": 1.0, + "content": "descent. In NIPS, 2017b.", + "type": "text", + "cross_page": true + } + ], + "index": 17, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional", + "type": "text", + "cross_page": true + } + ], + "index": 18, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 360, + 456, + 375 + ], + "spans": [ + { + "bbox": [ + 113, + 360, + 456, + 375 + ], + "score": 1.0, + "content": "generative adversarial networks. Technical Report arXiv:1511.06434, January 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 19, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 380, + 502, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 502, + 392 + ], + "score": 1.0, + "content": "R. Ranganath, J. Altosaar, D. Tran, and D. M. Blei. Operator variational inference. In NIPS, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 20, + "is_list_start_line": true + }, + { + "bbox": [ + 105, + 396, + 487, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 487, + 411 + ], + "score": 1.0, + "content": "D. J. Rezende and S. Mohamed. Variational inference with normalizing flows. In ICML, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 21, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "score": 1.0, + "content": "D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and approximate inference", + "type": "text", + "cross_page": true + } + ], + "index": 22, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 427, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 114, + 427, + 291, + 441 + ], + "score": 1.0, + "content": "in deep generative models. In ICML, 2014.", + "type": "text", + "cross_page": true + } + ], + "index": 23, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 446, + 411, + 459 + ], + "spans": [ + { + "bbox": [ + 104, + 446, + 411, + 459 + ], + "score": 1.0, + "content": "H. Risken. The Fokker-Planck equation. Springer-Verlag, New York, 1989.", + "type": "text", + "cross_page": true + } + ], + "index": 24, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 464, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 507, + 478 + ], + "score": 1.0, + "content": "T. Salimans, D. P. Kingma, and M. Welling. Markov chain Monte Carlo and variational inference:", + "type": "text", + "cross_page": true + } + ], + "index": 25, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 475, + 254, + 489 + ], + "spans": [ + { + "bbox": [ + 114, + 475, + 254, + 489 + ], + "score": 1.0, + "content": "Bridging the gap. In ICML, 2015.", + "type": "text", + "cross_page": true + } + ], + "index": 26, + "is_list_end_line": true + }, + { + "bbox": [ + 104, + 493, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 493, + 505, + 509 + ], + "score": 1.0, + "content": "T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques", + "type": "text", + "cross_page": true + } + ], + "index": 27, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 505, + 390, + 519 + ], + "spans": [ + { + "bbox": [ + 115, + 505, + 390, + 519 + ], + "score": 1.0, + "content": "for training GANs. Technical Report arXiv:1606.03498, June 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 28, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "score": 1.0, + "content": "Y. W. Teh, A. H. Thiery, and S. J. Vollmer. Consistency and fluctuations for stochastic gradient", + "type": "text", + "cross_page": true + } + ], + "index": 29, + "is_list_start_line": true + }, + { + "bbox": [ + 113, + 534, + 319, + 549 + ], + "spans": [ + { + "bbox": [ + 113, + 534, + 319, + 549 + ], + "score": 1.0, + "content": "Langevin dynamics. JMLR, 17(1):193–225, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 30, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 551, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 103, + 551, + 507, + 568 + ], + "score": 1.0, + "content": "J. M. Tomczak and M. Welling. Improving variational auto-encoders using Householder flow.", + "type": "text", + "cross_page": true + } + ], + "index": 31, + "is_list_start_line": true + }, + { + "bbox": [ + 115, + 564, + 334, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 334, + 578 + ], + "score": 1.0, + "content": "Technical Report arXiv:1611.09630, November 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 32, + "is_list_end_line": true + }, + { + "bbox": [ + 106, + 584, + 464, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 464, + 596 + ], + "score": 1.0, + "content": "D. Tran, R. Ranganath, and D. M. Blei. The variational gaussian process. In ICLR, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 33, + "is_list_start_line": true, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 601, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 507, + 615 + ], + "score": 1.0, + "content": "A. van den Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks. In ICML,", + "type": "text", + "cross_page": true + } + ], + "index": 34, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 612, + 144, + 627 + ], + "spans": [ + { + "bbox": [ + 114, + 612, + 144, + 627 + ], + "score": 1.0, + "content": "2016.", + "type": "text", + "cross_page": true + } + ], + "index": 35, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "S. J. Vollmer, K. C. Zygalakis, and Y. W. Teh. (exploration of the (Non-)asymptotic bias and variance", + "type": "text", + "cross_page": true + } + ], + "index": 36, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 642, + 375, + 655 + ], + "spans": [ + { + "bbox": [ + 114, + 642, + 375, + 655 + ], + "score": 1.0, + "content": "of stochastic gradient Langevin dynamics. JMLR, 1:1–48, 2016.", + "type": "text", + "cross_page": true + } + ], + "index": 37, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "score": 1.0, + "content": "D. Wang and Q. Liu. Learning to draw samples: With application to amortized MLE for generative", + "type": "text", + "cross_page": true + } + ], + "index": 38, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 672, + 306, + 686 + ], + "spans": [ + { + "bbox": [ + 114, + 672, + 306, + 686 + ], + "score": 1.0, + "content": "adversarial learning. In ICLR workshop, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 39, + "is_list_end_line": true + }, + { + "bbox": [ + 103, + 688, + 507, + 706 + ], + "spans": [ + { + "bbox": [ + 103, + 688, + 507, + 706 + ], + "score": 1.0, + "content": "M. Welling and Y. W. Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML,", + "type": "text", + "cross_page": true + } + ], + "index": 40, + "is_list_start_line": true + }, + { + "bbox": [ + 114, + 701, + 144, + 716 + ], + "spans": [ + { + "bbox": [ + 114, + 701, + 144, + 716 + ], + "score": 1.0, + "content": "2011.", + "type": "text", + "cross_page": true + } + ], + "index": 41, + "is_list_end_line": true + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "J. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative adversarial networks. In ICLR, 2017.", + "type": "text", + "cross_page": true + } + ], + "index": 42, + "is_list_start_line": true + } + ], + "index": 27.5, + "bbox_fs": [ + 104, + 239, + 507, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 105, + 62, + 507, + 735 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 506, + 96 + ], + "score": 1.0, + "content": "D. Kingma, T. P. Salimans, and M. Welling. Improving variational inference with inverse autoregres-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 219, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 219, + 107 + ], + "score": 1.0, + "content": "sive flow. In NIPS, 2016.", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 113, + 429, + 124 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 429, + 124 + ], + "score": 1.0, + "content": "D. P. Kingma and M. Welling. Auto-encoding variational Bayes. In ICLR, 2014.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 105, + 129, + 505, + 144 + ], + "score": 1.0, + "content": "Y. Li, R. E. Turner, and Q. Liu. Approximate inference with amortised MCMC. Technical Report", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 143, + 219, + 155 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 219, + 155 + ], + "score": 1.0, + "content": "arXiv:1702.08343, 2017.", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "spans": [ + { + "bbox": [ + 105, + 160, + 505, + 174 + ], + "score": 1.0, + "content": "Q. Liu and D. Wang. Stein variational gradient descent: A general purpose Bayesian inference", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 172, + 224, + 186 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 224, + 186 + ], + "score": 1.0, + "content": "algorithm. In NIPS, 2016.", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "spans": [ + { + "bbox": [ + 105, + 191, + 505, + 203 + ], + "score": 1.0, + "content": "J. C. Mattingly, A. M. Stuart, and M. V. Tretyakov. Construction of numerical time-average and", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 201, + 507, + 215 + ], + "spans": [ + { + "bbox": [ + 114, + 201, + 507, + 215 + ], + "score": 1.0, + "content": "stationary measures via Poisson equations. SIAM Journal on Numerical Analysis, 48(2):552–577,", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 212, + 143, + 226 + ], + "spans": [ + { + "bbox": [ + 115, + 212, + 143, + 226 + ], + "score": 1.0, + "content": "2010.", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 230, + 506, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 506, + 245 + ], + "score": 1.0, + "content": "S. Mohamed and B. Lakshminarayanan. Learning in implicit generative models. Technical Report", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 115, + 243, + 219, + 254 + ], + "spans": [ + { + "bbox": [ + 115, + 243, + 219, + 254 + ], + "score": 1.0, + "content": "arXiv:1610.03483, 2017.", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 260, + 507, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 507, + 274 + ], + "score": 1.0, + "content": "F. Otto. Dynamics of Labyrinthine pattern formation in magnetic fluids: A mean-field theory. Arch.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 271, + 282, + 285 + ], + "spans": [ + { + "bbox": [ + 115, + 271, + 282, + 285 + ], + "score": 1.0, + "content": "Rational Mech. Anal., pp. 63–103, 1998.", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 289, + 506, + 304 + ], + "score": 1.0, + "content": "Y. Pu, Z. Gan, R. Henao, C. Li, S. Han, and L. Carin. Stein variational autoencoder. Technical Report", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 115, + 301, + 223, + 313 + ], + "spans": [ + { + "bbox": [ + 115, + 301, + 223, + 313 + ], + "score": 1.0, + "content": "arXiv:1704.05155, 2017a.", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 318, + 506, + 333 + ], + "spans": [ + { + "bbox": [ + 105, + 318, + 506, + 333 + ], + "score": 1.0, + "content": "Y. Pu, Z. Gan, R. Henao, C. Li, S. Han, and L. Carin. Vae learning via stein variational gradient", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 114, + 330, + 220, + 344 + ], + "spans": [ + { + "bbox": [ + 114, + 330, + 220, + 344 + ], + "score": 1.0, + "content": "descent. In NIPS, 2017b.", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "spans": [ + { + "bbox": [ + 105, + 349, + 506, + 363 + ], + "score": 1.0, + "content": "A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 113, + 360, + 456, + 375 + ], + "spans": [ + { + "bbox": [ + 113, + 360, + 456, + 375 + ], + "score": 1.0, + "content": "generative adversarial networks. Technical Report arXiv:1511.06434, January 2016.", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 380, + 502, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 502, + 392 + ], + "score": 1.0, + "content": "R. Ranganath, J. Altosaar, D. Tran, and D. M. Blei. Operator variational inference. In NIPS, 2016.", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 396, + 487, + 411 + ], + "spans": [ + { + "bbox": [ + 105, + 396, + 487, + 411 + ], + "score": 1.0, + "content": "D. J. Rezende and S. Mohamed. Variational inference with normalizing flows. In ICML, 2015.", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 506, + 430 + ], + "score": 1.0, + "content": "D. J. Rezende, S. Mohamed, and D. Wierstra. Stochastic backpropagation and approximate inference", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 114, + 427, + 291, + 441 + ], + "spans": [ + { + "bbox": [ + 114, + 427, + 291, + 441 + ], + "score": 1.0, + "content": "in deep generative models. In ICML, 2014.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 104, + 446, + 411, + 459 + ], + "spans": [ + { + "bbox": [ + 104, + 446, + 411, + 459 + ], + "score": 1.0, + "content": "H. Risken. The Fokker-Planck equation. Springer-Verlag, New York, 1989.", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 464, + 507, + 478 + ], + "spans": [ + { + "bbox": [ + 105, + 464, + 507, + 478 + ], + "score": 1.0, + "content": "T. Salimans, D. P. Kingma, and M. Welling. Markov chain Monte Carlo and variational inference:", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 114, + 475, + 254, + 489 + ], + "spans": [ + { + "bbox": [ + 114, + 475, + 254, + 489 + ], + "score": 1.0, + "content": "Bridging the gap. In ICML, 2015.", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 493, + 505, + 509 + ], + "spans": [ + { + "bbox": [ + 104, + 493, + 505, + 509 + ], + "score": 1.0, + "content": "T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 505, + 390, + 519 + ], + "spans": [ + { + "bbox": [ + 115, + 505, + 390, + 519 + ], + "score": 1.0, + "content": "for training GANs. Technical Report arXiv:1606.03498, June 2016.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 538 + ], + "score": 1.0, + "content": "Y. W. Teh, A. H. Thiery, and S. J. Vollmer. Consistency and fluctuations for stochastic gradient", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 113, + 534, + 319, + 549 + ], + "spans": [ + { + "bbox": [ + 113, + 534, + 319, + 549 + ], + "score": 1.0, + "content": "Langevin dynamics. JMLR, 17(1):193–225, 2016.", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 103, + 551, + 507, + 568 + ], + "spans": [ + { + "bbox": [ + 103, + 551, + 507, + 568 + ], + "score": 1.0, + "content": "J. M. Tomczak and M. Welling. Improving variational auto-encoders using Householder flow.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 564, + 334, + 578 + ], + "spans": [ + { + "bbox": [ + 115, + 564, + 334, + 578 + ], + "score": 1.0, + "content": "Technical Report arXiv:1611.09630, November 2016.", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 584, + 464, + 596 + ], + "spans": [ + { + "bbox": [ + 106, + 584, + 464, + 596 + ], + "score": 1.0, + "content": "D. Tran, R. Ranganath, and D. M. Blei. The variational gaussian process. In ICLR, 2016.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 601, + 507, + 615 + ], + "spans": [ + { + "bbox": [ + 105, + 601, + 507, + 615 + ], + "score": 1.0, + "content": "A. van den Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks. In ICML,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 114, + 612, + 144, + 627 + ], + "spans": [ + { + "bbox": [ + 114, + 612, + 144, + 627 + ], + "score": 1.0, + "content": "2016.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "S. J. Vollmer, K. C. Zygalakis, and Y. W. Teh. (exploration of the (Non-)asymptotic bias and variance", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 114, + 642, + 375, + 655 + ], + "spans": [ + { + "bbox": [ + 114, + 642, + 375, + 655 + ], + "score": 1.0, + "content": "of stochastic gradient Langevin dynamics. JMLR, 1:1–48, 2016.", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "spans": [ + { + "bbox": [ + 105, + 662, + 505, + 676 + ], + "score": 1.0, + "content": "D. Wang and Q. Liu. Learning to draw samples: With application to amortized MLE for generative", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 114, + 672, + 306, + 686 + ], + "spans": [ + { + "bbox": [ + 114, + 672, + 306, + 686 + ], + "score": 1.0, + "content": "adversarial learning. In ICLR workshop, 2017.", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 103, + 688, + 507, + 706 + ], + "spans": [ + { + "bbox": [ + 103, + 688, + 507, + 706 + ], + "score": 1.0, + "content": "M. Welling and Y. W. Teh. Bayesian learning via stochastic gradient Langevin dynamics. In ICML,", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 114, + 701, + 144, + 716 + ], + "spans": [ + { + "bbox": [ + 114, + 701, + 144, + 716 + ], + "score": 1.0, + "content": "2011.", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "score": 1.0, + "content": "J. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative adversarial networks. In ICLR, 2017.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 21 + } + ], + "page_idx": 12, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "13", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "list", + "bbox": [ + 105, + 62, + 507, + 735 + ], + "lines": [], + "index": 21, + "bbox_fs": [ + 103, + 82, + 507, + 734 + ], + "lines_deleted": true + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 146, + 78, + 465, + 136 + ], + "lines": [ + { + "bbox": [ + 180, + 78, + 435, + 97 + ], + "spans": [ + { + "bbox": [ + 180, + 78, + 435, + 97 + ], + "score": 1.0, + "content": "SUPPLEMENTARY MATERIAL FOR:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 98, + 466, + 117 + ], + "spans": [ + { + "bbox": [ + 145, + 98, + 466, + 117 + ], + "score": 1.0, + "content": "CONTINUOUS-TIME FLOWS FOR EFFICIENT", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 158, + 118, + 452, + 136 + ], + "spans": [ + { + "bbox": [ + 158, + 118, + 452, + 136 + ], + "score": 1.0, + "content": "INFERENCE AND DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 184, + 284, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 183, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 286, + 199 + ], + "score": 1.0, + "content": "A ASSUMPTIONS OF THEOREM 2", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 470, + 222 + ], + "score": 1.0, + "content": "First, let us define the infinitesimal generator of the diffusion (2). Formally, the generator", + "type": "text" + }, + { + "bbox": [ + 470, + 210, + 478, + 219 + ], + "score": 0.82, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 427, + 233 + ], + "score": 1.0, + "content": "diffusion (2) is defined for any compactly supported twice differentiable function", + "type": "text" + }, + { + "bbox": [ + 427, + 220, + 481, + 232 + ], + "score": 0.94, + "content": "\\breve { f } : \\mathbf { R } ^ { L } \\mathbf { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 219, + 505, + 233 + ], + "score": 1.0, + "content": ", such", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 128, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 128, + 245 + ], + "score": 1.0, + "content": "that,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 246, + 492, + 274 + ], + "lines": [ + { + "bbox": [ + 117, + 246, + 492, + 274 + ], + "spans": [ + { + "bbox": [ + 117, + 246, + 492, + 274 + ], + "score": 0.94, + "content": "\\mathcal { L } f ( \\mathbf { Z } _ { t } ) \\triangleq \\operatorname* { l i m } _ { h 0 ^ { + } } \\frac { \\mathbb { E } [ f ( \\mathbf { Z } _ { t + h } ) ] - f ( \\mathbf { Z } _ { t } ) } { h } = ( F ( \\mathbf { Z } _ { t } ) \\cdot \\nabla + \\frac { 1 } { 2 } ( G ( \\mathbf { Z } _ { t } ) G ( \\mathbf { Z } _ { t } ) ^ { T } ) \\colon \\nabla \\nabla ^ { T } ) f ( \\mathbf { Z } _ { t } ) ,", + "type": "interline_equation", + "image_path": "836de84f757b84ca0f70a7f98051d6709f159c13ad0e0d33891b3dcfc5b9e778.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 117, + 246, + 492, + 274 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 280, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 133, + 295 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 280, + 181, + 292 + ], + "score": 0.41, + "content": "{ \\mathbf { a } } \\cdot { \\mathbf { b } } \\triangleq { \\mathbf { a } } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 278, + 192, + 295 + ], + "score": 1.0, + "content": "b,", + "type": "text" + }, + { + "bbox": [ + 192, + 279, + 267, + 293 + ], + "score": 0.72, + "content": "\\mathbf { A } : \\mathbf { B } \\triangleq { \\mathrm { t r } } ( \\mathbf { A } ^ { T } \\mathbf { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 278, + 270, + 295 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 271, + 280, + 306, + 291 + ], + "score": 0.86, + "content": "h \\to 0 ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 278, + 335, + 295 + ], + "score": 1.0, + "content": "means", + "type": "text" + }, + { + "bbox": [ + 336, + 282, + 343, + 291 + ], + "score": 0.82, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 278, + 505, + 295 + ], + "score": 1.0, + "content": "approaches zero along the positive real", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 291, + 128, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 128, + 304 + ], + "score": 1.0, + "content": "axis.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 106, + 308, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 346, + 321 + ], + "score": 1.0, + "content": "Given an ergodic diffusion (2) with an invariant measure", + "type": "text" + }, + { + "bbox": [ + 347, + 309, + 367, + 321 + ], + "score": 0.92, + "content": "\\rho ( \\mathbf { Z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 308, + 505, + 321 + ], + "score": 1.0, + "content": ", the posterior average is defined", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 122, + 335 + ], + "score": 1.0, + "content": "as:", + "type": "text" + }, + { + "bbox": [ + 122, + 320, + 210, + 334 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\bar { \\psi } \\triangleq \\int \\psi ( \\mathbf { Z } ) \\rho ( \\mathbf { Z } ) \\mathrm { d } \\mathbf { Z } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 320, + 305, + 335 + ], + "score": 1.0, + "content": "for some test function", + "type": "text" + }, + { + "bbox": [ + 305, + 321, + 327, + 334 + ], + "score": 0.92, + "content": "\\psi ( \\mathbf { Z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 320, + 506, + 335 + ], + "score": 1.0, + "content": "of interest. For a given numerical method", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 329, + 509, + 353 + ], + "spans": [ + { + "bbox": [ + 102, + 329, + 203, + 353 + ], + "score": 1.0, + "content": "with generated samples", + "type": "text" + }, + { + "bbox": [ + 203, + 334, + 236, + 348 + ], + "score": 0.92, + "content": "( \\mathbf { z } _ { k } ) _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 329, + 349, + 353 + ], + "score": 1.0, + "content": ", we use the sample average", + "type": "text" + }, + { + "bbox": [ + 350, + 334, + 358, + 347 + ], + "score": 0.87, + "content": "\\hat { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 329, + 402, + 353 + ], + "score": 1.0, + "content": "defined as", + "type": "text" + }, + { + "bbox": [ + 402, + 333, + 493, + 348 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\hat { \\psi } _ { K } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 329, + 509, + 353 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 347, + 449, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 158, + 363 + ], + "score": 1.0, + "content": "approximate", + "type": "text" + }, + { + "bbox": [ + 159, + 348, + 167, + 361 + ], + "score": 0.85, + "content": "\\bar { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 348, + 263, + 363 + ], + "score": 1.0, + "content": ". We define a functional", + "type": "text" + }, + { + "bbox": [ + 263, + 347, + 272, + 361 + ], + "score": 0.86, + "content": "\\tilde { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 348, + 449, + 363 + ], + "score": 1.0, + "content": "that solves the following Poisson Equation:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "interline_equation", + "bbox": [ + 260, + 365, + 351, + 381 + ], + "lines": [ + { + "bbox": [ + 260, + 365, + 351, + 381 + ], + "spans": [ + { + "bbox": [ + 260, + 365, + 351, + 381 + ], + "score": 0.93, + "content": "\\mathcal { L } \\tilde { \\psi } ( \\mathbf { z } _ { k } ) = \\psi ( \\mathbf { z } _ { k } ) - \\bar { \\psi }", + "type": "interline_equation", + "image_path": "27f14d4df4498051faa9ea9cfbe33c59a6117e2a35022a9cfac5565c4f78cc31.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 260, + 365, + 351, + 381 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 393, + 276, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 273, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 265, + 408 + ], + "score": 1.0, + "content": "We make the following assumptions on", + "type": "text" + }, + { + "bbox": [ + 266, + 393, + 273, + 406 + ], + "score": 0.86, + "content": "\\tilde { \\psi }", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15 + }, + { + "type": "text", + "bbox": [ + 106, + 409, + 506, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 507, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 172, + 423 + ], + "score": 1.0, + "content": "Assumption 1.", + "type": "text" + }, + { + "bbox": [ + 173, + 408, + 181, + 422 + ], + "score": 0.82, + "content": "\\tilde { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 408, + 358, + 423 + ], + "score": 1.0, + "content": "exists, and its up to 4rd-order derivatives,", + "type": "text" + }, + { + "bbox": [ + 358, + 409, + 379, + 422 + ], + "score": 0.92, + "content": "\\mathcal { D } ^ { k } \\tilde { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 408, + 495, + 423 + ], + "score": 1.0, + "content": ", are bounded by a function", + "type": "text" + }, + { + "bbox": [ + 495, + 411, + 503, + 420 + ], + "score": 0.73, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 408, + 507, + 423 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 421, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 125, + 436 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 125, + 421, + 198, + 435 + ], + "score": 0.91, + "content": "\\lVert \\mathcal { D } ^ { k } \\tilde { \\psi } \\rVert \\leq C _ { k } \\mathcal { V } ^ { p _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 421, + 216, + 436 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 216, + 423, + 288, + 434 + ], + "score": 0.63, + "content": "k = ( 0 , 1 , 2 , 3 , 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 421, + 293, + 436 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 293, + 423, + 342, + 435 + ], + "score": 0.78, + "content": "C _ { k } , p _ { k } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 421, + 482, + 436 + ], + "score": 1.0, + "content": ". Furthermore, the expectation of", + "type": "text" + }, + { + "bbox": [ + 483, + 424, + 491, + 433 + ], + "score": 0.75, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 421, + 506, + 436 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 432, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 107, + 434, + 127, + 446 + ], + "score": 0.91, + "content": "\\left\\{ { \\bf z } _ { k } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 432, + 179, + 449 + ], + "score": 1.0, + "content": "is bounded:", + "type": "text" + }, + { + "bbox": [ + 179, + 434, + 258, + 446 + ], + "score": 0.91, + "content": "\\mathrm { s u p } _ { l } \\mathbb { E } \\mathcal { V } ^ { p } ( { \\mathbf { z } } _ { k } ) < \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 432, + 280, + 449 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 280, + 435, + 288, + 444 + ], + "score": 0.8, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 432, + 368, + 449 + ], + "score": 1.0, + "content": "is smooth such that", + "type": "text" + }, + { + "bbox": [ + 369, + 434, + 505, + 447 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\operatorname* { s u p } _ { s \\in ( 0 , 1 ) } \\mathcal { V } ^ { p } \\left( s \\mathbf { z } + \\left( 1 - s \\right) \\mathbf { y } \\right) \\leq } \\end{array}", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 445, + 354, + 460 + ], + "spans": [ + { + "bbox": [ + 107, + 446, + 188, + 459 + ], + "score": 0.75, + "content": "C \\left( \\mathcal { V } ^ { p } \\left( \\mathbf { z } \\right) + \\mathcal { V } ^ { p } \\left( \\mathbf { y } \\right) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 445, + 194, + 460 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 194, + 447, + 285, + 459 + ], + "score": 0.79, + "content": "\\forall \\mathbf { z } , \\mathbf { y } , p \\leq \\operatorname* { m a x } \\{ 2 p _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 445, + 322, + 460 + ], + "score": 1.0, + "content": "for some", + "type": "text" + }, + { + "bbox": [ + 323, + 447, + 350, + 457 + ], + "score": 0.89, + "content": "C > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 445, + 354, + 460 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5 + }, + { + "type": "title", + "bbox": [ + 108, + 473, + 250, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 252, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 252, + 488 + ], + "score": 1.0, + "content": "B PROOFS FOR SECTION 3", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 105, + 498, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 201, + 513 + ], + "score": 1.0, + "content": "Sketch Proof of Lemma", + "type": "text" + }, + { + "bbox": [ + 202, + 500, + 207, + 509 + ], + "score": 0.37, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 497, + 506, + 513 + ], + "score": 1.0, + "content": ". First note that (5) in Lemma 1 corresponds to eq.13 in Jordan et al. (1998),", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 509, + 437, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 133, + 524 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 510, + 155, + 522 + ], + "score": 0.91, + "content": "F ( p )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 509, + 313, + 524 + ], + "score": 1.0, + "content": "in Jordan et al. (1998) is in the form of", + "type": "text" + }, + { + "bbox": [ + 313, + 510, + 378, + 522 + ], + "score": 0.91, + "content": "\\mathrm { K L } ( \\rho \\| p _ { \\pm } ( \\bar { \\bf x } , { \\bf z } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 509, + 437, + 524 + ], + "score": 1.0, + "content": "in our setting.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5 + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "Proposition 4.1 in Jordan et al. (1998) then proves that (5) has a unique solution. Theorem 5.1 in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 538, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 550 + ], + "score": 1.0, + "content": "Jordan et al. (1998) then guarantees that the solution of (5) approach the solution of the Fokker-Planck", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 549, + 308, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 206, + 560 + ], + "score": 1.0, + "content": "equation in (3), which is", + "type": "text" + }, + { + "bbox": [ + 206, + 550, + 219, + 560 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 549, + 277, + 560 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + }, + { + "bbox": [ + 277, + 549, + 304, + 559 + ], + "score": 0.91, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 549, + 308, + 560 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24 + }, + { + "type": "text", + "bbox": [ + 105, + 565, + 487, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 488, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 208, + 579 + ], + "score": 1.0, + "content": "Since this is true for each", + "type": "text" + }, + { + "bbox": [ + 208, + 566, + 215, + 576 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 564, + 258, + 579 + ], + "score": 1.0, + "content": "(thus each", + "type": "text" + }, + { + "bbox": [ + 258, + 567, + 263, + 576 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 564, + 274, + 579 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 274, + 567, + 284, + 577 + ], + "score": 0.82, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 564, + 360, + 579 + ], + "score": 1.0, + "content": "), we conclude that", + "type": "text" + }, + { + "bbox": [ + 360, + 566, + 399, + 577 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k } = \\rho _ { h k }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 564, + 456, + 579 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + }, + { + "bbox": [ + 456, + 566, + 484, + 576 + ], + "score": 0.89, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 564, + 488, + 579 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 506, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "To prove Theorem 2, we first need a convergence result about convergence to equilibrium in Wasser-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "stein distance for Fokker-Planck equations, which is presented in Bolley et al. (2012). Putting in our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 612, + 454, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 454, + 624 + ], + "score": 1.0, + "content": "setting, we can get the following lemma based on Corollary 2.4 in Bolley et al. (2012).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28 + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 253, + 639 + ], + "score": 1.0, + "content": "Lemma 5 (Bolley et al. (2012)). Let", + "type": "text" + }, + { + "bbox": [ + 253, + 628, + 266, + 638 + ], + "score": 0.79, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 626, + 351, + 639 + ], + "score": 1.0, + "content": "be the solution of the", + "type": "text" + }, + { + "bbox": [ + 351, + 627, + 365, + 637 + ], + "score": 0.44, + "content": "F P", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 626, + 445, + 639 + ], + "score": 1.0, + "content": "equation (3) at time", + "type": "text" + }, + { + "bbox": [ + 446, + 627, + 453, + 637 + ], + "score": 0.58, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 626, + 457, + 639 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 457, + 626, + 492, + 639 + ], + "score": 0.87, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 637, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 321, + 653 + ], + "score": 1.0, + "content": "the joint posterior distribution given x. Assume that", + "type": "text" + }, + { + "bbox": [ + 322, + 637, + 430, + 651 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\int \\rho _ { T } ( \\mathbf { z } ) p _ { \\pmb \\theta } ^ { - 1 } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } < \\infty } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 637, + 497, + 653 + ], + "score": 1.0, + "content": "and there exists", + "type": "text" + }, + { + "bbox": [ + 498, + 641, + 504, + 649 + ], + "score": 0.57, + "content": "a", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 649, + 379, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 143, + 665 + ], + "score": 1.0, + "content": "constant", + "type": "text" + }, + { + "bbox": [ + 143, + 654, + 152, + 664 + ], + "score": 0.75, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 653, + 192, + 665 + ], + "score": 1.0, + "content": "such tha", + "type": "text" + }, + { + "bbox": [ + 189, + 649, + 379, + 667 + ], + "score": 1.0, + "content": "t dW 22 (ρT ,pθ(x,z)) ≥ CW 22 (ρT , pθ(x, z)). Then", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31 + }, + { + "type": "interline_equation", + "bbox": [ + 216, + 671, + 395, + 686 + ], + "lines": [ + { + "bbox": [ + 216, + 671, + 395, + 686 + ], + "spans": [ + { + "bbox": [ + 216, + 671, + 395, + 686 + ], + "score": 0.87, + "content": "W _ { 2 } \\left( \\rho _ { T } , p ( { \\bf x } , { \\bf z } ) \\right) \\leq W _ { 2 } \\left( \\rho _ { 0 } , p ( { \\bf x } , { \\bf z } ) \\right) e ^ { - C T } .", + "type": "interline_equation", + "image_path": "9940691449e4abfa1d52b38b2247dbdb1018ebd321720f5c3284fedf7997a230.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 216, + 671, + 395, + 686 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 697, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 697, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 506, + 711 + ], + "score": 1.0, + "content": "We further need to borrow convergence results from Mattingly et al. (2010); Vollmer et al. (2016);", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "score": 1.0, + "content": "Chen et al. (2015) to characterize error bounds of a numerical integrator for the diffusion (2).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 414, + 734 + ], + "score": 1.0, + "content": "Specifically, the goal is to evaluate the posterior average of a test function", + "type": "text" + }, + { + "bbox": [ + 415, + 720, + 435, + 732 + ], + "score": 0.92, + "content": "\\psi ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 720, + 484, + 734 + ], + "score": 1.0, + "content": ", defined as", + "type": "text" + }, + { + "bbox": [ + 484, + 720, + 493, + 732 + ], + "score": 0.7, + "content": "\\bar { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 720, + 506, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + } + ], + "page_idx": 13, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "14", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 495, + 566, + 504, + 576 + ], + "lines": [ + { + "bbox": [ + 496, + 568, + 504, + 576 + ], + "spans": [ + { + "bbox": [ + 496, + 568, + 504, + 576 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 146, + 78, + 465, + 136 + ], + "lines": [ + { + "bbox": [ + 180, + 78, + 435, + 97 + ], + "spans": [ + { + "bbox": [ + 180, + 78, + 435, + 97 + ], + "score": 1.0, + "content": "SUPPLEMENTARY MATERIAL FOR:", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 98, + 466, + 117 + ], + "spans": [ + { + "bbox": [ + 145, + 98, + 466, + 117 + ], + "score": 1.0, + "content": "CONTINUOUS-TIME FLOWS FOR EFFICIENT", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 158, + 118, + 452, + 136 + ], + "spans": [ + { + "bbox": [ + 158, + 118, + 452, + 136 + ], + "score": 1.0, + "content": "INFERENCE AND DENSITY ESTIMATION", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 184, + 284, + 197 + ], + "lines": [ + { + "bbox": [ + 105, + 183, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 105, + 183, + 286, + 199 + ], + "score": 1.0, + "content": "A ASSUMPTIONS OF THEOREM 2", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 107, + 209, + 505, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 105, + 209, + 470, + 222 + ], + "score": 1.0, + "content": "First, let us define the infinitesimal generator of the diffusion (2). Formally, the generator", + "type": "text" + }, + { + "bbox": [ + 470, + 210, + 478, + 219 + ], + "score": 0.82, + "content": "\\mathcal { L }", + "type": "inline_equation" + }, + { + "bbox": [ + 478, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "of the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 219, + 505, + 233 + ], + "spans": [ + { + "bbox": [ + 105, + 219, + 427, + 233 + ], + "score": 1.0, + "content": "diffusion (2) is defined for any compactly supported twice differentiable function", + "type": "text" + }, + { + "bbox": [ + 427, + 220, + 481, + 232 + ], + "score": 0.94, + "content": "\\breve { f } : \\mathbf { R } ^ { L } \\mathbf { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 481, + 219, + 505, + 233 + ], + "score": 1.0, + "content": ", such", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 230, + 128, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 230, + 128, + 245 + ], + "score": 1.0, + "content": "that,", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 209, + 505, + 245 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 117, + 246, + 492, + 274 + ], + "lines": [ + { + "bbox": [ + 117, + 246, + 492, + 274 + ], + "spans": [ + { + "bbox": [ + 117, + 246, + 492, + 274 + ], + "score": 0.94, + "content": "\\mathcal { L } f ( \\mathbf { Z } _ { t } ) \\triangleq \\operatorname* { l i m } _ { h 0 ^ { + } } \\frac { \\mathbb { E } [ f ( \\mathbf { Z } _ { t + h } ) ] - f ( \\mathbf { Z } _ { t } ) } { h } = ( F ( \\mathbf { Z } _ { t } ) \\cdot \\nabla + \\frac { 1 } { 2 } ( G ( \\mathbf { Z } _ { t } ) G ( \\mathbf { Z } _ { t } ) ^ { T } ) \\colon \\nabla \\nabla ^ { T } ) f ( \\mathbf { Z } _ { t } ) ,", + "type": "interline_equation", + "image_path": "836de84f757b84ca0f70a7f98051d6709f159c13ad0e0d33891b3dcfc5b9e778.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 117, + 246, + 492, + 274 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "text", + "bbox": [ + 104, + 280, + 504, + 304 + ], + "lines": [ + { + "bbox": [ + 105, + 278, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 105, + 278, + 133, + 295 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 280, + 181, + 292 + ], + "score": 0.41, + "content": "{ \\mathbf { a } } \\cdot { \\mathbf { b } } \\triangleq { \\mathbf { a } } ^ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 278, + 192, + 295 + ], + "score": 1.0, + "content": "b,", + "type": "text" + }, + { + "bbox": [ + 192, + 279, + 267, + 293 + ], + "score": 0.72, + "content": "\\mathbf { A } : \\mathbf { B } \\triangleq { \\mathrm { t r } } ( \\mathbf { A } ^ { T } \\mathbf { B } )", + "type": "inline_equation" + }, + { + "bbox": [ + 267, + 278, + 270, + 295 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 271, + 280, + 306, + 291 + ], + "score": 0.86, + "content": "h \\to 0 ^ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 278, + 335, + 295 + ], + "score": 1.0, + "content": "means", + "type": "text" + }, + { + "bbox": [ + 336, + 282, + 343, + 291 + ], + "score": 0.82, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 278, + 505, + 295 + ], + "score": 1.0, + "content": "approaches zero along the positive real", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 291, + 128, + 304 + ], + "spans": [ + { + "bbox": [ + 105, + 291, + 128, + 304 + ], + "score": 1.0, + "content": "axis.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 105, + 278, + 505, + 304 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 308, + 505, + 361 + ], + "lines": [ + { + "bbox": [ + 106, + 308, + 505, + 321 + ], + "spans": [ + { + "bbox": [ + 106, + 308, + 346, + 321 + ], + "score": 1.0, + "content": "Given an ergodic diffusion (2) with an invariant measure", + "type": "text" + }, + { + "bbox": [ + 347, + 309, + 367, + 321 + ], + "score": 0.92, + "content": "\\rho ( \\mathbf { Z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 308, + 505, + 321 + ], + "score": 1.0, + "content": ", the posterior average is defined", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 320, + 506, + 335 + ], + "spans": [ + { + "bbox": [ + 105, + 320, + 122, + 335 + ], + "score": 1.0, + "content": "as:", + "type": "text" + }, + { + "bbox": [ + 122, + 320, + 210, + 334 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\bar { \\psi } \\triangleq \\int \\psi ( \\mathbf { Z } ) \\rho ( \\mathbf { Z } ) \\mathrm { d } \\mathbf { Z } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 210, + 320, + 305, + 335 + ], + "score": 1.0, + "content": "for some test function", + "type": "text" + }, + { + "bbox": [ + 305, + 321, + 327, + 334 + ], + "score": 0.92, + "content": "\\psi ( \\mathbf { Z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 328, + 320, + 506, + 335 + ], + "score": 1.0, + "content": "of interest. For a given numerical method", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 102, + 329, + 509, + 353 + ], + "spans": [ + { + "bbox": [ + 102, + 329, + 203, + 353 + ], + "score": 1.0, + "content": "with generated samples", + "type": "text" + }, + { + "bbox": [ + 203, + 334, + 236, + 348 + ], + "score": 0.92, + "content": "( \\mathbf { z } _ { k } ) _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 236, + 329, + 349, + 353 + ], + "score": 1.0, + "content": ", we use the sample average", + "type": "text" + }, + { + "bbox": [ + 350, + 334, + 358, + 347 + ], + "score": 0.87, + "content": "\\hat { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 358, + 329, + 402, + 353 + ], + "score": 1.0, + "content": "defined as", + "type": "text" + }, + { + "bbox": [ + 402, + 333, + 493, + 348 + ], + "score": 0.93, + "content": "\\begin{array} { r } { \\hat { \\psi } _ { K } = \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 494, + 329, + 509, + 353 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 347, + 449, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 158, + 363 + ], + "score": 1.0, + "content": "approximate", + "type": "text" + }, + { + "bbox": [ + 159, + 348, + 167, + 361 + ], + "score": 0.85, + "content": "\\bar { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 348, + 263, + 363 + ], + "score": 1.0, + "content": ". We define a functional", + "type": "text" + }, + { + "bbox": [ + 263, + 347, + 272, + 361 + ], + "score": 0.86, + "content": "\\tilde { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 348, + 449, + 363 + ], + "score": 1.0, + "content": "that solves the following Poisson Equation:", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 102, + 308, + 509, + 363 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 260, + 365, + 351, + 381 + ], + "lines": [ + { + "bbox": [ + 260, + 365, + 351, + 381 + ], + "spans": [ + { + "bbox": [ + 260, + 365, + 351, + 381 + ], + "score": 0.93, + "content": "\\mathcal { L } \\tilde { \\psi } ( \\mathbf { z } _ { k } ) = \\psi ( \\mathbf { z } _ { k } ) - \\bar { \\psi }", + "type": "interline_equation", + "image_path": "27f14d4df4498051faa9ea9cfbe33c59a6117e2a35022a9cfac5565c4f78cc31.jpg" + } + ] + } + ], + "index": 14, + "virtual_lines": [ + { + "bbox": [ + 260, + 365, + 351, + 381 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 393, + 276, + 406 + ], + "lines": [ + { + "bbox": [ + 105, + 392, + 273, + 408 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 265, + 408 + ], + "score": 1.0, + "content": "We make the following assumptions on", + "type": "text" + }, + { + "bbox": [ + 266, + 393, + 273, + 406 + ], + "score": 0.86, + "content": "\\tilde { \\psi }", + "type": "inline_equation" + } + ], + "index": 15 + } + ], + "index": 15, + "bbox_fs": [ + 105, + 392, + 273, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 409, + 506, + 459 + ], + "lines": [ + { + "bbox": [ + 105, + 408, + 507, + 423 + ], + "spans": [ + { + "bbox": [ + 105, + 408, + 172, + 423 + ], + "score": 1.0, + "content": "Assumption 1.", + "type": "text" + }, + { + "bbox": [ + 173, + 408, + 181, + 422 + ], + "score": 0.82, + "content": "\\tilde { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 408, + 358, + 423 + ], + "score": 1.0, + "content": "exists, and its up to 4rd-order derivatives,", + "type": "text" + }, + { + "bbox": [ + 358, + 409, + 379, + 422 + ], + "score": 0.92, + "content": "\\mathcal { D } ^ { k } \\tilde { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 379, + 408, + 495, + 423 + ], + "score": 1.0, + "content": ", are bounded by a function", + "type": "text" + }, + { + "bbox": [ + 495, + 411, + 503, + 420 + ], + "score": 0.73, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 503, + 408, + 507, + 423 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 421, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 421, + 125, + 436 + ], + "score": 1.0, + "content": "i.e.,", + "type": "text" + }, + { + "bbox": [ + 125, + 421, + 198, + 435 + ], + "score": 0.91, + "content": "\\lVert \\mathcal { D } ^ { k } \\tilde { \\psi } \\rVert \\leq C _ { k } \\mathcal { V } ^ { p _ { k } }", + "type": "inline_equation" + }, + { + "bbox": [ + 199, + 421, + 216, + 436 + ], + "score": 1.0, + "content": "for", + "type": "text" + }, + { + "bbox": [ + 216, + 423, + 288, + 434 + ], + "score": 0.63, + "content": "k = ( 0 , 1 , 2 , 3 , 4 )", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 421, + 293, + 436 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 293, + 423, + 342, + 435 + ], + "score": 0.78, + "content": "C _ { k } , p _ { k } > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 421, + 482, + 436 + ], + "score": 1.0, + "content": ". Furthermore, the expectation of", + "type": "text" + }, + { + "bbox": [ + 483, + 424, + 491, + 433 + ], + "score": 0.75, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 491, + 421, + 506, + 436 + ], + "score": 1.0, + "content": "on", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 107, + 432, + 505, + 449 + ], + "spans": [ + { + "bbox": [ + 107, + 434, + 127, + 446 + ], + "score": 0.91, + "content": "\\left\\{ { \\bf z } _ { k } \\right\\}", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 432, + 179, + 449 + ], + "score": 1.0, + "content": "is bounded:", + "type": "text" + }, + { + "bbox": [ + 179, + 434, + 258, + 446 + ], + "score": 0.91, + "content": "\\mathrm { s u p } _ { l } \\mathbb { E } \\mathcal { V } ^ { p } ( { \\mathbf { z } } _ { k } ) < \\infty", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 432, + 280, + 449 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 280, + 435, + 288, + 444 + ], + "score": 0.8, + "content": "\\nu", + "type": "inline_equation" + }, + { + "bbox": [ + 289, + 432, + 368, + 449 + ], + "score": 1.0, + "content": "is smooth such that", + "type": "text" + }, + { + "bbox": [ + 369, + 434, + 505, + 447 + ], + "score": 0.87, + "content": "\\begin{array} { r } { \\operatorname* { s u p } _ { s \\in ( 0 , 1 ) } \\mathcal { V } ^ { p } \\left( s \\mathbf { z } + \\left( 1 - s \\right) \\mathbf { y } \\right) \\leq } \\end{array}", + "type": "inline_equation" + } + ], + "index": 18 + }, + { + "bbox": [ + 107, + 445, + 354, + 460 + ], + "spans": [ + { + "bbox": [ + 107, + 446, + 188, + 459 + ], + "score": 0.75, + "content": "C \\left( \\mathcal { V } ^ { p } \\left( \\mathbf { z } \\right) + \\mathcal { V } ^ { p } \\left( \\mathbf { y } \\right) \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 189, + 445, + 194, + 460 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 194, + 447, + 285, + 459 + ], + "score": 0.79, + "content": "\\forall \\mathbf { z } , \\mathbf { y } , p \\leq \\operatorname* { m a x } \\{ 2 p _ { k } \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 445, + 322, + 460 + ], + "score": 1.0, + "content": "for some", + "type": "text" + }, + { + "bbox": [ + 323, + 447, + 350, + 457 + ], + "score": 0.89, + "content": "C > 0", + "type": "inline_equation" + }, + { + "bbox": [ + 350, + 445, + 354, + 460 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 17.5, + "bbox_fs": [ + 105, + 408, + 507, + 460 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 473, + 250, + 487 + ], + "lines": [ + { + "bbox": [ + 105, + 473, + 252, + 488 + ], + "spans": [ + { + "bbox": [ + 105, + 473, + 252, + 488 + ], + "score": 1.0, + "content": "B PROOFS FOR SECTION 3", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 105, + 498, + 505, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 497, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 201, + 513 + ], + "score": 1.0, + "content": "Sketch Proof of Lemma", + "type": "text" + }, + { + "bbox": [ + 202, + 500, + 207, + 509 + ], + "score": 0.37, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 208, + 497, + 506, + 513 + ], + "score": 1.0, + "content": ". First note that (5) in Lemma 1 corresponds to eq.13 in Jordan et al. (1998),", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 509, + 437, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 509, + 133, + 524 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 510, + 155, + 522 + ], + "score": 0.91, + "content": "F ( p )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 509, + 313, + 524 + ], + "score": 1.0, + "content": "in Jordan et al. (1998) is in the form of", + "type": "text" + }, + { + "bbox": [ + 313, + 510, + 378, + 522 + ], + "score": 0.91, + "content": "\\mathrm { K L } ( \\rho \\| p _ { \\pm } ( \\bar { \\bf x } , { \\bf z } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 378, + 509, + 437, + 524 + ], + "score": 1.0, + "content": "in our setting.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21.5, + "bbox_fs": [ + 105, + 497, + 506, + 524 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 526, + 504, + 560 + ], + "lines": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "spans": [ + { + "bbox": [ + 106, + 527, + 505, + 538 + ], + "score": 1.0, + "content": "Proposition 4.1 in Jordan et al. (1998) then proves that (5) has a unique solution. Theorem 5.1 in", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 538, + 506, + 550 + ], + "spans": [ + { + "bbox": [ + 105, + 538, + 506, + 550 + ], + "score": 1.0, + "content": "Jordan et al. (1998) then guarantees that the solution of (5) approach the solution of the Fokker-Planck", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 549, + 308, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 549, + 206, + 560 + ], + "score": 1.0, + "content": "equation in (3), which is", + "type": "text" + }, + { + "bbox": [ + 206, + 550, + 219, + 560 + ], + "score": 0.85, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 219, + 549, + 277, + 560 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + }, + { + "bbox": [ + 277, + 549, + 304, + 559 + ], + "score": 0.91, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 305, + 549, + 308, + 560 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 24, + "bbox_fs": [ + 105, + 527, + 506, + 560 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 565, + 487, + 578 + ], + "lines": [ + { + "bbox": [ + 105, + 564, + 488, + 579 + ], + "spans": [ + { + "bbox": [ + 105, + 564, + 208, + 579 + ], + "score": 1.0, + "content": "Since this is true for each", + "type": "text" + }, + { + "bbox": [ + 208, + 566, + 215, + 576 + ], + "score": 0.79, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 215, + 564, + 258, + 579 + ], + "score": 1.0, + "content": "(thus each", + "type": "text" + }, + { + "bbox": [ + 258, + 567, + 263, + 576 + ], + "score": 0.74, + "content": "t", + "type": "inline_equation" + }, + { + "bbox": [ + 264, + 564, + 274, + 579 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 274, + 567, + 284, + 577 + ], + "score": 0.82, + "content": "\\rho _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 564, + 360, + 579 + ], + "score": 1.0, + "content": "), we conclude that", + "type": "text" + }, + { + "bbox": [ + 360, + 566, + 399, + 577 + ], + "score": 0.92, + "content": "\\tilde { \\rho } _ { k } = \\rho _ { h k }", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 564, + 456, + 579 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + }, + { + "bbox": [ + 456, + 566, + 484, + 576 + ], + "score": 0.89, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 564, + 488, + 579 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 564, + 488, + 579 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 589, + 506, + 623 + ], + "lines": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "spans": [ + { + "bbox": [ + 106, + 590, + 506, + 602 + ], + "score": 1.0, + "content": "To prove Theorem 2, we first need a convergence result about convergence to equilibrium in Wasser-", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 106, + 601, + 505, + 613 + ], + "score": 1.0, + "content": "stein distance for Fokker-Planck equations, which is presented in Bolley et al. (2012). Putting in our", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 612, + 454, + 624 + ], + "spans": [ + { + "bbox": [ + 106, + 612, + 454, + 624 + ], + "score": 1.0, + "content": "setting, we can get the following lemma based on Corollary 2.4 in Bolley et al. (2012).", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 28, + "bbox_fs": [ + 106, + 590, + 506, + 624 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 626, + 505, + 667 + ], + "lines": [ + { + "bbox": [ + 105, + 626, + 505, + 639 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 253, + 639 + ], + "score": 1.0, + "content": "Lemma 5 (Bolley et al. (2012)). Let", + "type": "text" + }, + { + "bbox": [ + 253, + 628, + 266, + 638 + ], + "score": 0.79, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 266, + 626, + 351, + 639 + ], + "score": 1.0, + "content": "be the solution of the", + "type": "text" + }, + { + "bbox": [ + 351, + 627, + 365, + 637 + ], + "score": 0.44, + "content": "F P", + "type": "inline_equation" + }, + { + "bbox": [ + 365, + 626, + 445, + 639 + ], + "score": 1.0, + "content": "equation (3) at time", + "type": "text" + }, + { + "bbox": [ + 446, + 627, + 453, + 637 + ], + "score": 0.58, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 626, + 457, + 639 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 457, + 626, + 492, + 639 + ], + "score": 0.87, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 492, + 626, + 505, + 639 + ], + "score": 1.0, + "content": "be", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 637, + 504, + 653 + ], + "spans": [ + { + "bbox": [ + 105, + 637, + 321, + 653 + ], + "score": 1.0, + "content": "the joint posterior distribution given x. Assume that", + "type": "text" + }, + { + "bbox": [ + 322, + 637, + 430, + 651 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\int \\rho _ { T } ( \\mathbf { z } ) p _ { \\pmb \\theta } ^ { - 1 } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } < \\infty } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 430, + 637, + 497, + 653 + ], + "score": 1.0, + "content": "and there exists", + "type": "text" + }, + { + "bbox": [ + 498, + 641, + 504, + 649 + ], + "score": 0.57, + "content": "a", + "type": "inline_equation" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 649, + 379, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 143, + 665 + ], + "score": 1.0, + "content": "constant", + "type": "text" + }, + { + "bbox": [ + 143, + 654, + 152, + 664 + ], + "score": 0.75, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 152, + 653, + 192, + 665 + ], + "score": 1.0, + "content": "such tha", + "type": "text" + }, + { + "bbox": [ + 189, + 649, + 379, + 667 + ], + "score": 1.0, + "content": "t dW 22 (ρT ,pθ(x,z)) ≥ CW 22 (ρT , pθ(x, z)). Then", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 31, + "bbox_fs": [ + 105, + 626, + 505, + 667 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 216, + 671, + 395, + 686 + ], + "lines": [ + { + "bbox": [ + 216, + 671, + 395, + 686 + ], + "spans": [ + { + "bbox": [ + 216, + 671, + 395, + 686 + ], + "score": 0.87, + "content": "W _ { 2 } \\left( \\rho _ { T } , p ( { \\bf x } , { \\bf z } ) \\right) \\leq W _ { 2 } \\left( \\rho _ { 0 } , p ( { \\bf x } , { \\bf z } ) \\right) e ^ { - C T } .", + "type": "interline_equation", + "image_path": "9940691449e4abfa1d52b38b2247dbdb1018ebd321720f5c3284fedf7997a230.jpg" + } + ] + } + ], + "index": 33, + "virtual_lines": [ + { + "bbox": [ + 216, + 671, + 395, + 686 + ], + "spans": [], + "index": 33 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 697, + 506, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 697, + 506, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 697, + 506, + 711 + ], + "score": 1.0, + "content": "We further need to borrow convergence results from Mattingly et al. (2010); Vollmer et al. (2016);", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "spans": [ + { + "bbox": [ + 106, + 708, + 506, + 721 + ], + "score": 1.0, + "content": "Chen et al. (2015) to characterize error bounds of a numerical integrator for the diffusion (2).", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 720, + 506, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 414, + 734 + ], + "score": 1.0, + "content": "Specifically, the goal is to evaluate the posterior average of a test function", + "type": "text" + }, + { + "bbox": [ + 415, + 720, + 435, + 732 + ], + "score": 0.92, + "content": "\\psi ( \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 435, + 720, + 484, + 734 + ], + "score": 1.0, + "content": ", defined as", + "type": "text" + }, + { + "bbox": [ + 484, + 720, + 493, + 732 + ], + "score": 0.7, + "content": "\\bar { \\psi }", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 720, + 506, + 734 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 80, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 182, + 95 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\int \\psi ( \\mathbf { z } ) p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 182, + 80, + 449, + 97 + ], + "score": 1.0, + "content": ". When using a numerical integrator to solve (2) to get samples", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 450, + 81, + 485, + 95 + ], + "score": 0.93, + "content": "\\{ \\mathbf { z } _ { k } \\} _ { k = 1 } ^ { K }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 485, + 80, + 506, + 97 + ], + "score": 1.0, + "content": ", the", + "type": "text", + "cross_page": true + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 88, + 508, + 114 + ], + "spans": [ + { + "bbox": [ + 102, + 88, + 171, + 114 + ], + "score": 1.0, + "content": "sample average", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 172, + 94, + 263, + 110 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\hat { \\psi } _ { K } \\triangleq \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) } \\end{array}", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 263, + 88, + 508, + 114 + ], + "score": 1.0, + "content": "k=1 is used to approximate the posterior average. The accuracy", + "type": "text", + "cross_page": true + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 108, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 110, + 353, + 129 + ], + "score": 1.0, + "content": "is characterized by the mean square error (MSE) defined as:", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 353, + 108, + 413, + 131 + ], + "score": 0.93, + "content": "\\mathbb { E } \\left( \\hat { \\psi } _ { K } - \\bar { \\psi } \\right) ^ { 2 }", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 413, + 110, + 506, + 129 + ], + "score": 1.0, + "content": ". Lemma 6 derives the", + "type": "text", + "cross_page": true + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 128, + 187, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 187, + 140 + ], + "score": 1.0, + "content": "bound for the MSE.", + "type": "text", + "cross_page": true + } + ], + "index": 3 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 697, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 140 + ], + "lines": [ + { + "bbox": [ + 106, + 80, + 506, + 97 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 182, + 95 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\int \\psi ( \\mathbf { z } ) p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 182, + 80, + 449, + 97 + ], + "score": 1.0, + "content": ". When using a numerical integrator to solve (2) to get samples", + "type": "text" + }, + { + "bbox": [ + 450, + 81, + 485, + 95 + ], + "score": 0.93, + "content": "\\{ \\mathbf { z } _ { k } \\} _ { k = 1 } ^ { K }", + "type": "inline_equation" + }, + { + "bbox": [ + 485, + 80, + 506, + 97 + ], + "score": 1.0, + "content": ", the", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 88, + 508, + 114 + ], + "spans": [ + { + "bbox": [ + 102, + 88, + 171, + 114 + ], + "score": 1.0, + "content": "sample average", + "type": "text" + }, + { + "bbox": [ + 172, + 94, + 263, + 110 + ], + "score": 0.92, + "content": "\\begin{array} { r } { \\hat { \\psi } _ { K } \\triangleq \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 263, + 88, + 508, + 114 + ], + "score": 1.0, + "content": "k=1 is used to approximate the posterior average. The accuracy", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 104, + 108, + 506, + 131 + ], + "spans": [ + { + "bbox": [ + 104, + 110, + 353, + 129 + ], + "score": 1.0, + "content": "is characterized by the mean square error (MSE) defined as:", + "type": "text" + }, + { + "bbox": [ + 353, + 108, + 413, + 131 + ], + "score": 0.93, + "content": "\\mathbb { E } \\left( \\hat { \\psi } _ { K } - \\bar { \\psi } \\right) ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 413, + 110, + 506, + 129 + ], + "score": 1.0, + "content": ". Lemma 6 derives the", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 128, + 187, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 128, + 187, + 140 + ], + "score": 1.0, + "content": "bound for the MSE.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 325, + 156 + ], + "score": 1.0, + "content": "Lemma 6 (Vollmer et al. (2016)). Under Assumption", + "type": "text" + }, + { + "bbox": [ + 326, + 144, + 331, + 153 + ], + "score": 0.67, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 141, + 506, + 156 + ], + "score": 1.0, + "content": ", and for a 1st-order numerical intergrator,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 154, + 373, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 249, + 167 + ], + "score": 1.0, + "content": "the MSE is bounded, for a constant", + "type": "text" + }, + { + "bbox": [ + 250, + 154, + 258, + 164 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 154, + 321, + 167 + ], + "score": 1.0, + "content": "independent of", + "type": "text" + }, + { + "bbox": [ + 321, + 154, + 328, + 164 + ], + "score": 0.44, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 154, + 347, + 167 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 347, + 154, + 357, + 164 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 154, + 373, + 167 + ], + "score": 1.0, + "content": ", by", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 168, + 378, + 196 + ], + "lines": [ + { + "bbox": [ + 233, + 168, + 378, + 196 + ], + "spans": [ + { + "bbox": [ + 233, + 168, + 378, + 196 + ], + "score": 0.94, + "content": "\\mathbb { E } \\left( \\hat { \\psi } _ { K } - \\bar { \\psi } \\right) ^ { 2 } \\leq C \\left( \\frac { 1 } { h K } + h ^ { 2 } \\right) ~ .", + "type": "interline_equation", + "image_path": "f7c58fbb325ea18ca21fccaa787d7af868550507f1deb15800a82eedd4b7babf.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 233, + 168, + 378, + 196 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 504, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "Furthermore, except for the 2nd-order Wasserstein distance defined in Lemma 1, we define the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 214, + 422, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 368, + 228 + ], + "score": 1.0, + "content": "1st-order Wasserstein distance between two probability measures", + "type": "text" + }, + { + "bbox": [ + 368, + 217, + 380, + 227 + ], + "score": 0.85, + "content": "\\mu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 214, + 398, + 228 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 217, + 409, + 227 + ], + "score": 0.85, + "content": "\\mu _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 214, + 422, + 228 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "interline_equation", + "bbox": [ + 201, + 230, + 409, + 257 + ], + "lines": [ + { + "bbox": [ + 201, + 230, + 409, + 257 + ], + "spans": [ + { + "bbox": [ + 201, + 230, + 409, + 257 + ], + "score": 0.91, + "content": "W _ { 1 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) \\triangleq \\operatorname* { i n f } _ { p \\in { \\mathscr P } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) } \\int \\| { \\mathbf x } - { \\mathbf y } \\| _ { 2 } p ( \\mathrm { d } { \\mathbf x } , \\mathrm { d } { \\mathbf y } ) .", + "type": "interline_equation", + "image_path": "9524acac499dae0f55bc7ac6ae4eab25ac34cd1f411af9a2f8f32cb20096a344.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 201, + 230, + 409, + 257 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 396, + 280 + ], + "score": 1.0, + "content": "According to the Kantorovich-Rubinstein duality Arjovsky et al. (2017),", + "type": "text" + }, + { + "bbox": [ + 396, + 266, + 444, + 279 + ], + "score": 0.93, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 264, + 505, + 280 + ], + "score": 1.0, + "content": "is equivalently", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 277, + 167, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 167, + 289 + ], + "score": 1.0, + "content": "represented as", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 292, + 409, + 313 + ], + "lines": [ + { + "bbox": [ + 200, + 292, + 409, + 313 + ], + "spans": [ + { + "bbox": [ + 200, + 292, + 409, + 313 + ], + "score": 0.92, + "content": "W _ { 1 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) = \\underset { f \\in \\mathcal { L } _ { 1 } } { \\operatorname* { s u p } } \\ \\mathbb { E } _ { \\mathbf { z } \\sim \\mu _ { 1 } } \\left[ f ( \\mathbf { z } ) \\right] - \\mathbb { E } _ { \\mathbf { z } \\sim \\mu _ { 2 } } \\left[ f ( \\mathbf { z } ) \\right] \\ ,", + "type": "interline_equation", + "image_path": "393d61b8fbebc773842fe49e897304ed634efcc6132e9e0add233df1a844c541.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 200, + 292, + 409, + 313 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 348, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 349, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 133, + 331 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 319, + 145, + 330 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 317, + 294, + 331 + ], + "score": 1.0, + "content": "is the space of 1-Lipschitz functions", + "type": "text" + }, + { + "bbox": [ + 294, + 317, + 345, + 330 + ], + "score": 0.92, + "content": "f : \\mathbb { R } ^ { L } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 317, + 349, + 331 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 384, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 385, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 267, + 349 + ], + "score": 1.0, + "content": "We have the following relation between", + "type": "text" + }, + { + "bbox": [ + 267, + 335, + 315, + 348 + ], + "score": 0.93, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 333, + 333, + 349 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 333, + 335, + 381, + 348 + ], + "score": 0.94, + "content": "W _ { 2 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 333, + 385, + 349 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 109, + 350, + 505, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 385, + 363 + ], + "score": 1.0, + "content": "Lemma 7 (Givens & Shortt (1984)). We have for any two distributions", + "type": "text" + }, + { + "bbox": [ + 385, + 352, + 397, + 362 + ], + "score": 0.86, + "content": "\\mu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 348, + 415, + 363 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 415, + 351, + 426, + 362 + ], + "score": 0.84, + "content": "\\mu _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 348, + 445, + 363 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 446, + 350, + 505, + 362 + ], + "score": 0.93, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } ) \\leq", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 358, + 159, + 376 + ], + "spans": [ + { + "bbox": [ + 107, + 361, + 154, + 373 + ], + "score": 0.92, + "content": "W _ { 2 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 358, + 159, + 376 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 253, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 253, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 253, + 394 + ], + "score": 1.0, + "content": "Now it is ready to prove Theorem 2.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 404, + 506, + 438 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "score": 1.0, + "content": "Proof of Theorem 2. The idea is to simply decompose the MSE into two parts, with one part charac-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 415, + 507, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 436, + 429 + ], + "score": 1.0, + "content": "tering the MSE of the numerical method, the other part charactering the MSE of", + "type": "text" + }, + { + "bbox": [ + 436, + 417, + 449, + 427 + ], + "score": 0.86, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 415, + 467, + 429 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 468, + 416, + 502, + 428 + ], + "score": 0.93, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 415, + 507, + 429 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 427, + 356, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 356, + 439 + ], + "score": 1.0, + "content": "which consequentially can be bounded using Lemma 5 above.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 192, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 193, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 193, + 457 + ], + "score": 1.0, + "content": "Specifically, we have", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 460, + 498, + 733 + ], + "lines": [ + { + "bbox": [ + 112, + 460, + 498, + 733 + ], + "spans": [ + { + "bbox": [ + 112, + 460, + 498, + 733 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\quad \\mathrm { M S E ~ } \\left( \\left( \\frac { 1 } { N } , \\frac { 1 } { N } , \\rho _ { t } , \\eta _ { t } , \\phi , z ( \\phi ) \\right) + \\mathbb { E } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } } \\\\ & { = \\mathbb { E } \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } \\exp \\left( \\int \\phi ( \\mathrm { d } \\rho z ) \\mathrm { d } z \\mathrm { d } z \\right) \\right) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } - \\int \\psi ( \\mathrm { d } \\rho \\eta ) \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) ^ { - 1 } \\int \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } - \\int \\psi ( \\mathrm { d } \\rho z ) \\mathrm { d } \\rho z \\mathrm { d } z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { 2 } \\right) ^ { - 1 } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { - 1 } \\int \\psi \\mathrm { d } \\eta \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\mathrm { d } \\mathrm { \\rho } \\mathrm { \\rho } \\Bigg ) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\exp \\left( - \\int \\psi ( \\mathrm { d } \\rho z ) \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } + \\mathbb { E } \\left( \\int \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) } \\\\ & \\quad \\times \\mathbb { E } \\left( \\frac { 1 } \\end{array}", + "type": "interline_equation", + "image_path": "6b824cf02479b7d23d9a9bada5bb8519b6db673f0f18b58997a6977dc083a080.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 112, + 460, + 498, + 551.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 112, + 551.0, + 498, + 642.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 112, + 642.0, + 498, + 733.0 + ], + "spans": [], + "index": 24 + } + ] + } + ], + "page_idx": 14, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "15", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 81, + 506, + 140 + ], + "lines": [], + "index": 1.5, + "bbox_fs": [ + 102, + 80, + 508, + 140 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 105, + 142, + 505, + 165 + ], + "lines": [ + { + "bbox": [ + 105, + 141, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 105, + 141, + 325, + 156 + ], + "score": 1.0, + "content": "Lemma 6 (Vollmer et al. (2016)). Under Assumption", + "type": "text" + }, + { + "bbox": [ + 326, + 144, + 331, + 153 + ], + "score": 0.67, + "content": "^ { l }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 141, + 506, + 156 + ], + "score": 1.0, + "content": ", and for a 1st-order numerical intergrator,", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 154, + 373, + 167 + ], + "spans": [ + { + "bbox": [ + 106, + 154, + 249, + 167 + ], + "score": 1.0, + "content": "the MSE is bounded, for a constant", + "type": "text" + }, + { + "bbox": [ + 250, + 154, + 258, + 164 + ], + "score": 0.82, + "content": "C", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 154, + 321, + 167 + ], + "score": 1.0, + "content": "independent of", + "type": "text" + }, + { + "bbox": [ + 321, + 154, + 328, + 164 + ], + "score": 0.44, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 329, + 154, + 347, + 167 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 347, + 154, + 357, + 164 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 357, + 154, + 373, + 167 + ], + "score": 1.0, + "content": ", by", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 141, + 506, + 167 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 233, + 168, + 378, + 196 + ], + "lines": [ + { + "bbox": [ + 233, + 168, + 378, + 196 + ], + "spans": [ + { + "bbox": [ + 233, + 168, + 378, + 196 + ], + "score": 0.94, + "content": "\\mathbb { E } \\left( \\hat { \\psi } _ { K } - \\bar { \\psi } \\right) ^ { 2 } \\leq C \\left( \\frac { 1 } { h K } + h ^ { 2 } \\right) ~ .", + "type": "interline_equation", + "image_path": "f7c58fbb325ea18ca21fccaa787d7af868550507f1deb15800a82eedd4b7babf.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 233, + 168, + 378, + 196 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 204, + 504, + 227 + ], + "lines": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 505, + 216 + ], + "score": 1.0, + "content": "Furthermore, except for the 2nd-order Wasserstein distance defined in Lemma 1, we define the", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 214, + 422, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 368, + 228 + ], + "score": 1.0, + "content": "1st-order Wasserstein distance between two probability measures", + "type": "text" + }, + { + "bbox": [ + 368, + 217, + 380, + 227 + ], + "score": 0.85, + "content": "\\mu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 380, + 214, + 398, + 228 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 398, + 217, + 409, + 227 + ], + "score": 0.85, + "content": "\\mu _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 410, + 214, + 422, + 228 + ], + "score": 1.0, + "content": "as", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 203, + 505, + 228 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 201, + 230, + 409, + 257 + ], + "lines": [ + { + "bbox": [ + 201, + 230, + 409, + 257 + ], + "spans": [ + { + "bbox": [ + 201, + 230, + 409, + 257 + ], + "score": 0.91, + "content": "W _ { 1 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) \\triangleq \\operatorname* { i n f } _ { p \\in { \\mathscr P } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) } \\int \\| { \\mathbf x } - { \\mathbf y } \\| _ { 2 } p ( \\mathrm { d } { \\mathbf x } , \\mathrm { d } { \\mathbf y } ) .", + "type": "interline_equation", + "image_path": "9524acac499dae0f55bc7ac6ae4eab25ac34cd1f411af9a2f8f32cb20096a344.jpg" + } + ] + } + ], + "index": 9, + "virtual_lines": [ + { + "bbox": [ + 201, + 230, + 409, + 257 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 266, + 505, + 289 + ], + "lines": [ + { + "bbox": [ + 105, + 264, + 505, + 280 + ], + "spans": [ + { + "bbox": [ + 105, + 264, + 396, + 280 + ], + "score": 1.0, + "content": "According to the Kantorovich-Rubinstein duality Arjovsky et al. (2017),", + "type": "text" + }, + { + "bbox": [ + 396, + 266, + 444, + 279 + ], + "score": 0.93, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 444, + 264, + 505, + 280 + ], + "score": 1.0, + "content": "is equivalently", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 277, + 167, + 289 + ], + "spans": [ + { + "bbox": [ + 105, + 277, + 167, + 289 + ], + "score": 1.0, + "content": "represented as", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 264, + 505, + 289 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 200, + 292, + 409, + 313 + ], + "lines": [ + { + "bbox": [ + 200, + 292, + 409, + 313 + ], + "spans": [ + { + "bbox": [ + 200, + 292, + 409, + 313 + ], + "score": 0.92, + "content": "W _ { 1 } \\left( \\mu _ { 1 } , \\mu _ { 2 } \\right) = \\underset { f \\in \\mathcal { L } _ { 1 } } { \\operatorname* { s u p } } \\ \\mathbb { E } _ { \\mathbf { z } \\sim \\mu _ { 1 } } \\left[ f ( \\mathbf { z } ) \\right] - \\mathbb { E } _ { \\mathbf { z } \\sim \\mu _ { 2 } } \\left[ f ( \\mathbf { z } ) \\right] \\ ,", + "type": "interline_equation", + "image_path": "393d61b8fbebc773842fe49e897304ed634efcc6132e9e0add233df1a844c541.jpg" + } + ] + } + ], + "index": 12, + "virtual_lines": [ + { + "bbox": [ + 200, + 292, + 409, + 313 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 318, + 348, + 331 + ], + "lines": [ + { + "bbox": [ + 105, + 317, + 349, + 331 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 133, + 331 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 319, + 145, + 330 + ], + "score": 0.89, + "content": "\\mathcal { L } _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 146, + 317, + 294, + 331 + ], + "score": 1.0, + "content": "is the space of 1-Lipschitz functions", + "type": "text" + }, + { + "bbox": [ + 294, + 317, + 345, + 330 + ], + "score": 0.92, + "content": "f : \\mathbb { R } ^ { L } \\mathbb { R }", + "type": "inline_equation" + }, + { + "bbox": [ + 345, + 317, + 349, + 331 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 317, + 349, + 331 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 335, + 384, + 348 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 385, + 349 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 267, + 349 + ], + "score": 1.0, + "content": "We have the following relation between", + "type": "text" + }, + { + "bbox": [ + 267, + 335, + 315, + 348 + ], + "score": 0.93, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 333, + 333, + 349 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 333, + 335, + 381, + 348 + ], + "score": 0.94, + "content": "W _ { 2 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 381, + 333, + 385, + 349 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14, + "bbox_fs": [ + 105, + 333, + 385, + 349 + ] + }, + { + "type": "text", + "bbox": [ + 109, + 350, + 505, + 373 + ], + "lines": [ + { + "bbox": [ + 106, + 348, + 505, + 363 + ], + "spans": [ + { + "bbox": [ + 106, + 348, + 385, + 363 + ], + "score": 1.0, + "content": "Lemma 7 (Givens & Shortt (1984)). We have for any two distributions", + "type": "text" + }, + { + "bbox": [ + 385, + 352, + 397, + 362 + ], + "score": 0.86, + "content": "\\mu _ { 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 397, + 348, + 415, + 363 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 415, + 351, + 426, + 362 + ], + "score": 0.84, + "content": "\\mu _ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 427, + 348, + 445, + 363 + ], + "score": 1.0, + "content": "that", + "type": "text" + }, + { + "bbox": [ + 446, + 350, + 505, + 362 + ], + "score": 0.93, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } ) \\leq", + "type": "inline_equation" + } + ], + "index": 15 + }, + { + "bbox": [ + 107, + 358, + 159, + 376 + ], + "spans": [ + { + "bbox": [ + 107, + 361, + 154, + 373 + ], + "score": 0.92, + "content": "W _ { 2 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 155, + 358, + 159, + 376 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 15.5, + "bbox_fs": [ + 106, + 348, + 505, + 376 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 380, + 253, + 393 + ], + "lines": [ + { + "bbox": [ + 105, + 380, + 253, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 253, + 394 + ], + "score": 1.0, + "content": "Now it is ready to prove Theorem 2.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 380, + 253, + 394 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 404, + 506, + 438 + ], + "lines": [ + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "spans": [ + { + "bbox": [ + 105, + 404, + 506, + 417 + ], + "score": 1.0, + "content": "Proof of Theorem 2. The idea is to simply decompose the MSE into two parts, with one part charac-", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 415, + 507, + 429 + ], + "spans": [ + { + "bbox": [ + 105, + 415, + 436, + 429 + ], + "score": 1.0, + "content": "tering the MSE of the numerical method, the other part charactering the MSE of", + "type": "text" + }, + { + "bbox": [ + 436, + 417, + 449, + 427 + ], + "score": 0.86, + "content": "\\rho _ { T }", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 415, + 467, + 429 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 468, + 416, + 502, + 428 + ], + "score": 0.93, + "content": "p _ { \\pmb { \\theta } } ( \\mathbf { x } , \\mathbf { z } )", + "type": "inline_equation" + }, + { + "bbox": [ + 502, + 415, + 507, + 429 + ], + "score": 1.0, + "content": ",", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 427, + 356, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 427, + 356, + 439 + ], + "score": 1.0, + "content": "which consequentially can be bounded using Lemma 5 above.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19, + "bbox_fs": [ + 105, + 404, + 507, + 439 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 443, + 192, + 455 + ], + "lines": [ + { + "bbox": [ + 106, + 443, + 193, + 457 + ], + "spans": [ + { + "bbox": [ + 106, + 443, + 193, + 457 + ], + "score": 1.0, + "content": "Specifically, we have", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 443, + 193, + 457 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 112, + 460, + 498, + 733 + ], + "lines": [ + { + "bbox": [ + 112, + 460, + 498, + 733 + ], + "spans": [ + { + "bbox": [ + 112, + 460, + 498, + 733 + ], + "score": 0.93, + "content": "\\begin{array} { r l } & { \\quad \\mathrm { M S E ~ } \\left( \\left( \\frac { 1 } { N } , \\frac { 1 } { N } , \\rho _ { t } , \\eta _ { t } , \\phi , z ( \\phi ) \\right) + \\mathbb { E } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } } \\\\ & { = \\mathbb { E } \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } \\exp \\left( \\int \\phi ( \\mathrm { d } \\rho z ) \\mathrm { d } z \\mathrm { d } z \\right) \\right) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } - \\int \\psi ( \\mathrm { d } \\rho \\eta ) \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) ^ { - 1 } \\int \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\sqrt { \\alpha _ { i } ( z ) } - \\int \\psi ( \\mathrm { d } \\rho z ) \\mathrm { d } \\rho z \\mathrm { d } z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { 2 } \\right) ^ { - 1 } \\left( \\int \\sqrt { \\phi } \\eta \\mathrm { d } \\rho z \\mathrm { d } z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) ^ { - 1 } \\int \\psi \\mathrm { d } \\eta \\mathrm { d } \\rho \\mathrm { d } \\rho \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\mathrm { d } \\mathrm { \\rho } \\mathrm { \\rho } \\Bigg ) ^ { - 1 } } \\\\ & { \\quad \\times \\mathbb { E } \\left( \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\exp \\left( - \\int \\psi ( \\mathrm { d } \\rho z ) \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\right) \\right) ^ { - 1 } + \\mathbb { E } \\left( \\int \\eta \\mathrm { d } \\rho z \\mathrm { d } \\rho z \\mathrm { d } \\mathrm { d } \\rho z \\right) } \\\\ & \\quad \\times \\mathbb { E } \\left( \\frac { 1 } \\end{array}", + "type": "interline_equation", + "image_path": "6b824cf02479b7d23d9a9bada5bb8519b6db673f0f18b58997a6977dc083a080.jpg" + } + ] + } + ], + "index": 23, + "virtual_lines": [ + { + "bbox": [ + 112, + 460, + 498, + 551.0 + ], + "spans": [], + "index": 22 + }, + { + "bbox": [ + 112, + 551.0, + 498, + 642.0 + ], + "spans": [], + "index": 23 + }, + { + "bbox": [ + 112, + 642.0, + 498, + 733.0 + ], + "spans": [], + "index": 24 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 80, + 506, + 122 + ], + "lines": [ + { + "bbox": [ + 101, + 76, + 510, + 101 + ], + "spans": [ + { + "bbox": [ + 101, + 76, + 262, + 101 + ], + "score": 1.0, + "content": "where “(1)” follows by the fact that", + "type": "text" + }, + { + "bbox": [ + 262, + 80, + 456, + 100 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbb { E } \\left( \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) - \\int \\psi ( \\mathbf { z } ) p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } \\right) = 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 76, + 510, + 101 + ], + "score": 1.0, + "content": "Chen et al.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 273, + 112 + ], + "score": 1.0, + "content": "(2015); “(2)” follows by the definition of", + "type": "text" + }, + { + "bbox": [ + 273, + 100, + 321, + 111 + ], + "score": 0.89, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 97, + 506, + 112 + ], + "score": 1.0, + "content": "in (14) and the 1-Lipschitz assumption of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 108, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 158, + 123 + ], + "score": 1.0, + "content": "test function", + "type": "text" + }, + { + "bbox": [ + 159, + 110, + 166, + 122 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 108, + 450, + 123 + ], + "score": 1.0, + "content": "; “(3)” follows by Lemma 7; “(4)” follows by Lemma 5 and Lemma 6.", + "type": "text" + }, + { + "bbox": [ + 495, + 111, + 505, + 120 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "title", + "bbox": [ + 108, + 137, + 257, + 151 + ], + "lines": [ + { + "bbox": [ + 105, + 136, + 259, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 259, + 153 + ], + "score": 1.0, + "content": "C CONNECTION TO WGAN", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "We derive the upper bound of the maximum likelihood estimator, which connects MacGAN to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 173, + 448, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 157, + 187 + ], + "score": 1.0, + "content": "WGAN. Let", + "type": "text" + }, + { + "bbox": [ + 158, + 176, + 168, + 186 + ], + "score": 0.83, + "content": "p _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 173, + 448, + 187 + ], + "score": 1.0, + "content": "be the data distribution, rewrite our maximum likelihood objective as", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 191, + 453, + 226 + ], + "lines": [ + { + "bbox": [ + 159, + 191, + 453, + 226 + ], + "spans": [ + { + "bbox": [ + 159, + 191, + 453, + 226 + ], + "score": 0.94, + "content": "\\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) = \\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\bigg ( U ( \\mathbf { x } _ { i } ; \\theta ) - \\log \\int e ^ { U ( \\mathbf { x } ; \\theta ) } \\mathrm { d } \\mathbf { x } \\bigg ) \\enspace .", + "type": "interline_equation", + "image_path": "522d55a249fc4e1b448ac7a67251345bcdcfe6aeb4209828b88e1b4ac68959ed.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 159, + 191, + 453, + 202.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 159, + 202.66666666666666, + 453, + 214.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 159, + 214.33333333333331, + 453, + 225.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 452, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 452, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 452, + 244 + ], + "score": 1.0, + "content": "The above maximum likelihood estimator can be bounded with Jensen’s inequality as:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 248, + 469, + 329 + ], + "lines": [ + { + "bbox": [ + 143, + 248, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 143, + 248, + 469, + 329 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\displaystyle \\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\log \\int \\frac { e ^ { U ( \\mathbf { x } ; \\theta ) } } { q _ { \\phi } ( \\mathbf { x } ; \\omega ) } q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\mathrm { d } \\mathbf { x } } \\\\ & { \\displaystyle \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log \\frac { e ^ { U ( \\mathbf { x } ; \\theta ) } } { q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\right] } \\\\ & { \\displaystyle = \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "ca21dfad937e6480c739cfc6b3141731f22c027f012b367a238d3ec07c8925a1.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 143, + 248, + 469, + 275.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 143, + 275.0, + 469, + 302.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 143, + 302.0, + 469, + 329.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 506, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 460, + 352 + ], + "score": 1.0, + "content": "This results in the same objective form as WGAN except that our model does not restrict", + "type": "text" + }, + { + "bbox": [ + 461, + 339, + 493, + 351 + ], + "score": 0.91, + "content": "U ( \\mathbf { x } ; \\pmb { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 338, + 506, + 352 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 348, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 381, + 365 + ], + "score": 1.0, + "content": "be 1-Lipschitz functions and the objective has an extra constant term", + "type": "text" + }, + { + "bbox": [ + 381, + 350, + 483, + 363 + ], + "score": 0.85, + "content": "\\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 348, + 507, + 365 + ], + "score": 1.0, + "content": "w.r.t.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 360, + 119, + 373 + ], + "spans": [ + { + "bbox": [ + 107, + 362, + 114, + 371 + ], + "score": 0.67, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 360, + 119, + 373 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14 + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 224, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 226, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 226, + 393 + ], + "score": 1.0, + "content": "Now we prove Proposition 4.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16 + }, + { + "type": "text", + "bbox": [ + 105, + 402, + 460, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 461, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 461, + 416 + ], + "score": 1.0, + "content": "Proof of Proposition 4. First it is clear that the equality in (16) is achieved if and only if", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 420, + 366, + 436 + ], + "lines": [ + { + "bbox": [ + 244, + 420, + 366, + 436 + ], + "spans": [ + { + "bbox": [ + 244, + 420, + 366, + 436 + ], + "score": 0.91, + "content": "q _ { \\phi } ( \\mathbf { x } ; \\omega ) = p _ { \\theta } ( \\mathbf { x } ) \\propto e ^ { U ( \\mathbf { x } ; \\theta ) } .", + "type": "interline_equation", + "image_path": "a715f4a4f805e84f353ad665f3d014b1be7d8e23fcbccbde1a52bac791302396.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 244, + 420, + 366, + 436 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 335, + 461 + ], + "score": 1.0, + "content": "From the description in Section 4 and (16), we know that", + "type": "text" + }, + { + "bbox": [ + 335, + 448, + 343, + 457 + ], + "score": 0.82, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 445, + 360, + 461 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 360, + 448, + 368, + 459 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 445, + 506, + 461 + ], + "score": 1.0, + "content": "share the same objective function,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 458, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 287, + 471 + ], + "score": 1.0, + "content": "which is an upper bound of the MLE in (16).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5 + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 506, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "Furthermore, based on the property of continuous-time flows (or formally Theorem 2), we know that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 118, + 499 + ], + "score": 0.84, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 486, + 201, + 500 + ], + "score": 1.0, + "content": "is learned such that", + "type": "text" + }, + { + "bbox": [ + 201, + 488, + 239, + 499 + ], + "score": 0.91, + "content": "q _ { \\phi } p _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 486, + 300, + 500 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + }, + { + "bbox": [ + 301, + 486, + 330, + 496 + ], + "score": 0.91, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "(or alternatively, we could achieve this by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "using a decreasing-step-size sequence in a numerical method, as proved in Chen et al. (2015)). When", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 107, + 510, + 141, + 521 + ], + "score": 0.9, + "content": "q _ { \\phi } = p _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 509, + 355, + 520 + ], + "score": 1.0, + "content": ", the equality in (16) is achieved, leading to the MLE.", + "type": "text" + }, + { + "bbox": [ + 494, + 509, + 505, + 519 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5 + }, + { + "type": "title", + "bbox": [ + 108, + 536, + 272, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 274, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 274, + 550 + ], + "score": 1.0, + "content": "D ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 106, + 560, + 350, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 351, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 351, + 574 + ], + "score": 1.0, + "content": "D.1 CALCULATING THE TESTING ELBO FOR MACVAE", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 209, + 594 + ], + "score": 1.0, + "content": "We follow the method in", + "type": "text" + }, + { + "bbox": [ + 209, + 583, + 222, + 592 + ], + "score": 0.64, + "content": "\\mathrm { P u }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 581, + 443, + 594 + ], + "score": 1.0, + "content": "et al. (2017a) for calculating the ELBO for a test data", + "type": "text" + }, + { + "bbox": [ + 443, + 584, + 455, + 593 + ], + "score": 0.87, + "content": "\\mathbf { x } _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 581, + 505, + 594 + ], + "score": 1.0, + "content": ". First, after", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 592, + 486, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 287, + 605 + ], + "score": 1.0, + "content": "distilling the CTF into the inference network", + "type": "text" + }, + { + "bbox": [ + 287, + 595, + 298, + 606 + ], + "score": 0.86, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 592, + 486, + 605 + ], + "score": 1.0, + "content": ", we have that the ELBO can be represented as", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5 + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 610, + 405, + 625 + ], + "lines": [ + { + "bbox": [ + 205, + 610, + 405, + 625 + ], + "spans": [ + { + "bbox": [ + 205, + 610, + 405, + 625 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\log p ( \\mathbf { x } _ { * } ) \\geq \\mathbb { E } _ { q _ { \\phi } } \\left[ \\log p _ { \\theta } ( \\mathbf { x } _ { * } , \\mathbf { z } _ { * } ) \\right] - \\mathbb { E } _ { q _ { \\phi } } \\left[ \\log q _ { \\phi } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "329c76880e24ab3a6b461a553d09af3719f11c498d3ceb4059e4246856c1648a.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 205, + 610, + 405, + 625 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 507, + 688 + ], + "lines": [ + { + "bbox": [ + 104, + 627, + 508, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 293, + 648 + ], + "score": 1.0, + "content": "The expectation is approximated with samples", + "type": "text" + }, + { + "bbox": [ + 293, + 631, + 331, + 645 + ], + "score": 0.94, + "content": "\\{ \\mathbf { z } _ { * j } \\} _ { j = 1 } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 627, + 352, + 648 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 352, + 632, + 422, + 644 + ], + "score": 0.92, + "content": "\\mathbf { z } _ { \\ast j } = f _ { \\phi } ( \\mathbf { x } _ { \\ast } , \\zeta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 627, + 443, + 648 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 443, + 632, + 489, + 644 + ], + "score": 0.92, + "content": "\\zeta _ { j } \\sim q _ { 0 } ( \\zeta )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 627, + 508, + 648 + ], + "score": 1.0, + "content": "the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 241, + 656 + ], + "score": 1.0, + "content": "standard isotropic normal. Here", + "type": "text" + }, + { + "bbox": [ + 241, + 643, + 253, + 655 + ], + "score": 0.89, + "content": "f _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "represents the deep neural network in the inference network.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 129, + 667 + ], + "score": 1.0, + "content": "Note", + "type": "text" + }, + { + "bbox": [ + 130, + 654, + 158, + 666 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } _ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "is not readily obtained. To evaluate it, we use the density transformation formula:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 665, + 242, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 242, + 688 + ], + "score": 0.65, + "content": "\\begin{array} { r } { q _ { \\phi } \\mathopen { } \\mathclose \\bgroup \\left( \\mathbf { z } _ { * } \\aftergroup \\egroup \\right) \\stackrel { } { = } q _ { 0 } \\mathopen { } \\mathclose \\bgroup \\left( \\zeta \\aftergroup \\egroup \\right) \\mathopen { } \\mathclose \\bgroup \\left| \\mathrm { d e t } \\frac { \\partial f _ { \\phi } \\mathopen { } \\mathclose \\bgroup \\left( \\mathbf { x } _ { * } , \\zeta \\aftergroup \\egroup \\right) } { \\partial \\zeta } \\aftergroup \\egroup \\right| ^ { - 1 } . } \\end{array}", + "type": "inline_equation", + "image_path": "98b9e8f080555c6c4ccd855fc1712d146c58a3f52f706db79c3aa8365a071601.jpg" + } + ], + "index": 33 + } + ], + "index": 31.5 + }, + { + "type": "title", + "bbox": [ + 107, + 699, + 248, + 711 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 249, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 249, + 712 + ], + "score": 1.0, + "content": "D.2 NETWORK ARCHITECTURE", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 369, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 370, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 370, + 734 + ], + "score": 1.0, + "content": "The architecture of the generator of MacGAN is given in Table 1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35 + } + ], + "page_idx": 15, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "16", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 80, + 506, + 122 + ], + "lines": [ + { + "bbox": [ + 101, + 76, + 510, + 101 + ], + "spans": [ + { + "bbox": [ + 101, + 76, + 262, + 101 + ], + "score": 1.0, + "content": "where “(1)” follows by the fact that", + "type": "text" + }, + { + "bbox": [ + 262, + 80, + 456, + 100 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\mathbb { E } \\left( \\frac { 1 } { K } \\sum _ { k = 1 } ^ { K } \\psi ( \\mathbf { z } _ { k } ) - \\int \\psi ( \\mathbf { z } ) p _ { \\theta } ( \\mathbf { x } , \\mathbf { z } ) \\mathrm { d } \\mathbf { z } \\right) = 0 } \\end{array}", + "type": "inline_equation" + }, + { + "bbox": [ + 456, + 76, + 510, + 101 + ], + "score": 1.0, + "content": "Chen et al.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 97, + 506, + 112 + ], + "spans": [ + { + "bbox": [ + 105, + 97, + 273, + 112 + ], + "score": 1.0, + "content": "(2015); “(2)” follows by the definition of", + "type": "text" + }, + { + "bbox": [ + 273, + 100, + 321, + 111 + ], + "score": 0.89, + "content": "W _ { 1 } ( \\mu _ { 1 } , \\mu _ { 2 } )", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 97, + 506, + 112 + ], + "score": 1.0, + "content": "in (14) and the 1-Lipschitz assumption of the", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 108, + 505, + 123 + ], + "spans": [ + { + "bbox": [ + 105, + 108, + 158, + 123 + ], + "score": 1.0, + "content": "test function", + "type": "text" + }, + { + "bbox": [ + 159, + 110, + 166, + 122 + ], + "score": 0.84, + "content": "\\psi", + "type": "inline_equation" + }, + { + "bbox": [ + 167, + 108, + 450, + 123 + ], + "score": 1.0, + "content": "; “(3)” follows by Lemma 7; “(4)” follows by Lemma 5 and Lemma 6.", + "type": "text" + }, + { + "bbox": [ + 495, + 111, + 505, + 120 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 101, + 76, + 510, + 123 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 137, + 257, + 151 + ], + "lines": [ + { + "bbox": [ + 105, + 136, + 259, + 153 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 259, + 153 + ], + "score": 1.0, + "content": "C CONNECTION TO WGAN", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + }, + { + "type": "text", + "bbox": [ + 105, + 162, + 505, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 176 + ], + "score": 1.0, + "content": "We derive the upper bound of the maximum likelihood estimator, which connects MacGAN to", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 173, + 448, + 187 + ], + "spans": [ + { + "bbox": [ + 105, + 173, + 157, + 187 + ], + "score": 1.0, + "content": "WGAN. Let", + "type": "text" + }, + { + "bbox": [ + 158, + 176, + 168, + 186 + ], + "score": 0.83, + "content": "p _ { r }", + "type": "inline_equation" + }, + { + "bbox": [ + 169, + 173, + 448, + 187 + ], + "score": 1.0, + "content": "be the data distribution, rewrite our maximum likelihood objective as", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 105, + 162, + 505, + 187 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 159, + 191, + 453, + 226 + ], + "lines": [ + { + "bbox": [ + 159, + 191, + 453, + 226 + ], + "spans": [ + { + "bbox": [ + 159, + 191, + 453, + 226 + ], + "score": 0.94, + "content": "\\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) = \\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\bigg ( U ( \\mathbf { x } _ { i } ; \\theta ) - \\log \\int e ^ { U ( \\mathbf { x } ; \\theta ) } \\mathrm { d } \\mathbf { x } \\bigg ) \\enspace .", + "type": "interline_equation", + "image_path": "522d55a249fc4e1b448ac7a67251345bcdcfe6aeb4209828b88e1b4ac68959ed.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 159, + 191, + 453, + 202.66666666666666 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 159, + 202.66666666666666, + 453, + 214.33333333333331 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 159, + 214.33333333333331, + 453, + 225.99999999999997 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 230, + 452, + 243 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 452, + 244 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 452, + 244 + ], + "score": 1.0, + "content": "The above maximum likelihood estimator can be bounded with Jensen’s inequality as:", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 105, + 228, + 452, + 244 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 143, + 248, + 469, + 329 + ], + "lines": [ + { + "bbox": [ + 143, + 248, + 469, + 329 + ], + "spans": [ + { + "bbox": [ + 143, + 248, + 469, + 329 + ], + "score": 0.94, + "content": "\\begin{array} { r l } & { \\displaystyle \\operatorname* { m a x } \\frac { 1 } { N } \\sum _ { i = 1 } ^ { N } \\log p _ { \\theta } ( \\mathbf { x } _ { i } ) \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\log \\int \\frac { e ^ { U ( \\mathbf { x } ; \\theta ) } } { q _ { \\phi } ( \\mathbf { x } ; \\omega ) } q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\mathrm { d } \\mathbf { x } } \\\\ & { \\displaystyle \\leq \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log \\frac { e ^ { U ( \\mathbf { x } ; \\theta ) } } { q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\right] } \\\\ & { \\displaystyle = \\operatorname* { m a x } \\mathbb { E } _ { \\mathbf { x } \\sim p _ { r } } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ U ( \\mathbf { x } ; \\theta ) \\right] - \\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "ca21dfad937e6480c739cfc6b3141731f22c027f012b367a238d3ec07c8925a1.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 143, + 248, + 469, + 275.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 143, + 275.0, + 469, + 302.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 143, + 302.0, + 469, + 329.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 338, + 506, + 372 + ], + "lines": [ + { + "bbox": [ + 105, + 338, + 506, + 352 + ], + "spans": [ + { + "bbox": [ + 105, + 338, + 460, + 352 + ], + "score": 1.0, + "content": "This results in the same objective form as WGAN except that our model does not restrict", + "type": "text" + }, + { + "bbox": [ + 461, + 339, + 493, + 351 + ], + "score": 0.91, + "content": "U ( \\mathbf { x } ; \\pmb { \\theta } )", + "type": "inline_equation" + }, + { + "bbox": [ + 493, + 338, + 506, + 352 + ], + "score": 1.0, + "content": "to", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 104, + 348, + 507, + 365 + ], + "spans": [ + { + "bbox": [ + 104, + 348, + 381, + 365 + ], + "score": 1.0, + "content": "be 1-Lipschitz functions and the objective has an extra constant term", + "type": "text" + }, + { + "bbox": [ + 381, + 350, + 483, + 363 + ], + "score": 0.85, + "content": "\\mathbb { E } _ { \\mathbf { x } \\sim q _ { \\phi } ( \\mathbf { x } ; \\omega ) } \\left[ \\log q _ { \\phi } ( \\mathbf { x } ; \\omega ) \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 348, + 507, + 365 + ], + "score": 1.0, + "content": "w.r.t.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 107, + 360, + 119, + 373 + ], + "spans": [ + { + "bbox": [ + 107, + 362, + 114, + 371 + ], + "score": 0.67, + "content": "\\pmb \\theta", + "type": "inline_equation" + }, + { + "bbox": [ + 114, + 360, + 119, + 373 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14, + "bbox_fs": [ + 104, + 338, + 507, + 373 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 378, + 224, + 389 + ], + "lines": [ + { + "bbox": [ + 105, + 376, + 226, + 393 + ], + "spans": [ + { + "bbox": [ + 105, + 376, + 226, + 393 + ], + "score": 1.0, + "content": "Now we prove Proposition 4.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 376, + 226, + 393 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 402, + 460, + 415 + ], + "lines": [ + { + "bbox": [ + 105, + 402, + 461, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 461, + 416 + ], + "score": 1.0, + "content": "Proof of Proposition 4. First it is clear that the equality in (16) is achieved if and only if", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17, + "bbox_fs": [ + 105, + 402, + 461, + 416 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 244, + 420, + 366, + 436 + ], + "lines": [ + { + "bbox": [ + 244, + 420, + 366, + 436 + ], + "spans": [ + { + "bbox": [ + 244, + 420, + 366, + 436 + ], + "score": 0.91, + "content": "q _ { \\phi } ( \\mathbf { x } ; \\omega ) = p _ { \\theta } ( \\mathbf { x } ) \\propto e ^ { U ( \\mathbf { x } ; \\theta ) } .", + "type": "interline_equation", + "image_path": "a715f4a4f805e84f353ad665f3d014b1be7d8e23fcbccbde1a52bac791302396.jpg" + } + ] + } + ], + "index": 18, + "virtual_lines": [ + { + "bbox": [ + 244, + 420, + 366, + 436 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 447, + 505, + 470 + ], + "lines": [ + { + "bbox": [ + 105, + 445, + 506, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 445, + 335, + 461 + ], + "score": 1.0, + "content": "From the description in Section 4 and (16), we know that", + "type": "text" + }, + { + "bbox": [ + 335, + 448, + 343, + 457 + ], + "score": 0.82, + "content": "\\pmb { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 343, + 445, + 360, + 461 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 360, + 448, + 368, + 459 + ], + "score": 0.86, + "content": "\\phi", + "type": "inline_equation" + }, + { + "bbox": [ + 369, + 445, + 506, + 461 + ], + "score": 1.0, + "content": "share the same objective function,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 458, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 106, + 458, + 287, + 471 + ], + "score": 1.0, + "content": "which is an upper bound of the MLE in (16).", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 19.5, + "bbox_fs": [ + 105, + 445, + 506, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 475, + 506, + 520 + ], + "lines": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "spans": [ + { + "bbox": [ + 105, + 475, + 506, + 487 + ], + "score": 1.0, + "content": "Furthermore, based on the property of continuous-time flows (or formally Theorem 2), we know that", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 486, + 506, + 500 + ], + "spans": [ + { + "bbox": [ + 106, + 488, + 118, + 499 + ], + "score": 0.84, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 118, + 486, + 201, + 500 + ], + "score": 1.0, + "content": "is learned such that", + "type": "text" + }, + { + "bbox": [ + 201, + 488, + 239, + 499 + ], + "score": 0.91, + "content": "q _ { \\phi } p _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 240, + 486, + 300, + 500 + ], + "score": 1.0, + "content": "in the limit of", + "type": "text" + }, + { + "bbox": [ + 301, + 486, + 330, + 496 + ], + "score": 0.91, + "content": "h 0", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 486, + 506, + 500 + ], + "score": 1.0, + "content": "(or alternatively, we could achieve this by", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 506, + 510 + ], + "score": 1.0, + "content": "using a decreasing-step-size sequence in a numerical method, as proved in Chen et al. (2015)). When", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 509, + 505, + 521 + ], + "spans": [ + { + "bbox": [ + 107, + 510, + 141, + 521 + ], + "score": 0.9, + "content": "q _ { \\phi } = p _ { \\theta }", + "type": "inline_equation" + }, + { + "bbox": [ + 141, + 509, + 355, + 520 + ], + "score": 1.0, + "content": ", the equality in (16) is achieved, leading to the MLE.", + "type": "text" + }, + { + "bbox": [ + 494, + 509, + 505, + 519 + ], + "score": 0.999, + "content": "□", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22.5, + "bbox_fs": [ + 105, + 475, + 506, + 521 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 536, + 272, + 549 + ], + "lines": [ + { + "bbox": [ + 106, + 536, + 274, + 550 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 274, + 550 + ], + "score": 1.0, + "content": "D ADDITIONAL EXPERIMENTS", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 106, + 560, + 350, + 573 + ], + "lines": [ + { + "bbox": [ + 106, + 561, + 351, + 574 + ], + "spans": [ + { + "bbox": [ + 106, + 561, + 351, + 574 + ], + "score": 1.0, + "content": "D.1 CALCULATING THE TESTING ELBO FOR MACVAE", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 581, + 505, + 605 + ], + "lines": [ + { + "bbox": [ + 105, + 581, + 505, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 581, + 209, + 594 + ], + "score": 1.0, + "content": "We follow the method in", + "type": "text" + }, + { + "bbox": [ + 209, + 583, + 222, + 592 + ], + "score": 0.64, + "content": "\\mathrm { P u }", + "type": "inline_equation" + }, + { + "bbox": [ + 222, + 581, + 443, + 594 + ], + "score": 1.0, + "content": "et al. (2017a) for calculating the ELBO for a test data", + "type": "text" + }, + { + "bbox": [ + 443, + 584, + 455, + 593 + ], + "score": 0.87, + "content": "\\mathbf { x } _ { * }", + "type": "inline_equation" + }, + { + "bbox": [ + 455, + 581, + 505, + 594 + ], + "score": 1.0, + "content": ". First, after", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 592, + 486, + 606 + ], + "spans": [ + { + "bbox": [ + 106, + 592, + 287, + 605 + ], + "score": 1.0, + "content": "distilling the CTF into the inference network", + "type": "text" + }, + { + "bbox": [ + 287, + 595, + 298, + 606 + ], + "score": 0.86, + "content": "q _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 298, + 592, + 486, + 605 + ], + "score": 1.0, + "content": ", we have that the ELBO can be represented as", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 27.5, + "bbox_fs": [ + 105, + 581, + 505, + 606 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 205, + 610, + 405, + 625 + ], + "lines": [ + { + "bbox": [ + 205, + 610, + 405, + 625 + ], + "spans": [ + { + "bbox": [ + 205, + 610, + 405, + 625 + ], + "score": 0.9, + "content": "\\begin{array} { r } { \\log p ( \\mathbf { x } _ { * } ) \\geq \\mathbb { E } _ { q _ { \\phi } } \\left[ \\log p _ { \\theta } ( \\mathbf { x } _ { * } , \\mathbf { z } _ { * } ) \\right] - \\mathbb { E } _ { q _ { \\phi } } \\left[ \\log q _ { \\phi } \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "329c76880e24ab3a6b461a553d09af3719f11c498d3ceb4059e4246856c1648a.jpg" + } + ] + } + ], + "index": 29, + "virtual_lines": [ + { + "bbox": [ + 205, + 610, + 405, + 625 + ], + "spans": [], + "index": 29 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 631, + 507, + 688 + ], + "lines": [ + { + "bbox": [ + 104, + 627, + 508, + 648 + ], + "spans": [ + { + "bbox": [ + 104, + 627, + 293, + 648 + ], + "score": 1.0, + "content": "The expectation is approximated with samples", + "type": "text" + }, + { + "bbox": [ + 293, + 631, + 331, + 645 + ], + "score": 0.94, + "content": "\\{ \\mathbf { z } _ { * j } \\} _ { j = 1 } ^ { M }", + "type": "inline_equation" + }, + { + "bbox": [ + 331, + 627, + 352, + 648 + ], + "score": 1.0, + "content": "with", + "type": "text" + }, + { + "bbox": [ + 352, + 632, + 422, + 644 + ], + "score": 0.92, + "content": "\\mathbf { z } _ { \\ast j } = f _ { \\phi } ( \\mathbf { x } _ { \\ast } , \\zeta _ { j } )", + "type": "inline_equation" + }, + { + "bbox": [ + 423, + 627, + 443, + 648 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 443, + 632, + 489, + 644 + ], + "score": 0.92, + "content": "\\zeta _ { j } \\sim q _ { 0 } ( \\zeta )", + "type": "inline_equation" + }, + { + "bbox": [ + 489, + 627, + 508, + 648 + ], + "score": 1.0, + "content": "the", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 642, + 506, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 642, + 241, + 656 + ], + "score": 1.0, + "content": "standard isotropic normal. Here", + "type": "text" + }, + { + "bbox": [ + 241, + 643, + 253, + 655 + ], + "score": 0.89, + "content": "f _ { \\phi }", + "type": "inline_equation" + }, + { + "bbox": [ + 253, + 642, + 506, + 656 + ], + "score": 1.0, + "content": "represents the deep neural network in the inference network.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 653, + 506, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 653, + 129, + 667 + ], + "score": 1.0, + "content": "Note", + "type": "text" + }, + { + "bbox": [ + 130, + 654, + 158, + 666 + ], + "score": 0.93, + "content": "q _ { \\phi } ( \\mathbf { z } _ { * } )", + "type": "inline_equation" + }, + { + "bbox": [ + 158, + 653, + 506, + 667 + ], + "score": 1.0, + "content": "is not readily obtained. To evaluate it, we use the density transformation formula:", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 665, + 242, + 688 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 242, + 688 + ], + "score": 0.65, + "content": "\\begin{array} { r } { q _ { \\phi } \\mathopen { } \\mathclose \\bgroup \\left( \\mathbf { z } _ { * } \\aftergroup \\egroup \\right) \\stackrel { } { = } q _ { 0 } \\mathopen { } \\mathclose \\bgroup \\left( \\zeta \\aftergroup \\egroup \\right) \\mathopen { } \\mathclose \\bgroup \\left| \\mathrm { d e t } \\frac { \\partial f _ { \\phi } \\mathopen { } \\mathclose \\bgroup \\left( \\mathbf { x } _ { * } , \\zeta \\aftergroup \\egroup \\right) } { \\partial \\zeta } \\aftergroup \\egroup \\right| ^ { - 1 } . } \\end{array}", + "type": "inline_equation", + "image_path": "98b9e8f080555c6c4ccd855fc1712d146c58a3f52f706db79c3aa8365a071601.jpg" + } + ], + "index": 33 + } + ], + "index": 31.5, + "bbox_fs": [ + 104, + 627, + 508, + 688 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 699, + 248, + 711 + ], + "lines": [ + { + "bbox": [ + 106, + 700, + 249, + 712 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 249, + 712 + ], + "score": 1.0, + "content": "D.2 NETWORK ARCHITECTURE", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 34 + }, + { + "type": "text", + "bbox": [ + 106, + 720, + 369, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 719, + 370, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 719, + 370, + 734 + ], + "score": 1.0, + "content": "The architecture of the generator of MacGAN is given in Table 1.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 719, + 370, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 101, + 467, + 168 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 211, + 90, + 401, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 210, + 89, + 401, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 89, + 401, + 102 + ], + "score": 1.0, + "content": "Table 1: Architecture of generator in MacGAN", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 144, + 101, + 467, + 168 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 144, + 101, + 467, + 168 + ], + "spans": [ + { + "bbox": [ + 144, + 101, + 467, + 168 + ], + "score": 0.972, + "html": "
Output SizeArchitecture
100×1100 ×10Linear,BN,ReLU
256×8×8512 × 4 × 4 deconv,256 5 × 5 kernels,ReLU, strike 2, BN
128 ×16 ×16256 × 8 × 8 deconv,128 5 × 5 kernels,ReLU, strike 2,BN
3× 32 × 32128 × 16 × 16 deconv,3 5 × 5 kernels,Tanh,strike 2
", + "type": "table", + "image_path": "0358d199f954b07c4ac42f553c3ad035695782e54e3e210720e39bd698999fc2.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 144, + 101, + 467, + 123.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 144, + 123.33333333333333, + 467, + 145.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 144, + 145.66666666666666, + 467, + 168.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 106, + 178, + 505, + 589 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 178, + 505, + 589 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 589 + ], + "score": 0.963, + "type": "image", + "image_path": "baf33b33adb73feefa1f183846c66ce80219a0cb8d10bbd7a1726cb544553764.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 106, + 178, + 505, + 315.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 315.0, + 505, + 452.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 452.0, + 505, + 589.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 113, + 594, + 494, + 606 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 593, + 496, + 608 + ], + "spans": [ + { + "bbox": [ + 114, + 593, + 496, + 608 + ], + "score": 1.0, + "content": "Figure 7: Generated images for MNIST datasets with MacGAN (top) and SteinGAN (bottom).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "title", + "bbox": [ + 107, + 628, + 230, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 628, + 231, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 231, + 641 + ], + "score": 1.0, + "content": "D.3 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 340, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 342, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 342, + 663 + ], + "score": 1.0, + "content": "Additional experimental results are given in Figure 7 – 12.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9 + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 337, + 688 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 339, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 339, + 689 + ], + "score": 1.0, + "content": "D.4 ROBUSTNESS OF THE DISCRETIZATION STEPSIZE", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 295, + 712 + ], + "score": 1.0, + "content": "To test the impact of the discretization stepsize", + "type": "text" + }, + { + "bbox": [ + 295, + 700, + 302, + 709 + ], + "score": 0.84, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "in (6), following SteinGAN Feng et al. (2017), we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "test MacGAN on the MNIST dataset, where ee use a simple Gaussian-Bernoulli Restricted Boltzmann", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "Machines as the energy-based model. We adopt the annealed importance sampling method to evaluate", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12 + } + ], + "page_idx": 16, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 14, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 101, + 467, + 168 + ], + "blocks": [ + { + "type": "table_caption", + "bbox": [ + 211, + 90, + 401, + 101 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 210, + 89, + 401, + 102 + ], + "spans": [ + { + "bbox": [ + 210, + 89, + 401, + 102 + ], + "score": 1.0, + "content": "Table 1: Architecture of generator in MacGAN", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "table_body", + "bbox": [ + 144, + 101, + 467, + 168 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 144, + 101, + 467, + 168 + ], + "spans": [ + { + "bbox": [ + 144, + 101, + 467, + 168 + ], + "score": 0.972, + "html": "
Output SizeArchitecture
100×1100 ×10Linear,BN,ReLU
256×8×8512 × 4 × 4 deconv,256 5 × 5 kernels,ReLU, strike 2, BN
128 ×16 ×16256 × 8 × 8 deconv,128 5 × 5 kernels,ReLU, strike 2,BN
3× 32 × 32128 × 16 × 16 deconv,3 5 × 5 kernels,Tanh,strike 2
", + "type": "table", + "image_path": "0358d199f954b07c4ac42f553c3ad035695782e54e3e210720e39bd698999fc2.jpg" + } + ] + } + ], + "index": 2, + "virtual_lines": [ + { + "bbox": [ + 144, + 101, + 467, + 123.33333333333333 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 144, + 123.33333333333333, + 467, + 145.66666666666666 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 144, + 145.66666666666666, + 467, + 168.0 + ], + "spans": [], + "index": 3 + } + ] + } + ], + "index": 1.0 + }, + { + "type": "image", + "bbox": [ + 106, + 178, + 505, + 589 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 106, + 178, + 505, + 589 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 178, + 505, + 589 + ], + "spans": [ + { + "bbox": [ + 106, + 178, + 505, + 589 + ], + "score": 0.963, + "type": "image", + "image_path": "baf33b33adb73feefa1f183846c66ce80219a0cb8d10bbd7a1726cb544553764.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 106, + 178, + 505, + 315.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 106, + 315.0, + 505, + 452.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 452.0, + 505, + 589.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 113, + 594, + 494, + 606 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 593, + 496, + 608 + ], + "spans": [ + { + "bbox": [ + 114, + 593, + 496, + 608 + ], + "score": 1.0, + "content": "Figure 7: Generated images for MNIST datasets with MacGAN (top) and SteinGAN (bottom).", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + }, + { + "type": "title", + "bbox": [ + 107, + 628, + 230, + 640 + ], + "lines": [ + { + "bbox": [ + 106, + 628, + 231, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 628, + 231, + 641 + ], + "score": 1.0, + "content": "D.3 ADDITIONAL RESULTS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 107, + 650, + 340, + 662 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 342, + 663 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 342, + 663 + ], + "score": 1.0, + "content": "Additional experimental results are given in Figure 7 – 12.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 9, + "bbox_fs": [ + 106, + 649, + 342, + 663 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 677, + 337, + 688 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 339, + 689 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 339, + 689 + ], + "score": 1.0, + "content": "D.4 ROBUSTNESS OF THE DISCRETIZATION STEPSIZE", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + }, + { + "type": "text", + "bbox": [ + 108, + 698, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 295, + 712 + ], + "score": 1.0, + "content": "To test the impact of the discretization stepsize", + "type": "text" + }, + { + "bbox": [ + 295, + 700, + 302, + 709 + ], + "score": 0.84, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "in (6), following SteinGAN Feng et al. (2017), we", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "test MacGAN on the MNIST dataset, where ee use a simple Gaussian-Bernoulli Restricted Boltzmann", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 505, + 733 + ], + "score": 1.0, + "content": "Machines as the energy-based model. We adopt the annealed importance sampling method to evaluate", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12, + "bbox_fs": [ + 105, + 698, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 80, + 504, + 479 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 80, + 504, + 479 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 504, + 479 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 504, + 479 + ], + "score": 0.971, + "type": "image", + "image_path": "df70e18f19fb102d117a9f4a20fddc9375221a3993e0003772e47a29cb6ee7c3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 80, + 504, + 213.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 213.0, + 504, + 346.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 346.0, + 504, + 479.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 175, + 486, + 434, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 175, + 485, + 435, + 499 + ], + "spans": [ + { + "bbox": [ + 175, + 485, + 435, + 499 + ], + "score": 1.0, + "content": "Figure 8: Generated images for CelebA datasets with MacGAN.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 108, + 520, + 506, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 276, + 534 + ], + "score": 1.0, + "content": "log-likelihoods Feng et al. (2017). We vary", + "type": "text" + }, + { + "bbox": [ + 277, + 522, + 284, + 531 + ], + "score": 0.78, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 520, + 294, + 534 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 295, + 521, + 503, + 533 + ], + "score": 0.88, + "content": "\\{ 6 e - 4 , 2 . 4 e - 3 , 3 . 6 e - 3 , 6 e - 3 , 1 e - 2 , 1 . 5 e - 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 504, + 520, + 506, + 534 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "score": 1.0, + "content": "The trend of log-likelihoods is plotted in Figure 13. We can see that log-likelihoods do not change a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 543, + 402, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 390, + 555 + ], + "score": 1.0, + "content": "lot within the chosen stepsize interval, demonstrating the robustness of", + "type": "text" + }, + { + "bbox": [ + 391, + 543, + 398, + 553 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 543, + 402, + 555 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 166, + 569, + 445, + 709 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 166, + 569, + 445, + 709 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 166, + 569, + 445, + 709 + ], + "spans": [ + { + "bbox": [ + 166, + 569, + 445, + 709 + ], + "score": 0.968, + "type": "image", + "image_path": "510b1ba5679995516cf70d0e8a807c5b979fddd698b18ac7c1a42b5722b649e0.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 166, + 569, + 445, + 615.6666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 166, + 615.6666666666666, + 445, + 662.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 166, + 662.3333333333333, + 445, + 708.9999999999999 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 148, + 718, + 462, + 730 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 147, + 717, + 464, + 731 + ], + "spans": [ + { + "bbox": [ + 147, + 717, + 464, + 731 + ], + "score": 1.0, + "content": "Figure 13: Log-likelihoods vs discretization stepsize for MacGAN on MNIST.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + } + ], + "page_idx": 17, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 763 + ], + "score": 1.0, + "content": "18", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 80, + 504, + 479 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 80, + 504, + 479 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 80, + 504, + 479 + ], + "spans": [ + { + "bbox": [ + 107, + 80, + 504, + 479 + ], + "score": 0.971, + "type": "image", + "image_path": "df70e18f19fb102d117a9f4a20fddc9375221a3993e0003772e47a29cb6ee7c3.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 80, + 504, + 213.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 213.0, + 504, + 346.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 346.0, + 504, + 479.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 175, + 486, + 434, + 498 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 175, + 485, + 435, + 499 + ], + "spans": [ + { + "bbox": [ + 175, + 485, + 435, + 499 + ], + "score": 1.0, + "content": "Figure 8: Generated images for CelebA datasets with MacGAN.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 108, + 520, + 506, + 554 + ], + "lines": [ + { + "bbox": [ + 106, + 520, + 506, + 534 + ], + "spans": [ + { + "bbox": [ + 106, + 520, + 276, + 534 + ], + "score": 1.0, + "content": "log-likelihoods Feng et al. (2017). We vary", + "type": "text" + }, + { + "bbox": [ + 277, + 522, + 284, + 531 + ], + "score": 0.78, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 284, + 520, + 294, + 534 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 295, + 521, + 503, + 533 + ], + "score": 0.88, + "content": "\\{ 6 e - 4 , 2 . 4 e - 3 , 3 . 6 e - 3 , 6 e - 3 , 1 e - 2 , 1 . 5 e - 2 \\}", + "type": "inline_equation" + }, + { + "bbox": [ + 504, + 520, + 506, + 534 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "spans": [ + { + "bbox": [ + 105, + 531, + 506, + 545 + ], + "score": 1.0, + "content": "The trend of log-likelihoods is plotted in Figure 13. We can see that log-likelihoods do not change a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 543, + 402, + 555 + ], + "spans": [ + { + "bbox": [ + 106, + 543, + 390, + 555 + ], + "score": 1.0, + "content": "lot within the chosen stepsize interval, demonstrating the robustness of", + "type": "text" + }, + { + "bbox": [ + 391, + 543, + 398, + 553 + ], + "score": 0.81, + "content": "h", + "type": "inline_equation" + }, + { + "bbox": [ + 398, + 543, + 402, + 555 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5, + "bbox_fs": [ + 105, + 520, + 506, + 555 + ] + }, + { + "type": "image", + "bbox": [ + 166, + 569, + 445, + 709 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 166, + 569, + 445, + 709 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 166, + 569, + 445, + 709 + ], + "spans": [ + { + "bbox": [ + 166, + 569, + 445, + 709 + ], + "score": 0.968, + "type": "image", + "image_path": "510b1ba5679995516cf70d0e8a807c5b979fddd698b18ac7c1a42b5722b649e0.jpg" + } + ] + } + ], + "index": 8, + "virtual_lines": [ + { + "bbox": [ + 166, + 569, + 445, + 615.6666666666666 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 166, + 615.6666666666666, + 445, + 662.3333333333333 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 166, + 662.3333333333333, + 445, + 708.9999999999999 + ], + "spans": [], + "index": 9 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 148, + 718, + 462, + 730 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 147, + 717, + 464, + 731 + ], + "spans": [ + { + "bbox": [ + 147, + 717, + 464, + 731 + ], + "score": 1.0, + "content": "Figure 13: Log-likelihoods vs discretization stepsize for MacGAN on MNIST.", + "type": "text" + } + ], + "index": 10 + } + ], + "index": 10 + } + ], + "index": 9.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 83, + 504, + 282 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 83, + 504, + 282 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 83, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 504, + 282 + ], + "score": 0.969, + "type": "image", + "image_path": "2590ba66def78acb17923aea4b076627cdc952e917fb7aa8f72cc66ce1929962.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 83, + 504, + 149.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 149.33333333333331, + 504, + 215.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 215.66666666666663, + 504, + 281.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 170, + 289, + 440, + 301 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 288, + 442, + 303 + ], + "spans": [ + { + "bbox": [ + 169, + 288, + 442, + 303 + ], + "score": 1.0, + "content": "Figure 9: Generated images for CIFAR-10 datasets with MacGAN.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 310, + 505, + 709 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 310, + 505, + 709 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 310, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 107, + 310, + 505, + 709 + ], + "score": 0.972, + "type": "image", + "image_path": "20b4b0a19beab48e103ccf648183466dd297f9eee0a29d1029a2884965d2eb58.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 310, + 505, + 443.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 443.0, + 505, + 576.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 576.0, + 505, + 709.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 716, + 438, + 728 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 171, + 716, + 439, + 730 + ], + "spans": [ + { + "bbox": [ + 171, + 716, + 439, + 730 + ], + "score": 1.0, + "content": "Figure 10: Generated images for CelebA datasets with SteinGAN.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + } + ], + "page_idx": 18, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "19", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 83, + 504, + 282 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 83, + 504, + 282 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 83, + 504, + 282 + ], + "spans": [ + { + "bbox": [ + 107, + 83, + 504, + 282 + ], + "score": 0.969, + "type": "image", + "image_path": "2590ba66def78acb17923aea4b076627cdc952e917fb7aa8f72cc66ce1929962.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 83, + 504, + 149.33333333333331 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 149.33333333333331, + 504, + 215.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 215.66666666666663, + 504, + 281.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 170, + 289, + 440, + 301 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 169, + 288, + 442, + 303 + ], + "spans": [ + { + "bbox": [ + 169, + 288, + 442, + 303 + ], + "score": 1.0, + "content": "Figure 9: Generated images for CIFAR-10 datasets with MacGAN.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "image", + "bbox": [ + 107, + 310, + 505, + 709 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 310, + 505, + 709 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 107, + 310, + 505, + 709 + ], + "spans": [ + { + "bbox": [ + 107, + 310, + 505, + 709 + ], + "score": 0.972, + "type": "image", + "image_path": "20b4b0a19beab48e103ccf648183466dd297f9eee0a29d1029a2884965d2eb58.jpg" + } + ] + } + ], + "index": 5, + "virtual_lines": [ + { + "bbox": [ + 107, + 310, + 505, + 443.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 107, + 443.0, + 505, + 576.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 107, + 576.0, + 505, + 709.0 + ], + "spans": [], + "index": 6 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 172, + 716, + 438, + 728 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 171, + 716, + 439, + 730 + ], + "spans": [ + { + "bbox": [ + 171, + 716, + 439, + 730 + ], + "score": 1.0, + "content": "Figure 10: Generated images for CelebA datasets with SteinGAN.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 7 + } + ], + "index": 6.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 296, + 505, + 495 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 296, + 505, + 495 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 296, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 107, + 296, + 505, + 495 + ], + "score": 0.975, + "type": "image", + "image_path": "dafbac1d644b9153e4d3341ddb10e46b6d37a0d08e2929035e46e43910c12d6f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 296, + 505, + 362.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 362.3333333333333, + 505, + 428.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 428.66666666666663, + 505, + 494.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 166, + 502, + 444, + 515 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 166, + 502, + 444, + 516 + ], + "spans": [ + { + "bbox": [ + 166, + 502, + 444, + 516 + ], + "score": 1.0, + "content": "Figure 11: Generated images for CIFAR-10 datasets with SteinGAN.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ], + "page_idx": 19, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 764 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 15 + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 106, + 26, + 308, + 38 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 296, + 505, + 495 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 107, + 296, + 505, + 495 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 296, + 505, + 495 + ], + "spans": [ + { + "bbox": [ + 107, + 296, + 505, + 495 + ], + "score": 0.975, + "type": "image", + "image_path": "dafbac1d644b9153e4d3341ddb10e46b6d37a0d08e2929035e46e43910c12d6f.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 107, + 296, + 505, + 362.3333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 107, + 362.3333333333333, + 505, + 428.66666666666663 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 107, + 428.66666666666663, + 505, + 494.99999999999994 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 166, + 502, + 444, + 515 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 166, + 502, + 444, + 516 + ], + "spans": [ + { + "bbox": [ + 166, + 502, + 444, + 516 + ], + "score": 1.0, + "content": "Figure 11: Generated images for CIFAR-10 datasets with SteinGAN.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 76, + 465, + 715 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 76, + 465, + 715 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 76, + 465, + 715 + ], + "spans": [ + { + "bbox": [ + 146, + 76, + 465, + 715 + ], + "score": 0.958, + "type": "image", + "image_path": "dfc2e36b0e28db379525f7f5e2f26eecca5dd5a6bc6132ffa3a69178104ec0d7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 146, + 76, + 465, + 289.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 146, + 289.0, + 465, + 502.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 502.0, + 465, + 715.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 723, + 502, + 747 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 723, + 503, + 735 + ], + "spans": [ + { + "bbox": [ + 107, + 723, + 330, + 735 + ], + "score": 1.0, + "content": "Figure 12: Generated images with a random walk on the", + "type": "text" + }, + { + "bbox": [ + 330, + 726, + 338, + 734 + ], + "score": 0.72, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 723, + 503, + 735 + ], + "score": 1.0, + "content": "space for CelebA datasets with MacGAN,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 734, + 249, + 748 + ], + "spans": [ + { + "bbox": [ + 106, + 734, + 245, + 747 + ], + "score": 0.87, + "content": "\\omega _ { t } \\overline { { = \\omega _ { t - 1 } + 0 . 0 2 \\times \\mathrm { r a n d } ( [ - 1 , 1 ] ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 734, + 249, + 748 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ], + "page_idx": 20, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 25, + 309, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 25, + 309, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2018", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 761 + ], + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 764 + ], + "score": 1.0, + "content": "21", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 146, + 76, + 465, + 715 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 146, + 76, + 465, + 715 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 146, + 76, + 465, + 715 + ], + "spans": [ + { + "bbox": [ + 146, + 76, + 465, + 715 + ], + "score": 0.958, + "type": "image", + "image_path": "dfc2e36b0e28db379525f7f5e2f26eecca5dd5a6bc6132ffa3a69178104ec0d7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 146, + 76, + 465, + 289.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 146, + 289.0, + 465, + 502.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 146, + 502.0, + 465, + 715.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 108, + 723, + 502, + 747 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 107, + 723, + 503, + 735 + ], + "spans": [ + { + "bbox": [ + 107, + 723, + 330, + 735 + ], + "score": 1.0, + "content": "Figure 12: Generated images with a random walk on the", + "type": "text" + }, + { + "bbox": [ + 330, + 726, + 338, + 734 + ], + "score": 0.72, + "content": "\\omega", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 723, + 503, + 735 + ], + "score": 1.0, + "content": "space for CelebA datasets with MacGAN,", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 734, + 249, + 748 + ], + "spans": [ + { + "bbox": [ + 106, + 734, + 245, + 747 + ], + "score": 0.87, + "content": "\\omega _ { t } \\overline { { = \\omega _ { t - 1 } + 0 . 0 2 \\times \\mathrm { r a n d } ( [ - 1 , 1 ] ) } }", + "type": "inline_equation" + }, + { + "bbox": [ + 246, + 734, + 249, + 748 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/rkgZaT4tDr/rkgZaT4tDr.md b/parse/train/rkgZaT4tDr/rkgZaT4tDr.md new file mode 100644 index 0000000000000000000000000000000000000000..354a4f8aa3272ea3be862899dca30a6d6cefd1fa --- /dev/null +++ b/parse/train/rkgZaT4tDr/rkgZaT4tDr.md @@ -0,0 +1,257 @@ +# PARAMETERIZED ACTION REINFORCEMENT LEARNING FOR INVERTED INDEX MATCH PLAN GENERATION + +Anonymous authors Paper under double-blind review + +# ABSTRACT + +Match plan generation in the inverted index at Microsoft Bing is used to be based on hand-crafted rules. We formulate the generation process as a Parameterized Action MDP with sharing parameters and purpose a reinforcement learning algorithm on such formulation. We combine deterministic policy learning on discrete and continuous action spaces and several recent advances in deep reinforcement learning. For exploring in the parameterized action space, the agent outputs softmax values for discrete actions and applies Parameter Space Noise on policy network to unify the exploration direction in both spaces. We apply prioritized recurrent replay on match plan sequences and pad short match plans. We also use invertible value function rescaling and $n$ -step return to stabilize the training. The agent is evaluated on our environment and some benchmarks. It outperforms the well-designed production match plan and beats the baselines on the benchmarks. + +# 1 INTRODUCTION + +Using machine learning to optimize and accelerate software and hardware systems is an emerging field in the past few years (Mirhoseini et al. (2017); Rosset et al. (2018)). The process of decisionmaking in those systems is usually hand-crafted by human engineers, which is less explored to be automatically done by learning algorithms. A promising direction is to formulate such sequential decision-making problems as Reinforcement Learning (RL) problems, such as search plan generation problem in inverted indexes (Rosset et al. (2018)). + +The inverted index is a specialized data structure that is commonly used in search engines, including Microsoft $\mathrm { B i n g ^ { 1 } }$ . An inverted index provides an inverted mapping from term to documents. Each term has a posting list which contains all the (document, location) pairs that the term appears. By combining the posting lists of the terms in a user query, the initial document candidates are generated (Witten et al. (1999); Zobel & Moffat (2006)). + +In Bing, documents are scanned with specified match plans, either predefined or generated in realtime. A match plan has a sequence of rewrites (match rules, e.g., it treats the query as a phrase which should appear in the document exactly as it is), where each rewrite can be controlled by several quotas (stopping criteria). There are several types of rewrites, and all these types share the same continuous quotas. A pair of a rewrite type and the quotas forms a single action which determines if a document will be a ranking candidate. + +We formulate the generation problem as a RL problem. The state consists of system runtime signals and semantic embeddings of queries. The action space is called parameterized or discretecontinuous hybrid, where an action has a discrete action and continuous action-parameters. The reward is weighted by result quality and query latency. It is similar to Parameterized Action RL (Masson et al. (2016)) (PARL), while our setting requires all actions to share same parameters. + +Previously, the match plans are predefined manually for each query. It hardly utilizes the rich information in the state to dynamically adjust the rewrites and quotas for specific query online. Rosset et al. (2018) tries to automatically generate match plans using tabular Q-learning with discretized state space and predefined action-parameters. They learn the generation policy for a specific query class each time and solve it only in discretized spaces with tabular methods. We extend the generation process to the general case, such that the match plans are fully parameterized and learned from scratch without any predefined knowledge (e.g., limited match rules). + +![](images/7459112fc13a2ca51b6133f349d4373e932db4f4ab389ec3ae4eb649b28ac3c6.jpg) +Figure 1: Match plan example: for the query “Reinforcement Learning”, the search engine firstly gets a document posting list for each term from the inverted index, and then scans the document candidates following a serial of rewrites (match rules) according to the match plan. + +In this paper, we purpose a parameterized action RL algorithm that learns to act on parameterized action space environments, such as the match plan generation process. We introduce normalized softmax values of discrete actions to form a categorical distributions to enable gradients backpropagation. We also combine several recent advances in RL to accelerate and stabilize the training. We investigate parameter space noise (Plappert et al. (2017)) on parameters of the policy for unifying directions in exploring the structured action space. We explore recurrent deterministic policies (Heess et al. (2015)) with prioritized replay buffer (Schaul et al. (2015)) on sequence, due to the inherent nature of partial observability and sparse feedback in our setting. We also use $n$ -step return and invertible value function rescaling (Kapturowski et al. (2018)) for further stability. + +We present an agent to integrate these techniques and evaluate on offline training on a query dataset collected from Bing search. When training on the dataset with uniformly sampling, the agent outperforms the currently deployed production match plan, which is generated based on well-designed hand-crafted rules. We study the performance of training on the dataset or only with several selected complicated queries individually, as well as ablation studies for the various components. We test on a few existing PARL benchmarks, where our agent beats our baseline methods and performs the state-of-the-art results. + +# 2 BACKGROUND + +# 2.1 REINFORCEMENT LEARNING + +We address the problem using reinforcement learning (RL) framework with a parameterized action space. An agent interacts with an environment to maximize the accumulated reward with discount factor $\gamma \in \ [ 0 , 1 )$ . We model the environment with a discrete-time Partially Observable Markov Decision Process (POMDP). A POMDP is given by a tuple $( S , \mathcal { A } , T , R , \Omega , \dot { \mathcal { O } } )$ , and the underlying Markov Decision Process (MDP) is defined by $( S , { \mathcal { A } } , T , R )$ , where $s$ is the state space, $\mathcal { A }$ the parameterized action space, $T$ the transition function $\mathcal { T } : \mathcal { S } \times \mathcal { S } \times \mathcal { A } \mathbb { R } _ { + }$ , and $R : S \times \mathcal { A } \mathbb { R }$ is the reward function. The set of observations is given by $\Omega$ and the the observation function mapping underlying states to probability distributions over observations is given by $\mathcal { O }$ . + +Specifically, the environment in the match plan generation has a parameterized action space. In Masson et al. (2016); Bester et al. (2019), such formulation in fully observable settings is referred as Parameterized Action MDP (PAMDP), where the action space is denoted as + +$$ +\mathcal { A } = \bigcup _ { k \in \mathcal { A } _ { d } } \left\{ ( k , x _ { k } ) | x _ { k } \in \mathcal { X } _ { k } \right\} , +$$ + +where each discrete action $a \in \mathcal { A } _ { d } = [ K ]$ has a corresponding continuous action-parameter space $\mathcal { X } _ { a }$ . However, our setting requires a slightly different formulation: + +$$ +\mathcal { A } = \{ ( k , x ) | k \in \mathcal { A } _ { d } , x \in \mathcal { X } \} = \mathcal { A } _ { d } \times \mathcal { X } , +$$ + +where the discrete action space (rewrite) $\mathbf { \mathcal { A } } _ { d }$ share the action-parameter space $\mathcal { X }$ . Such formulation results in a disentangled action space between discrete and continuous actions which a class of parameterized action (P-DQN (Xiong et al. (2018)) and MP-DQN (Bester et al. (2019))) may not be trivially applicable to. Considering the notational simplicity in partially observable environment (POMDP), we do not explicitly denote it when there is no ambiguity in action settings. + +# 2.2 FORMULATION OF MATCH PLAN GENERATION + +In Rosset et al. (2018), the action-parameters are predefined for the match plan generation, where the action space is then a set of discrete rewrites (actions). We fully parameterize a match plan to allow the agent to generate any valid plans. + +State. At each time step, the agent receives a state with two parts. First part has selected run-time system signals from the inverted index system. Since the query is uniformly sampled in different episode, we include some statistical features and semantic embeddings of queries to allow the agent to identify different queries and then generate corresponding match plans. Such statistical features are used in the current production system, such as the length of a query. All signals and embeddings are normalized to a reasonable range around $[ - 1 , 1 ]$ based on empirical estimations. + +Action. The agent selects a parameterized action $a _ { t }$ including a rewrite id and its allocated quotas at each step. There are 29 types of rewrites in $\mathbf { \mathcal { A } } _ { d }$ and also a 5-dimensional quota parameter space $\mathcal { X }$ for each rewrite in int64 type. However, the range of each quota is empirically set to valid values and then normalized to $[ - 1 , 1 ]$ . All outputted quotas are effective for the current step. + +There is also a special action to note: stop (by agent). Normally, the environment will return a terminal signal, but it happens in extreme cases, such as low system resources. To allow a better balance of latency and performance, we allow the agent to choose stop or not as another type of discrete action, which is the same level as other 29 rewrites. + +Reward. We use two criterions in the design: latency and performance. For latency consideration, we use a signal Seek Count that is constant for a same query in different runs with same match plan, instead of executed time. We weight the Ranking Scores of top five returned documents as an indicator of performance with weights [0.4, 0.2, 0.2, 0.1, 0.1]. When there are less than five documents returned, the scores of missing documents are treated as a minimum possible ranking score. The final scalar reward is weighted by these two objectives to balance them. We do not use the division in Rosset et al. (2018), since the loss surfaces may become more non-convex for optimizing. + +We also have punishments (negative rewards) for some special cases. One is to assign a punishment when the agent selects to illegally stop at the first step or some unsupported rewrites (for some special queries). However, unsupported rewrites do not necessarily cause a terminal state, since the production system has same hand-crafted match plans for a class of queries, in which the unsupported rewrites are simply omitted. + +# 3 METHOD + +For generating match plans, we introduce a method that works on the aforementioned formulation with slightly different parameterized action space than standard PAMDPs in the literature. We start from the intuition with deterministic policy learning (Silver et al. (2014)) and how it relates to parameterized actions. PAMDP often refers to the formulation that each individual discrete action $k \in \mathcal { A } _ { d } = [ K ]$ has a corresponding continuous action-parameter space. Thus, there are $K$ continuous action-parameter spaces corresponding to $K$ discrete action (Xiong et al. (2018)). In our shared-parameter PAMDP, the Bellman equation incorporated both discrete action $k$ and continuous action-parameters $x$ is given by: + +$$ +\begin{array} { r } { Q \left( s , k , x \right) = \underset { r , s ^ { \prime } } { \mathbb { E } } \left[ r + \gamma \underset { k ^ { \prime } } { \operatorname* { m a x } } \ \underset { x ^ { \prime } \in \mathcal { X } } { \operatorname* { s u p } } Q \left( s ^ { \prime } , k ^ { \prime } , x ^ { \prime } \right) \vert s , k , x \right] . } \end{array} +$$ + +P-DQN (Xiong et al. (2018)) tackles PAMDP by incorporating multiple action-parameter policies $x _ { k } ( s ; \theta _ { x } ) : { \cal { S } } { \mathcal { X } } _ { k }$ for each action $k$ and update them with $\begin{array} { r } { \operatorname* { m a x } _ { k ^ { \prime } } \operatorname* { s u p } _ { x _ { k ^ { \prime } } \in \mathcal { X } _ { k ^ { \prime } } } Q \left( s ^ { \prime } , k ^ { \prime } , x _ { k ^ { \prime } } \right) } \end{array}$ simultaneously. However, since all parameters $x$ are shared for actions $k$ in our case, we do not require multiple policies. The policy loss in P-DQN (Xiong et al. (2018)) given by + +$$ +L _ { x } \left( \theta _ { x } \right) = \underset { s \sim D } { \mathbb { E } } \left[ - \sum _ { k = 1 } ^ { K } Q \left( s , k , x _ { k } \left( s ; \theta _ { x } \right) ; \theta _ { Q } \right) \right] +$$ + +![](images/908f0f5be906ce40a46112737a183c610852cc851433f4f94d61674bb007f08e.jpg) +Figure 2: The diagram of the process. + +naturally degenerates to a single policy in next case, where the gradient $\nabla L _ { x } \left( \theta _ { x } \right)$ exactly is the deterministic policy gradient (DPG, Silver et al. (2014)). MP-DQN (Bester et al. (2019)) states a problem regarding to the erroneous gradients of Q-network $\nabla L _ { Q }$ in P-DQN caused by multiple action-parameter policies backpropagating gradients at the same time. However, it does not exist in such single network situation. This is a desired architecture since separating multiple actionparameter policies loses the knowledge that they share same potential meanings. + +Following this intuition, we directly parameterize the policy with two heads $\mu : \mathcal { S } \to \mathbb { R } ^ { K } \times \mathcal { X }$ with $\theta _ { \mu }$ $\mathrm { ~ ` ~ } _ { \mu } \mathrm { ~ a s : ~ } k _ { t } ^ { s o f t } , x _ { t } = a _ { t } ^ { s o f t } = \mu ( s _ { t } ; \theta _ { \mu } )$ , where $k _ { t } ^ { s o f t }$ and $a _ { t } ^ { s o f t }$ refers to action-parameters or actions with softmax values. One head outputs the sof tmax values for all discrete actions. For another head, it outputs normalized action-parameter values with hyperbolic tangent $( t a n h )$ activation after the continuous head. The environment will denormalize it to a valid range. This corresponds to the Squashing Gradients method for bounded-continuous action spaces in Hausknecht & Stone (2015). + +Exploration with Parameter Space Noise and action sampling. In DQN (Mnih et al. (2015)), the behavioral policy is given by $\epsilon$ -greedy strategy for exploration purpose, where an agent randomly selects a discrete action with probability $\epsilon < 1$ . In DDPG (Lillicrap et al. (2015)), although the policy learned is deterministic, to explore the action space, the behavioral policy needs to be different because of the off-policy nature. It uses action space noise such as uncorrelated Gaussian noise or correlated Ornstein-Uhlenbeck process (Lillicrap et al. (2015)). + +Directly combining such two exploration strategies is straightforward. However, it may induce a problem that the exploration in two action spaces $\mathcal { A } = \mathcal { A } _ { d } \times \mathcal { X }$ in different paces. For example, the match plan may need another rewrite with larger quotas in a step. However, if the exploration strategy is to use $\epsilon$ -greedy for a rewrite $k$ and Gaussian noise for its quotas $x$ , the agent may require more samples to discover the potential rewards (positive documents). We instead use parameter space noise (Plappert et al. (2017)) on parameterized action space to tackle such issue. + +We denote the discrete and continuous action heads as $( \pi _ { k } ^ { s o f t } ( s ) , \pi _ { x } ( s ) ) = \mu ( s ; \theta _ { \mu } ) .$ . To compute the distance $d ( \pi , \widetilde { \pi } ) = D _ { \mathrm { K L } } ( \pi \| \widetilde { \pi } )$ of non-perturbed and perturbed policies $\mu ( s ; \theta _ { \mu } ) , \widetilde { \mu } ( s ; \theta _ { \mu } )$ , we e e euse weighted sum of the distance of discrete and continuous actions. For the continuous actions $\widetilde { \pi } _ { x } ( s )$ , the distance is given by $\mathbb { E } _ { s } \left[ \left( \pi _ { x } ( s ) _ { i } - \widetilde { \pi } _ { x } ( s ) _ { i } \right) ^ { 2 } \right]$ to estimate KL-divergence empirically. For e ethe discrete actions, we use outputted softmax probabilities to compute The state and action pairs are sampled from a replay memory. The varianc $D _ { \mathrm { K L } } ( \pi _ { k } ^ { s o f t } ( s ) \Vert \widetilde { \pi } _ { k } ^ { s o f t } ( s ) )$ $\sigma$ after a policy update based on the distance and threhold $\delta$ (Plappert et al. (2017)). The policy and target networks with layer normalization (Ba et al. (2016)) are perturbed per episode. + +Prioritized replay and recurrent policies. We use recurrent architecture to obtain the underlying system state of the POMDP for both value and policy networks (Heess et al. (2015)). To avoid recurrent state staleness (Kapturowski et al. (2018)), we store and replay a sequence of $( s , k ^ { s o f t } , x , r )$ with softmax values into a replay memory. Since a match plan is usually short, we set a maximum length and pad shorter episodes with zero or randomly sampled states and actions. + +The sampling from a replay memory can be prioritized with a probability $p _ { i }$ proportional to TDerrors (Schaul et al. (2015)) to increase the sample efficiency in such structured action space. To prioritize the transitions with well-matched documents, we slightly modify the sampling strategy. With a half probability, the agent uses regular prioritization, otherwise partitions the memory to multiple bins and retrieve a transition (sequence) with max reward from each bin. Our strategy provides a more efficient and balanced exploration strategy of the evaluation. + +Policy update. In the update, we use clipped double $Q$ -learning and delayed policy update (Fujimoto et al. (2018)) to stabilizing the training. After the agent samples a batch of (sequences), it perturbs the target policy network and computes the perturbed actions for target policy smoothing. The parameter space noise variance $\sigma$ is then updated with the sampled states and actions. + +The policy is still deterministic since it is learned in off-policy and only the exploration involves stochasticity. The update is given by deterministic policy gradients theorem (Silver et al. (2014)): + +$$ +\nabla L _ { \mu } \left( \theta _ { \mu } \right) = \nabla _ { s \sim D } \left[ Q \left( s , \mu ( s ; \theta _ { \mu } ) ; \theta _ { Q } \right) \right] = \nabla \frac { 1 } { \left| D \right| } \sum _ { s \in D } Q \left( s , k ^ { s o f t } , x ; \theta _ { Q } \right) , +$$ + +where $D$ is a replay memory. We found this policy architecture is similar to PA-DDPG (Hausknecht $\&$ Stone (2015)), while it does not including sof tmax activation for discrete actions and other advanced techniques. We also build upon other useful techniques such as invertible value function rescaling $h ( x ) = \mathrm { { s i g n } } ( x ) ( \sqrt { | x | } + 1 - 1 ) + \epsilon x$ (Kapturowski et al. (2018)): + +$$ +y = h \left( \sum _ { k = 0 } ^ { n - 1 } r _ { t + k } \gamma ^ { k } + \gamma ^ { n } h ^ { - 1 } \left( Q \left( s _ { t + n } , a ^ { \ast } ; \theta _ { Q } ^ { - } \right) \right) \right) , +$$ + +where $a ^ { * } = ( \operatorname* { m a x } k _ { t } ^ { s o f t } , x _ { t } ) = ( k _ { t } , x _ { t } )$ is the greedy action, $y$ the target for updating Q-network, and $\theta _ { Q } ^ { - }$ denotes the target Q-network. The policy network outputs differentiable softmax values and is updated by the gradients backpropagated from the Q-network: + +$$ +\nabla L _ { Q } \left( \theta _ { Q } \right) = \nabla \mathbb { E } \left[ \frac { 1 } { 2 } \left( y - Q \left( s , k ^ { s o f t } , x ; \theta _ { Q } \right) \right) ^ { 2 } \right] , +$$ + +where the expectation is took over samples from a memory. + +The pseudocode for the algorithm with more details can be found in the Appendix. + +# 4 EXPERIMENTS + +In this section, we apply the purposed agent on the inverted index match plan generation problem. We create a dataset which contains a set of queries and corresponding query embeddings. We perform various ablation study on how each component interacts and the benefits of them. We also test on other Paramterized Action RL benchmarking baselines. + +# 4.1 EXPERIMENT SETTINGS + +For match plan generation, we experiment on a dataset that has about 100,000 queries sampled from Bing search log. In the current production system, each query is classified to a predefined query classes online based on a set of rules which are related to statistical features of the query. Each query class has some hand-crafted rules which outputs a match plan for the classified query. We use each query’s production match plan as baseline. The match plan for each query is generated by the hand-crafted rules and does not change over each running. We only skip a few special queries that do not have embeddings or need additional operations beyond match plans. + +The generated match plan is evaluated by the delta values of both Ranking Score and Seek Count and overall reward. We provide learning curves of delta values of evaluation rewards. Note that the reward are not symmetrical around 0, since the ranking scores has a minimum value and is sparse since a few queries are hard to find good documents and will be assigned a very low score. It may significantly pull down the average rewards shown in the curves, thus we present histograms for fair comparison. For the benchmarks, we report the results on evaluation reward curves during tuning. + +# .2 MATCH PLAN GENERATION PERFORMANCE EVALUATION + +We first visualize the query embeddings in Figure 3 using UMAP (McInnes et al. (2018)) to demonstrate our learned models. We evaluate 6,000 queries and draw the colormap based on the difference between evaluation reward and production baseline reward. Although most queries perform similar to the well-designed production rules, there are some clusters of queries and some patterns exist. We notice that there is a main cluster at the center (about $( - 7 , - 1 ) )$ ) that gathers most queries on which our agent does not perform well. That cluster contains some random inputs from the users, which may have some typos. For the queries that the agent outperforms the baseline, we find they are quite scattered. We guess the reason is that the baseline with hand-crafted rules do not consider the embeddings, thus it is less affected by embeddings. This may suggest us to look for more informative embeddings of queries to better distinguish some clustered queries. + +![](images/1b912e4438fe0c4576fc52e07dd81917a7cf3148fcc9fac32a9a96a543cd4db8.jpg) +Figure 3: Visualization of query embeddings. + +![](images/5ea9b9c596510988cdcbdc7742ad1ba606263fb7aae0db82cd463ab8f11bfb27.jpg) +Figure 4: Different techniques. + +![](images/6f60df92f6ce68b3bc8a05d04a9fa957d6cb5bab0fe7d16ef5692b5aaa72f95a.jpg) +Figure 5: Delta Reward Distribution in Evaluation Phrase. + +In the plots shown in Figure 6, we visualize the distribution of delta rewards, ranking scores and seek count values comparing to the well-designed production rules. For most of the queries the agent learns promising match plan without any prior knowledge just based on the reward signals. There are also some hard queries to find a pattern that have poor reward (about $- 1 0 0 )$ and usually lies in the main cluster and around a few more small clusters. Other than some meaningless queries around the center, they also include some non-English queries, such as French and Chinese, which are possibly clustered around their centroid and are challenging to learn the policy for all of them. + +Reward design. We consider different weights to trade-off between the query latency from the input and the quality of the returned documents. The quality is evaluated by Ranking Scores of top five documents. We found that it is sparse since the good documents are hard to match. When we weight more on ranking score, the agent tends to stop the search early. + +We also investigate a few punishments and try to use less of them to avoid manual design. We found the agent learns some common patterns guided by the reward and punishments. It tries to avoid ”stop” at the first step because we set a huge punishment on such invalid stop. This punishment is necessary since it prevents the agent stops at first step for better value than trying more steps but get no documents matched. In other words, the agent is encouraged to explore different rewrites instead of using empty match plans. Without such punishment, the performance significantly drops since it is hard to explore good documents at first. We also punish the agent to avoid unsupported rewrites, since the production system uses designed rules for each class of the queries and simply omits a rewrite without any useful feedback. We test punishing repeated rewrites, however, it helps seek count (more efficient) but may harm ranking scores. + +Table 1: Improved Scores of Different Noise Compared to RNN Gaussian Noise. Ranking Score and Seek Count are scaled to match Overall $=$ RankingScore−SeekCount for easy comparison. + +
OverallRanking Score (Quality)Seek Count (Efficiency)
OU Noise+1.00+1.20-0.20
Param Noise+1.69+0.82-0.87
+ +Table 2: Improved Scores of Different Model Compared to MLP + +
OverallRanking Score (Quality)Seek Count (Efficiency)
MLP+Prioritized+14.54+8.56-5.98
RNN+Prioritized+32.63+14.44-18.19
RNN + Prioritized Sample+32.67+15.29-17.38
+ +# 4.3 ABLATION STUDY + +We compare different exploration noise, including action space noise and parameter space noise. For the action space noise, the noise on discrete and continuous actions are applied separately. We test Gaussian and OrnsteinUhlenbeck noise on continuous actions, while the discrete actions only use $\epsilon$ -greedy. To keep the comparison fair, we also test $\epsilon$ -greedy strategy on continuous actions by uniformly sampling a point with probability $\epsilon$ . The parameter noise is directly applied on policy network $\widetilde \mu ( s ; \theta _ { \mu } )$ . We found the parameter space noise performs more stable than all types of action espace noise. The uniform sampling on both discrete and continous action with $\epsilon$ -greedy failed for sometimes, thus we did not include it in comparison. We give the relative improvement in Table 1 compared to RNN with Gaussian Noise. The results show that parameter space noise has best performance overall. + +The results in Figure 6(a) and Table 2 show that RNN got significantly better performance than MLP with or without parameterization on our environment. It indicates that the environment is highly non-Markov. We guess one obvious possibility is that the system signals of the state space cannot include all the information, while recurrent networks try to extract latent state from the history. Note that each episode will sample one query, thus the query embedding is constant for all steps in one episode. Another possible reason is that some rewrites may not be supported by special queries or have too small change for the state signals. + +We compare different parameterization methods on recurrent networks. The results are shown in Figure 4(b) and Table 2. We empirically found that our modified strategy using reward bin is more stable from the beginning. It is possible that the agent repeatedly replays not only good experience on matching documents, but also learns to avoid punishment we set, as we expect. In general prioritized replay, the sampling probability is just based on TD-error and may be biased to worse samples since value function may not update towards possible results. + +# 4.4 BENCHMARKING GAMES + +We experiment on Platform-v0 and Goal-v0 from Bester et al. (2019). Note that, in all these games, each discrete action has a separate continuous action-parameter space. Our algorithm is designed for shared action-parameters and does not utilize such prior, thus we do not compare with P-DQN-style algorithms. We apply n-step return since the episode in these games is much longer. We assume the environment is fully observable and do not use recurrent networks in comparison. + +Hyperparameters. We examine the games with different combinations of hyperparameters, since they are easy to parallize on each training nodes without the need to connect to an production environment emulator. With NNI, the learning rates for value and policy network are set to log-uniform in $[ 3 \times 1 0 ^ { - 4 } , 3 \times 1 0 ^ { - 3 } ]$ and $[ 1 \times 1 0 ^ { - 4 } , 1 \times \mathsf { \bar { 1 } 0 ^ { - 3 } } ]$ . The $\alpha$ in parameterized replay is set to log-uniform in [0.2, 1.0]. The action noise threshold $\delta$ in parameter space noise is set to log-uniform between [0.05, 0.8]. Tn soft parameter update, the $\tau$ is set to log-uniform in $[ 1 \times 1 0 ^ { - 3 } , 1 \times 1 0 ^ { - 2 } ]$ . + +In Platform-v0 and Goal-v0, we found the agent is not very sensitive to the range we set, such as $\alpha , \delta$ and $\tau$ . The $\delta$ values of top $20 \%$ trials vary between [0.05, 0.3], while $\alpha$ and $\tau$ values are evenly scattered in the defined range. In these settings, usually the agent prefers slightly larger value learning rate than policy learning rate. + +
Average Eval ReturnOurPA-DDPG
Platform-v00.95730.3113
Goal-v034.20-6.208
+ +Table 3: Average evaluation results (the average of all training rewards and final evaluation reward) on benchmarks Platform-v0 and Goal-v0 with PA-DDPG (Hausknecht & Stone (2015)), MP-DQN (Bester et al. (2019)) and P-DQN (Xiong et al. (2018)). We use reported number from the papers, while last two methods report another metric on Goal-v0. + +# 5 RELATED WORK + +While the aforementioned algorithms and techniques work on discrete or continuous action spaces, it is not trivial to apply them on parameterized action space, since such discrete-continuous hybrid action space is hard to parameterized by a single distribution. A related series of work is to combine DDPG and DQN to optimize Q-value function on parameter actions. There are two classes of methods that belong to them: PA-DDPG (Hausknecht & Stone (2015)) based on DDPG and PDQN (Xiong et al. (2018)) based on DQN. They are Q-Learning-based methods which select the best action by maximizing the Q-value function on discrete action space or learning a deterministic policy outputting best continuous action. However, they use different strategies to combine discrete and continuous actions. Bester et al. (2019) (MP-DQN) extends P-DQN to tackle the problem that Q-value is a function of the joint action-parameter vector $Q ( s ^ { \prime } , k ^ { \prime } , { \bf x } ^ { \tilde { Q } } ( s ^ { \prime } ) )$ in normal PAMDP, which may results in fault gradients. However, such problem does not exist in our slightly modified setting, since the action-parameter space $\mathcal { X }$ in the match plan generation is inherently defined to be shared for each $k \in \mathcal { A } _ { d }$ . Masson et al. (2016) purposes a method to iteratively optimizing discrete and continuous actions by alternating between them. Another perspective (Klimek et al. (2017); Wei et al. (2018); Fu et al. (2019)) for a parameterized action space is to regard it as a twohierarchy action space, where an agent selects discrete action first and continuous parameter later. However, we do not consider this direction in current scheme because we share the same parameters for all discrete action in not very large scale. Therefore, the hierarchical methods may not bring a significant performance gain. + +# 6 DISCUSSIONS + +In this paper, we present a parameterized action RL match plan generation method which extends the plan generation to the general case without any predefined knowledge. Key to address the problem are normalized softmax values of discrete actions to enable gradients backpropagation, parameter space noise on parameters of the policy for unifying the exploration direction in both discrete and continuous spaces, and recurrent deterministic policies with prioritized replay buffer to accelerate and stabilize the training. Our algorithm can be applied to not only the match plan generation environment, but also other similar parameteried action environments. The experiment results demonstrate our method outperforms the well-designed hand-crafted rules in Bing and serveral baseline results in some existing PARL benchmarks. In this paper, we mainly discuss about offline training procedure. In the future, we plan to apply learned policy to the production environment. + +# REFERENCES + +Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. + +Craig J Bester, Steven D James, and George D Konidaris. Multi-pass q-networks for deep reinforcement learning with parameterised action spaces. arXiv preprint arXiv:1905.04388, 2019. + +Haotian Fu, Hongyao Tang, Jianye Hao, Zihan Lei, Yingfeng Chen, and Changjie Fan. Deep multi-agent reinforcement learning with discrete-continuous hybrid action spaces. arXiv preprint arXiv:1903.04959, 2019. + +Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in actor-critic methods. arXiv preprint arXiv:1802.09477, 2018. + +Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. + +Nicolas Heess, Jonathan J Hunt, Timothy P Lillicrap, and David Silver. Memory-based control with recurrent neural networks. arXiv preprint arXiv:1512.04455, 2015. + +Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. + +Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent experience replay in distributed reinforcement learning. 2018. + +Maciej Klimek, Henryk Michalewski, Piotr Mi, et al. Hierarchical reinforcement learning with parameters. In Conference on Robot Learning, pp. 301–313, 2017. + +Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. + +Warwick Masson, Pravesh Ranchod, and George Konidaris. Reinforcement learning with parameterized actions. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. + +Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. + +Azalia Mirhoseini, Hieu Pham, Quoc V Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean. Device placement optimization with reinforcement learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 2430–2439. JMLR. org, 2017. + +Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529, 2015. + +Matthias Plappert, Rein Houthooft, Prafulla Dhariwal, Szymon Sidor, Richard Y Chen, Xi Chen, Tamim Asfour, Pieter Abbeel, and Marcin Andrychowicz. Parameter space noise for exploration. arXiv preprint arXiv:1706.01905, 2017. + +Corby Rosset, Damien Jose, Gargi Ghosh, Bhaskar Mitra, and Saurabh Tiwary. Optimizing query evaluations using reinforcement learning for web search. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, pp. 1193–1196. ACM, 2018. + +Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. + +David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. 2014. + +Ermo Wei, Drew Wicke, and Sean Luke. Hierarchical approaches for reinforcement learning in parameterized action space. In 2018 AAAI Spring Symposium Series, 2018. + +Ian H Witten, Ian H Witten, Alistair Moffat, Timothy C Bell, Timothy C Bell, and Timothy C Bell. Managing gigabytes: compressing and indexing documents and images. Morgan Kaufmann, 1999. + +Jiechao Xiong, Qing Wang, Zhuoran Yang, Peng Sun, Lei Han, Yang Zheng, Haobo Fu, Tong Zhang, Ji Liu, and Han Liu. Parametrized deep q-networks learning: Reinforcement learning with discrete-continuous hybrid action space. arXiv preprint arXiv:1810.06394, 2018. + +Justin Zobel and Alistair Moffat. Inverted files for text search engines. ACM computing surveys (CSUR), 38(2):6, 2006. + +# A APPENDIX + +# A.1 PSEUDOCODE + +Algorithm pseudocode for the algorithm is provided in Algorithm 1 which includes all aforementioned details. + +# Algorithm 1 + +
Input: Empty replay buffer D, init parameter noise std δparam, action noise threshold δ Initialize policy parameters 0,value parameters Φ ←0Q for each episode do Perturb policy parameters θ ← θ +(O,Oparam) and target policy parameters θtarget Observe state St, output a disturbed action embedding A =p(st；0） Compute executing action at by taking max over k
Execute parameterized action at in the environment server Observe next state St+1,reward rt, done signal dt denoting if St+1 is terminal t,rt, St+1,dt) to the buffer D If St+1 is terminal, reset to an initial state so for each update if update-condition do
Sample a minibatch from prioritized replay buffer D with specific priorities Compute and transform target actions with disturbed target policy network
Update parameter noise std Oparam using empirical distance d(μ, μ)
Compute targets for TD-error with min Q-value in the twin Q-networks Update Q-network parameters using gradient descent
if policy update frequency then
Update policy network parameters θ with Update target parameters with polyak averaging end if
+ +# A.2 FURTHER EXPERIMENTAL DETAILS + +Hyperparameter search. We use Microsoft $\mathrm { N N I } ^ { 2 }$ and OpenPAI3 to search hyperparameters. The final metric to report to NNI is set to the sum of average training reward and final evaluation reward (repeated 1,000 times) final_metric $=$ (avg_reward+eval_reward)/2. The intermediate metric is set to evaluation reward per 1,000 episodes. We also use the early stop assessor. Each GPU server node connects to a production environment emulator with ethernet. + +Model architecture. Both policy and value networks use two fully connected layers with 512 hidden units and a LSTM layer (Hochreiter & Schmidhuber (1997)). Each layer also uses a layer normalization (LayerNorm, Ba et al. (2016)) as suggested by Plappert et al. (2017) in consideration of stability for noise applied on parameters, and follows a ReLU activation. Both output heads of the policy network has a hidden layer with sof tmax or tanh activations. + +# A.3 FURTHER ENVIRONMENTAL DETAILS + +Accumulated values. Note that, for accumulated values in received states, rewards and outputted action-parameters, we use the difference (delta values) from the last step. For states and rewards, it is $s _ { t } = s _ { t } ^ { \prime } - s _ { t - 1 }$ , $r _ { t } = r _ { t } ^ { \prime } - r _ { t - 1 }$ , where where $s ^ { \prime }$ and $r ^ { \prime }$ denote raw state and reward. For actions, the agent outputs raw action-parameter output $a ^ { \prime }$ , and the emulator converts it to accumulated value $a _ { t } = a _ { t } ^ { \prime } + a _ { t - 1 }$ . + +![](images/6d04d136c67be4f2400a72fd7e2f02c247f2d8e0e0954f2f1782df4d09b152ff.jpg) +Figure 6: Best evaluation learning curves on both environments during tuning. The agent achieves maximum possible reward on both environments (50 and 1). + +Inverted index in Bing. In the inverted index system in Bing, there are two key steps: (i) the system rewrites the query to enlarge candidate set with specified match plan (or search plan), (ii) it returns candidates with top ranking scores. We just refer to the math plan part regarding to generation, but not user-input content. In the match plan generation, the goal is to generate the optimal search plan (policy) for each query. + +Closed-loop and open-loop control. The production match plan is generated online in open-loop (feedforward) without taking runtime system signals into consideration for the latency and implementation consideration. However, RL is closed-loop which takes feedback from system signals to make decisions. This may increase the online overhead, but it can be tackled by converting the reflective policy to a shooting-style action sequence by predicting with learned transition dynamics. In the implementation, the environment emulator receives an action sequence $a _ { 0 } , . . . , a _ { t }$ and return $s _ { t + 1 }$ to simulate the open-loop style. + +# A.4 MORE TRAINING FOR BENCHMARKS + +We provide the best evaluation learning curves on both environments during tuning. \ No newline at end of file diff --git a/parse/train/rkgZaT4tDr/rkgZaT4tDr_content_list.json b/parse/train/rkgZaT4tDr/rkgZaT4tDr_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c39a7082efa28f08270dd26097424daf5bc932f --- /dev/null +++ b/parse/train/rkgZaT4tDr/rkgZaT4tDr_content_list.json @@ -0,0 +1,1397 @@ +[ + { + "type": "text", + "text": "PARAMETERIZED ACTION REINFORCEMENT LEARNING FOR INVERTED INDEX MATCH PLAN GENERATION ", + "text_level": 1, + "bbox": [ + 176, + 98, + 821, + 146 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anonymous authors Paper under double-blind review ", + "bbox": [ + 183, + 171, + 398, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ABSTRACT ", + "text_level": 1, + "bbox": [ + 454, + 234, + 544, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Match plan generation in the inverted index at Microsoft Bing is used to be based on hand-crafted rules. We formulate the generation process as a Parameterized Action MDP with sharing parameters and purpose a reinforcement learning algorithm on such formulation. We combine deterministic policy learning on discrete and continuous action spaces and several recent advances in deep reinforcement learning. For exploring in the parameterized action space, the agent outputs softmax values for discrete actions and applies Parameter Space Noise on policy network to unify the exploration direction in both spaces. We apply prioritized recurrent replay on match plan sequences and pad short match plans. We also use invertible value function rescaling and $n$ -step return to stabilize the training. The agent is evaluated on our environment and some benchmarks. It outperforms the well-designed production match plan and beats the baselines on the benchmarks. ", + "bbox": [ + 233, + 267, + 764, + 434 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 INTRODUCTION ", + "text_level": 1, + "bbox": [ + 176, + 462, + 334, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Using machine learning to optimize and accelerate software and hardware systems is an emerging field in the past few years (Mirhoseini et al. (2017); Rosset et al. (2018)). The process of decisionmaking in those systems is usually hand-crafted by human engineers, which is less explored to be automatically done by learning algorithms. A promising direction is to formulate such sequential decision-making problems as Reinforcement Learning (RL) problems, such as search plan generation problem in inverted indexes (Rosset et al. (2018)). ", + "bbox": [ + 174, + 494, + 825, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The inverted index is a specialized data structure that is commonly used in search engines, including Microsoft $\\mathrm { B i n g ^ { 1 } }$ . An inverted index provides an inverted mapping from term to documents. Each term has a posting list which contains all the (document, location) pairs that the term appears. By combining the posting lists of the terms in a user query, the initial document candidates are generated (Witten et al. (1999); Zobel & Moffat (2006)). ", + "bbox": [ + 174, + 584, + 823, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In Bing, documents are scanned with specified match plans, either predefined or generated in realtime. A match plan has a sequence of rewrites (match rules, e.g., it treats the query as a phrase which should appear in the document exactly as it is), where each rewrite can be controlled by several quotas (stopping criteria). There are several types of rewrites, and all these types share the same continuous quotas. A pair of a rewrite type and the quotas forms a single action which determines if a document will be a ranking candidate. ", + "bbox": [ + 174, + 661, + 823, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We formulate the generation problem as a RL problem. The state consists of system runtime signals and semantic embeddings of queries. The action space is called parameterized or discretecontinuous hybrid, where an action has a discrete action and continuous action-parameters. The reward is weighted by result quality and query latency. It is similar to Parameterized Action RL (Masson et al. (2016)) (PARL), while our setting requires all actions to share same parameters. ", + "bbox": [ + 174, + 752, + 823, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previously, the match plans are predefined manually for each query. It hardly utilizes the rich information in the state to dynamically adjust the rewrites and quotas for specific query online. Rosset et al. (2018) tries to automatically generate match plans using tabular Q-learning with discretized state space and predefined action-parameters. They learn the generation policy for a specific query class each time and solve it only in discretized spaces with tabular methods. We extend the generation process to the general case, such that the match plans are fully parameterized and learned from scratch without any predefined knowledge (e.g., limited match rules). ", + "bbox": [ + 174, + 828, + 823, + 898 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7459112fc13a2ca51b6133f349d4373e932db4f4ab389ec3ae4eb649b28ac3c6.jpg", + "image_caption": [ + "Figure 1: Match plan example: for the query “Reinforcement Learning”, the search engine firstly gets a document posting list for each term from the inverted index, and then scans the document candidates following a serial of rewrites (match rules) according to the match plan. " + ], + "image_footnote": [], + "bbox": [ + 310, + 108, + 687, + 184 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 171, + 263, + 821, + 292 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we purpose a parameterized action RL algorithm that learns to act on parameterized action space environments, such as the match plan generation process. We introduce normalized softmax values of discrete actions to form a categorical distributions to enable gradients backpropagation. We also combine several recent advances in RL to accelerate and stabilize the training. We investigate parameter space noise (Plappert et al. (2017)) on parameters of the policy for unifying directions in exploring the structured action space. We explore recurrent deterministic policies (Heess et al. (2015)) with prioritized replay buffer (Schaul et al. (2015)) on sequence, due to the inherent nature of partial observability and sparse feedback in our setting. We also use $n$ -step return and invertible value function rescaling (Kapturowski et al. (2018)) for further stability. ", + "bbox": [ + 173, + 299, + 825, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present an agent to integrate these techniques and evaluate on offline training on a query dataset collected from Bing search. When training on the dataset with uniformly sampling, the agent outperforms the currently deployed production match plan, which is generated based on well-designed hand-crafted rules. We study the performance of training on the dataset or only with several selected complicated queries individually, as well as ablation studies for the various components. We test on a few existing PARL benchmarks, where our agent beats our baseline methods and performs the state-of-the-art results. ", + "bbox": [ + 173, + 431, + 825, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2 BACKGROUND ", + "text_level": 1, + "bbox": [ + 176, + 547, + 326, + 565 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1 REINFORCEMENT LEARNING ", + "text_level": 1, + "bbox": [ + 176, + 579, + 415, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We address the problem using reinforcement learning (RL) framework with a parameterized action space. An agent interacts with an environment to maximize the accumulated reward with discount factor $\\gamma \\in \\ [ 0 , 1 )$ . We model the environment with a discrete-time Partially Observable Markov Decision Process (POMDP). A POMDP is given by a tuple $( S , \\mathcal { A } , T , R , \\Omega , \\dot { \\mathcal { O } } )$ , and the underlying Markov Decision Process (MDP) is defined by $( S , { \\mathcal { A } } , T , R )$ , where $s$ is the state space, $\\mathcal { A }$ the parameterized action space, $T$ the transition function $\\mathcal { T } : \\mathcal { S } \\times \\mathcal { S } \\times \\mathcal { A } \\mathbb { R } _ { + }$ , and $R : S \\times \\mathcal { A } \\mathbb { R }$ is the reward function. The set of observations is given by $\\Omega$ and the the observation function mapping underlying states to probability distributions over observations is given by $\\mathcal { O }$ . ", + "bbox": [ + 173, + 604, + 825, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, the environment in the match plan generation has a parameterized action space. In Masson et al. (2016); Bester et al. (2019), such formulation in fully observable settings is referred as Parameterized Action MDP (PAMDP), where the action space is denoted as ", + "bbox": [ + 174, + 723, + 825, + 765 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/76ea0d7a9eb735ac9b64468a264943b9eec3b340c391396f2eb432de507a6511.jpg", + "text": "$$\n\\mathcal { A } = \\bigcup _ { k \\in \\mathcal { A } _ { d } } \\left\\{ ( k , x _ { k } ) | x _ { k } \\in \\mathcal { X } _ { k } \\right\\} ,\n$$", + "text_format": "latex", + "bbox": [ + 395, + 768, + 599, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where each discrete action $a \\in \\mathcal { A } _ { d } = [ K ]$ has a corresponding continuous action-parameter space $\\mathcal { X } _ { a }$ . However, our setting requires a slightly different formulation: ", + "bbox": [ + 173, + 804, + 823, + 833 + ], + "page_idx": 1 + }, + { + "type": "equation", + "img_path": "images/243001bdfb8447abcdd89869de2d2453c03e6261b0ef47f965d0d6494f259a02.jpg", + "text": "$$\n\\mathcal { A } = \\{ ( k , x ) | k \\in \\mathcal { A } _ { d } , x \\in \\mathcal { X } \\} = \\mathcal { A } _ { d } \\times \\mathcal { X } ,\n$$", + "text_format": "latex", + "bbox": [ + 359, + 835, + 635, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the discrete action space (rewrite) $\\mathbf { \\mathcal { A } } _ { d }$ share the action-parameter space $\\mathcal { X }$ . Such formulation results in a disentangled action space between discrete and continuous actions which a class of parameterized action (P-DQN (Xiong et al. (2018)) and MP-DQN (Bester et al. (2019))) may not be trivially applicable to. Considering the notational simplicity in partially observable environment (POMDP), we do not explicitly denote it when there is no ambiguity in action settings. ", + "bbox": [ + 174, + 854, + 825, + 924 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2 FORMULATION OF MATCH PLAN GENERATION ", + "text_level": 1, + "bbox": [ + 173, + 103, + 539, + 118 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Rosset et al. (2018), the action-parameters are predefined for the match plan generation, where the action space is then a set of discrete rewrites (actions). We fully parameterize a match plan to allow the agent to generate any valid plans. ", + "bbox": [ + 176, + 128, + 823, + 171 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State. At each time step, the agent receives a state with two parts. First part has selected run-time system signals from the inverted index system. Since the query is uniformly sampled in different episode, we include some statistical features and semantic embeddings of queries to allow the agent to identify different queries and then generate corresponding match plans. Such statistical features are used in the current production system, such as the length of a query. All signals and embeddings are normalized to a reasonable range around $[ - 1 , 1 ]$ based on empirical estimations. ", + "bbox": [ + 173, + 188, + 825, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Action. The agent selects a parameterized action $a _ { t }$ including a rewrite id and its allocated quotas at each step. There are 29 types of rewrites in $\\mathbf { \\mathcal { A } } _ { d }$ and also a 5-dimensional quota parameter space $\\mathcal { X }$ for each rewrite in int64 type. However, the range of each quota is empirically set to valid values and then normalized to $[ - 1 , 1 ]$ . All outputted quotas are effective for the current step. ", + "bbox": [ + 174, + 286, + 825, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There is also a special action to note: stop (by agent). Normally, the environment will return a terminal signal, but it happens in extreme cases, such as low system resources. To allow a better balance of latency and performance, we allow the agent to choose stop or not as another type of discrete action, which is the same level as other 29 rewrites. ", + "bbox": [ + 174, + 349, + 825, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reward. We use two criterions in the design: latency and performance. For latency consideration, we use a signal Seek Count that is constant for a same query in different runs with same match plan, instead of executed time. We weight the Ranking Scores of top five returned documents as an indicator of performance with weights [0.4, 0.2, 0.2, 0.1, 0.1]. When there are less than five documents returned, the scores of missing documents are treated as a minimum possible ranking score. The final scalar reward is weighted by these two objectives to balance them. We do not use the division in Rosset et al. (2018), since the loss surfaces may become more non-convex for optimizing. ", + "bbox": [ + 174, + 420, + 825, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also have punishments (negative rewards) for some special cases. One is to assign a punishment when the agent selects to illegally stop at the first step or some unsupported rewrites (for some special queries). However, unsupported rewrites do not necessarily cause a terminal state, since the production system has same hand-crafted match plans for a class of queries, in which the unsupported rewrites are simply omitted. ", + "bbox": [ + 174, + 525, + 825, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3 METHOD ", + "text_level": 1, + "bbox": [ + 174, + 614, + 282, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For generating match plans, we introduce a method that works on the aforementioned formulation with slightly different parameterized action space than standard PAMDPs in the literature. We start from the intuition with deterministic policy learning (Silver et al. (2014)) and how it relates to parameterized actions. PAMDP often refers to the formulation that each individual discrete action $k \\in \\mathcal { A } _ { d } = [ K ]$ has a corresponding continuous action-parameter space. Thus, there are $K$ continuous action-parameter spaces corresponding to $K$ discrete action (Xiong et al. (2018)). In our shared-parameter PAMDP, the Bellman equation incorporated both discrete action $k$ and continuous action-parameters $x$ is given by: ", + "bbox": [ + 173, + 646, + 825, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/4baf652d557a9988b78e9d6b9f883817be7d2229b913710040c6d0deac8aada0.jpg", + "text": "$$\n\\begin{array} { r } { Q \\left( s , k , x \\right) = \\underset { r , s ^ { \\prime } } { \\mathbb { E } } \\left[ r + \\gamma \\underset { k ^ { \\prime } } { \\operatorname* { m a x } } \\ \\underset { x ^ { \\prime } \\in \\mathcal { X } } { \\operatorname* { s u p } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x ^ { \\prime } \\right) \\vert s , k , x \\right] . } \\end{array}\n$$", + "text_format": "latex", + "bbox": [ + 308, + 775, + 687, + 810 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "P-DQN (Xiong et al. (2018)) tackles PAMDP by incorporating multiple action-parameter policies $x _ { k } ( s ; \\theta _ { x } ) : { \\cal { S } } { \\mathcal { X } } _ { k }$ for each action $k$ and update them with $\\begin{array} { r } { \\operatorname* { m a x } _ { k ^ { \\prime } } \\operatorname* { s u p } _ { x _ { k ^ { \\prime } } \\in \\mathcal { X } _ { k ^ { \\prime } } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x _ { k ^ { \\prime } } \\right) } \\end{array}$ simultaneously. However, since all parameters $x$ are shared for actions $k$ in our case, we do not require multiple policies. The policy loss in P-DQN (Xiong et al. (2018)) given by ", + "bbox": [ + 173, + 819, + 825, + 878 + ], + "page_idx": 2 + }, + { + "type": "equation", + "img_path": "images/38a0ce4a63eabe51e761e09dd8728767079ede8cf87f9111a5e7ab29f0c3a281.jpg", + "text": "$$\nL _ { x } \\left( \\theta _ { x } \\right) = \\underset { s \\sim D } { \\mathbb { E } } \\left[ - \\sum _ { k = 1 } ^ { K } Q \\left( s , k , x _ { k } \\left( s ; \\theta _ { x } \\right) ; \\theta _ { Q } \\right) \\right]\n$$", + "text_format": "latex", + "bbox": [ + 339, + 885, + 656, + 928 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/908f0f5be906ce40a46112737a183c610852cc851433f4f94d61674bb007f08e.jpg", + "image_caption": [ + "Figure 2: The diagram of the process. " + ], + "image_footnote": [], + "bbox": [ + 284, + 98, + 704, + 164 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "naturally degenerates to a single policy in next case, where the gradient $\\nabla L _ { x } \\left( \\theta _ { x } \\right)$ exactly is the deterministic policy gradient (DPG, Silver et al. (2014)). MP-DQN (Bester et al. (2019)) states a problem regarding to the erroneous gradients of Q-network $\\nabla L _ { Q }$ in P-DQN caused by multiple action-parameter policies backpropagating gradients at the same time. However, it does not exist in such single network situation. This is a desired architecture since separating multiple actionparameter policies loses the knowledge that they share same potential meanings. ", + "bbox": [ + 174, + 228, + 825, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following this intuition, we directly parameterize the policy with two heads $\\mu : \\mathcal { S } \\to \\mathbb { R } ^ { K } \\times \\mathcal { X }$ with $\\theta _ { \\mu }$ $\\mathrm { ~ ` ~ } _ { \\mu } \\mathrm { ~ a s : ~ } k _ { t } ^ { s o f t } , x _ { t } = a _ { t } ^ { s o f t } = \\mu ( s _ { t } ; \\theta _ { \\mu } )$ , where $k _ { t } ^ { s o f t }$ and $a _ { t } ^ { s o f t }$ refers to action-parameters or actions with softmax values. One head outputs the sof tmax values for all discrete actions. For another head, it outputs normalized action-parameter values with hyperbolic tangent $( t a n h )$ activation after the continuous head. The environment will denormalize it to a valid range. This corresponds to the Squashing Gradients method for bounded-continuous action spaces in Hausknecht & Stone (2015). ", + "bbox": [ + 173, + 319, + 825, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exploration with Parameter Space Noise and action sampling. In DQN (Mnih et al. (2015)), the behavioral policy is given by $\\epsilon$ -greedy strategy for exploration purpose, where an agent randomly selects a discrete action with probability $\\epsilon < 1$ . In DDPG (Lillicrap et al. (2015)), although the policy learned is deterministic, to explore the action space, the behavioral policy needs to be different because of the off-policy nature. It uses action space noise such as uncorrelated Gaussian noise or correlated Ornstein-Uhlenbeck process (Lillicrap et al. (2015)). ", + "bbox": [ + 173, + 430, + 825, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Directly combining such two exploration strategies is straightforward. However, it may induce a problem that the exploration in two action spaces $\\mathcal { A } = \\mathcal { A } _ { d } \\times \\mathcal { X }$ in different paces. For example, the match plan may need another rewrite with larger quotas in a step. However, if the exploration strategy is to use $\\epsilon$ -greedy for a rewrite $k$ and Gaussian noise for its quotas $x$ , the agent may require more samples to discover the potential rewards (positive documents). We instead use parameter space noise (Plappert et al. (2017)) on parameterized action space to tackle such issue. ", + "bbox": [ + 173, + 521, + 825, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the discrete and continuous action heads as $( \\pi _ { k } ^ { s o f t } ( s ) , \\pi _ { x } ( s ) ) = \\mu ( s ; \\theta _ { \\mu } ) .$ . To compute the distance $d ( \\pi , \\widetilde { \\pi } ) = D _ { \\mathrm { K L } } ( \\pi \\| \\widetilde { \\pi } )$ of non-perturbed and perturbed policies $\\mu ( s ; \\theta _ { \\mu } ) , \\widetilde { \\mu } ( s ; \\theta _ { \\mu } )$ , we e e euse weighted sum of the distance of discrete and continuous actions. For the continuous actions $\\widetilde { \\pi } _ { x } ( s )$ , the distance is given by $\\mathbb { E } _ { s } \\left[ \\left( \\pi _ { x } ( s ) _ { i } - \\widetilde { \\pi } _ { x } ( s ) _ { i } \\right) ^ { 2 } \\right]$ to estimate KL-divergence empirically. For e ethe discrete actions, we use outputted softmax probabilities to compute The state and action pairs are sampled from a replay memory. The varianc $D _ { \\mathrm { K L } } ( \\pi _ { k } ^ { s o f t } ( s ) \\Vert \\widetilde { \\pi } _ { k } ^ { s o f t } ( s ) )$ $\\sigma$ after a policy update based on the distance and threhold $\\delta$ (Plappert et al. (2017)). The policy and target networks with layer normalization (Ba et al. (2016)) are perturbed per episode. ", + "bbox": [ + 173, + 612, + 825, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prioritized replay and recurrent policies. We use recurrent architecture to obtain the underlying system state of the POMDP for both value and policy networks (Heess et al. (2015)). To avoid recurrent state staleness (Kapturowski et al. (2018)), we store and replay a sequence of $( s , k ^ { s o f t } , x , r )$ with softmax values into a replay memory. Since a match plan is usually short, we set a maximum length and pad shorter episodes with zero or randomly sampled states and actions. ", + "bbox": [ + 174, + 762, + 825, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sampling from a replay memory can be prioritized with a probability $p _ { i }$ proportional to TDerrors (Schaul et al. (2015)) to increase the sample efficiency in such structured action space. To prioritize the transitions with well-matched documents, we slightly modify the sampling strategy. With a half probability, the agent uses regular prioritization, otherwise partitions the memory to multiple bins and retrieve a transition (sequence) with max reward from each bin. Our strategy provides a more efficient and balanced exploration strategy of the evaluation. ", + "bbox": [ + 173, + 840, + 825, + 924 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Policy update. In the update, we use clipped double $Q$ -learning and delayed policy update (Fujimoto et al. (2018)) to stabilizing the training. After the agent samples a batch of (sequences), it perturbs the target policy network and computes the perturbed actions for target policy smoothing. The parameter space noise variance $\\sigma$ is then updated with the sampled states and actions. ", + "bbox": [ + 173, + 103, + 823, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The policy is still deterministic since it is learned in off-policy and only the exploration involves stochasticity. The update is given by deterministic policy gradients theorem (Silver et al. (2014)): ", + "bbox": [ + 171, + 166, + 823, + 195 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/3562b5ab1968d6fce4dd58258e2c12c671f3fadb5d89d26146c9a6856a9e24a7.jpg", + "text": "$$\n\\nabla L _ { \\mu } \\left( \\theta _ { \\mu } \\right) = \\nabla _ { s \\sim D } \\left[ Q \\left( s , \\mu ( s ; \\theta _ { \\mu } ) ; \\theta _ { Q } \\right) \\right] = \\nabla \\frac { 1 } { \\left| D \\right| } \\sum _ { s \\in D } Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 253, + 200, + 743, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ is a replay memory. We found this policy architecture is similar to PA-DDPG (Hausknecht $\\&$ Stone (2015)), while it does not including sof tmax activation for discrete actions and other advanced techniques. We also build upon other useful techniques such as invertible value function rescaling $h ( x ) = \\mathrm { { s i g n } } ( x ) ( \\sqrt { | x | } + 1 - 1 ) + \\epsilon x$ (Kapturowski et al. (2018)): ", + "bbox": [ + 174, + 243, + 825, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/fce58c5a07a0c3febf35ee414e9539f1e4ea20fd2393bb65923f45a0fdcb7751.jpg", + "text": "$$\ny = h \\left( \\sum _ { k = 0 } ^ { n - 1 } r _ { t + k } \\gamma ^ { k } + \\gamma ^ { n } h ^ { - 1 } \\left( Q \\left( s _ { t + n } , a ^ { \\ast } ; \\theta _ { Q } ^ { - } \\right) \\right) \\right) ,\n$$", + "text_format": "latex", + "bbox": [ + 318, + 308, + 678, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $a ^ { * } = ( \\operatorname* { m a x } k _ { t } ^ { s o f t } , x _ { t } ) = ( k _ { t } , x _ { t } )$ is the greedy action, $y$ the target for updating Q-network, and $\\theta _ { Q } ^ { - }$ denotes the target Q-network. The policy network outputs differentiable softmax values and is updated by the gradients backpropagated from the Q-network: ", + "bbox": [ + 174, + 358, + 825, + 405 + ], + "page_idx": 4 + }, + { + "type": "equation", + "img_path": "images/d951c0efa2af92e644a3128d1ed4e8b66e025fb64e2e4c6797ccd5e0e6c726fd.jpg", + "text": "$$\n\\nabla L _ { Q } \\left( \\theta _ { Q } \\right) = \\nabla \\mathbb { E } \\left[ \\frac { 1 } { 2 } \\left( y - Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) \\right) ^ { 2 } \\right] ,\n$$", + "text_format": "latex", + "bbox": [ + 331, + 411, + 666, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the expectation is took over samples from a memory. ", + "bbox": [ + 176, + 450, + 565, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pseudocode for the algorithm with more details can be found in the Appendix. ", + "bbox": [ + 174, + 472, + 715, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4 EXPERIMENTS ", + "text_level": 1, + "bbox": [ + 176, + 506, + 326, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we apply the purposed agent on the inverted index match plan generation problem. We create a dataset which contains a set of queries and corresponding query embeddings. We perform various ablation study on how each component interacts and the benefits of them. We also test on other Paramterized Action RL benchmarking baselines. ", + "bbox": [ + 174, + 537, + 825, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1 EXPERIMENT SETTINGS ", + "text_level": 1, + "bbox": [ + 174, + 611, + 382, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For match plan generation, we experiment on a dataset that has about 100,000 queries sampled from Bing search log. In the current production system, each query is classified to a predefined query classes online based on a set of rules which are related to statistical features of the query. Each query class has some hand-crafted rules which outputs a match plan for the classified query. We use each query’s production match plan as baseline. The match plan for each query is generated by the hand-crafted rules and does not change over each running. We only skip a few special queries that do not have embeddings or need additional operations beyond match plans. ", + "bbox": [ + 174, + 637, + 825, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generated match plan is evaluated by the delta values of both Ranking Score and Seek Count and overall reward. We provide learning curves of delta values of evaluation rewards. Note that the reward are not symmetrical around 0, since the ranking scores has a minimum value and is sparse since a few queries are hard to find good documents and will be assigned a very low score. It may significantly pull down the average rewards shown in the curves, thus we present histograms for fair comparison. For the benchmarks, we report the results on evaluation reward curves during tuning. ", + "bbox": [ + 174, + 741, + 825, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": ".2 MATCH PLAN GENERATION PERFORMANCE EVALUATION ", + "text_level": 1, + "bbox": [ + 173, + 840, + 614, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first visualize the query embeddings in Figure 3 using UMAP (McInnes et al. (2018)) to demonstrate our learned models. We evaluate 6,000 queries and draw the colormap based on the difference between evaluation reward and production baseline reward. Although most queries perform similar to the well-designed production rules, there are some clusters of queries and some patterns exist. We notice that there is a main cluster at the center (about $( - 7 , - 1 ) )$ ) that gathers most queries on which our agent does not perform well. That cluster contains some random inputs from the users, which may have some typos. For the queries that the agent outperforms the baseline, we find they are quite scattered. We guess the reason is that the baseline with hand-crafted rules do not consider the embeddings, thus it is less affected by embeddings. This may suggest us to look for more informative embeddings of queries to better distinguish some clustered queries. ", + "bbox": [ + 176, + 867, + 823, + 924 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1b912e4438fe0c4576fc52e07dd81917a7cf3148fcc9fac32a9a96a543cd4db8.jpg", + "image_caption": [ + "Figure 3: Visualization of query embeddings. " + ], + "image_footnote": [], + "bbox": [ + 179, + 103, + 547, + 316 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ea9b9c596510988cdcbdc7742ad1ba606263fb7aae0db82cd463ab8f11bfb27.jpg", + "image_caption": [ + "Figure 4: Different techniques. " + ], + "image_footnote": [], + "bbox": [ + 602, + 109, + 774, + 327 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f60df92f6ce68b3bc8a05d04a9fa957d6cb5bab0fe7d16ef5692b5aaa72f95a.jpg", + "image_caption": [ + "Figure 5: Delta Reward Distribution in Evaluation Phrase. " + ], + "image_footnote": [], + "bbox": [ + 183, + 382, + 808, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 570, + 825, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the plots shown in Figure 6, we visualize the distribution of delta rewards, ranking scores and seek count values comparing to the well-designed production rules. For most of the queries the agent learns promising match plan without any prior knowledge just based on the reward signals. There are also some hard queries to find a pattern that have poor reward (about $- 1 0 0 )$ and usually lies in the main cluster and around a few more small clusters. Other than some meaningless queries around the center, they also include some non-English queries, such as French and Chinese, which are possibly clustered around their centroid and are challenging to learn the policy for all of them. ", + "bbox": [ + 174, + 661, + 825, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reward design. We consider different weights to trade-off between the query latency from the input and the quality of the returned documents. The quality is evaluated by Ranking Scores of top five documents. We found that it is sparse since the good documents are hard to match. When we weight more on ranking score, the agent tends to stop the search early. ", + "bbox": [ + 174, + 777, + 823, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also investigate a few punishments and try to use less of them to avoid manual design. We found the agent learns some common patterns guided by the reward and punishments. It tries to avoid ”stop” at the first step because we set a huge punishment on such invalid stop. This punishment is necessary since it prevents the agent stops at first step for better value than trying more steps but get no documents matched. In other words, the agent is encouraged to explore different rewrites instead of using empty match plans. Without such punishment, the performance significantly drops since it is hard to explore good documents at first. We also punish the agent to avoid unsupported rewrites, since the production system uses designed rules for each class of the queries and simply omits a rewrite without any useful feedback. We test punishing repeated rewrites, however, it helps seek count (more efficient) but may harm ranking scores. ", + "bbox": [ + 174, + 840, + 825, + 924 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e3a1464f0b6cd2d71b46e2251154e906a70ddf2e498fb4263bde557589df6bb7.jpg", + "table_caption": [ + "Table 1: Improved Scores of Different Noise Compared to RNN Gaussian Noise. Ranking Score and Seek Count are scaled to match Overall $=$ RankingScore−SeekCount for easy comparison. " + ], + "table_footnote": [], + "table_body": "
OverallRanking Score (Quality)Seek Count (Efficiency)
OU Noise+1.00+1.20-0.20
Param Noise+1.69+0.82-0.87
", + "bbox": [ + 232, + 101, + 763, + 160 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7320f38c5eca22be88d47fec5b5f6b8be13097a72a46698ca9f2671a3e8e7c6c.jpg", + "table_caption": [ + "Table 2: Improved Scores of Different Model Compared to MLP " + ], + "table_footnote": [], + "table_body": "
OverallRanking Score (Quality)Seek Count (Efficiency)
MLP+Prioritized+14.54+8.56-5.98
RNN+Prioritized+32.63+14.44-18.19
RNN + Prioritized Sample+32.67+15.29-17.38
", + "bbox": [ + 187, + 222, + 810, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "", + "bbox": [ + 174, + 358, + 825, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3 ABLATION STUDY ", + "text_level": 1, + "bbox": [ + 176, + 438, + 339, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare different exploration noise, including action space noise and parameter space noise. For the action space noise, the noise on discrete and continuous actions are applied separately. We test Gaussian and OrnsteinUhlenbeck noise on continuous actions, while the discrete actions only use $\\epsilon$ -greedy. To keep the comparison fair, we also test $\\epsilon$ -greedy strategy on continuous actions by uniformly sampling a point with probability $\\epsilon$ . The parameter noise is directly applied on policy network $\\widetilde \\mu ( s ; \\theta _ { \\mu } )$ . We found the parameter space noise performs more stable than all types of action espace noise. The uniform sampling on both discrete and continous action with $\\epsilon$ -greedy failed for sometimes, thus we did not include it in comparison. We give the relative improvement in Table 1 compared to RNN with Gaussian Noise. The results show that parameter space noise has best performance overall. ", + "bbox": [ + 174, + 467, + 825, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Figure 6(a) and Table 2 show that RNN got significantly better performance than MLP with or without parameterization on our environment. It indicates that the environment is highly non-Markov. We guess one obvious possibility is that the system signals of the state space cannot include all the information, while recurrent networks try to extract latent state from the history. Note that each episode will sample one query, thus the query embedding is constant for all steps in one episode. Another possible reason is that some rewrites may not be supported by special queries or have too small change for the state signals. ", + "bbox": [ + 174, + 613, + 825, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare different parameterization methods on recurrent networks. The results are shown in Figure 4(b) and Table 2. We empirically found that our modified strategy using reward bin is more stable from the beginning. It is possible that the agent repeatedly replays not only good experience on matching documents, but also learns to avoid punishment we set, as we expect. In general prioritized replay, the sampling probability is just based on TD-error and may be biased to worse samples since value function may not update towards possible results. ", + "bbox": [ + 174, + 717, + 825, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4 BENCHMARKING GAMES ", + "text_level": 1, + "bbox": [ + 176, + 824, + 388, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We experiment on Platform-v0 and Goal-v0 from Bester et al. (2019). Note that, in all these games, each discrete action has a separate continuous action-parameter space. Our algorithm is designed for shared action-parameters and does not utilize such prior, thus we do not compare with P-DQN-style algorithms. We apply n-step return since the episode in these games is much longer. We assume the environment is fully observable and do not use recurrent networks in comparison. ", + "bbox": [ + 174, + 853, + 823, + 922 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters. We examine the games with different combinations of hyperparameters, since they are easy to parallize on each training nodes without the need to connect to an production environment emulator. With NNI, the learning rates for value and policy network are set to log-uniform in $[ 3 \\times 1 0 ^ { - 4 } , 3 \\times 1 0 ^ { - 3 } ]$ and $[ 1 \\times 1 0 ^ { - 4 } , 1 \\times \\mathsf { \\bar { 1 } 0 ^ { - 3 } } ]$ . The $\\alpha$ in parameterized replay is set to log-uniform in [0.2, 1.0]. The action noise threshold $\\delta$ in parameter space noise is set to log-uniform between [0.05, 0.8]. Tn soft parameter update, the $\\tau$ is set to log-uniform in $[ 1 \\times 1 0 ^ { - 3 } , 1 \\times 1 0 ^ { - 2 } ]$ . ", + "bbox": [ + 174, + 103, + 825, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Platform-v0 and Goal-v0, we found the agent is not very sensitive to the range we set, such as $\\alpha , \\delta$ and $\\tau$ . The $\\delta$ values of top $20 \\%$ trials vary between [0.05, 0.3], while $\\alpha$ and $\\tau$ values are evenly scattered in the defined range. In these settings, usually the agent prefers slightly larger value learning rate than policy learning rate. ", + "bbox": [ + 174, + 194, + 825, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/43fb6a7fec969297f5ab873d3137dea0d0470694b6859f537189d55399ea27e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Average Eval ReturnOurPA-DDPG
Platform-v00.95730.3113
Goal-v034.20-6.208
", + "bbox": [ + 346, + 266, + 648, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3: Average evaluation results (the average of all training rewards and final evaluation reward) on benchmarks Platform-v0 and Goal-v0 with PA-DDPG (Hausknecht & Stone (2015)), MP-DQN (Bester et al. (2019)) and P-DQN (Xiong et al. (2018)). We use reported number from the papers, while last two methods report another metric on Goal-v0. ", + "bbox": [ + 173, + 340, + 825, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5 RELATED WORK ", + "text_level": 1, + "bbox": [ + 176, + 430, + 344, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While the aforementioned algorithms and techniques work on discrete or continuous action spaces, it is not trivial to apply them on parameterized action space, since such discrete-continuous hybrid action space is hard to parameterized by a single distribution. A related series of work is to combine DDPG and DQN to optimize Q-value function on parameter actions. There are two classes of methods that belong to them: PA-DDPG (Hausknecht & Stone (2015)) based on DDPG and PDQN (Xiong et al. (2018)) based on DQN. They are Q-Learning-based methods which select the best action by maximizing the Q-value function on discrete action space or learning a deterministic policy outputting best continuous action. However, they use different strategies to combine discrete and continuous actions. Bester et al. (2019) (MP-DQN) extends P-DQN to tackle the problem that Q-value is a function of the joint action-parameter vector $Q ( s ^ { \\prime } , k ^ { \\prime } , { \\bf x } ^ { \\tilde { Q } } ( s ^ { \\prime } ) )$ in normal PAMDP, which may results in fault gradients. However, such problem does not exist in our slightly modified setting, since the action-parameter space $\\mathcal { X }$ in the match plan generation is inherently defined to be shared for each $k \\in \\mathcal { A } _ { d }$ . Masson et al. (2016) purposes a method to iteratively optimizing discrete and continuous actions by alternating between them. Another perspective (Klimek et al. (2017); Wei et al. (2018); Fu et al. (2019)) for a parameterized action space is to regard it as a twohierarchy action space, where an agent selects discrete action first and continuous parameter later. However, we do not consider this direction in current scheme because we share the same parameters for all discrete action in not very large scale. Therefore, the hierarchical methods may not bring a significant performance gain. ", + "bbox": [ + 174, + 463, + 825, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6 DISCUSSIONS ", + "text_level": 1, + "bbox": [ + 174, + 751, + 320, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a parameterized action RL match plan generation method which extends the plan generation to the general case without any predefined knowledge. Key to address the problem are normalized softmax values of discrete actions to enable gradients backpropagation, parameter space noise on parameters of the policy for unifying the exploration direction in both discrete and continuous spaces, and recurrent deterministic policies with prioritized replay buffer to accelerate and stabilize the training. Our algorithm can be applied to not only the match plan generation environment, but also other similar parameteried action environments. The experiment results demonstrate our method outperforms the well-designed hand-crafted rules in Bing and serveral baseline results in some existing PARL benchmarks. In this paper, we mainly discuss about offline training procedure. In the future, we plan to apply learned policy to the production environment. ", + "bbox": [ + 174, + 784, + 825, + 924 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "REFERENCES ", + "text_level": 1, + "bbox": [ + 174, + 102, + 287, + 118 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. ", + "bbox": [ + 173, + 126, + 821, + 155 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Craig J Bester, Steven D James, and George D Konidaris. Multi-pass q-networks for deep reinforcement learning with parameterised action spaces. arXiv preprint arXiv:1905.04388, 2019. ", + "bbox": [ + 173, + 162, + 821, + 193 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Haotian Fu, Hongyao Tang, Jianye Hao, Zihan Lei, Yingfeng Chen, and Changjie Fan. Deep multi-agent reinforcement learning with discrete-continuous hybrid action spaces. arXiv preprint arXiv:1903.04959, 2019. ", + "bbox": [ + 173, + 199, + 823, + 242 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in actor-critic methods. arXiv preprint arXiv:1802.09477, 2018. ", + "bbox": [ + 171, + 251, + 823, + 280 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space. arXiv preprint arXiv:1511.04143, 2015. ", + "bbox": [ + 174, + 289, + 823, + 318 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nicolas Heess, Jonathan J Hunt, Timothy P Lillicrap, and David Silver. Memory-based control with recurrent neural networks. arXiv preprint arXiv:1512.04455, 2015. ", + "bbox": [ + 173, + 325, + 823, + 356 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8): 1735–1780, 1997. ", + "bbox": [ + 174, + 363, + 823, + 392 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent experience replay in distributed reinforcement learning. 2018. ", + "bbox": [ + 173, + 400, + 823, + 430 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Maciej Klimek, Henryk Michalewski, Piotr Mi, et al. Hierarchical reinforcement learning with parameters. In Conference on Robot Learning, pp. 301–313, 2017. ", + "bbox": [ + 173, + 438, + 823, + 468 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv preprint arXiv:1509.02971, 2015. ", + "bbox": [ + 176, + 474, + 821, + 518 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Warwick Masson, Pravesh Ranchod, and George Konidaris. Reinforcement learning with parameterized actions. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. ", + "bbox": [ + 169, + 526, + 821, + 556 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. ", + "bbox": [ + 169, + 564, + 823, + 594 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Azalia Mirhoseini, Hieu Pham, Quoc V Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean. Device placement optimization with reinforcement learning. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 2430–2439. JMLR. org, 2017. ", + "bbox": [ + 173, + 601, + 823, + 659 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529, 2015. ", + "bbox": [ + 173, + 666, + 823, + 710 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Matthias Plappert, Rein Houthooft, Prafulla Dhariwal, Szymon Sidor, Richard Y Chen, Xi Chen, Tamim Asfour, Pieter Abbeel, and Marcin Andrychowicz. Parameter space noise for exploration. arXiv preprint arXiv:1706.01905, 2017. ", + "bbox": [ + 174, + 717, + 823, + 761 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Corby Rosset, Damien Jose, Gargi Ghosh, Bhaskar Mitra, and Saurabh Tiwary. Optimizing query evaluations using reinforcement learning for web search. In The 41st International ACM SIGIR Conference on Research & Development in Information Retrieval, pp. 1193–1196. ACM, 2018. ", + "bbox": [ + 171, + 768, + 823, + 813 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv preprint arXiv:1511.05952, 2015. ", + "bbox": [ + 173, + 820, + 823, + 849 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller. Deterministic policy gradient algorithms. 2014. ", + "bbox": [ + 174, + 858, + 821, + 887 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ermo Wei, Drew Wicke, and Sean Luke. Hierarchical approaches for reinforcement learning in parameterized action space. In 2018 AAAI Spring Symposium Series, 2018. ", + "bbox": [ + 174, + 895, + 821, + 924 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Ian H Witten, Ian H Witten, Alistair Moffat, Timothy C Bell, Timothy C Bell, and Timothy C Bell. Managing gigabytes: compressing and indexing documents and images. Morgan Kaufmann, 1999. ", + "bbox": [ + 174, + 103, + 823, + 145 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Jiechao Xiong, Qing Wang, Zhuoran Yang, Peng Sun, Lei Han, Yang Zheng, Haobo Fu, Tong Zhang, Ji Liu, and Han Liu. Parametrized deep q-networks learning: Reinforcement learning with discrete-continuous hybrid action space. arXiv preprint arXiv:1810.06394, 2018. ", + "bbox": [ + 174, + 155, + 823, + 196 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "Justin Zobel and Alistair Moffat. Inverted files for text search engines. ACM computing surveys (CSUR), 38(2):6, 2006. ", + "bbox": [ + 171, + 207, + 823, + 234 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "A APPENDIX ", + "text_level": 1, + "bbox": [ + 176, + 102, + 299, + 117 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A.1 PSEUDOCODE ", + "text_level": 1, + "bbox": [ + 176, + 133, + 315, + 148 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Algorithm pseudocode for the algorithm is provided in Algorithm 1 which includes all aforementioned details. ", + "bbox": [ + 176, + 161, + 823, + 189 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Algorithm 1 ", + "text_level": 1, + "bbox": [ + 174, + 204, + 259, + 218 + ], + "page_idx": 10 + }, + { + "type": "table", + "img_path": "images/6805768dc1374f48ea940804c4f043f9541f672a807ee8463e41885f8a35b386.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input: Empty replay buffer D, init parameter noise std δparam, action noise threshold δ Initialize policy parameters 0,value parameters Φ ←0Q for each episode do Perturb policy parameters θ ← θ +(O,Oparam) and target policy parameters θtarget Observe state St, output a disturbed action embedding A =p(st；0） Compute executing action at by taking max over k
Execute parameterized action at in the environment server Observe next state St+1,reward rt, done signal dt denoting if St+1 is terminal t,rt, St+1,dt) to the buffer D If St+1 is terminal, reset to an initial state so for each update if update-condition do
Sample a minibatch from prioritized replay buffer D with specific priorities Compute and transform target actions with disturbed target policy network
Update parameter noise std Oparam using empirical distance d(μ, μ)
Compute targets for TD-error with min Q-value in the twin Q-networks Update Q-network parameters using gradient descent
if policy update frequency then
Update policy network parameters θ with Update target parameters with polyak averaging end if
", + "bbox": [ + 173, + 219, + 808, + 554 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A.2 FURTHER EXPERIMENTAL DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 583, + 464, + 598 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Hyperparameter search. We use Microsoft $\\mathrm { N N I } ^ { 2 }$ and OpenPAI3 to search hyperparameters. The final metric to report to NNI is set to the sum of average training reward and final evaluation reward (repeated 1,000 times) final_metric $=$ (avg_reward+eval_reward)/2. The intermediate metric is set to evaluation reward per 1,000 episodes. We also use the early stop assessor. Each GPU server node connects to a production environment emulator with ethernet. ", + "bbox": [ + 173, + 609, + 825, + 680 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Model architecture. Both policy and value networks use two fully connected layers with 512 hidden units and a LSTM layer (Hochreiter & Schmidhuber (1997)). Each layer also uses a layer normalization (LayerNorm, Ba et al. (2016)) as suggested by Plappert et al. (2017) in consideration of stability for noise applied on parameters, and follows a ReLU activation. Both output heads of the policy network has a hidden layer with sof tmax or tanh activations. ", + "bbox": [ + 174, + 696, + 825, + 767 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "A.3 FURTHER ENVIRONMENTAL DETAILS ", + "text_level": 1, + "bbox": [ + 176, + 785, + 477, + 800 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Accumulated values. Note that, for accumulated values in received states, rewards and outputted action-parameters, we use the difference (delta values) from the last step. For states and rewards, it is $s _ { t } = s _ { t } ^ { \\prime } - s _ { t - 1 }$ , $r _ { t } = r _ { t } ^ { \\prime } - r _ { t - 1 }$ , where where $s ^ { \\prime }$ and $r ^ { \\prime }$ denote raw state and reward. For actions, the agent outputs raw action-parameter output $a ^ { \\prime }$ , and the emulator converts it to accumulated value $a _ { t } = a _ { t } ^ { \\prime } + a _ { t - 1 }$ . ", + "bbox": [ + 174, + 813, + 825, + 883 + ], + "page_idx": 10 + }, + { + "type": "image", + "img_path": "images/6d04d136c67be4f2400a72fd7e2f02c247f2d8e0e0954f2f1782df4d09b152ff.jpg", + "image_caption": [ + "Figure 6: Best evaluation learning curves on both environments during tuning. The agent achieves maximum possible reward on both environments (50 and 1). " + ], + "image_footnote": [], + "bbox": [ + 218, + 106, + 779, + 297 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Inverted index in Bing. In the inverted index system in Bing, there are two key steps: (i) the system rewrites the query to enlarge candidate set with specified match plan (or search plan), (ii) it returns candidates with top ranking scores. We just refer to the math plan part regarding to generation, but not user-input content. In the match plan generation, the goal is to generate the optimal search plan (policy) for each query. ", + "bbox": [ + 174, + 366, + 825, + 435 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "Closed-loop and open-loop control. The production match plan is generated online in open-loop (feedforward) without taking runtime system signals into consideration for the latency and implementation consideration. However, RL is closed-loop which takes feedback from system signals to make decisions. This may increase the online overhead, but it can be tackled by converting the reflective policy to a shooting-style action sequence by predicting with learned transition dynamics. In the implementation, the environment emulator receives an action sequence $a _ { 0 } , . . . , a _ { t }$ and return $s _ { t + 1 }$ to simulate the open-loop style. ", + "bbox": [ + 174, + 450, + 825, + 549 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "A.4 MORE TRAINING FOR BENCHMARKS ", + "text_level": 1, + "bbox": [ + 176, + 565, + 473, + 579 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "We provide the best evaluation learning curves on both environments during tuning. ", + "bbox": [ + 173, + 592, + 720, + 606 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/parse/train/rkgZaT4tDr/rkgZaT4tDr_middle.json b/parse/train/rkgZaT4tDr/rkgZaT4tDr_middle.json new file mode 100644 index 0000000000000000000000000000000000000000..464196a8654a6d767ef3a6ff6252dc75a4ab9d2c --- /dev/null +++ b/parse/train/rkgZaT4tDr/rkgZaT4tDr_middle.json @@ -0,0 +1,28757 @@ +{ + "pdf_info": [ + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "PARAMETERIZED ACTION REINFORCEMENT LEARN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 505, + 118 + ], + "score": 1.0, + "content": "ING FOR INVERTED INDEX MATCH PLAN GENERATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 224 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 224 + ], + "score": 1.0, + "content": "Match plan generation in the inverted index at Microsoft Bing is used to be based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 470, + 235 + ], + "score": 1.0, + "content": "on hand-crafted rules. We formulate the generation process as a Parameterized", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "score": 1.0, + "content": "Action MDP with sharing parameters and purpose a reinforcement learning algo-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "rithm on such formulation. We combine deterministic policy learning on discrete", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "and continuous action spaces and several recent advances in deep reinforcement", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "score": 1.0, + "content": "learning. For exploring in the parameterized action space, the agent outputs soft-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "score": 1.0, + "content": "max values for discrete actions and applies Parameter Space Noise on policy net-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 142, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "work to unify the exploration direction in both spaces. We apply prioritized re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "score": 1.0, + "content": "current replay on match plan sequences and pad short match plans. We also use", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 298, + 323 + ], + "score": 1.0, + "content": "invertible value function rescaling and", + "type": "text" + }, + { + "bbox": [ + 298, + 313, + 305, + 321 + ], + "score": 0.44, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "-step return to stabilize the training. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "agent is evaluated on our environment and some benchmarks. It outperforms the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 467, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 467, + 345 + ], + "score": 1.0, + "content": "well-designed production match plan and beats the baselines on the benchmarks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 366, + 205, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 208, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 208, + 382 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 406 + ], + "score": 1.0, + "content": "Using machine learning to optimize and accelerate software and hardware systems is an emerging", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "field in the past few years (Mirhoseini et al. (2017); Rosset et al. (2018)). The process of decision-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "making in those systems is usually hand-crafted by human engineers, which is less explored to be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "automatically done by learning algorithms. A promising direction is to formulate such sequential", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "decision-making problems as Reinforcement Learning (RL) problems, such as search plan genera-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 446, + 327, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 327, + 459 + ], + "score": 1.0, + "content": "tion problem in inverted indexes (Rosset et al. (2018)).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5 + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 504, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "The inverted index is a specialized data structure that is commonly used in search engines, including", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 148, + 487 + ], + "score": 1.0, + "content": "Microsoft", + "type": "text" + }, + { + "bbox": [ + 149, + 474, + 173, + 486 + ], + "score": 0.41, + "content": "\\mathrm { B i n g ^ { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 474, + 505, + 487 + ], + "score": 1.0, + "content": ". An inverted index provides an inverted mapping from term to documents. Each", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "term has a posting list which contains all the (document, location) pairs that the term appears. By", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "combining the posting lists of the terms in a user query, the initial document candidates are generated", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 507, + 293, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 293, + 520 + ], + "score": 1.0, + "content": "(Witten et al. (1999); Zobel & Moffat (2006)).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "In Bing, documents are scanned with specified match plans, either predefined or generated in real-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "time. A match plan has a sequence of rewrites (match rules, e.g., it treats the query as a phrase", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "score": 1.0, + "content": "which should appear in the document exactly as it is), where each rewrite can be controlled by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "several quotas (stopping criteria). There are several types of rewrites, and all these types share", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "the same continuous quotas. A pair of a rewrite type and the quotas forms a single action which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 323, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 323, + 592 + ], + "score": 1.0, + "content": "determines if a document will be a ranking candidate.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "We formulate the generation problem as a RL problem. The state consists of system runtime sig-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "nals and semantic embeddings of queries. The action space is called parameterized or discrete-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "continuous hybrid, where an action has a discrete action and continuous action-parameters. The", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "reward is weighted by result quality and query latency. It is similar to Parameterized Action RL", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 640, + 487, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 487, + 653 + ], + "score": 1.0, + "content": "(Masson et al. (2016)) (PARL), while our setting requires all actions to share same parameters.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "score": 1.0, + "content": "Previously, the match plans are predefined manually for each query. It hardly utilizes the rich infor-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "mation in the state to dynamically adjust the rewrites and quotas for specific query online. Rosset", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "et al. (2018) tries to automatically generate match plans using tabular Q-learning with discretized", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 704 + ], + "score": 1.0, + "content": "state space and predefined action-parameters. They learn the generation policy for a specific query", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "class each time and solve it only in discretized spaces with tabular methods. We extend the genera-", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42 + } + ], + "page_idx": 0, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 120, + 722, + 204, + 731 + ], + "lines": [ + { + "bbox": [ + 119, + 720, + 205, + 734 + ], + "spans": [ + { + "bbox": [ + 119, + 720, + 205, + 734 + ], + "score": 1.0, + "content": "1https://www.bing.com", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "1", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 78, + 503, + 116 + ], + "lines": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "spans": [ + { + "bbox": [ + 105, + 78, + 505, + 98 + ], + "score": 1.0, + "content": "PARAMETERIZED ACTION REINFORCEMENT LEARN-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 99, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 106, + 99, + 505, + 118 + ], + "score": 1.0, + "content": "ING FOR INVERTED INDEX MATCH PLAN GENERATION", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 0.5 + }, + { + "type": "text", + "bbox": [ + 112, + 136, + 244, + 157 + ], + "lines": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "spans": [ + { + "bbox": [ + 113, + 136, + 201, + 147 + ], + "score": 1.0, + "content": "Anonymous authors", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 146, + 245, + 159 + ], + "score": 1.0, + "content": "Paper under double-blind review", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 112, + 136, + 245, + 159 + ] + }, + { + "type": "title", + "bbox": [ + 278, + 186, + 333, + 199 + ], + "lines": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "spans": [ + { + "bbox": [ + 277, + 186, + 335, + 200 + ], + "score": 1.0, + "content": "ABSTRACT", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 143, + 212, + 468, + 344 + ], + "lines": [ + { + "bbox": [ + 141, + 212, + 469, + 224 + ], + "spans": [ + { + "bbox": [ + 141, + 212, + 469, + 224 + ], + "score": 1.0, + "content": "Match plan generation in the inverted index at Microsoft Bing is used to be based", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 223, + 470, + 235 + ], + "spans": [ + { + "bbox": [ + 141, + 223, + 470, + 235 + ], + "score": 1.0, + "content": "on hand-crafted rules. We formulate the generation process as a Parameterized", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "spans": [ + { + "bbox": [ + 141, + 233, + 469, + 247 + ], + "score": 1.0, + "content": "Action MDP with sharing parameters and purpose a reinforcement learning algo-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "spans": [ + { + "bbox": [ + 141, + 245, + 469, + 257 + ], + "score": 1.0, + "content": "rithm on such formulation. We combine deterministic policy learning on discrete", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 470, + 268 + ], + "score": 1.0, + "content": "and continuous action spaces and several recent advances in deep reinforcement", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "spans": [ + { + "bbox": [ + 141, + 267, + 469, + 280 + ], + "score": 1.0, + "content": "learning. For exploring in the parameterized action space, the agent outputs soft-", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "spans": [ + { + "bbox": [ + 141, + 278, + 469, + 290 + ], + "score": 1.0, + "content": "max values for discrete actions and applies Parameter Space Noise on policy net-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 142, + 289, + 469, + 301 + ], + "spans": [ + { + "bbox": [ + 142, + 289, + 469, + 301 + ], + "score": 1.0, + "content": "work to unify the exploration direction in both spaces. We apply prioritized re-", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "spans": [ + { + "bbox": [ + 141, + 300, + 470, + 312 + ], + "score": 1.0, + "content": "current replay on match plan sequences and pad short match plans. We also use", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 141, + 311, + 469, + 323 + ], + "spans": [ + { + "bbox": [ + 141, + 311, + 298, + 323 + ], + "score": 1.0, + "content": "invertible value function rescaling and", + "type": "text" + }, + { + "bbox": [ + 298, + 313, + 305, + 321 + ], + "score": 0.44, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 306, + 311, + 469, + 323 + ], + "score": 1.0, + "content": "-step return to stabilize the training. The", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "spans": [ + { + "bbox": [ + 141, + 322, + 469, + 334 + ], + "score": 1.0, + "content": "agent is evaluated on our environment and some benchmarks. It outperforms the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 141, + 333, + 467, + 345 + ], + "spans": [ + { + "bbox": [ + 141, + 333, + 467, + 345 + ], + "score": 1.0, + "content": "well-designed production match plan and beats the baselines on the benchmarks.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 10.5, + "bbox_fs": [ + 141, + 212, + 470, + 345 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 366, + 205, + 379 + ], + "lines": [ + { + "bbox": [ + 105, + 365, + 208, + 382 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 208, + 382 + ], + "score": 1.0, + "content": "1 INTRODUCTION", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 392, + 505, + 457 + ], + "lines": [ + { + "bbox": [ + 105, + 390, + 505, + 406 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 406 + ], + "score": 1.0, + "content": "Using machine learning to optimize and accelerate software and hardware systems is an emerging", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 402, + 505, + 415 + ], + "score": 1.0, + "content": "field in the past few years (Mirhoseini et al. (2017); Rosset et al. (2018)). The process of decision-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "making in those systems is usually hand-crafted by human engineers, which is less explored to be", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 505, + 438 + ], + "score": 1.0, + "content": "automatically done by learning algorithms. A promising direction is to formulate such sequential", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "decision-making problems as Reinforcement Learning (RL) problems, such as search plan genera-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 446, + 327, + 459 + ], + "spans": [ + { + "bbox": [ + 105, + 446, + 327, + 459 + ], + "score": 1.0, + "content": "tion problem in inverted indexes (Rosset et al. (2018)).", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 20.5, + "bbox_fs": [ + 105, + 390, + 505, + 459 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 463, + 504, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "spans": [ + { + "bbox": [ + 105, + 462, + 506, + 477 + ], + "score": 1.0, + "content": "The inverted index is a specialized data structure that is commonly used in search engines, including", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 474, + 505, + 487 + ], + "spans": [ + { + "bbox": [ + 106, + 474, + 148, + 487 + ], + "score": 1.0, + "content": "Microsoft", + "type": "text" + }, + { + "bbox": [ + 149, + 474, + 173, + 486 + ], + "score": 0.41, + "content": "\\mathrm { B i n g ^ { 1 } }", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 474, + 505, + 487 + ], + "score": 1.0, + "content": ". An inverted index provides an inverted mapping from term to documents. Each", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "spans": [ + { + "bbox": [ + 105, + 484, + 505, + 499 + ], + "score": 1.0, + "content": "term has a posting list which contains all the (document, location) pairs that the term appears. By", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 106, + 496, + 506, + 509 + ], + "score": 1.0, + "content": "combining the posting lists of the terms in a user query, the initial document candidates are generated", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 507, + 293, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 507, + 293, + 520 + ], + "score": 1.0, + "content": "(Witten et al. (1999); Zobel & Moffat (2006)).", + "type": "text" + } + ], + "index": 28 + } + ], + "index": 26, + "bbox_fs": [ + 105, + 462, + 506, + 520 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 504, + 590 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "In Bing, documents are scanned with specified match plans, either predefined or generated in real-", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 506, + 548 + ], + "score": 1.0, + "content": "time. A match plan has a sequence of rewrites (match rules, e.g., it treats the query as a phrase", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "spans": [ + { + "bbox": [ + 105, + 545, + 506, + 560 + ], + "score": 1.0, + "content": "which should appear in the document exactly as it is), where each rewrite can be controlled by", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 558, + 505, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 569 + ], + "score": 1.0, + "content": "several quotas (stopping criteria). There are several types of rewrites, and all these types share", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 506, + 581 + ], + "score": 1.0, + "content": "the same continuous quotas. A pair of a rewrite type and the quotas forms a single action which", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 578, + 323, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 323, + 592 + ], + "score": 1.0, + "content": "determines if a document will be a ranking candidate.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31.5, + "bbox_fs": [ + 105, + 524, + 506, + 592 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 596, + 504, + 651 + ], + "lines": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "spans": [ + { + "bbox": [ + 106, + 595, + 505, + 609 + ], + "score": 1.0, + "content": "We formulate the generation problem as a RL problem. The state consists of system runtime sig-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "spans": [ + { + "bbox": [ + 106, + 606, + 505, + 619 + ], + "score": 1.0, + "content": "nals and semantic embeddings of queries. The action space is called parameterized or discrete-", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "spans": [ + { + "bbox": [ + 105, + 618, + 505, + 631 + ], + "score": 1.0, + "content": "continuous hybrid, where an action has a discrete action and continuous action-parameters. The", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "spans": [ + { + "bbox": [ + 106, + 629, + 505, + 641 + ], + "score": 1.0, + "content": "reward is weighted by result quality and query latency. It is similar to Parameterized Action RL", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 640, + 487, + 653 + ], + "spans": [ + { + "bbox": [ + 106, + 640, + 487, + 653 + ], + "score": 1.0, + "content": "(Masson et al. (2016)) (PARL), while our setting requires all actions to share same parameters.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 37, + "bbox_fs": [ + 105, + 595, + 505, + 653 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 656, + 504, + 712 + ], + "lines": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "spans": [ + { + "bbox": [ + 105, + 656, + 505, + 670 + ], + "score": 1.0, + "content": "Previously, the match plans are predefined manually for each query. It hardly utilizes the rich infor-", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "spans": [ + { + "bbox": [ + 105, + 667, + 506, + 681 + ], + "score": 1.0, + "content": "mation in the state to dynamically adjust the rewrites and quotas for specific query online. Rosset", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 678, + 506, + 692 + ], + "spans": [ + { + "bbox": [ + 105, + 678, + 506, + 692 + ], + "score": 1.0, + "content": "et al. (2018) tries to automatically generate match plans using tabular Q-learning with discretized", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 688, + 505, + 704 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 704 + ], + "score": 1.0, + "content": "state space and predefined action-parameters. They learn the generation policy for a specific query", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 505, + 713 + ], + "score": 1.0, + "content": "class each time and solve it only in discretized spaces with tabular methods. We extend the genera-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "tion process to the general case, such that the match plans are fully parameterized and learned from", + "type": "text", + "cross_page": true + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 221, + 385, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 385, + 234 + ], + "score": 1.0, + "content": "scratch without any predefined knowledge (e.g., limited match rules).", + "type": "text", + "cross_page": true + } + ], + "index": 8 + } + ], + "index": 42, + "bbox_fs": [ + 105, + 656, + 506, + 713 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 190, + 86, + 421, + 146 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 190, + 86, + 421, + 146 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 190, + 86, + 421, + 146 + ], + "spans": [ + { + "bbox": [ + 190, + 86, + 421, + 146 + ], + "score": 0.96, + "type": "image", + "image_path": "7459112fc13a2ca51b6133f349d4373e932db4f4ab389ec3ae4eb649b28ac3c6.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 86, + 421, + 101.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 190, + 101.0, + 421, + 116.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 190, + 116.0, + 421, + 131.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 190, + 131.0, + 421, + 146.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 162, + 503, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "score": 1.0, + "content": "Figure 1: Match plan example: for the query “Reinforcement Learning”, the search engine firstly gets a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 172, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 184 + ], + "score": 1.0, + "content": "document posting list for each term from the inverted index, and then scans the document candidates following", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 182, + 329, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 329, + 194 + ], + "score": 1.0, + "content": "a serial of rewrites (match rules) according to the match plan.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 105, + 209, + 503, + 232 + ], + "lines": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "spans": [ + { + "bbox": [ + 106, + 209, + 505, + 222 + ], + "score": 1.0, + "content": "tion process to the general case, such that the match plans are fully parameterized and learned from", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 221, + 385, + 234 + ], + "spans": [ + { + "bbox": [ + 106, + 221, + 385, + 234 + ], + "score": 1.0, + "content": "scratch without any predefined knowledge (e.g., limited match rules).", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 7.5 + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "score": 1.0, + "content": "In this paper, we purpose a parameterized action RL algorithm that learns to act on parameterized", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "action space environments, such as the match plan generation process. We introduce normalized", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "score": 1.0, + "content": "softmax values of discrete actions to form a categorical distributions to enable gradients backprop-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 269, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 285 + ], + "score": 1.0, + "content": "agation. We also combine several recent advances in RL to accelerate and stabilize the training.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "We investigate parameter space noise (Plappert et al. (2017)) on parameters of the policy for unify-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 293, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 305 + ], + "score": 1.0, + "content": "ing directions in exploring the structured action space. We explore recurrent deterministic policies", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "score": 1.0, + "content": "(Heess et al. (2015)) with prioritized replay buffer (Schaul et al. (2015)) on sequence, due to the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 313, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 451, + 328 + ], + "score": 1.0, + "content": "inherent nature of partial observability and sparse feedback in our setting. We also use", + "type": "text" + }, + { + "bbox": [ + 451, + 317, + 457, + 325 + ], + "score": 0.49, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 313, + 506, + 328 + ], + "score": 1.0, + "content": "-step return", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 325, + 453, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 453, + 339 + ], + "score": 1.0, + "content": "and invertible value function rescaling (Kapturowski et al. (2018)) for further stability.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "We present an agent to integrate these techniques and evaluate on offline training on a query dataset", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 354, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 366 + ], + "score": 1.0, + "content": "collected from Bing search. When training on the dataset with uniformly sampling, the agent out-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "score": 1.0, + "content": "performs the currently deployed production match plan, which is generated based on well-designed", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "hand-crafted rules. We study the performance of training on the dataset or only with several selected", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "complicated queries individually, as well as ablation studies for the various components. We test", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "on a few existing PARL benchmarks, where our agent beats our baseline methods and performs the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 409, + 199, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 199, + 420 + ], + "score": 1.0, + "content": "state-of-the-art results.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 434, + 200, + 448 + ], + "lines": [ + { + "bbox": [ + 104, + 433, + 202, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 433, + 202, + 451 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 108, + 459, + 254, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 256, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 256, + 472 + ], + "score": 1.0, + "content": "2.1 REINFORCEMENT LEARNING", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "We address the problem using reinforcement learning (RL) framework with a parameterized action", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 491, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 502 + ], + "score": 1.0, + "content": "space. An agent interacts with an environment to maximize the accumulated reward with discount", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 133, + 514 + ], + "score": 1.0, + "content": "factor", + "type": "text" + }, + { + "bbox": [ + 133, + 502, + 177, + 514 + ], + "score": 0.92, + "content": "\\gamma \\in \\ [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 502, + 505, + 514 + ], + "score": 1.0, + "content": ". We model the environment with a discrete-time Partially Observable Markov", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 511, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 348, + 526 + ], + "score": 1.0, + "content": "Decision Process (POMDP). A POMDP is given by a tuple", + "type": "text" + }, + { + "bbox": [ + 348, + 513, + 423, + 524 + ], + "score": 0.92, + "content": "( S , \\mathcal { A } , T , R , \\Omega , \\dot { \\mathcal { O } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 511, + 505, + 526 + ], + "score": 1.0, + "content": ", and the underlying", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 306, + 536 + ], + "score": 1.0, + "content": "Markov Decision Process (MDP) is defined by", + "type": "text" + }, + { + "bbox": [ + 306, + 524, + 356, + 535 + ], + "score": 0.9, + "content": "( S , { \\mathcal { A } } , T , R )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 524, + 390, + 536 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 390, + 525, + 398, + 534 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 524, + 478, + 536 + ], + "score": 1.0, + "content": "is the state space,", + "type": "text" + }, + { + "bbox": [ + 479, + 524, + 488, + 534 + ], + "score": 0.77, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 535, + 504, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 221, + 546 + ], + "score": 1.0, + "content": "parameterized action space,", + "type": "text" + }, + { + "bbox": [ + 221, + 535, + 230, + 545 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 535, + 322, + 546 + ], + "score": 1.0, + "content": "the transition function", + "type": "text" + }, + { + "bbox": [ + 322, + 535, + 415, + 546 + ], + "score": 0.9, + "content": "\\mathcal { T } : \\mathcal { S } \\times \\mathcal { S } \\times \\mathcal { A } \\mathbb { R } _ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 535, + 436, + 546 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 436, + 535, + 504, + 545 + ], + "score": 0.91, + "content": "R : S \\times \\mathcal { A } \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 546, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 358, + 558 + ], + "score": 1.0, + "content": "is the reward function. The set of observations is given by", + "type": "text" + }, + { + "bbox": [ + 358, + 546, + 366, + 556 + ], + "score": 0.77, + "content": "\\Omega", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 546, + 506, + 558 + ], + "score": 1.0, + "content": "and the the observation function", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 557, + 454, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 441, + 569 + ], + "score": 1.0, + "content": "mapping underlying states to probability distributions over observations is given by", + "type": "text" + }, + { + "bbox": [ + 441, + 557, + 450, + 567 + ], + "score": 0.79, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 557, + 454, + 569 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5 + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "Specifically, the environment in the match plan generation has a parameterized action space. In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Masson et al. (2016); Bester et al. (2019), such formulation in fully observable settings is referred", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 596, + 423, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 423, + 608 + ], + "score": 1.0, + "content": "as Parameterized Action MDP (PAMDP), where the action space is denoted as", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36 + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 609, + 367, + 636 + ], + "lines": [ + { + "bbox": [ + 242, + 609, + 367, + 636 + ], + "spans": [ + { + "bbox": [ + 242, + 609, + 367, + 636 + ], + "score": 0.93, + "content": "\\mathcal { A } = \\bigcup _ { k \\in \\mathcal { A } _ { d } } \\left\\{ ( k , x _ { k } ) | x _ { k } \\in \\mathcal { X } _ { k } \\right\\} ,", + "type": "interline_equation", + "image_path": "76ea0d7a9eb735ac9b64468a264943b9eec3b340c391396f2eb432de507a6511.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 609, + 367, + 622.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 242, + 622.5, + 367, + 636.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 216, + 651 + ], + "score": 1.0, + "content": "where each discrete action", + "type": "text" + }, + { + "bbox": [ + 216, + 637, + 278, + 649 + ], + "score": 0.93, + "content": "a \\in \\mathcal { A } _ { d } = [ K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 636, + 505, + 651 + ], + "score": 1.0, + "content": "has a corresponding continuous action-parameter space", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 649, + 371, + 661 + ], + "spans": [ + { + "bbox": [ + 107, + 650, + 119, + 660 + ], + "score": 0.86, + "content": "\\mathcal { X } _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 649, + 371, + 661 + ], + "score": 1.0, + "content": ". However, our setting requires a slightly different formulation:", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 662, + 389, + 675 + ], + "lines": [ + { + "bbox": [ + 220, + 662, + 389, + 675 + ], + "spans": [ + { + "bbox": [ + 220, + 662, + 389, + 675 + ], + "score": 0.92, + "content": "\\mathcal { A } = \\{ ( k , x ) | k \\in \\mathcal { A } _ { d } , x \\in \\mathcal { X } \\} = \\mathcal { A } _ { d } \\times \\mathcal { X } ,", + "type": "interline_equation", + "image_path": "243001bdfb8447abcdd89869de2d2453c03e6261b0ef47f965d0d6494f259a02.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 220, + 662, + 389, + 675 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 270, + 690 + ], + "score": 1.0, + "content": "where the discrete action space (rewrite)", + "type": "text" + }, + { + "bbox": [ + 271, + 677, + 284, + 688 + ], + "score": 0.88, + "content": "\\mathbf { \\mathcal { A } } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 677, + 418, + 690 + ], + "score": 1.0, + "content": "share the action-parameter space", + "type": "text" + }, + { + "bbox": [ + 419, + 677, + 428, + 687 + ], + "score": 0.78, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". Such formulation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "results in a disentangled action space between discrete and continuous actions which a class of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "parameterized action (P-DQN (Xiong et al. (2018)) and MP-DQN (Bester et al. (2019))) may not", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "be trivially applicable to. Considering the notational simplicity in partially observable environment", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 454, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 454, + 734 + ], + "score": 1.0, + "content": "(POMDP), we do not explicitly denote it when there is no ambiguity in action settings.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45 + } + ], + "page_idx": 1, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 763 + ], + "score": 1.0, + "content": "2", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 190, + 86, + 421, + 146 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 190, + 86, + 421, + 146 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 190, + 86, + 421, + 146 + ], + "spans": [ + { + "bbox": [ + 190, + 86, + 421, + 146 + ], + "score": 0.96, + "type": "image", + "image_path": "7459112fc13a2ca51b6133f349d4373e932db4f4ab389ec3ae4eb649b28ac3c6.jpg" + } + ] + } + ], + "index": 1.5, + "virtual_lines": [ + { + "bbox": [ + 190, + 86, + 421, + 101.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 190, + 101.0, + 421, + 116.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 190, + 116.0, + 421, + 131.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 190, + 131.0, + 421, + 146.0 + ], + "spans": [], + "index": 3 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 106, + 162, + 503, + 193 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 506, + 174 + ], + "score": 1.0, + "content": "Figure 1: Match plan example: for the query “Reinforcement Learning”, the search engine firstly gets a", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 172, + 505, + 184 + ], + "spans": [ + { + "bbox": [ + 105, + 172, + 505, + 184 + ], + "score": 1.0, + "content": "document posting list for each term from the inverted index, and then scans the document candidates following", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 182, + 329, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 182, + 329, + 194 + ], + "score": 1.0, + "content": "a serial of rewrites (match rules) according to the match plan.", + "type": "text" + } + ], + "index": 6 + } + ], + "index": 5 + } + ], + "index": 3.25 + }, + { + "type": "text", + "bbox": [ + 105, + 209, + 503, + 232 + ], + "lines": [], + "index": 7.5, + "bbox_fs": [ + 106, + 209, + 505, + 234 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 106, + 237, + 505, + 337 + ], + "lines": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 237, + 505, + 251 + ], + "score": 1.0, + "content": "In this paper, we purpose a parameterized action RL algorithm that learns to act on parameterized", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 261 + ], + "score": 1.0, + "content": "action space environments, such as the match plan generation process. We introduce normalized", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 505, + 274 + ], + "score": 1.0, + "content": "softmax values of discrete actions to form a categorical distributions to enable gradients backprop-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 269, + 505, + 285 + ], + "spans": [ + { + "bbox": [ + 105, + 269, + 505, + 285 + ], + "score": 1.0, + "content": "agation. We also combine several recent advances in RL to accelerate and stabilize the training.", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 295 + ], + "score": 1.0, + "content": "We investigate parameter space noise (Plappert et al. (2017)) on parameters of the policy for unify-", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 293, + 505, + 305 + ], + "spans": [ + { + "bbox": [ + 106, + 293, + 505, + 305 + ], + "score": 1.0, + "content": "ing directions in exploring the structured action space. We explore recurrent deterministic policies", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "spans": [ + { + "bbox": [ + 106, + 304, + 505, + 316 + ], + "score": 1.0, + "content": "(Heess et al. (2015)) with prioritized replay buffer (Schaul et al. (2015)) on sequence, due to the", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 313, + 506, + 328 + ], + "spans": [ + { + "bbox": [ + 105, + 313, + 451, + 328 + ], + "score": 1.0, + "content": "inherent nature of partial observability and sparse feedback in our setting. We also use", + "type": "text" + }, + { + "bbox": [ + 451, + 317, + 457, + 325 + ], + "score": 0.49, + "content": "n", + "type": "inline_equation" + }, + { + "bbox": [ + 458, + 313, + 506, + 328 + ], + "score": 1.0, + "content": "-step return", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 325, + 453, + 339 + ], + "spans": [ + { + "bbox": [ + 105, + 325, + 453, + 339 + ], + "score": 1.0, + "content": "and invertible value function rescaling (Kapturowski et al. (2018)) for further stability.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 237, + 506, + 339 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 342, + 505, + 419 + ], + "lines": [ + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "spans": [ + { + "bbox": [ + 106, + 342, + 505, + 355 + ], + "score": 1.0, + "content": "We present an agent to integrate these techniques and evaluate on offline training on a query dataset", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 354, + 506, + 366 + ], + "spans": [ + { + "bbox": [ + 106, + 354, + 506, + 366 + ], + "score": 1.0, + "content": "collected from Bing search. When training on the dataset with uniformly sampling, the agent out-", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 365, + 505, + 376 + ], + "score": 1.0, + "content": "performs the currently deployed production match plan, which is generated based on well-designed", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "spans": [ + { + "bbox": [ + 105, + 375, + 506, + 388 + ], + "score": 1.0, + "content": "hand-crafted rules. We study the performance of training on the dataset or only with several selected", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "spans": [ + { + "bbox": [ + 106, + 386, + 505, + 398 + ], + "score": 1.0, + "content": "complicated queries individually, as well as ablation studies for the various components. We test", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 105, + 397, + 505, + 410 + ], + "score": 1.0, + "content": "on a few existing PARL benchmarks, where our agent beats our baseline methods and performs the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 409, + 199, + 420 + ], + "spans": [ + { + "bbox": [ + 106, + 409, + 199, + 420 + ], + "score": 1.0, + "content": "state-of-the-art results.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 342, + 506, + 420 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 434, + 200, + 448 + ], + "lines": [ + { + "bbox": [ + 104, + 433, + 202, + 451 + ], + "spans": [ + { + "bbox": [ + 104, + 433, + 202, + 451 + ], + "score": 1.0, + "content": "2 BACKGROUND", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 25 + }, + { + "type": "title", + "bbox": [ + 108, + 459, + 254, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 459, + 256, + 472 + ], + "spans": [ + { + "bbox": [ + 106, + 459, + 256, + 472 + ], + "score": 1.0, + "content": "2.1 REINFORCEMENT LEARNING", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 26 + }, + { + "type": "text", + "bbox": [ + 106, + 479, + 505, + 568 + ], + "lines": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 505, + 493 + ], + "score": 1.0, + "content": "We address the problem using reinforcement learning (RL) framework with a parameterized action", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 491, + 506, + 502 + ], + "spans": [ + { + "bbox": [ + 105, + 491, + 506, + 502 + ], + "score": 1.0, + "content": "space. An agent interacts with an environment to maximize the accumulated reward with discount", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 502, + 505, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 502, + 133, + 514 + ], + "score": 1.0, + "content": "factor", + "type": "text" + }, + { + "bbox": [ + 133, + 502, + 177, + 514 + ], + "score": 0.92, + "content": "\\gamma \\in \\ [ 0 , 1 )", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 502, + 505, + 514 + ], + "score": 1.0, + "content": ". We model the environment with a discrete-time Partially Observable Markov", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 511, + 505, + 526 + ], + "spans": [ + { + "bbox": [ + 105, + 511, + 348, + 526 + ], + "score": 1.0, + "content": "Decision Process (POMDP). A POMDP is given by a tuple", + "type": "text" + }, + { + "bbox": [ + 348, + 513, + 423, + 524 + ], + "score": 0.92, + "content": "( S , \\mathcal { A } , T , R , \\Omega , \\dot { \\mathcal { O } } )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 511, + 505, + 526 + ], + "score": 1.0, + "content": ", and the underlying", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 524, + 505, + 536 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 306, + 536 + ], + "score": 1.0, + "content": "Markov Decision Process (MDP) is defined by", + "type": "text" + }, + { + "bbox": [ + 306, + 524, + 356, + 535 + ], + "score": 0.9, + "content": "( S , { \\mathcal { A } } , T , R )", + "type": "inline_equation" + }, + { + "bbox": [ + 356, + 524, + 390, + 536 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 390, + 525, + 398, + 534 + ], + "score": 0.81, + "content": "s", + "type": "inline_equation" + }, + { + "bbox": [ + 399, + 524, + 478, + 536 + ], + "score": 1.0, + "content": "is the state space,", + "type": "text" + }, + { + "bbox": [ + 479, + 524, + 488, + 534 + ], + "score": 0.77, + "content": "\\mathcal { A }", + "type": "inline_equation" + }, + { + "bbox": [ + 488, + 524, + 505, + 536 + ], + "score": 1.0, + "content": "the", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 535, + 504, + 546 + ], + "spans": [ + { + "bbox": [ + 105, + 535, + 221, + 546 + ], + "score": 1.0, + "content": "parameterized action space,", + "type": "text" + }, + { + "bbox": [ + 221, + 535, + 230, + 545 + ], + "score": 0.76, + "content": "T", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 535, + 322, + 546 + ], + "score": 1.0, + "content": "the transition function", + "type": "text" + }, + { + "bbox": [ + 322, + 535, + 415, + 546 + ], + "score": 0.9, + "content": "\\mathcal { T } : \\mathcal { S } \\times \\mathcal { S } \\times \\mathcal { A } \\mathbb { R } _ { + }", + "type": "inline_equation" + }, + { + "bbox": [ + 415, + 535, + 436, + 546 + ], + "score": 1.0, + "content": ", and", + "type": "text" + }, + { + "bbox": [ + 436, + 535, + 504, + 545 + ], + "score": 0.91, + "content": "R : S \\times \\mathcal { A } \\mathbb { R }", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 546, + 506, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 358, + 558 + ], + "score": 1.0, + "content": "is the reward function. The set of observations is given by", + "type": "text" + }, + { + "bbox": [ + 358, + 546, + 366, + 556 + ], + "score": 0.77, + "content": "\\Omega", + "type": "inline_equation" + }, + { + "bbox": [ + 367, + 546, + 506, + 558 + ], + "score": 1.0, + "content": "and the the observation function", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 557, + 454, + 569 + ], + "spans": [ + { + "bbox": [ + 105, + 557, + 441, + 569 + ], + "score": 1.0, + "content": "mapping underlying states to probability distributions over observations is given by", + "type": "text" + }, + { + "bbox": [ + 441, + 557, + 450, + 567 + ], + "score": 0.79, + "content": "\\mathcal { O }", + "type": "inline_equation" + }, + { + "bbox": [ + 451, + 557, + 454, + 569 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 30.5, + "bbox_fs": [ + 105, + 479, + 506, + 569 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 573, + 505, + 606 + ], + "lines": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 505, + 586 + ], + "score": 1.0, + "content": "Specifically, the environment in the match plan generation has a parameterized action space. In", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "spans": [ + { + "bbox": [ + 106, + 585, + 505, + 597 + ], + "score": 1.0, + "content": "Masson et al. (2016); Bester et al. (2019), such formulation in fully observable settings is referred", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 106, + 596, + 423, + 608 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 423, + 608 + ], + "score": 1.0, + "content": "as Parameterized Action MDP (PAMDP), where the action space is denoted as", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 36, + "bbox_fs": [ + 106, + 573, + 505, + 608 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 242, + 609, + 367, + 636 + ], + "lines": [ + { + "bbox": [ + 242, + 609, + 367, + 636 + ], + "spans": [ + { + "bbox": [ + 242, + 609, + 367, + 636 + ], + "score": 0.93, + "content": "\\mathcal { A } = \\bigcup _ { k \\in \\mathcal { A } _ { d } } \\left\\{ ( k , x _ { k } ) | x _ { k } \\in \\mathcal { X } _ { k } \\right\\} ,", + "type": "interline_equation", + "image_path": "76ea0d7a9eb735ac9b64468a264943b9eec3b340c391396f2eb432de507a6511.jpg" + } + ] + } + ], + "index": 38.5, + "virtual_lines": [ + { + "bbox": [ + 242, + 609, + 367, + 622.5 + ], + "spans": [], + "index": 38 + }, + { + "bbox": [ + 242, + 622.5, + 367, + 636.0 + ], + "spans": [], + "index": 39 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 637, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 106, + 636, + 505, + 651 + ], + "spans": [ + { + "bbox": [ + 106, + 636, + 216, + 651 + ], + "score": 1.0, + "content": "where each discrete action", + "type": "text" + }, + { + "bbox": [ + 216, + 637, + 278, + 649 + ], + "score": 0.93, + "content": "a \\in \\mathcal { A } _ { d } = [ K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 636, + 505, + 651 + ], + "score": 1.0, + "content": "has a corresponding continuous action-parameter space", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 107, + 649, + 371, + 661 + ], + "spans": [ + { + "bbox": [ + 107, + 650, + 119, + 660 + ], + "score": 0.86, + "content": "\\mathcal { X } _ { a }", + "type": "inline_equation" + }, + { + "bbox": [ + 120, + 649, + 371, + 661 + ], + "score": 1.0, + "content": ". However, our setting requires a slightly different formulation:", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 636, + 505, + 661 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 220, + 662, + 389, + 675 + ], + "lines": [ + { + "bbox": [ + 220, + 662, + 389, + 675 + ], + "spans": [ + { + "bbox": [ + 220, + 662, + 389, + 675 + ], + "score": 0.92, + "content": "\\mathcal { A } = \\{ ( k , x ) | k \\in \\mathcal { A } _ { d } , x \\in \\mathcal { X } \\} = \\mathcal { A } _ { d } \\times \\mathcal { X } ,", + "type": "interline_equation", + "image_path": "243001bdfb8447abcdd89869de2d2453c03e6261b0ef47f965d0d6494f259a02.jpg" + } + ] + } + ], + "index": 42, + "virtual_lines": [ + { + "bbox": [ + 220, + 662, + 389, + 675 + ], + "spans": [], + "index": 42 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 677, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 677, + 270, + 690 + ], + "score": 1.0, + "content": "where the discrete action space (rewrite)", + "type": "text" + }, + { + "bbox": [ + 271, + 677, + 284, + 688 + ], + "score": 0.88, + "content": "\\mathbf { \\mathcal { A } } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 285, + 677, + 418, + 690 + ], + "score": 1.0, + "content": "share the action-parameter space", + "type": "text" + }, + { + "bbox": [ + 419, + 677, + 428, + 687 + ], + "score": 0.78, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 429, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ". Such formulation", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "results in a disentangled action space between discrete and continuous actions which a class of", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "parameterized action (P-DQN (Xiong et al. (2018)) and MP-DQN (Bester et al. (2019))) may not", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 505, + 722 + ], + "score": 1.0, + "content": "be trivially applicable to. Considering the notational simplicity in partially observable environment", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 720, + 454, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 454, + 734 + ], + "score": 1.0, + "content": "(POMDP), we do not explicitly denote it when there is no ambiguity in action settings.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 45, + "bbox_fs": [ + 105, + 677, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 330, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 331, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 331, + 95 + ], + "score": 1.0, + "content": "2.2 FORMULATION OF MATCH PLAN GENERATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "In Rosset et al. (2018), the action-parameters are predefined for the match plan generation, where", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "the action space is then a set of discrete rewrites (actions). We fully parameterize a match plan to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 125, + 281, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 281, + 137 + ], + "score": 1.0, + "content": "allow the agent to generate any valid plans.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2 + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "State. At each time step, the agent receives a state with two parts. First part has selected run-time", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "system signals from the inverted index system. Since the query is uniformly sampled in different", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "episode, we include some statistical features and semantic embeddings of queries to allow the agent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "score": 1.0, + "content": "to identify different queries and then generate corresponding match plans. Such statistical features", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "are used in the current production system, such as the length of a query. All signals and embeddings", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 203, + 444, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 285, + 216 + ], + "score": 1.0, + "content": "are normalized to a reasonable range around", + "type": "text" + }, + { + "bbox": [ + 286, + 204, + 314, + 215 + ], + "score": 0.89, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 203, + 444, + 216 + ], + "score": 1.0, + "content": "based on empirical estimations.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 310, + 240 + ], + "score": 1.0, + "content": "Action. The agent selects a parameterized action", + "type": "text" + }, + { + "bbox": [ + 310, + 229, + 320, + 238 + ], + "score": 0.86, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 227, + 505, + 240 + ], + "score": 1.0, + "content": "including a rewrite id and its allocated quotas", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 294, + 251 + ], + "score": 1.0, + "content": "at each step. There are 29 types of rewrites in", + "type": "text" + }, + { + "bbox": [ + 294, + 239, + 308, + 249 + ], + "score": 0.89, + "content": "\\mathbf { \\mathcal { A } } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "and also a 5-dimensional quota parameter space", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 250, + 116, + 259 + ], + "score": 0.8, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 248, + 505, + 262 + ], + "score": 1.0, + "content": "for each rewrite in int64 type. However, the range of each quota is empirically set to valid values", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 260, + 450, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 200, + 273 + ], + "score": 1.0, + "content": "and then normalized to", + "type": "text" + }, + { + "bbox": [ + 201, + 260, + 229, + 272 + ], + "score": 0.86, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 260, + 450, + 273 + ], + "score": 1.0, + "content": ". All outputted quotas are effective for the current step.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5 + }, + { + "type": "text", + "bbox": [ + 107, + 277, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 507, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 507, + 290 + ], + "score": 1.0, + "content": "There is also a special action to note: stop (by agent). Normally, the environment will return a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "score": 1.0, + "content": "terminal signal, but it happens in extreme cases, such as low system resources. To allow a better", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "balance of latency and performance, we allow the agent to choose stop or not as another type of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 310, + 348, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 348, + 322 + ], + "score": 1.0, + "content": "discrete action, which is the same level as other 29 rewrites.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5 + }, + { + "type": "text", + "bbox": [ + 107, + 333, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "Reward. We use two criterions in the design: latency and performance. For latency consideration,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "we use a signal Seek Count that is constant for a same query in different runs with same match plan,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "instead of executed time. We weight the Ranking Scores of top five returned documents as an indi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 367, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 378 + ], + "score": 1.0, + "content": "cator of performance with weights [0.4, 0.2, 0.2, 0.1, 0.1]. When there are less than five documents", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 378, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "returned, the scores of missing documents are treated as a minimum possible ranking score. The", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "score": 1.0, + "content": "final scalar reward is weighted by these two objectives to balance them. We do not use the division", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 399, + 477, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 477, + 413 + ], + "score": 1.0, + "content": "in Rosset et al. (2018), since the loss surfaces may become more non-convex for optimizing.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "We also have punishments (negative rewards) for some special cases. One is to assign a punish-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "ment when the agent selects to illegally stop at the first step or some unsupported rewrites (for some", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "special queries). However, unsupported rewrites do not necessarily cause a terminal state, since the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "production system has same hand-crafted match plans for a class of queries, in which the unsup-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 460, + 248, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 248, + 473 + ], + "score": 1.0, + "content": "ported rewrites are simply omitted.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 487, + 173, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 174, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 174, + 504 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "For generating match plans, we introduce a method that works on the aforementioned formulation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "with slightly different parameterized action space than standard PAMDPs in the literature. We start", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "score": 1.0, + "content": "from the intuition with deterministic policy learning (Silver et al. (2014)) and how it relates to pa-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "rameterized actions. PAMDP often refers to the formulation that each individual discrete action", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 172, + 569 + ], + "score": 0.92, + "content": "k \\in \\mathcal { A } _ { d } = [ K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 556, + 472, + 570 + ], + "score": 1.0, + "content": "has a corresponding continuous action-parameter space. Thus, there are", + "type": "text" + }, + { + "bbox": [ + 473, + 557, + 483, + 567 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "con-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 310, + 580 + ], + "score": 1.0, + "content": "tinuous action-parameter spaces corresponding to", + "type": "text" + }, + { + "bbox": [ + 311, + 568, + 321, + 578 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "discrete action (Xiong et al. (2018)). In our", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 434, + 592 + ], + "score": 1.0, + "content": "shared-parameter PAMDP, the Bellman equation incorporated both discrete action", + "type": "text" + }, + { + "bbox": [ + 434, + 580, + 441, + 589 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "and continuous", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 589, + 237, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 180, + 603 + ], + "score": 1.0, + "content": "action-parameters", + "type": "text" + }, + { + "bbox": [ + 181, + 592, + 187, + 600 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 589, + 237, + 603 + ], + "score": 1.0, + "content": "is given by:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5 + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 614, + 421, + 642 + ], + "lines": [ + { + "bbox": [ + 189, + 614, + 421, + 642 + ], + "spans": [ + { + "bbox": [ + 189, + 614, + 421, + 642 + ], + "score": 0.93, + "content": "\\begin{array} { r } { Q \\left( s , k , x \\right) = \\underset { r , s ^ { \\prime } } { \\mathbb { E } } \\left[ r + \\gamma \\underset { k ^ { \\prime } } { \\operatorname* { m a x } } \\ \\underset { x ^ { \\prime } \\in \\mathcal { X } } { \\operatorname* { s u p } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x ^ { \\prime } \\right) \\vert s , k , x \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "4baf652d557a9988b78e9d6b9f883817be7d2229b913710040c6d0deac8aada0.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 614, + 421, + 628.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 189, + 628.0, + 421, + 642.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 649, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "P-DQN (Xiong et al. (2018)) tackles PAMDP by incorporating multiple action-parameter policies", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 659, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 107, + 661, + 200, + 673 + ], + "score": 0.9, + "content": "x _ { k } ( s ; \\theta _ { x } ) : { \\cal { S } } { \\mathcal { X } } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 659, + 268, + 677 + ], + "score": 1.0, + "content": "for each action", + "type": "text" + }, + { + "bbox": [ + 269, + 662, + 276, + 671 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 659, + 372, + 677 + ], + "score": 1.0, + "content": "and update them with", + "type": "text" + }, + { + "bbox": [ + 373, + 661, + 504, + 675 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\operatorname* { m a x } _ { k ^ { \\prime } } \\operatorname* { s u p } _ { x _ { k ^ { \\prime } } \\in \\mathcal { X } _ { k ^ { \\prime } } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x _ { k ^ { \\prime } } \\right) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 300, + 685 + ], + "score": 1.0, + "content": "simultaneously. However, since all parameters", + "type": "text" + }, + { + "bbox": [ + 300, + 676, + 307, + 683 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 673, + 401, + 685 + ], + "score": 1.0, + "content": "are shared for actions", + "type": "text" + }, + { + "bbox": [ + 401, + 674, + 408, + 683 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "in our case, we do not", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 685, + 437, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 437, + 696 + ], + "score": 1.0, + "content": "require multiple policies. The policy loss in P-DQN (Xiong et al. (2018)) given by", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5 + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 701, + 402, + 735 + ], + "lines": [ + { + "bbox": [ + 208, + 701, + 402, + 735 + ], + "spans": [ + { + "bbox": [ + 208, + 701, + 402, + 735 + ], + "score": 0.94, + "content": "L _ { x } \\left( \\theta _ { x } \\right) = \\underset { s \\sim D } { \\mathbb { E } } \\left[ - \\sum _ { k = 1 } ^ { K } Q \\left( s , k , x _ { k } \\left( s ; \\theta _ { x } \\right) ; \\theta _ { Q } \\right) \\right]", + "type": "interline_equation", + "image_path": "38a0ce4a63eabe51e761e09dd8728767079ede8cf87f9111a5e7ab29f0c3a281.jpg" + } + ] + } + ], + "index": 45.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 701, + 402, + 718.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 208, + 718.0, + 402, + 735.0 + ], + "spans": [], + "index": 46 + } + ] + } + ], + "page_idx": 2, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 760 + ], + "lines": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "spans": [ + { + "bbox": [ + 301, + 750, + 310, + 762 + ], + "score": 1.0, + "content": "3", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 106, + 82, + 330, + 94 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 331, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 331, + 95 + ], + "score": 1.0, + "content": "2.2 FORMULATION OF MATCH PLAN GENERATION", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 108, + 102, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "spans": [ + { + "bbox": [ + 105, + 102, + 505, + 115 + ], + "score": 1.0, + "content": "In Rosset et al. (2018), the action-parameters are predefined for the match plan generation, where", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "spans": [ + { + "bbox": [ + 106, + 114, + 505, + 126 + ], + "score": 1.0, + "content": "the action space is then a set of discrete rewrites (actions). We fully parameterize a match plan to", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 125, + 281, + 137 + ], + "spans": [ + { + "bbox": [ + 106, + 125, + 281, + 137 + ], + "score": 1.0, + "content": "allow the agent to generate any valid plans.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2, + "bbox_fs": [ + 105, + 102, + 505, + 137 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 149, + 505, + 215 + ], + "lines": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "spans": [ + { + "bbox": [ + 106, + 149, + 505, + 161 + ], + "score": 1.0, + "content": "State. At each time step, the agent receives a state with two parts. First part has selected run-time", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 159, + 506, + 173 + ], + "score": 1.0, + "content": "system signals from the inverted index system. Since the query is uniformly sampled in different", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "spans": [ + { + "bbox": [ + 105, + 171, + 506, + 183 + ], + "score": 1.0, + "content": "episode, we include some statistical features and semantic embeddings of queries to allow the agent", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 506, + 194 + ], + "score": 1.0, + "content": "to identify different queries and then generate corresponding match plans. Such statistical features", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 505, + 206 + ], + "score": 1.0, + "content": "are used in the current production system, such as the length of a query. All signals and embeddings", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 203, + 444, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 203, + 285, + 216 + ], + "score": 1.0, + "content": "are normalized to a reasonable range around", + "type": "text" + }, + { + "bbox": [ + 286, + 204, + 314, + 215 + ], + "score": 0.89, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 315, + 203, + 444, + 216 + ], + "score": 1.0, + "content": "based on empirical estimations.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 149, + 506, + 216 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 227, + 505, + 272 + ], + "lines": [ + { + "bbox": [ + 105, + 227, + 505, + 240 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 310, + 240 + ], + "score": 1.0, + "content": "Action. The agent selects a parameterized action", + "type": "text" + }, + { + "bbox": [ + 310, + 229, + 320, + 238 + ], + "score": 0.86, + "content": "a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 321, + 227, + 505, + 240 + ], + "score": 1.0, + "content": "including a rewrite id and its allocated quotas", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 238, + 505, + 251 + ], + "spans": [ + { + "bbox": [ + 105, + 238, + 294, + 251 + ], + "score": 1.0, + "content": "at each step. There are 29 types of rewrites in", + "type": "text" + }, + { + "bbox": [ + 294, + 239, + 308, + 249 + ], + "score": 0.89, + "content": "\\mathbf { \\mathcal { A } } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 238, + 505, + 251 + ], + "score": 1.0, + "content": "and also a 5-dimensional quota parameter space", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 250, + 116, + 259 + ], + "score": 0.8, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 117, + 248, + 505, + 262 + ], + "score": 1.0, + "content": "for each rewrite in int64 type. However, the range of each quota is empirically set to valid values", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 260, + 450, + 273 + ], + "spans": [ + { + "bbox": [ + 105, + 260, + 200, + 273 + ], + "score": 1.0, + "content": "and then normalized to", + "type": "text" + }, + { + "bbox": [ + 201, + 260, + 229, + 272 + ], + "score": 0.86, + "content": "[ - 1 , 1 ]", + "type": "inline_equation" + }, + { + "bbox": [ + 230, + 260, + 450, + 273 + ], + "score": 1.0, + "content": ". All outputted quotas are effective for the current step.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11.5, + "bbox_fs": [ + 105, + 227, + 505, + 273 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 277, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 105, + 276, + 507, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 276, + 507, + 290 + ], + "score": 1.0, + "content": "There is also a special action to note: stop (by agent). Normally, the environment will return a", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 105, + 288, + 506, + 301 + ], + "score": 1.0, + "content": "terminal signal, but it happens in extreme cases, such as low system resources. To allow a better", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "spans": [ + { + "bbox": [ + 105, + 299, + 506, + 312 + ], + "score": 1.0, + "content": "balance of latency and performance, we allow the agent to choose stop or not as another type of", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 310, + 348, + 322 + ], + "spans": [ + { + "bbox": [ + 106, + 310, + 348, + 322 + ], + "score": 1.0, + "content": "discrete action, which is the same level as other 29 rewrites.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 15.5, + "bbox_fs": [ + 105, + 276, + 507, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 333, + 505, + 411 + ], + "lines": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "spans": [ + { + "bbox": [ + 105, + 333, + 505, + 347 + ], + "score": 1.0, + "content": "Reward. We use two criterions in the design: latency and performance. For latency consideration,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "spans": [ + { + "bbox": [ + 106, + 345, + 505, + 357 + ], + "score": 1.0, + "content": "we use a signal Seek Count that is constant for a same query in different runs with same match plan,", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "spans": [ + { + "bbox": [ + 105, + 355, + 505, + 368 + ], + "score": 1.0, + "content": "instead of executed time. We weight the Ranking Scores of top five returned documents as an indi-", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 367, + 505, + 378 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 378 + ], + "score": 1.0, + "content": "cator of performance with weights [0.4, 0.2, 0.2, 0.1, 0.1]. When there are less than five documents", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 378, + 505, + 389 + ], + "spans": [ + { + "bbox": [ + 106, + 378, + 505, + 389 + ], + "score": 1.0, + "content": "returned, the scores of missing documents are treated as a minimum possible ranking score. The", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "spans": [ + { + "bbox": [ + 106, + 389, + 505, + 400 + ], + "score": 1.0, + "content": "final scalar reward is weighted by these two objectives to balance them. We do not use the division", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 399, + 477, + 413 + ], + "spans": [ + { + "bbox": [ + 105, + 399, + 477, + 413 + ], + "score": 1.0, + "content": "in Rosset et al. (2018), since the loss surfaces may become more non-convex for optimizing.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 21, + "bbox_fs": [ + 105, + 333, + 505, + 413 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 416, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "spans": [ + { + "bbox": [ + 106, + 416, + 505, + 429 + ], + "score": 1.0, + "content": "We also have punishments (negative rewards) for some special cases. One is to assign a punish-", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "spans": [ + { + "bbox": [ + 106, + 428, + 505, + 440 + ], + "score": 1.0, + "content": "ment when the agent selects to illegally stop at the first step or some unsupported rewrites (for some", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 439, + 505, + 451 + ], + "score": 1.0, + "content": "special queries). However, unsupported rewrites do not necessarily cause a terminal state, since the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "spans": [ + { + "bbox": [ + 105, + 448, + 505, + 464 + ], + "score": 1.0, + "content": "production system has same hand-crafted match plans for a class of queries, in which the unsup-", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 460, + 248, + 473 + ], + "spans": [ + { + "bbox": [ + 105, + 460, + 248, + 473 + ], + "score": 1.0, + "content": "ported rewrites are simply omitted.", + "type": "text" + } + ], + "index": 29 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 416, + 505, + 473 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 487, + 173, + 500 + ], + "lines": [ + { + "bbox": [ + 105, + 486, + 174, + 504 + ], + "spans": [ + { + "bbox": [ + 105, + 486, + 174, + 504 + ], + "score": 1.0, + "content": "3 METHOD", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 30 + }, + { + "type": "text", + "bbox": [ + 106, + 512, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "spans": [ + { + "bbox": [ + 105, + 513, + 505, + 525 + ], + "score": 1.0, + "content": "For generating match plans, we introduce a method that works on the aforementioned formulation", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 523, + 506, + 537 + ], + "score": 1.0, + "content": "with slightly different parameterized action space than standard PAMDPs in the literature. We start", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "spans": [ + { + "bbox": [ + 105, + 534, + 505, + 548 + ], + "score": 1.0, + "content": "from the intuition with deterministic policy learning (Silver et al. (2014)) and how it relates to pa-", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 546, + 505, + 558 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 558 + ], + "score": 1.0, + "content": "rameterized actions. PAMDP often refers to the formulation that each individual discrete action", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 556, + 506, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 172, + 569 + ], + "score": 0.92, + "content": "k \\in \\mathcal { A } _ { d } = [ K ]", + "type": "inline_equation" + }, + { + "bbox": [ + 173, + 556, + 472, + 570 + ], + "score": 1.0, + "content": "has a corresponding continuous action-parameter space. Thus, there are", + "type": "text" + }, + { + "bbox": [ + 473, + 557, + 483, + 567 + ], + "score": 0.8, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 483, + 556, + 506, + 570 + ], + "score": 1.0, + "content": "con-", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 568, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 310, + 580 + ], + "score": 1.0, + "content": "tinuous action-parameter spaces corresponding to", + "type": "text" + }, + { + "bbox": [ + 311, + 568, + 321, + 578 + ], + "score": 0.81, + "content": "K", + "type": "inline_equation" + }, + { + "bbox": [ + 322, + 568, + 505, + 580 + ], + "score": 1.0, + "content": "discrete action (Xiong et al. (2018)). In our", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 578, + 506, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 578, + 434, + 592 + ], + "score": 1.0, + "content": "shared-parameter PAMDP, the Bellman equation incorporated both discrete action", + "type": "text" + }, + { + "bbox": [ + 434, + 580, + 441, + 589 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 578, + 506, + 592 + ], + "score": 1.0, + "content": "and continuous", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 589, + 237, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 589, + 180, + 603 + ], + "score": 1.0, + "content": "action-parameters", + "type": "text" + }, + { + "bbox": [ + 181, + 592, + 187, + 600 + ], + "score": 0.78, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 187, + 589, + 237, + 603 + ], + "score": 1.0, + "content": "is given by:", + "type": "text" + } + ], + "index": 38 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 513, + 506, + 603 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 189, + 614, + 421, + 642 + ], + "lines": [ + { + "bbox": [ + 189, + 614, + 421, + 642 + ], + "spans": [ + { + "bbox": [ + 189, + 614, + 421, + 642 + ], + "score": 0.93, + "content": "\\begin{array} { r } { Q \\left( s , k , x \\right) = \\underset { r , s ^ { \\prime } } { \\mathbb { E } } \\left[ r + \\gamma \\underset { k ^ { \\prime } } { \\operatorname* { m a x } } \\ \\underset { x ^ { \\prime } \\in \\mathcal { X } } { \\operatorname* { s u p } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x ^ { \\prime } \\right) \\vert s , k , x \\right] . } \\end{array}", + "type": "interline_equation", + "image_path": "4baf652d557a9988b78e9d6b9f883817be7d2229b913710040c6d0deac8aada0.jpg" + } + ] + } + ], + "index": 39.5, + "virtual_lines": [ + { + "bbox": [ + 189, + 614, + 421, + 628.0 + ], + "spans": [], + "index": 39 + }, + { + "bbox": [ + 189, + 628.0, + 421, + 642.0 + ], + "spans": [], + "index": 40 + } + ] + }, + { + "type": "text", + "bbox": [ + 106, + 649, + 505, + 696 + ], + "lines": [ + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 505, + 663 + ], + "score": 1.0, + "content": "P-DQN (Xiong et al. (2018)) tackles PAMDP by incorporating multiple action-parameter policies", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 107, + 659, + 504, + 677 + ], + "spans": [ + { + "bbox": [ + 107, + 661, + 200, + 673 + ], + "score": 0.9, + "content": "x _ { k } ( s ; \\theta _ { x } ) : { \\cal { S } } { \\mathcal { X } } _ { k }", + "type": "inline_equation" + }, + { + "bbox": [ + 200, + 659, + 268, + 677 + ], + "score": 1.0, + "content": "for each action", + "type": "text" + }, + { + "bbox": [ + 269, + 662, + 276, + 671 + ], + "score": 0.82, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 276, + 659, + 372, + 677 + ], + "score": 1.0, + "content": "and update them with", + "type": "text" + }, + { + "bbox": [ + 373, + 661, + 504, + 675 + ], + "score": 0.88, + "content": "\\begin{array} { r } { \\operatorname* { m a x } _ { k ^ { \\prime } } \\operatorname* { s u p } _ { x _ { k ^ { \\prime } } \\in \\mathcal { X } _ { k ^ { \\prime } } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x _ { k ^ { \\prime } } \\right) } \\end{array}", + "type": "inline_equation" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 673, + 505, + 685 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 300, + 685 + ], + "score": 1.0, + "content": "simultaneously. However, since all parameters", + "type": "text" + }, + { + "bbox": [ + 300, + 676, + 307, + 683 + ], + "score": 0.76, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 308, + 673, + 401, + 685 + ], + "score": 1.0, + "content": "are shared for actions", + "type": "text" + }, + { + "bbox": [ + 401, + 674, + 408, + 683 + ], + "score": 0.78, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 408, + 673, + 505, + 685 + ], + "score": 1.0, + "content": "in our case, we do not", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 685, + 437, + 696 + ], + "spans": [ + { + "bbox": [ + 106, + 685, + 437, + 696 + ], + "score": 1.0, + "content": "require multiple policies. The policy loss in P-DQN (Xiong et al. (2018)) given by", + "type": "text" + } + ], + "index": 44 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 649, + 505, + 696 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 208, + 701, + 402, + 735 + ], + "lines": [ + { + "bbox": [ + 208, + 701, + 402, + 735 + ], + "spans": [ + { + "bbox": [ + 208, + 701, + 402, + 735 + ], + "score": 0.94, + "content": "L _ { x } \\left( \\theta _ { x } \\right) = \\underset { s \\sim D } { \\mathbb { E } } \\left[ - \\sum _ { k = 1 } ^ { K } Q \\left( s , k , x _ { k } \\left( s ; \\theta _ { x } \\right) ; \\theta _ { Q } \\right) \\right]", + "type": "interline_equation", + "image_path": "38a0ce4a63eabe51e761e09dd8728767079ede8cf87f9111a5e7ab29f0c3a281.jpg" + } + ] + } + ], + "index": 45.5, + "virtual_lines": [ + { + "bbox": [ + 208, + 701, + 402, + 718.0 + ], + "spans": [], + "index": 45 + }, + { + "bbox": [ + 208, + 718.0, + 402, + 735.0 + ], + "spans": [], + "index": 46 + } + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 174, + 78, + 431, + 130 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 174, + 78, + 431, + 130 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 78, + 431, + 130 + ], + "spans": [ + { + "bbox": [ + 174, + 78, + 431, + 130 + ], + "score": 0.965, + "type": "image", + "image_path": "908f0f5be906ce40a46112737a183c610852cc851433f4f94d61674bb007f08e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 174, + 78, + 431, + 95.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 174, + 95.33333333333333, + 431, + 112.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 174, + 112.66666666666666, + 431, + 130.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 235, + 143, + 375, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 234, + 141, + 377, + 156 + ], + "spans": [ + { + "bbox": [ + 234, + 141, + 377, + 156 + ], + "score": 1.0, + "content": "Figure 2: The diagram of the process.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 405, + 194 + ], + "score": 1.0, + "content": "naturally degenerates to a single policy in next case, where the gradient", + "type": "text" + }, + { + "bbox": [ + 406, + 181, + 446, + 193 + ], + "score": 0.92, + "content": "\\nabla L _ { x } \\left( \\theta _ { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "exactly is the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "deterministic policy gradient (DPG, Silver et al. (2014)). MP-DQN (Bester et al. (2019)) states a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 354, + 216 + ], + "score": 1.0, + "content": "problem regarding to the erroneous gradients of Q-network", + "type": "text" + }, + { + "bbox": [ + 354, + 204, + 377, + 216 + ], + "score": 0.91, + "content": "\\nabla L _ { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "in P-DQN caused by multiple", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "action-parameter policies backpropagating gradients at the same time. However, it does not exist", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "in such single network situation. This is a desired architecture since separating multiple action-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 236, + 430, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 430, + 250 + ], + "score": 1.0, + "content": "parameter policies loses the knowledge that they share same potential meanings.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5 + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 104, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 410, + 266 + ], + "score": 1.0, + "content": "Following this intuition, we directly parameterize the policy with two heads", + "type": "text" + }, + { + "bbox": [ + 411, + 252, + 483, + 265 + ], + "score": 0.92, + "content": "\\mu : \\mathcal { S } \\to \\mathbb { R } ^ { K } \\times \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 260, + 507, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 118, + 278 + ], + "score": 0.83, + "content": "\\theta _ { \\mu }", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 264, + 258, + 278 + ], + "score": 0.91, + "content": "\\mathrm { ~ ` ~ } _ { \\mu } \\mathrm { ~ a s : ~ } k _ { t } ^ { s o f t } , x _ { t } = a _ { t } ^ { s o f t } = \\mu ( s _ { t } ; \\theta _ { \\mu } )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 260, + 288, + 283 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 289, + 264, + 311, + 278 + ], + "score": 0.92, + "content": "k _ { t } ^ { s o f t }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 260, + 329, + 283 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 330, + 264, + 352, + 278 + ], + "score": 0.93, + "content": "a _ { t } ^ { s o f t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 260, + 507, + 283 + ], + "score": 1.0, + "content": "refers to action-parameters or actions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "score": 1.0, + "content": "with softmax values. One head outputs the sof tmax values for all discrete actions. For another", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 413, + 300 + ], + "score": 1.0, + "content": "head, it outputs normalized action-parameter values with hyperbolic tangent", + "type": "text" + }, + { + "bbox": [ + 414, + 288, + 442, + 299 + ], + "score": 0.4, + "content": "( t a n h )", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "activation after", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "the continuous head. The environment will denormalize it to a valid range. This corresponds to the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "Squashing Gradients method for bounded-continuous action spaces in Hausknecht & Stone (2015).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 106, + 341, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "Exploration with Parameter Space Noise and action sampling. In DQN (Mnih et al. (2015)),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 236, + 366 + ], + "score": 1.0, + "content": "the behavioral policy is given by", + "type": "text" + }, + { + "bbox": [ + 236, + 355, + 241, + 362 + ], + "score": 0.62, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 352, + 505, + 366 + ], + "score": 1.0, + "content": "-greedy strategy for exploration purpose, where an agent randomly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 274, + 376 + ], + "score": 1.0, + "content": "selects a discrete action with probability", + "type": "text" + }, + { + "bbox": [ + 275, + 364, + 302, + 374 + ], + "score": 0.89, + "content": "\\epsilon < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 363, + 505, + 376 + ], + "score": 1.0, + "content": ". In DDPG (Lillicrap et al. (2015)), although the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "policy learned is deterministic, to explore the action space, the behavioral policy needs to be different", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "because of the off-policy nature. It uses action space noise such as uncorrelated Gaussian noise or", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 397, + 361, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 361, + 408 + ], + "score": 1.0, + "content": "correlated Ornstein-Uhlenbeck process (Lillicrap et al. (2015)).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 106, + 413, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "Directly combining such two exploration strategies is straightforward. However, it may induce a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 310, + 437 + ], + "score": 1.0, + "content": "problem that the exploration in two action spaces", + "type": "text" + }, + { + "bbox": [ + 311, + 424, + 370, + 435 + ], + "score": 0.92, + "content": "\\mathcal { A } = \\mathcal { A } _ { d } \\times \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "in different paces. For example,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "the match plan may need another rewrite with larger quotas in a step. However, if the exploration", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 175, + 459 + ], + "score": 1.0, + "content": "strategy is to use", + "type": "text" + }, + { + "bbox": [ + 176, + 448, + 181, + 456 + ], + "score": 0.66, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 446, + 263, + 459 + ], + "score": 1.0, + "content": "-greedy for a rewrite", + "type": "text" + }, + { + "bbox": [ + 264, + 447, + 271, + 456 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 446, + 405, + 459 + ], + "score": 1.0, + "content": "and Gaussian noise for its quotas", + "type": "text" + }, + { + "bbox": [ + 405, + 448, + 412, + 456 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 446, + 505, + 459 + ], + "score": 1.0, + "content": ", the agent may require", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "more samples to discover the potential rewards (positive documents). We instead use parameter", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 468, + 454, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 454, + 481 + ], + "score": 1.0, + "content": "space noise (Plappert et al. (2017)) on parameterized action space to tackle such issue.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5 + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 103, + 481, + 508, + 505 + ], + "spans": [ + { + "bbox": [ + 103, + 481, + 330, + 505 + ], + "score": 1.0, + "content": "We denote the discrete and continuous action heads as", + "type": "text" + }, + { + "bbox": [ + 331, + 485, + 450, + 499 + ], + "score": 0.92, + "content": "( \\pi _ { k } ^ { s o f t } ( s ) , \\pi _ { x } ( s ) ) = \\mu ( s ; \\theta _ { \\mu } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 481, + 508, + 505 + ], + "score": 1.0, + "content": ". To compute", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 497, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 158, + 512 + ], + "score": 1.0, + "content": "the distance", + "type": "text" + }, + { + "bbox": [ + 158, + 498, + 248, + 510 + ], + "score": 0.92, + "content": "d ( \\pi , \\widetilde { \\pi } ) = D _ { \\mathrm { K L } } ( \\pi \\| \\widetilde { \\pi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 497, + 415, + 512 + ], + "score": 1.0, + "content": "of non-perturbed and perturbed policies", + "type": "text" + }, + { + "bbox": [ + 415, + 498, + 486, + 510 + ], + "score": 0.92, + "content": "\\mu ( s ; \\theta _ { \\mu } ) , \\widetilde { \\mu } ( s ; \\theta _ { \\mu } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 497, + 505, + 512 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "e e euse weighted sum of the distance of discrete and continuous actions. For the continuous actions", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 519, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 131, + 536 + ], + "score": 0.91, + "content": "\\widetilde { \\pi } _ { x } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 522, + 229, + 538 + ], + "score": 1.0, + "content": ", the distance is given by", + "type": "text" + }, + { + "bbox": [ + 230, + 519, + 329, + 540 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { s } \\left[ \\left( \\pi _ { x } ( s ) _ { i } - \\widetilde { \\pi } _ { x } ( s ) _ { i } \\right) ^ { 2 } \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 522, + 506, + 538 + ], + "score": 1.0, + "content": "to estimate KL-divergence empirically. For", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 101, + 529, + 501, + 565 + ], + "spans": [ + { + "bbox": [ + 101, + 529, + 399, + 565 + ], + "score": 1.0, + "content": "e ethe discrete actions, we use outputted softmax probabilities to compute The state and action pairs are sampled from a replay memory. The varianc", + "type": "text" + }, + { + "bbox": [ + 399, + 538, + 501, + 553 + ], + "score": 0.93, + "content": "D _ { \\mathrm { K L } } ( \\pi _ { k } ^ { s o f t } ( s ) \\Vert \\widetilde { \\pi } _ { k } ^ { s o f t } ( s ) )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 406, + 554, + 413, + 561 + ], + "spans": [ + { + "bbox": [ + 406, + 554, + 413, + 561 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 335, + 575 + ], + "score": 1.0, + "content": "after a policy update based on the distance and threhold", + "type": "text" + }, + { + "bbox": [ + 335, + 563, + 342, + 572 + ], + "score": 0.73, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "(Plappert et al. (2017)). The policy and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 573, + 448, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 448, + 586 + ], + "score": 1.0, + "content": "target networks with layer normalization (Ba et al. (2016)) are perturbed per episode.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5 + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Prioritized replay and recurrent policies. We use recurrent architecture to obtain the underlying", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "system state of the POMDP for both value and policy networks (Heess et al. (2015)). To avoid re-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 624, + 504, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 445, + 641 + ], + "score": 1.0, + "content": "current state staleness (Kapturowski et al. (2018)), we store and replay a sequence of", + "type": "text" + }, + { + "bbox": [ + 445, + 626, + 504, + 639 + ], + "score": 0.91, + "content": "( s , k ^ { s o f t } , x , r )", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "with softmax values into a replay memory. Since a match plan is usually short, we set a maximum", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 649, + 437, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 437, + 661 + ], + "score": 1.0, + "content": "length and pad shorter episodes with zero or randomly sampled states and actions.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 410, + 678 + ], + "score": 1.0, + "content": "The sampling from a replay memory can be prioritized with a probability", + "type": "text" + }, + { + "bbox": [ + 411, + 668, + 420, + 677 + ], + "score": 0.83, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "proportional to TD-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "errors (Schaul et al. (2015)) to increase the sample efficiency in such structured action space. To", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "prioritize the transitions with well-matched documents, we slightly modify the sampling strategy.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "With a half probability, the agent uses regular prioritization, otherwise partitions the memory to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 708, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 724 + ], + "score": 1.0, + "content": "multiple bins and retrieve a transition (sequence) with max reward from each bin. Our strategy", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 721, + 415, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 415, + 733 + ], + "score": 1.0, + "content": "provides a more efficient and balanced exploration strategy of the evaluation.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5 + } + ], + "page_idx": 3, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 106, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 759 + ], + "lines": [] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 174, + 78, + 431, + 130 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 174, + 78, + 431, + 130 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 174, + 78, + 431, + 130 + ], + "spans": [ + { + "bbox": [ + 174, + 78, + 431, + 130 + ], + "score": 0.965, + "type": "image", + "image_path": "908f0f5be906ce40a46112737a183c610852cc851433f4f94d61674bb007f08e.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 174, + 78, + 431, + 95.33333333333333 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 174, + 95.33333333333333, + 431, + 112.66666666666666 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 174, + 112.66666666666666, + 431, + 130.0 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 235, + 143, + 375, + 154 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 234, + 141, + 377, + 156 + ], + "spans": [ + { + "bbox": [ + 234, + 141, + 377, + 156 + ], + "score": 1.0, + "content": "Figure 2: The diagram of the process.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 3 + } + ], + "index": 2.0 + }, + { + "type": "text", + "bbox": [ + 107, + 181, + 505, + 248 + ], + "lines": [ + { + "bbox": [ + 105, + 181, + 505, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 181, + 405, + 194 + ], + "score": 1.0, + "content": "naturally degenerates to a single policy in next case, where the gradient", + "type": "text" + }, + { + "bbox": [ + 406, + 181, + 446, + 193 + ], + "score": 0.92, + "content": "\\nabla L _ { x } \\left( \\theta _ { x } \\right)", + "type": "inline_equation" + }, + { + "bbox": [ + 446, + 181, + 505, + 194 + ], + "score": 1.0, + "content": "exactly is the", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "spans": [ + { + "bbox": [ + 105, + 192, + 506, + 205 + ], + "score": 1.0, + "content": "deterministic policy gradient (DPG, Silver et al. (2014)). MP-DQN (Bester et al. (2019)) states a", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 204, + 505, + 216 + ], + "spans": [ + { + "bbox": [ + 105, + 204, + 354, + 216 + ], + "score": 1.0, + "content": "problem regarding to the erroneous gradients of Q-network", + "type": "text" + }, + { + "bbox": [ + 354, + 204, + 377, + 216 + ], + "score": 0.91, + "content": "\\nabla L _ { Q }", + "type": "inline_equation" + }, + { + "bbox": [ + 377, + 204, + 505, + 216 + ], + "score": 1.0, + "content": "in P-DQN caused by multiple", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "spans": [ + { + "bbox": [ + 105, + 214, + 505, + 227 + ], + "score": 1.0, + "content": "action-parameter policies backpropagating gradients at the same time. However, it does not exist", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "spans": [ + { + "bbox": [ + 105, + 225, + 505, + 238 + ], + "score": 1.0, + "content": "in such single network situation. This is a desired architecture since separating multiple action-", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 236, + 430, + 250 + ], + "spans": [ + { + "bbox": [ + 105, + 236, + 430, + 250 + ], + "score": 1.0, + "content": "parameter policies loses the knowledge that they share same potential meanings.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 181, + 506, + 250 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 253, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 104, + 252, + 506, + 266 + ], + "spans": [ + { + "bbox": [ + 104, + 252, + 410, + 266 + ], + "score": 1.0, + "content": "Following this intuition, we directly parameterize the policy with two heads", + "type": "text" + }, + { + "bbox": [ + 411, + 252, + 483, + 265 + ], + "score": 0.92, + "content": "\\mu : \\mathcal { S } \\to \\mathbb { R } ^ { K } \\times \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 484, + 252, + 506, + 266 + ], + "score": 1.0, + "content": "with", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 260, + 507, + 283 + ], + "spans": [ + { + "bbox": [ + 106, + 266, + 118, + 278 + ], + "score": 0.83, + "content": "\\theta _ { \\mu }", + "type": "inline_equation" + }, + { + "bbox": [ + 111, + 264, + 258, + 278 + ], + "score": 0.91, + "content": "\\mathrm { ~ ` ~ } _ { \\mu } \\mathrm { ~ a s : ~ } k _ { t } ^ { s o f t } , x _ { t } = a _ { t } ^ { s o f t } = \\mu ( s _ { t } ; \\theta _ { \\mu } )", + "type": "inline_equation" + }, + { + "bbox": [ + 258, + 260, + 288, + 283 + ], + "score": 1.0, + "content": ", where", + "type": "text" + }, + { + "bbox": [ + 289, + 264, + 311, + 278 + ], + "score": 0.92, + "content": "k _ { t } ^ { s o f t }", + "type": "inline_equation" + }, + { + "bbox": [ + 311, + 260, + 329, + 283 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 330, + 264, + 352, + 278 + ], + "score": 0.93, + "content": "a _ { t } ^ { s o f t }", + "type": "inline_equation" + }, + { + "bbox": [ + 352, + 260, + 507, + 283 + ], + "score": 1.0, + "content": "refers to action-parameters or actions", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "spans": [ + { + "bbox": [ + 105, + 275, + 505, + 290 + ], + "score": 1.0, + "content": "with softmax values. One head outputs the sof tmax values for all discrete actions. For another", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 287, + 505, + 300 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 413, + 300 + ], + "score": 1.0, + "content": "head, it outputs normalized action-parameter values with hyperbolic tangent", + "type": "text" + }, + { + "bbox": [ + 414, + 288, + 442, + 299 + ], + "score": 0.4, + "content": "( t a n h )", + "type": "inline_equation" + }, + { + "bbox": [ + 442, + 287, + 505, + 300 + ], + "score": 1.0, + "content": "activation after", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "spans": [ + { + "bbox": [ + 105, + 298, + 505, + 311 + ], + "score": 1.0, + "content": "the continuous head. The environment will denormalize it to a valid range. This corresponds to the", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 505, + 322 + ], + "score": 1.0, + "content": "Squashing Gradients method for bounded-continuous action spaces in Hausknecht & Stone (2015).", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 12.5, + "bbox_fs": [ + 104, + 252, + 507, + 322 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 341, + 505, + 408 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 505, + 354 + ], + "score": 1.0, + "content": "Exploration with Parameter Space Noise and action sampling. In DQN (Mnih et al. (2015)),", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 352, + 505, + 366 + ], + "spans": [ + { + "bbox": [ + 105, + 352, + 236, + 366 + ], + "score": 1.0, + "content": "the behavioral policy is given by", + "type": "text" + }, + { + "bbox": [ + 236, + 355, + 241, + 362 + ], + "score": 0.62, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 241, + 352, + 505, + 366 + ], + "score": 1.0, + "content": "-greedy strategy for exploration purpose, where an agent randomly", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 363, + 505, + 376 + ], + "spans": [ + { + "bbox": [ + 105, + 363, + 274, + 376 + ], + "score": 1.0, + "content": "selects a discrete action with probability", + "type": "text" + }, + { + "bbox": [ + 275, + 364, + 302, + 374 + ], + "score": 0.89, + "content": "\\epsilon < 1", + "type": "inline_equation" + }, + { + "bbox": [ + 302, + 363, + 505, + 376 + ], + "score": 1.0, + "content": ". In DDPG (Lillicrap et al. (2015)), although the", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "spans": [ + { + "bbox": [ + 105, + 374, + 506, + 387 + ], + "score": 1.0, + "content": "policy learned is deterministic, to explore the action space, the behavioral policy needs to be different", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "spans": [ + { + "bbox": [ + 105, + 385, + 506, + 398 + ], + "score": 1.0, + "content": "because of the off-policy nature. It uses action space noise such as uncorrelated Gaussian noise or", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 397, + 361, + 408 + ], + "spans": [ + { + "bbox": [ + 106, + 397, + 361, + 408 + ], + "score": 1.0, + "content": "correlated Ornstein-Uhlenbeck process (Lillicrap et al. (2015)).", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 340, + 506, + 408 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 413, + 505, + 479 + ], + "lines": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 106, + 414, + 505, + 425 + ], + "score": 1.0, + "content": "Directly combining such two exploration strategies is straightforward. However, it may induce a", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 424, + 505, + 437 + ], + "spans": [ + { + "bbox": [ + 105, + 424, + 310, + 437 + ], + "score": 1.0, + "content": "problem that the exploration in two action spaces", + "type": "text" + }, + { + "bbox": [ + 311, + 424, + 370, + 435 + ], + "score": 0.92, + "content": "\\mathcal { A } = \\mathcal { A } _ { d } \\times \\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 370, + 424, + 505, + 437 + ], + "score": 1.0, + "content": "in different paces. For example,", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "spans": [ + { + "bbox": [ + 106, + 435, + 505, + 448 + ], + "score": 1.0, + "content": "the match plan may need another rewrite with larger quotas in a step. However, if the exploration", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 446, + 505, + 459 + ], + "spans": [ + { + "bbox": [ + 106, + 446, + 175, + 459 + ], + "score": 1.0, + "content": "strategy is to use", + "type": "text" + }, + { + "bbox": [ + 176, + 448, + 181, + 456 + ], + "score": 0.66, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 181, + 446, + 263, + 459 + ], + "score": 1.0, + "content": "-greedy for a rewrite", + "type": "text" + }, + { + "bbox": [ + 264, + 447, + 271, + 456 + ], + "score": 0.81, + "content": "k", + "type": "inline_equation" + }, + { + "bbox": [ + 271, + 446, + 405, + 459 + ], + "score": 1.0, + "content": "and Gaussian noise for its quotas", + "type": "text" + }, + { + "bbox": [ + 405, + 448, + 412, + 456 + ], + "score": 0.75, + "content": "x", + "type": "inline_equation" + }, + { + "bbox": [ + 412, + 446, + 505, + 459 + ], + "score": 1.0, + "content": ", the agent may require", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "spans": [ + { + "bbox": [ + 105, + 457, + 506, + 470 + ], + "score": 1.0, + "content": "more samples to discover the potential rewards (positive documents). We instead use parameter", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 468, + 454, + 481 + ], + "spans": [ + { + "bbox": [ + 105, + 468, + 454, + 481 + ], + "score": 1.0, + "content": "space noise (Plappert et al. (2017)) on parameterized action space to tackle such issue.", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 414, + 506, + 481 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 485, + 505, + 585 + ], + "lines": [ + { + "bbox": [ + 103, + 481, + 508, + 505 + ], + "spans": [ + { + "bbox": [ + 103, + 481, + 330, + 505 + ], + "score": 1.0, + "content": "We denote the discrete and continuous action heads as", + "type": "text" + }, + { + "bbox": [ + 331, + 485, + 450, + 499 + ], + "score": 0.92, + "content": "( \\pi _ { k } ^ { s o f t } ( s ) , \\pi _ { x } ( s ) ) = \\mu ( s ; \\theta _ { \\mu } ) .", + "type": "inline_equation" + }, + { + "bbox": [ + 450, + 481, + 508, + 505 + ], + "score": 1.0, + "content": ". To compute", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 497, + 505, + 512 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 158, + 512 + ], + "score": 1.0, + "content": "the distance", + "type": "text" + }, + { + "bbox": [ + 158, + 498, + 248, + 510 + ], + "score": 0.92, + "content": "d ( \\pi , \\widetilde { \\pi } ) = D _ { \\mathrm { K L } } ( \\pi \\| \\widetilde { \\pi } )", + "type": "inline_equation" + }, + { + "bbox": [ + 249, + 497, + 415, + 512 + ], + "score": 1.0, + "content": "of non-perturbed and perturbed policies", + "type": "text" + }, + { + "bbox": [ + 415, + 498, + 486, + 510 + ], + "score": 0.92, + "content": "\\mu ( s ; \\theta _ { \\mu } ) , \\widetilde { \\mu } ( s ; \\theta _ { \\mu } )", + "type": "inline_equation" + }, + { + "bbox": [ + 486, + 497, + 505, + 512 + ], + "score": 1.0, + "content": ", we", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 508, + 505, + 522 + ], + "spans": [ + { + "bbox": [ + 106, + 508, + 505, + 522 + ], + "score": 1.0, + "content": "e e euse weighted sum of the distance of discrete and continuous actions. For the continuous actions", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 106, + 519, + 506, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 523, + 131, + 536 + ], + "score": 0.91, + "content": "\\widetilde { \\pi } _ { x } ( s )", + "type": "inline_equation" + }, + { + "bbox": [ + 131, + 522, + 229, + 538 + ], + "score": 1.0, + "content": ", the distance is given by", + "type": "text" + }, + { + "bbox": [ + 230, + 519, + 329, + 540 + ], + "score": 0.94, + "content": "\\mathbb { E } _ { s } \\left[ \\left( \\pi _ { x } ( s ) _ { i } - \\widetilde { \\pi } _ { x } ( s ) _ { i } \\right) ^ { 2 } \\right]", + "type": "inline_equation" + }, + { + "bbox": [ + 330, + 522, + 506, + 538 + ], + "score": 1.0, + "content": "to estimate KL-divergence empirically. For", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 101, + 529, + 501, + 565 + ], + "spans": [ + { + "bbox": [ + 101, + 529, + 399, + 565 + ], + "score": 1.0, + "content": "e ethe discrete actions, we use outputted softmax probabilities to compute The state and action pairs are sampled from a replay memory. The varianc", + "type": "text" + }, + { + "bbox": [ + 399, + 538, + 501, + 553 + ], + "score": 0.93, + "content": "D _ { \\mathrm { K L } } ( \\pi _ { k } ^ { s o f t } ( s ) \\Vert \\widetilde { \\pi } _ { k } ^ { s o f t } ( s ) )", + "type": "inline_equation" + } + ], + "index": 32 + }, + { + "bbox": [ + 406, + 554, + 413, + 561 + ], + "spans": [ + { + "bbox": [ + 406, + 554, + 413, + 561 + ], + "score": 0.78, + "content": "\\sigma", + "type": "inline_equation" + } + ], + "index": 33 + }, + { + "bbox": [ + 106, + 562, + 505, + 575 + ], + "spans": [ + { + "bbox": [ + 106, + 562, + 335, + 575 + ], + "score": 1.0, + "content": "after a policy update based on the distance and threhold", + "type": "text" + }, + { + "bbox": [ + 335, + 563, + 342, + 572 + ], + "score": 0.73, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 342, + 562, + 505, + 575 + ], + "score": 1.0, + "content": "(Plappert et al. (2017)). The policy and", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 573, + 448, + 586 + ], + "spans": [ + { + "bbox": [ + 106, + 573, + 448, + 586 + ], + "score": 1.0, + "content": "target networks with layer normalization (Ba et al. (2016)) are perturbed per episode.", + "type": "text" + } + ], + "index": 35 + } + ], + "index": 31.5, + "bbox_fs": [ + 101, + 481, + 508, + 586 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 604, + 505, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "spans": [ + { + "bbox": [ + 105, + 604, + 505, + 618 + ], + "score": 1.0, + "content": "Prioritized replay and recurrent policies. We use recurrent architecture to obtain the underlying", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "system state of the POMDP for both value and policy networks (Heess et al. (2015)). To avoid re-", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 104, + 624, + 504, + 641 + ], + "spans": [ + { + "bbox": [ + 104, + 624, + 445, + 641 + ], + "score": 1.0, + "content": "current state staleness (Kapturowski et al. (2018)), we store and replay a sequence of", + "type": "text" + }, + { + "bbox": [ + 445, + 626, + 504, + 639 + ], + "score": 0.91, + "content": "( s , k ^ { s o f t } , x , r )", + "type": "inline_equation" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 638, + 506, + 651 + ], + "score": 1.0, + "content": "with softmax values into a replay memory. Since a match plan is usually short, we set a maximum", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 649, + 437, + 661 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 437, + 661 + ], + "score": 1.0, + "content": "length and pad shorter episodes with zero or randomly sampled states and actions.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 38, + "bbox_fs": [ + 104, + 604, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 410, + 678 + ], + "score": 1.0, + "content": "The sampling from a replay memory can be prioritized with a probability", + "type": "text" + }, + { + "bbox": [ + 411, + 668, + 420, + 677 + ], + "score": 0.83, + "content": "p _ { i }", + "type": "inline_equation" + }, + { + "bbox": [ + 421, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "proportional to TD-", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 689 + ], + "score": 1.0, + "content": "errors (Schaul et al. (2015)) to increase the sample efficiency in such structured action space. To", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 104, + 687, + 506, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 687, + 506, + 702 + ], + "score": 1.0, + "content": "prioritize the transitions with well-matched documents, we slightly modify the sampling strategy.", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 505, + 712 + ], + "score": 1.0, + "content": "With a half probability, the agent uses regular prioritization, otherwise partitions the memory to", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 708, + 505, + 724 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 724 + ], + "score": 1.0, + "content": "multiple bins and retrieve a transition (sequence) with max reward from each bin. Our strategy", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 721, + 415, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 415, + 733 + ], + "score": 1.0, + "content": "provides a more efficient and balanced exploration strategy of the evaluation.", + "type": "text" + } + ], + "index": 46 + } + ], + "index": 43.5, + "bbox_fs": [ + 104, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 329, + 95 + ], + "score": 1.0, + "content": "Policy update. In the update, we use clipped double", + "type": "text" + }, + { + "bbox": [ + 329, + 83, + 338, + 94 + ], + "score": 0.56, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "-learning and delayed policy update (Fu-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 104, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "jimoto et al. (2018)) to stabilizing the training. After the agent samples a batch of (sequences), it", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "perturbs the target policy network and computes the perturbed actions for target policy smoothing.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 468, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 250, + 128 + ], + "score": 1.0, + "content": "The parameter space noise variance", + "type": "text" + }, + { + "bbox": [ + 251, + 117, + 258, + 125 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 115, + 468, + 128 + ], + "score": 1.0, + "content": "is then updated with the sampled states and actions.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 105, + 132, + 504, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "The policy is still deterministic since it is learned in off-policy and only the exploration involves", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 496, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 496, + 156 + ], + "score": 1.0, + "content": "stochasticity. The update is given by deterministic policy gradients theorem (Silver et al. (2014)):", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5 + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 159, + 455, + 189 + ], + "lines": [ + { + "bbox": [ + 155, + 159, + 455, + 189 + ], + "spans": [ + { + "bbox": [ + 155, + 159, + 455, + 189 + ], + "score": 0.92, + "content": "\\nabla L _ { \\mu } \\left( \\theta _ { \\mu } \\right) = \\nabla _ { s \\sim D } \\left[ Q \\left( s , \\mu ( s ; \\theta _ { \\mu } ) ; \\theta _ { Q } \\right) \\right] = \\nabla \\frac { 1 } { \\left| D \\right| } \\sum _ { s \\in D } Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) ,", + "type": "interline_equation", + "image_path": "3562b5ab1968d6fce4dd58258e2c12c671f3fadb5d89d26146c9a6856a9e24a7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 155, + 159, + 455, + 169.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 169.0, + 455, + 179.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 179.0, + 455, + 189.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 133, + 205 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 194, + 143, + 203 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 194, + 505, + 205 + ], + "score": 1.0, + "content": "is a replay memory. We found this policy architecture is similar to PA-DDPG (Hausknecht", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 116, + 214 + ], + "score": 0.32, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "Stone (2015)), while it does not including sof tmax activation for discrete actions and other", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "advanced techniques. We also build upon other useful techniques such as invertible value function", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 411, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 146, + 241 + ], + "score": 1.0, + "content": "rescaling", + "type": "text" + }, + { + "bbox": [ + 146, + 226, + 295, + 241 + ], + "score": 0.92, + "content": "h ( x ) = \\mathrm { { s i g n } } ( x ) ( \\sqrt { | x | } + 1 - 1 ) + \\epsilon x", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 227, + 411, + 241 + ], + "score": 1.0, + "content": "(Kapturowski et al. (2018)):", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 244, + 415, + 279 + ], + "lines": [ + { + "bbox": [ + 195, + 244, + 415, + 279 + ], + "spans": [ + { + "bbox": [ + 195, + 244, + 415, + 279 + ], + "score": 0.93, + "content": "y = h \\left( \\sum _ { k = 0 } ^ { n - 1 } r _ { t + k } \\gamma ^ { k } + \\gamma ^ { n } h ^ { - 1 } \\left( Q \\left( s _ { t + n } , a ^ { \\ast } ; \\theta _ { Q } ^ { - } \\right) \\right) \\right) ,", + "type": "interline_equation", + "image_path": "fce58c5a07a0c3febf35ee414e9539f1e4ea20fd2393bb65923f45a0fdcb7751.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 195, + 244, + 415, + 261.5 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 195, + 261.5, + 415, + 279.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 102, + 282, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 102, + 282, + 133, + 302 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 284, + 268, + 298 + ], + "score": 0.93, + "content": "a ^ { * } = ( \\operatorname* { m a x } k _ { t } ^ { s o f t } , x _ { t } ) = ( k _ { t } , x _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 282, + 355, + 302 + ], + "score": 1.0, + "content": "is the greedy action,", + "type": "text" + }, + { + "bbox": [ + 356, + 288, + 362, + 298 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 282, + 505, + 302 + ], + "score": 1.0, + "content": "the target for updating Q-network,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 123, + 312 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 298, + 136, + 311 + ], + "score": 0.89, + "content": "\\theta _ { Q } ^ { - }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "denotes the target Q-network. The policy network outputs differentiable softmax values and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 366, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 366, + 323 + ], + "score": 1.0, + "content": "is updated by the gradients backpropagated from the Q-network:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16 + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 326, + 408, + 353 + ], + "lines": [ + { + "bbox": [ + 203, + 326, + 408, + 353 + ], + "spans": [ + { + "bbox": [ + 203, + 326, + 408, + 353 + ], + "score": 0.93, + "content": "\\nabla L _ { Q } \\left( \\theta _ { Q } \\right) = \\nabla \\mathbb { E } \\left[ \\frac { 1 } { 2 } \\left( y - Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) \\right) ^ { 2 } \\right] ,", + "type": "interline_equation", + "image_path": "d951c0efa2af92e644a3128d1ed4e8b66e025fb64e2e4c6797ccd5e0e6c726fd.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 326, + 408, + 339.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 203, + 339.5, + 408, + 353.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 357, + 346, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 347, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 347, + 372 + ], + "score": 1.0, + "content": "where the expectation is took over samples from a memory.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20 + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 438, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 439, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 439, + 387 + ], + "score": 1.0, + "content": "The pseudocode for the algorithm with more details can be found in the Appendix.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21 + }, + { + "type": "title", + "bbox": [ + 108, + 401, + 200, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 201, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 201, + 416 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "In this section, we apply the purposed agent on the inverted index match plan generation problem.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "We create a dataset which contains a set of queries and corresponding query embeddings. We", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "perform various ablation study on how each component interacts and the benefits of them. We also", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 459, + 358, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 358, + 471 + ], + "score": 1.0, + "content": "test on other Paramterized Action RL benchmarking baselines.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5 + }, + { + "type": "title", + "bbox": [ + 107, + 484, + 234, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 235, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 235, + 496 + ], + "score": 1.0, + "content": "4.1 EXPERIMENT SETTINGS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "For match plan generation, we experiment on a dataset that has about 100,000 queries sampled from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "Bing search log. In the current production system, each query is classified to a predefined query", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 526, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 539 + ], + "score": 1.0, + "content": "classes online based on a set of rules which are related to statistical features of the query. Each", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 506, + 551 + ], + "score": 1.0, + "content": "query class has some hand-crafted rules which outputs a match plan for the classified query. We use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "each query’s production match plan as baseline. The match plan for each query is generated by the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 558, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 573 + ], + "score": 1.0, + "content": "hand-crafted rules and does not change over each running. We only skip a few special queries that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 570, + 409, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 409, + 583 + ], + "score": 1.0, + "content": "do not have embeddings or need additional operations beyond match plans.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31 + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "The generated match plan is evaluated by the delta values of both Ranking Score and Seek Count", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "and overall reward. We provide learning curves of delta values of evaluation rewards. Note that the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "reward are not symmetrical around 0, since the ranking scores has a minimum value and is sparse", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "since a few queries are hard to find good documents and will be assigned a very low score. It may", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "significantly pull down the average rewards shown in the curves, thus we present histograms for fair", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 641, + 500, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 500, + 656 + ], + "score": 1.0, + "content": "comparison. For the benchmarks, we report the results on evaluation reward curves during tuning.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5 + }, + { + "type": "title", + "bbox": [ + 106, + 666, + 376, + 678 + ], + "lines": [ + { + "bbox": [ + 111, + 667, + 376, + 680 + ], + "spans": [ + { + "bbox": [ + 111, + 667, + 376, + 680 + ], + "score": 1.0, + "content": ".2 MATCH PLAN GENERATION PERFORMANCE EVALUATION", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "We first visualize the query embeddings in Figure 3 using UMAP (McInnes et al. (2018)) to demon-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "strate our learned models. We evaluate 6,000 queries and draw the colormap based on the difference", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "between evaluation reward and production baseline reward. Although most queries perform similar", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "to the well-designed production rules, there are some clusters of queries and some patterns exist. We", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 43.5 + } + ], + "page_idx": 4, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 763 + ], + "score": 1.0, + "content": "5", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 106, + 82, + 504, + 127 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 329, + 95 + ], + "score": 1.0, + "content": "Policy update. In the update, we use clipped double", + "type": "text" + }, + { + "bbox": [ + 329, + 83, + 338, + 94 + ], + "score": 0.56, + "content": "Q", + "type": "inline_equation" + }, + { + "bbox": [ + 338, + 81, + 505, + 95 + ], + "score": 1.0, + "content": "-learning and delayed policy update (Fu-", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 104, + 94, + 506, + 106 + ], + "spans": [ + { + "bbox": [ + 104, + 94, + 506, + 106 + ], + "score": 1.0, + "content": "jimoto et al. (2018)) to stabilizing the training. After the agent samples a batch of (sequences), it", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "spans": [ + { + "bbox": [ + 105, + 105, + 505, + 118 + ], + "score": 1.0, + "content": "perturbs the target policy network and computes the perturbed actions for target policy smoothing.", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 115, + 468, + 128 + ], + "spans": [ + { + "bbox": [ + 106, + 115, + 250, + 128 + ], + "score": 1.0, + "content": "The parameter space noise variance", + "type": "text" + }, + { + "bbox": [ + 251, + 117, + 258, + 125 + ], + "score": 0.77, + "content": "\\sigma", + "type": "inline_equation" + }, + { + "bbox": [ + 259, + 115, + 468, + 128 + ], + "score": 1.0, + "content": "is then updated with the sampled states and actions.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 1.5, + "bbox_fs": [ + 104, + 81, + 506, + 128 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 132, + 504, + 155 + ], + "lines": [ + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "spans": [ + { + "bbox": [ + 106, + 132, + 505, + 144 + ], + "score": 1.0, + "content": "The policy is still deterministic since it is learned in off-policy and only the exploration involves", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 144, + 496, + 156 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 496, + 156 + ], + "score": 1.0, + "content": "stochasticity. The update is given by deterministic policy gradients theorem (Silver et al. (2014)):", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4.5, + "bbox_fs": [ + 106, + 132, + 505, + 156 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 155, + 159, + 455, + 189 + ], + "lines": [ + { + "bbox": [ + 155, + 159, + 455, + 189 + ], + "spans": [ + { + "bbox": [ + 155, + 159, + 455, + 189 + ], + "score": 0.92, + "content": "\\nabla L _ { \\mu } \\left( \\theta _ { \\mu } \\right) = \\nabla _ { s \\sim D } \\left[ Q \\left( s , \\mu ( s ; \\theta _ { \\mu } ) ; \\theta _ { Q } \\right) \\right] = \\nabla \\frac { 1 } { \\left| D \\right| } \\sum _ { s \\in D } Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) ,", + "type": "interline_equation", + "image_path": "3562b5ab1968d6fce4dd58258e2c12c671f3fadb5d89d26146c9a6856a9e24a7.jpg" + } + ] + } + ], + "index": 7, + "virtual_lines": [ + { + "bbox": [ + 155, + 159, + 455, + 169.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 155, + 169.0, + 455, + 179.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 155, + 179.0, + 455, + 189.0 + ], + "spans": [], + "index": 8 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 193, + 505, + 240 + ], + "lines": [ + { + "bbox": [ + 106, + 194, + 505, + 205 + ], + "spans": [ + { + "bbox": [ + 106, + 194, + 133, + 205 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 133, + 194, + 143, + 203 + ], + "score": 0.83, + "content": "D", + "type": "inline_equation" + }, + { + "bbox": [ + 143, + 194, + 505, + 205 + ], + "score": 1.0, + "content": "is a replay memory. We found this policy architecture is similar to PA-DDPG (Hausknecht", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 107, + 203, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 116, + 214 + ], + "score": 0.32, + "content": "\\&", + "type": "inline_equation" + }, + { + "bbox": [ + 116, + 203, + 506, + 217 + ], + "score": 1.0, + "content": "Stone (2015)), while it does not including sof tmax activation for discrete actions and other", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 216, + 505, + 228 + ], + "score": 1.0, + "content": "advanced techniques. We also build upon other useful techniques such as invertible value function", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 226, + 411, + 241 + ], + "spans": [ + { + "bbox": [ + 105, + 227, + 146, + 241 + ], + "score": 1.0, + "content": "rescaling", + "type": "text" + }, + { + "bbox": [ + 146, + 226, + 295, + 241 + ], + "score": 0.92, + "content": "h ( x ) = \\mathrm { { s i g n } } ( x ) ( \\sqrt { | x | } + 1 - 1 ) + \\epsilon x", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 227, + 411, + 241 + ], + "score": 1.0, + "content": "(Kapturowski et al. (2018)):", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 194, + 506, + 241 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 195, + 244, + 415, + 279 + ], + "lines": [ + { + "bbox": [ + 195, + 244, + 415, + 279 + ], + "spans": [ + { + "bbox": [ + 195, + 244, + 415, + 279 + ], + "score": 0.93, + "content": "y = h \\left( \\sum _ { k = 0 } ^ { n - 1 } r _ { t + k } \\gamma ^ { k } + \\gamma ^ { n } h ^ { - 1 } \\left( Q \\left( s _ { t + n } , a ^ { \\ast } ; \\theta _ { Q } ^ { - } \\right) \\right) \\right) ,", + "type": "interline_equation", + "image_path": "fce58c5a07a0c3febf35ee414e9539f1e4ea20fd2393bb65923f45a0fdcb7751.jpg" + } + ] + } + ], + "index": 13.5, + "virtual_lines": [ + { + "bbox": [ + 195, + 244, + 415, + 261.5 + ], + "spans": [], + "index": 13 + }, + { + "bbox": [ + 195, + 261.5, + 415, + 279.0 + ], + "spans": [], + "index": 14 + } + ] + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 321 + ], + "lines": [ + { + "bbox": [ + 102, + 282, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 102, + 282, + 133, + 302 + ], + "score": 1.0, + "content": "where", + "type": "text" + }, + { + "bbox": [ + 134, + 284, + 268, + 298 + ], + "score": 0.93, + "content": "a ^ { * } = ( \\operatorname* { m a x } k _ { t } ^ { s o f t } , x _ { t } ) = ( k _ { t } , x _ { t } )", + "type": "inline_equation" + }, + { + "bbox": [ + 268, + 282, + 355, + 302 + ], + "score": 1.0, + "content": "is the greedy action,", + "type": "text" + }, + { + "bbox": [ + 356, + 288, + 362, + 298 + ], + "score": 0.8, + "content": "y", + "type": "inline_equation" + }, + { + "bbox": [ + 363, + 282, + 505, + 302 + ], + "score": 1.0, + "content": "the target for updating Q-network,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 297, + 505, + 312 + ], + "spans": [ + { + "bbox": [ + 106, + 297, + 123, + 312 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 123, + 298, + 136, + 311 + ], + "score": 0.89, + "content": "\\theta _ { Q } ^ { - }", + "type": "inline_equation" + }, + { + "bbox": [ + 136, + 297, + 505, + 312 + ], + "score": 1.0, + "content": "denotes the target Q-network. The policy network outputs differentiable softmax values and", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 309, + 366, + 323 + ], + "spans": [ + { + "bbox": [ + 105, + 309, + 366, + 323 + ], + "score": 1.0, + "content": "is updated by the gradients backpropagated from the Q-network:", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16, + "bbox_fs": [ + 102, + 282, + 505, + 323 + ] + }, + { + "type": "interline_equation", + "bbox": [ + 203, + 326, + 408, + 353 + ], + "lines": [ + { + "bbox": [ + 203, + 326, + 408, + 353 + ], + "spans": [ + { + "bbox": [ + 203, + 326, + 408, + 353 + ], + "score": 0.93, + "content": "\\nabla L _ { Q } \\left( \\theta _ { Q } \\right) = \\nabla \\mathbb { E } \\left[ \\frac { 1 } { 2 } \\left( y - Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) \\right) ^ { 2 } \\right] ,", + "type": "interline_equation", + "image_path": "d951c0efa2af92e644a3128d1ed4e8b66e025fb64e2e4c6797ccd5e0e6c726fd.jpg" + } + ] + } + ], + "index": 18.5, + "virtual_lines": [ + { + "bbox": [ + 203, + 326, + 408, + 339.5 + ], + "spans": [], + "index": 18 + }, + { + "bbox": [ + 203, + 339.5, + 408, + 353.0 + ], + "spans": [], + "index": 19 + } + ] + }, + { + "type": "text", + "bbox": [ + 108, + 357, + 346, + 369 + ], + "lines": [ + { + "bbox": [ + 105, + 354, + 347, + 372 + ], + "spans": [ + { + "bbox": [ + 105, + 354, + 347, + 372 + ], + "score": 1.0, + "content": "where the expectation is took over samples from a memory.", + "type": "text" + } + ], + "index": 20 + } + ], + "index": 20, + "bbox_fs": [ + 105, + 354, + 347, + 372 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 374, + 438, + 386 + ], + "lines": [ + { + "bbox": [ + 106, + 373, + 439, + 387 + ], + "spans": [ + { + "bbox": [ + 106, + 373, + 439, + 387 + ], + "score": 1.0, + "content": "The pseudocode for the algorithm with more details can be found in the Appendix.", + "type": "text" + } + ], + "index": 21 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 373, + 439, + 387 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 401, + 200, + 414 + ], + "lines": [ + { + "bbox": [ + 105, + 401, + 201, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 201, + 416 + ], + "score": 1.0, + "content": "4 EXPERIMENTS", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 22 + }, + { + "type": "text", + "bbox": [ + 107, + 426, + 505, + 471 + ], + "lines": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "spans": [ + { + "bbox": [ + 105, + 426, + 505, + 439 + ], + "score": 1.0, + "content": "In this section, we apply the purposed agent on the inverted index match plan generation problem.", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "spans": [ + { + "bbox": [ + 105, + 437, + 505, + 451 + ], + "score": 1.0, + "content": "We create a dataset which contains a set of queries and corresponding query embeddings. We", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 505, + 461 + ], + "score": 1.0, + "content": "perform various ablation study on how each component interacts and the benefits of them. We also", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 459, + 358, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 459, + 358, + 471 + ], + "score": 1.0, + "content": "test on other Paramterized Action RL benchmarking baselines.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 24.5, + "bbox_fs": [ + 105, + 426, + 505, + 471 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 484, + 234, + 495 + ], + "lines": [ + { + "bbox": [ + 106, + 484, + 235, + 496 + ], + "spans": [ + { + "bbox": [ + 106, + 484, + 235, + 496 + ], + "score": 1.0, + "content": "4.1 EXPERIMENT SETTINGS", + "type": "text" + } + ], + "index": 27 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 505, + 505, + 582 + ], + "lines": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "spans": [ + { + "bbox": [ + 105, + 505, + 505, + 517 + ], + "score": 1.0, + "content": "For match plan generation, we experiment on a dataset that has about 100,000 queries sampled from", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 105, + 515, + 505, + 529 + ], + "score": 1.0, + "content": "Bing search log. In the current production system, each query is classified to a predefined query", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 526, + 505, + 539 + ], + "spans": [ + { + "bbox": [ + 105, + 526, + 505, + 539 + ], + "score": 1.0, + "content": "classes online based on a set of rules which are related to statistical features of the query. Each", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 104, + 537, + 506, + 551 + ], + "spans": [ + { + "bbox": [ + 104, + 537, + 506, + 551 + ], + "score": 1.0, + "content": "query class has some hand-crafted rules which outputs a match plan for the classified query. We use", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "spans": [ + { + "bbox": [ + 105, + 548, + 505, + 561 + ], + "score": 1.0, + "content": "each query’s production match plan as baseline. The match plan for each query is generated by the", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 558, + 505, + 573 + ], + "spans": [ + { + "bbox": [ + 105, + 558, + 505, + 573 + ], + "score": 1.0, + "content": "hand-crafted rules and does not change over each running. We only skip a few special queries that", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 570, + 409, + 583 + ], + "spans": [ + { + "bbox": [ + 105, + 570, + 409, + 583 + ], + "score": 1.0, + "content": "do not have embeddings or need additional operations beyond match plans.", + "type": "text" + } + ], + "index": 34 + } + ], + "index": 31, + "bbox_fs": [ + 104, + 505, + 506, + 583 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 587, + 505, + 653 + ], + "lines": [ + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "spans": [ + { + "bbox": [ + 106, + 587, + 506, + 600 + ], + "score": 1.0, + "content": "The generated match plan is evaluated by the delta values of both Ranking Score and Seek Count", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 598, + 505, + 611 + ], + "score": 1.0, + "content": "and overall reward. We provide learning curves of delta values of evaluation rewards. Note that the", + "type": "text" + } + ], + "index": 36 + }, + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "spans": [ + { + "bbox": [ + 105, + 609, + 506, + 622 + ], + "score": 1.0, + "content": "reward are not symmetrical around 0, since the ranking scores has a minimum value and is sparse", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "spans": [ + { + "bbox": [ + 105, + 619, + 505, + 633 + ], + "score": 1.0, + "content": "since a few queries are hard to find good documents and will be assigned a very low score. It may", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 631, + 506, + 645 + ], + "score": 1.0, + "content": "significantly pull down the average rewards shown in the curves, thus we present histograms for fair", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 641, + 500, + 656 + ], + "spans": [ + { + "bbox": [ + 105, + 641, + 500, + 656 + ], + "score": 1.0, + "content": "comparison. For the benchmarks, we report the results on evaluation reward curves during tuning.", + "type": "text" + } + ], + "index": 40 + } + ], + "index": 37.5, + "bbox_fs": [ + 105, + 587, + 506, + 656 + ] + }, + { + "type": "title", + "bbox": [ + 106, + 666, + 376, + 678 + ], + "lines": [ + { + "bbox": [ + 111, + 667, + 376, + 680 + ], + "spans": [ + { + "bbox": [ + 111, + 667, + 376, + 680 + ], + "score": 1.0, + "content": ".2 MATCH PLAN GENERATION PERFORMANCE EVALUATION", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 41 + }, + { + "type": "text", + "bbox": [ + 108, + 687, + 504, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "We first visualize the query embeddings in Figure 3 using UMAP (McInnes et al. (2018)) to demon-", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 698, + 506, + 712 + ], + "score": 1.0, + "content": "strate our learned models. We evaluate 6,000 queries and draw the colormap based on the difference", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 106, + 710, + 505, + 722 + ], + "score": 1.0, + "content": "between evaluation reward and production baseline reward. Although most queries perform similar", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 720, + 506, + 733 + ], + "score": 1.0, + "content": "to the well-designed production rules, there are some clusters of queries and some patterns exist. We", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 321, + 465 + ], + "score": 1.0, + "content": "notice that there is a main cluster at the center (about", + "type": "text", + "cross_page": true + }, + { + "bbox": [ + 321, + 452, + 361, + 464 + ], + "score": 0.92, + "content": "( - 7 , - 1 ) )", + "type": "inline_equation", + "cross_page": true + }, + { + "bbox": [ + 361, + 452, + 505, + 465 + ], + "score": 1.0, + "content": ") that gathers most queries on which", + "type": "text", + "cross_page": true + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "our agent does not perform well. That cluster contains some random inputs from the users, which", + "type": "text", + "cross_page": true + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 475, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 504, + 486 + ], + "score": 1.0, + "content": "may have some typos. For the queries that the agent outperforms the baseline, we find they are quite", + "type": "text", + "cross_page": true + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 486, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 504, + 497 + ], + "score": 1.0, + "content": "scattered. We guess the reason is that the baseline with hand-crafted rules do not consider the em-", + "type": "text", + "cross_page": true + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "beddings, thus it is less affected by embeddings. This may suggest us to look for more informative", + "type": "text", + "cross_page": true + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 506, + 376, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 376, + 520 + ], + "score": 1.0, + "content": "embeddings of queries to better distinguish some clustered queries.", + "type": "text", + "cross_page": true + } + ], + "index": 25 + } + ], + "index": 43.5, + "bbox_fs": [ + 105, + 688, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 82, + 335, + 251 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 82, + 335, + 251 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 82, + 335, + 251 + ], + "spans": [ + { + "bbox": [ + 110, + 82, + 335, + 251 + ], + "score": 0.95, + "type": "image", + "image_path": "1b912e4438fe0c4576fc52e07dd81917a7cf3148fcc9fac32a9a96a543cd4db8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 110, + 82, + 335, + 95.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 95.0, + 335, + 108.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 108.0, + 335, + 121.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 121.0, + 335, + 134.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 134.0, + 335, + 147.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 147.0, + 335, + 160.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 110, + 160.0, + 335, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 110, + 173.0, + 335, + 186.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 110, + 186.0, + 335, + 199.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 199.0, + 335, + 212.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 212.0, + 335, + 225.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 110, + 225.0, + 335, + 238.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 110, + 238.0, + 335, + 251.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 272, + 297, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 270, + 298, + 286 + ], + "spans": [ + { + "bbox": [ + 114, + 270, + 298, + 286 + ], + "score": 1.0, + "content": "Figure 3: Visualization of query embeddings.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 10.0 + }, + { + "type": "image", + "bbox": [ + 369, + 87, + 474, + 259 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 369, + 87, + 474, + 259 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 369, + 87, + 474, + 259 + ], + "spans": [ + { + "bbox": [ + 369, + 87, + 474, + 259 + ], + "score": 0.948, + "type": "image", + "image_path": "5ea9b9c596510988cdcbdc7742ad1ba606263fb7aae0db82cd463ab8f11bfb27.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 369, + 87, + 474, + 259 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 362, + 272, + 487, + 283 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 361, + 271, + 487, + 285 + ], + "spans": [ + { + "bbox": [ + 361, + 271, + 487, + 285 + ], + "score": 1.0, + "content": "Figure 4: Different techniques.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + }, + { + "type": "image", + "bbox": [ + 112, + 303, + 495, + 402 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 303, + 495, + 402 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 303, + 495, + 402 + ], + "spans": [ + { + "bbox": [ + 112, + 303, + 495, + 402 + ], + "score": 0.948, + "type": "image", + "image_path": "6f60df92f6ce68b3bc8a05d04a9fa957d6cb5bab0fe7d16ef5692b5aaa72f95a.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 112, + 303, + 495, + 336.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 112, + 336.0, + 495, + 369.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 112, + 369.0, + 495, + 402.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 198, + 418, + 412, + 430 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 198, + 417, + 414, + 431 + ], + "spans": [ + { + "bbox": [ + 198, + 417, + 414, + 431 + ], + "score": 1.0, + "content": "Figure 5: Delta Reward Distribution in Evaluation Phrase.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 18.0 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 518 + ], + "lines": [ + { + "bbox": [ + 105, + 452, + 505, + 465 + ], + "spans": [ + { + "bbox": [ + 105, + 452, + 321, + 465 + ], + "score": 1.0, + "content": "notice that there is a main cluster at the center (about", + "type": "text" + }, + { + "bbox": [ + 321, + 452, + 361, + 464 + ], + "score": 0.92, + "content": "( - 7 , - 1 ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 361, + 452, + 505, + 465 + ], + "score": 1.0, + "content": ") that gathers most queries on which", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "spans": [ + { + "bbox": [ + 105, + 463, + 505, + 475 + ], + "score": 1.0, + "content": "our agent does not perform well. That cluster contains some random inputs from the users, which", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 106, + 475, + 504, + 486 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 504, + 486 + ], + "score": 1.0, + "content": "may have some typos. For the queries that the agent outperforms the baseline, we find they are quite", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 486, + 504, + 497 + ], + "spans": [ + { + "bbox": [ + 106, + 486, + 504, + 497 + ], + "score": 1.0, + "content": "scattered. We guess the reason is that the baseline with hand-crafted rules do not consider the em-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 105, + 495, + 506, + 509 + ], + "score": 1.0, + "content": "beddings, thus it is less affected by embeddings. This may suggest us to look for more informative", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 506, + 376, + 520 + ], + "spans": [ + { + "bbox": [ + 106, + 506, + 376, + 520 + ], + "score": 1.0, + "content": "embeddings of queries to better distinguish some clustered queries.", + "type": "text" + } + ], + "index": 25 + } + ], + "index": 22.5 + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "In the plots shown in Figure 6, we visualize the distribution of delta rewards, ranking scores and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "seek count values comparing to the well-designed production rules. For most of the queries the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "agent learns promising match plan without any prior knowledge just based on the reward signals.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 428, + 570 + ], + "score": 1.0, + "content": "There are also some hard queries to find a pattern that have poor reward (about", + "type": "text" + }, + { + "bbox": [ + 428, + 557, + 453, + 568 + ], + "score": 0.57, + "content": "- 1 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "and usually", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "lies in the main cluster and around a few more small clusters. Other than some meaningless queries", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "around the center, they also include some non-English queries, such as French and Chinese, which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 499, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 499, + 603 + ], + "score": 1.0, + "content": "are possibly clustered around their centroid and are challenging to learn the policy for all of them.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29 + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "Reward design. We consider different weights to trade-off between the query latency from the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "input and the quality of the returned documents. The quality is evaluated by Ranking Scores of top", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "score": 1.0, + "content": "five documents. We found that it is sparse since the good documents are hard to match. When we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 649, + 388, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 388, + 661 + ], + "score": 1.0, + "content": "weight more on ranking score, the agent tends to stop the search early.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5 + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We also investigate a few punishments and try to use less of them to avoid manual design. We found", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "the agent learns some common patterns guided by the reward and punishments. It tries to avoid", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "”stop” at the first step because we set a huge punishment on such invalid stop. This punishment is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "necessary since it prevents the agent stops at first step for better value than trying more steps but", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "get no documents matched. In other words, the agent is encouraged to explore different rewrites", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "instead of using empty match plans. Without such punishment, the performance significantly drops", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 39.5 + } + ], + "page_idx": 5, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 302, + 752, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "6", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 82, + 335, + 251 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 110, + 82, + 335, + 251 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 110, + 82, + 335, + 251 + ], + "spans": [ + { + "bbox": [ + 110, + 82, + 335, + 251 + ], + "score": 0.95, + "type": "image", + "image_path": "1b912e4438fe0c4576fc52e07dd81917a7cf3148fcc9fac32a9a96a543cd4db8.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 110, + 82, + 335, + 95.0 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 110, + 95.0, + 335, + 108.0 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 110, + 108.0, + 335, + 121.0 + ], + "spans": [], + "index": 2 + }, + { + "bbox": [ + 110, + 121.0, + 335, + 134.0 + ], + "spans": [], + "index": 3 + }, + { + "bbox": [ + 110, + 134.0, + 335, + 147.0 + ], + "spans": [], + "index": 4 + }, + { + "bbox": [ + 110, + 147.0, + 335, + 160.0 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 110, + 160.0, + 335, + 173.0 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 110, + 173.0, + 335, + 186.0 + ], + "spans": [], + "index": 7 + }, + { + "bbox": [ + 110, + 186.0, + 335, + 199.0 + ], + "spans": [], + "index": 8 + }, + { + "bbox": [ + 110, + 199.0, + 335, + 212.0 + ], + "spans": [], + "index": 9 + }, + { + "bbox": [ + 110, + 212.0, + 335, + 225.0 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 110, + 225.0, + 335, + 238.0 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 110, + 238.0, + 335, + 251.0 + ], + "spans": [], + "index": 12 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 115, + 272, + 297, + 284 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 114, + 270, + 298, + 286 + ], + "spans": [ + { + "bbox": [ + 114, + 270, + 298, + 286 + ], + "score": 1.0, + "content": "Figure 3: Visualization of query embeddings.", + "type": "text" + } + ], + "index": 14 + } + ], + "index": 14 + } + ], + "index": 10.0 + }, + { + "type": "image", + "bbox": [ + 369, + 87, + 474, + 259 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 369, + 87, + 474, + 259 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 369, + 87, + 474, + 259 + ], + "spans": [ + { + "bbox": [ + 369, + 87, + 474, + 259 + ], + "score": 0.948, + "type": "image", + "image_path": "5ea9b9c596510988cdcbdc7742ad1ba606263fb7aae0db82cd463ab8f11bfb27.jpg" + } + ] + } + ], + "index": 13, + "virtual_lines": [ + { + "bbox": [ + 369, + 87, + 474, + 259 + ], + "spans": [], + "index": 13 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 362, + 272, + 487, + 283 + ], + "group_id": 2, + "lines": [ + { + "bbox": [ + 361, + 271, + 487, + 285 + ], + "spans": [ + { + "bbox": [ + 361, + 271, + 487, + 285 + ], + "score": 1.0, + "content": "Figure 4: Different techniques.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 15 + } + ], + "index": 14.0 + }, + { + "type": "image", + "bbox": [ + 112, + 303, + 495, + 402 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 112, + 303, + 495, + 402 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 112, + 303, + 495, + 402 + ], + "spans": [ + { + "bbox": [ + 112, + 303, + 495, + 402 + ], + "score": 0.948, + "type": "image", + "image_path": "6f60df92f6ce68b3bc8a05d04a9fa957d6cb5bab0fe7d16ef5692b5aaa72f95a.jpg" + } + ] + } + ], + "index": 17, + "virtual_lines": [ + { + "bbox": [ + 112, + 303, + 495, + 336.0 + ], + "spans": [], + "index": 16 + }, + { + "bbox": [ + 112, + 336.0, + 495, + 369.0 + ], + "spans": [], + "index": 17 + }, + { + "bbox": [ + 112, + 369.0, + 495, + 402.0 + ], + "spans": [], + "index": 18 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 198, + 418, + 412, + 430 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 198, + 417, + 414, + 431 + ], + "spans": [ + { + "bbox": [ + 198, + 417, + 414, + 431 + ], + "score": 1.0, + "content": "Figure 5: Delta Reward Distribution in Evaluation Phrase.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + } + ], + "index": 18.0 + }, + { + "type": "text", + "bbox": [ + 107, + 452, + 505, + 518 + ], + "lines": [], + "index": 22.5, + "bbox_fs": [ + 105, + 452, + 506, + 520 + ], + "lines_deleted": true + }, + { + "type": "text", + "bbox": [ + 107, + 524, + 505, + 601 + ], + "lines": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "spans": [ + { + "bbox": [ + 105, + 524, + 505, + 537 + ], + "score": 1.0, + "content": "In the plots shown in Figure 6, we visualize the distribution of delta rewards, ranking scores and", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "spans": [ + { + "bbox": [ + 106, + 536, + 505, + 547 + ], + "score": 1.0, + "content": "seek count values comparing to the well-designed production rules. For most of the queries the", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 546, + 505, + 559 + ], + "score": 1.0, + "content": "agent learns promising match plan without any prior knowledge just based on the reward signals.", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 557, + 505, + 570 + ], + "spans": [ + { + "bbox": [ + 106, + 557, + 428, + 570 + ], + "score": 1.0, + "content": "There are also some hard queries to find a pattern that have poor reward (about", + "type": "text" + }, + { + "bbox": [ + 428, + 557, + 453, + 568 + ], + "score": 0.57, + "content": "- 1 0 0 )", + "type": "inline_equation" + }, + { + "bbox": [ + 454, + 557, + 505, + 570 + ], + "score": 1.0, + "content": "and usually", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "spans": [ + { + "bbox": [ + 105, + 567, + 505, + 580 + ], + "score": 1.0, + "content": "lies in the main cluster and around a few more small clusters. Other than some meaningless queries", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 591 + ], + "score": 1.0, + "content": "around the center, they also include some non-English queries, such as French and Chinese, which", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 590, + 499, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 590, + 499, + 603 + ], + "score": 1.0, + "content": "are possibly clustered around their centroid and are challenging to learn the policy for all of them.", + "type": "text" + } + ], + "index": 32 + } + ], + "index": 29, + "bbox_fs": [ + 105, + 524, + 505, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 616, + 504, + 660 + ], + "lines": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "spans": [ + { + "bbox": [ + 105, + 615, + 505, + 628 + ], + "score": 1.0, + "content": "Reward design. We consider different weights to trade-off between the query latency from the", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "spans": [ + { + "bbox": [ + 105, + 626, + 506, + 640 + ], + "score": 1.0, + "content": "input and the quality of the returned documents. The quality is evaluated by Ranking Scores of top", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "spans": [ + { + "bbox": [ + 106, + 639, + 506, + 650 + ], + "score": 1.0, + "content": "five documents. We found that it is sparse since the good documents are hard to match. When we", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 649, + 388, + 661 + ], + "spans": [ + { + "bbox": [ + 105, + 649, + 388, + 661 + ], + "score": 1.0, + "content": "weight more on ranking score, the agent tends to stop the search early.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 34.5, + "bbox_fs": [ + 105, + 615, + 506, + 661 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 666, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 665, + 505, + 678 + ], + "score": 1.0, + "content": "We also investigate a few punishments and try to use less of them to avoid manual design. We found", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "the agent learns some common patterns guided by the reward and punishments. It tries to avoid", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "”stop” at the first step because we set a huge punishment on such invalid stop. This punishment is", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 700, + 505, + 711 + ], + "score": 1.0, + "content": "necessary since it prevents the agent stops at first step for better value than trying more steps but", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 709, + 506, + 722 + ], + "score": 1.0, + "content": "get no documents matched. In other words, the agent is encouraged to explore different rewrites", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 505, + 733 + ], + "score": 1.0, + "content": "instead of using empty match plans. Without such punishment, the performance significantly drops", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "since it is hard to explore good documents at first. We also punish the agent to avoid unsupported", + "type": "text", + "cross_page": true + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "score": 1.0, + "content": "rewrites, since the production system uses designed rules for each class of the queries and simply", + "type": "text", + "cross_page": true + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "omits a rewrite without any useful feedback. We test punishing repeated rewrites, however, it helps", + "type": "text", + "cross_page": true + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 317, + 336, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 336, + 329 + ], + "score": 1.0, + "content": "seek count (more efficient) but may harm ranking scores.", + "type": "text", + "cross_page": true + } + ], + "index": 12 + } + ], + "index": 39.5, + "bbox_fs": [ + 105, + 665, + 506, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 80, + 467, + 127 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 142, + 80, + 467, + 127 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 80, + 467, + 127 + ], + "spans": [ + { + "bbox": [ + 142, + 80, + 467, + 127 + ], + "score": 0.975, + "html": "
OverallRanking Score (Quality)Seek Count (Efficiency)
OU Noise+1.00+1.20-0.20
Param Noise+1.69+0.82-0.87
", + "type": "table", + "image_path": "e3a1464f0b6cd2d71b46e2251154e906a70ddf2e498fb4263bde557589df6bb7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 142, + 80, + 467, + 95.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 142, + 95.66666666666667, + 467, + 111.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 142, + 111.33333333333334, + 467, + 127.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 139, + 505, + 162 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 138, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 152 + ], + "score": 1.0, + "content": "Table 1: Improved Scores of Different Noise Compared to RNN Gaussian Noise. Ranking Score", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 283, + 163 + ], + "score": 1.0, + "content": "and Seek Count are scaled to match Overall", + "type": "text" + }, + { + "bbox": [ + 283, + 151, + 296, + 160 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 149, + 505, + 163 + ], + "score": 1.0, + "content": "RankingScore−SeekCount for easy comparison.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 115, + 176, + 496, + 234 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 115, + 176, + 496, + 234 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 176, + 496, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 176, + 496, + 234 + ], + "score": 0.978, + "html": "
OverallRanking Score (Quality)Seek Count (Efficiency)
MLP+Prioritized+14.54+8.56-5.98
RNN+Prioritized+32.63+14.44-18.19
RNN + Prioritized Sample+32.67+15.29-17.38
", + "type": "table", + "image_path": "7320f38c5eca22be88d47fec5b5f6b8be13097a72a46698ca9f2671a3e8e7c6c.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 115, + 176, + 496, + 195.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 115, + 195.33333333333334, + 496, + 214.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 214.66666666666669, + 496, + 234.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 174, + 246, + 435, + 259 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 175, + 246, + 436, + 261 + ], + "spans": [ + { + "bbox": [ + 175, + 246, + 436, + 261 + ], + "score": 1.0, + "content": "Table 2: Improved Scores of Different Model Compared to MLP", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 328 + ], + "lines": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "spans": [ + { + "bbox": [ + 106, + 284, + 505, + 296 + ], + "score": 1.0, + "content": "since it is hard to explore good documents at first. We also punish the agent to avoid unsupported", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 294, + 505, + 308 + ], + "score": 1.0, + "content": "rewrites, since the production system uses designed rules for each class of the queries and simply", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "spans": [ + { + "bbox": [ + 106, + 307, + 505, + 318 + ], + "score": 1.0, + "content": "omits a rewrite without any useful feedback. We test punishing repeated rewrites, however, it helps", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 317, + 336, + 329 + ], + "spans": [ + { + "bbox": [ + 105, + 317, + 336, + 329 + ], + "score": 1.0, + "content": "seek count (more efficient) but may harm ranking scores.", + "type": "text" + } + ], + "index": 12 + } + ], + "index": 10.5 + }, + { + "type": "title", + "bbox": [ + 108, + 347, + 208, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 210, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 210, + 360 + ], + "score": 1.0, + "content": "4.3 ABLATION STUDY", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "We compare different exploration noise, including action space noise and parameter space noise.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "For the action space noise, the noise on discrete and continuous actions are applied separately. We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "test Gaussian and OrnsteinUhlenbeck noise on continuous actions, while the discrete actions only", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 123, + 416 + ], + "score": 1.0, + "content": "use", + "type": "text" + }, + { + "bbox": [ + 123, + 405, + 128, + 414 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 403, + 330, + 416 + ], + "score": 1.0, + "content": "-greedy. To keep the comparison fair, we also test", + "type": "text" + }, + { + "bbox": [ + 331, + 405, + 336, + 414 + ], + "score": 0.61, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "-greedy strategy on continuous actions by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 289, + 427 + ], + "score": 1.0, + "content": "uniformly sampling a point with probability", + "type": "text" + }, + { + "bbox": [ + 289, + 416, + 295, + 424 + ], + "score": 0.41, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 414, + 505, + 427 + ], + "score": 1.0, + "content": ". The parameter noise is directly applied on policy", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 141, + 438 + ], + "score": 1.0, + "content": "network", + "type": "text" + }, + { + "bbox": [ + 142, + 425, + 175, + 438 + ], + "score": 0.93, + "content": "\\widetilde \\mu ( s ; \\theta _ { \\mu } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 425, + 505, + 438 + ], + "score": 1.0, + "content": ". We found the parameter space noise performs more stable than all types of action", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 428, + 449 + ], + "score": 1.0, + "content": "espace noise. The uniform sampling on both discrete and continous action with", + "type": "text" + }, + { + "bbox": [ + 428, + 438, + 433, + 446 + ], + "score": 0.66, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "-greedy failed for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "sometimes, thus we did not include it in comparison. We give the relative improvement in Table", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "1 compared to RNN with Gaussian Noise. The results show that parameter space noise has best", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 470, + 191, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 191, + 480 + ], + "score": 1.0, + "content": "performance overall.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "The results in Figure 6(a) and Table 2 show that RNN got significantly better performance than MLP", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 497, + 504, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 504, + 508 + ], + "score": 1.0, + "content": "with or without parameterization on our environment. It indicates that the environment is highly", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "non-Markov. We guess one obvious possibility is that the system signals of the state space cannot", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "include all the information, while recurrent networks try to extract latent state from the history. Note", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "that each episode will sample one query, thus the query embedding is constant for all steps in one", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "episode. Another possible reason is that some rewrites may not be supported by special queries or", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 552, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 279, + 564 + ], + "score": 1.0, + "content": "have too small change for the state signals.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27 + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "We compare different parameterization methods on recurrent networks. The results are shown in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "Figure 4(b) and Table 2. We empirically found that our modified strategy using reward bin is more", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "stable from the beginning. It is possible that the agent repeatedly replays not only good experience", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 602, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 613 + ], + "score": 1.0, + "content": "on matching documents, but also learns to avoid punishment we set, as we expect. In general", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "score": 1.0, + "content": "prioritized replay, the sampling probability is just based on TD-error and may be biased to worse", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 623, + 388, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 388, + 636 + ], + "score": 1.0, + "content": "samples since value function may not update towards possible results.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5 + }, + { + "type": "title", + "bbox": [ + 108, + 653, + 238, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 240, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 240, + 666 + ], + "score": 1.0, + "content": "4.4 BENCHMARKING GAMES", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "We experiment on Platform-v0 and Goal-v0 from Bester et al. (2019). Note that, in all these games,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "each discrete action has a separate continuous action-parameter space. Our algorithm is designed for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "shared action-parameters and does not utilize such prior, thus we do not compare with P-DQN-style", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "algorithms. We apply n-step return since the episode in these games is much longer. We assume the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 434, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 434, + 733 + ], + "score": 1.0, + "content": "environment is fully observable and do not use recurrent networks in comparison.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40 + } + ], + "page_idx": 6, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 303, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 750, + 309, + 762 + ], + "score": 1.0, + "content": "7", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "table", + "bbox": [ + 142, + 80, + 467, + 127 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 142, + 80, + 467, + 127 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 142, + 80, + 467, + 127 + ], + "spans": [ + { + "bbox": [ + 142, + 80, + 467, + 127 + ], + "score": 0.975, + "html": "
OverallRanking Score (Quality)Seek Count (Efficiency)
OU Noise+1.00+1.20-0.20
Param Noise+1.69+0.82-0.87
", + "type": "table", + "image_path": "e3a1464f0b6cd2d71b46e2251154e906a70ddf2e498fb4263bde557589df6bb7.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 142, + 80, + 467, + 95.66666666666667 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 142, + 95.66666666666667, + 467, + 111.33333333333334 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 142, + 111.33333333333334, + 467, + 127.00000000000001 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 107, + 139, + 505, + 162 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 138, + 505, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 138, + 505, + 152 + ], + "score": 1.0, + "content": "Table 1: Improved Scores of Different Noise Compared to RNN Gaussian Noise. Ranking Score", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 149, + 505, + 163 + ], + "spans": [ + { + "bbox": [ + 105, + 149, + 283, + 163 + ], + "score": 1.0, + "content": "and Seek Count are scaled to match Overall", + "type": "text" + }, + { + "bbox": [ + 283, + 151, + 296, + 160 + ], + "score": 0.57, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 296, + 149, + 505, + 163 + ], + "score": 1.0, + "content": "RankingScore−SeekCount for easy comparison.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "table", + "bbox": [ + 115, + 176, + 496, + 234 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 115, + 176, + 496, + 234 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 115, + 176, + 496, + 234 + ], + "spans": [ + { + "bbox": [ + 115, + 176, + 496, + 234 + ], + "score": 0.978, + "html": "
OverallRanking Score (Quality)Seek Count (Efficiency)
MLP+Prioritized+14.54+8.56-5.98
RNN+Prioritized+32.63+14.44-18.19
RNN + Prioritized Sample+32.67+15.29-17.38
", + "type": "table", + "image_path": "7320f38c5eca22be88d47fec5b5f6b8be13097a72a46698ca9f2671a3e8e7c6c.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 115, + 176, + 496, + 195.33333333333334 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 115, + 195.33333333333334, + 496, + 214.66666666666669 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 115, + 214.66666666666669, + 496, + 234.00000000000003 + ], + "spans": [], + "index": 7 + } + ] + }, + { + "type": "table_caption", + "bbox": [ + 174, + 246, + 435, + 259 + ], + "group_id": 1, + "lines": [ + { + "bbox": [ + 175, + 246, + 436, + 261 + ], + "spans": [ + { + "bbox": [ + 175, + 246, + 436, + 261 + ], + "score": 1.0, + "content": "Table 2: Improved Scores of Different Model Compared to MLP", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + } + ], + "index": 7.0 + }, + { + "type": "text", + "bbox": [ + 107, + 284, + 505, + 328 + ], + "lines": [], + "index": 10.5, + "bbox_fs": [ + 105, + 284, + 505, + 329 + ], + "lines_deleted": true + }, + { + "type": "title", + "bbox": [ + 108, + 347, + 208, + 358 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 210, + 360 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 210, + 360 + ], + "score": 1.0, + "content": "4.3 ABLATION STUDY", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 13 + }, + { + "type": "text", + "bbox": [ + 107, + 370, + 505, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "spans": [ + { + "bbox": [ + 105, + 369, + 505, + 383 + ], + "score": 1.0, + "content": "We compare different exploration noise, including action space noise and parameter space noise.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "spans": [ + { + "bbox": [ + 105, + 381, + 505, + 394 + ], + "score": 1.0, + "content": "For the action space noise, the noise on discrete and continuous actions are applied separately. We", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "spans": [ + { + "bbox": [ + 105, + 392, + 505, + 405 + ], + "score": 1.0, + "content": "test Gaussian and OrnsteinUhlenbeck noise on continuous actions, while the discrete actions only", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 105, + 403, + 505, + 416 + ], + "spans": [ + { + "bbox": [ + 105, + 403, + 123, + 416 + ], + "score": 1.0, + "content": "use", + "type": "text" + }, + { + "bbox": [ + 123, + 405, + 128, + 414 + ], + "score": 0.64, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 128, + 403, + 330, + 416 + ], + "score": 1.0, + "content": "-greedy. To keep the comparison fair, we also test", + "type": "text" + }, + { + "bbox": [ + 331, + 405, + 336, + 414 + ], + "score": 0.61, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 336, + 403, + 505, + 416 + ], + "score": 1.0, + "content": "-greedy strategy on continuous actions by", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 105, + 414, + 505, + 427 + ], + "spans": [ + { + "bbox": [ + 105, + 414, + 289, + 427 + ], + "score": 1.0, + "content": "uniformly sampling a point with probability", + "type": "text" + }, + { + "bbox": [ + 289, + 416, + 295, + 424 + ], + "score": 0.41, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 295, + 414, + 505, + 427 + ], + "score": 1.0, + "content": ". The parameter noise is directly applied on policy", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 425, + 505, + 438 + ], + "spans": [ + { + "bbox": [ + 105, + 425, + 141, + 438 + ], + "score": 1.0, + "content": "network", + "type": "text" + }, + { + "bbox": [ + 142, + 425, + 175, + 438 + ], + "score": 0.93, + "content": "\\widetilde \\mu ( s ; \\theta _ { \\mu } )", + "type": "inline_equation" + }, + { + "bbox": [ + 176, + 425, + 505, + 438 + ], + "score": 1.0, + "content": ". We found the parameter space noise performs more stable than all types of action", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 435, + 506, + 449 + ], + "spans": [ + { + "bbox": [ + 105, + 435, + 428, + 449 + ], + "score": 1.0, + "content": "espace noise. The uniform sampling on both discrete and continous action with", + "type": "text" + }, + { + "bbox": [ + 428, + 438, + 433, + 446 + ], + "score": 0.66, + "content": "\\epsilon", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 435, + 506, + 449 + ], + "score": 1.0, + "content": "-greedy failed for", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "sometimes, thus we did not include it in comparison. We give the relative improvement in Table", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "spans": [ + { + "bbox": [ + 105, + 458, + 506, + 471 + ], + "score": 1.0, + "content": "1 compared to RNN with Gaussian Noise. The results show that parameter space noise has best", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 470, + 191, + 480 + ], + "spans": [ + { + "bbox": [ + 105, + 470, + 191, + 480 + ], + "score": 1.0, + "content": "performance overall.", + "type": "text" + } + ], + "index": 23 + } + ], + "index": 18.5, + "bbox_fs": [ + 105, + 369, + 506, + 480 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 486, + 505, + 563 + ], + "lines": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "spans": [ + { + "bbox": [ + 105, + 485, + 505, + 498 + ], + "score": 1.0, + "content": "The results in Figure 6(a) and Table 2 show that RNN got significantly better performance than MLP", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 106, + 497, + 504, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 497, + 504, + 508 + ], + "score": 1.0, + "content": "with or without parameterization on our environment. It indicates that the environment is highly", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "spans": [ + { + "bbox": [ + 105, + 508, + 505, + 520 + ], + "score": 1.0, + "content": "non-Markov. We guess one obvious possibility is that the system signals of the state space cannot", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "spans": [ + { + "bbox": [ + 105, + 519, + 505, + 531 + ], + "score": 1.0, + "content": "include all the information, while recurrent networks try to extract latent state from the history. Note", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 105, + 529, + 505, + 542 + ], + "score": 1.0, + "content": "that each episode will sample one query, thus the query embedding is constant for all steps in one", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "spans": [ + { + "bbox": [ + 105, + 541, + 505, + 554 + ], + "score": 1.0, + "content": "episode. Another possible reason is that some rewrites may not be supported by special queries or", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 106, + 552, + 279, + 564 + ], + "spans": [ + { + "bbox": [ + 106, + 552, + 279, + 564 + ], + "score": 1.0, + "content": "have too small change for the state signals.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 27, + "bbox_fs": [ + 105, + 485, + 505, + 564 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 505, + 635 + ], + "lines": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "spans": [ + { + "bbox": [ + 106, + 568, + 505, + 581 + ], + "score": 1.0, + "content": "We compare different parameterization methods on recurrent networks. The results are shown in", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "spans": [ + { + "bbox": [ + 105, + 579, + 505, + 592 + ], + "score": 1.0, + "content": "Figure 4(b) and Table 2. We empirically found that our modified strategy using reward bin is more", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "spans": [ + { + "bbox": [ + 105, + 591, + 505, + 603 + ], + "score": 1.0, + "content": "stable from the beginning. It is possible that the agent repeatedly replays not only good experience", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 602, + 505, + 613 + ], + "spans": [ + { + "bbox": [ + 105, + 602, + 505, + 613 + ], + "score": 1.0, + "content": "on matching documents, but also learns to avoid punishment we set, as we expect. In general", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "spans": [ + { + "bbox": [ + 105, + 612, + 506, + 626 + ], + "score": 1.0, + "content": "prioritized replay, the sampling probability is just based on TD-error and may be biased to worse", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 105, + 623, + 388, + 636 + ], + "spans": [ + { + "bbox": [ + 105, + 623, + 388, + 636 + ], + "score": 1.0, + "content": "samples since value function may not update towards possible results.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 33.5, + "bbox_fs": [ + 105, + 568, + 506, + 636 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 653, + 238, + 665 + ], + "lines": [ + { + "bbox": [ + 105, + 652, + 240, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 652, + 240, + 666 + ], + "score": 1.0, + "content": "4.4 BENCHMARKING GAMES", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 676, + 504, + 731 + ], + "lines": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 106, + 676, + 505, + 690 + ], + "score": 1.0, + "content": "We experiment on Platform-v0 and Goal-v0 from Bester et al. (2019). Note that, in all these games,", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 106, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "each discrete action has a separate continuous action-parameter space. Our algorithm is designed for", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "spans": [ + { + "bbox": [ + 106, + 699, + 505, + 711 + ], + "score": 1.0, + "content": "shared action-parameters and does not utilize such prior, thus we do not compare with P-DQN-style", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "spans": [ + { + "bbox": [ + 106, + 709, + 505, + 723 + ], + "score": 1.0, + "content": "algorithms. We apply n-step return since the episode in these games is much longer. We assume the", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 721, + 434, + 733 + ], + "spans": [ + { + "bbox": [ + 106, + 721, + 434, + 733 + ], + "score": 1.0, + "content": "environment is fully observable and do not use recurrent networks in comparison.", + "type": "text" + } + ], + "index": 42 + } + ], + "index": 40, + "bbox_fs": [ + 106, + 676, + 505, + 733 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Hyperparameters. We examine the games with different combinations of hyperparameters, since", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "they are easy to parallize on each training nodes without the need to connect to an production envi-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "ronment emulator. With NNI, the learning rates for value and policy network are set to log-uniform", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 113, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 113, + 117, + 129 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 115, + 198, + 127 + ], + "score": 0.58, + "content": "[ 3 \\times 1 0 ^ { - 4 } , 3 \\times 1 0 ^ { - 3 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 113, + 215, + 129 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 216, + 115, + 296, + 127 + ], + "score": 0.89, + "content": "[ 1 \\times 1 0 ^ { - 4 } , 1 \\times \\mathsf { \\bar { 1 } 0 ^ { - 3 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 113, + 318, + 129 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 318, + 117, + 326, + 125 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 113, + 505, + 129 + ], + "score": 1.0, + "content": "in parameterized replay is set to log-uniform", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 271, + 139 + ], + "score": 1.0, + "content": "in [0.2, 1.0]. The action noise threshold", + "type": "text" + }, + { + "bbox": [ + 271, + 127, + 278, + 136 + ], + "score": 0.71, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "in parameter space noise is set to log-uniform between", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 465, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 272, + 150 + ], + "score": 1.0, + "content": "[0.05, 0.8]. Tn soft parameter update, the", + "type": "text" + }, + { + "bbox": [ + 272, + 139, + 279, + 147 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 136, + 374, + 150 + ], + "score": 1.0, + "content": "is set to log-uniform in", + "type": "text" + }, + { + "bbox": [ + 374, + 137, + 460, + 149 + ], + "score": 0.91, + "content": "[ 1 \\times 1 0 ^ { - 3 } , 1 \\times 1 0 ^ { - 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 136, + 465, + 150 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5 + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "In Platform-v0 and Goal-v0, we found the agent is not very sensitive to the range we set, such", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 118, + 177 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 118, + 166, + 137, + 176 + ], + "score": 0.25, + "content": "\\alpha , \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 165, + 156, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 167, + 163, + 175 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 165, + 188, + 177 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 188, + 166, + 195, + 175 + ], + "score": 0.76, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 165, + 252, + 177 + ], + "score": 1.0, + "content": "values of top", + "type": "text" + }, + { + "bbox": [ + 252, + 165, + 272, + 176 + ], + "score": 0.86, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 165, + 425, + 177 + ], + "score": 1.0, + "content": "trials vary between [0.05, 0.3], while", + "type": "text" + }, + { + "bbox": [ + 426, + 167, + 434, + 175 + ], + "score": 0.78, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 165, + 453, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 453, + 167, + 460, + 175 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "values are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 504, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 504, + 189 + ], + "score": 1.0, + "content": "evenly scattered in the defined range. In these settings, usually the agent prefers slightly larger value", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 261, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 261, + 200 + ], + "score": 1.0, + "content": "learning rate than policy learning rate.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5 + }, + { + "type": "table", + "bbox": [ + 212, + 211, + 397, + 257 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 212, + 211, + 397, + 257 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 211, + 397, + 257 + ], + "spans": [ + { + "bbox": [ + 212, + 211, + 397, + 257 + ], + "score": 0.972, + "html": "
Average Eval ReturnOurPA-DDPG
Platform-v00.95730.3113
Goal-v034.20-6.208
", + "type": "table", + "image_path": "43fb6a7fec969297f5ab873d3137dea0d0470694b6859f537189d55399ea27e8.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 212, + 211, + 397, + 226.33333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 212, + 226.33333333333334, + 397, + 241.66666666666669 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 212, + 241.66666666666669, + 397, + 257.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 270, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "Table 3: Average evaluation results (the average of all training rewards and final evaluation reward)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "on benchmarks Platform-v0 and Goal-v0 with PA-DDPG (Hausknecht & Stone (2015)), MP-DQN", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "(Bester et al. (2019)) and P-DQN (Xiong et al. (2018)). We use reported number from the papers,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 303, + 337, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 337, + 315 + ], + "score": 1.0, + "content": "while last two methods report another metric on Goal-v0.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5 + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 211, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 213, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 213, + 356 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "While the aforementioned algorithms and techniques work on discrete or continuous action spaces,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "it is not trivial to apply them on parameterized action space, since such discrete-continuous hybrid", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "action space is hard to parameterized by a single distribution. A related series of work is to combine", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "DDPG and DQN to optimize Q-value function on parameter actions. There are two classes of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "methods that belong to them: PA-DDPG (Hausknecht & Stone (2015)) based on DDPG and P-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "DQN (Xiong et al. (2018)) based on DQN. They are Q-Learning-based methods which select the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "best action by maximizing the Q-value function on discrete action space or learning a deterministic", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 444, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 458 + ], + "score": 1.0, + "content": "policy outputting best continuous action. However, they use different strategies to combine discrete", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "and continuous actions. Bester et al. (2019) (MP-DQN) extends P-DQN to tackle the problem", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 465, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 356, + 480 + ], + "score": 1.0, + "content": "that Q-value is a function of the joint action-parameter vector", + "type": "text" + }, + { + "bbox": [ + 356, + 466, + 424, + 479 + ], + "score": 0.92, + "content": "Q ( s ^ { \\prime } , k ^ { \\prime } , { \\bf x } ^ { \\tilde { Q } } ( s ^ { \\prime } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 465, + 506, + 480 + ], + "score": 1.0, + "content": "in normal PAMDP,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "which may results in fault gradients. However, such problem does not exist in our slightly modified", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 275, + 501 + ], + "score": 1.0, + "content": "setting, since the action-parameter space", + "type": "text" + }, + { + "bbox": [ + 275, + 489, + 285, + 498 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 489, + 506, + 501 + ], + "score": 1.0, + "content": "in the match plan generation is inherently defined to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 497, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 187, + 514 + ], + "score": 1.0, + "content": "be shared for each", + "type": "text" + }, + { + "bbox": [ + 187, + 500, + 224, + 511 + ], + "score": 0.92, + "content": "k \\in \\mathcal { A } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 497, + 506, + 514 + ], + "score": 1.0, + "content": ". Masson et al. (2016) purposes a method to iteratively optimizing", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "discrete and continuous actions by alternating between them. Another perspective (Klimek et al.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 520, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 535 + ], + "score": 1.0, + "content": "(2017); Wei et al. (2018); Fu et al. (2019)) for a parameterized action space is to regard it as a two-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "hierarchy action space, where an agent selects discrete action first and continuous parameter later.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "However, we do not consider this direction in current scheme because we share the same parameters", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 553, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 568 + ], + "score": 1.0, + "content": "for all discrete action in not very large scale. Therefore, the hierarchical methods may not bring a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 566, + 226, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 226, + 579 + ], + "score": 1.0, + "content": "significant performance gain.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 27 + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 196, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 198, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 198, + 611 + ], + "score": 1.0, + "content": "6 DISCUSSIONS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "In this paper, we present a parameterized action RL match plan generation method which extends the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "plan generation to the general case without any predefined knowledge. Key to address the problem", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "are normalized softmax values of discrete actions to enable gradients backpropagation, parameter", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "space noise on parameters of the policy for unifying the exploration direction in both discrete and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 667, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 678 + ], + "score": 1.0, + "content": "continuous spaces, and recurrent deterministic policies with prioritized replay buffer to accelerate", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "and stabilize the training. Our algorithm can be applied to not only the match plan generation envi-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "ronment, but also other similar parameteried action environments. The experiment results demon-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "strate our method outperforms the well-designed hand-crafted rules in Bing and serveral baseline", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "results in some existing PARL benchmarks. In this paper, we mainly discuss about offline training", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 459, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 459, + 734 + ], + "score": 1.0, + "content": "procedure. In the future, we plan to apply learned policy to the production environment.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 42.5 + } + ], + "page_idx": 7, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 308, + 760 + ], + "lines": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 309, + 761 + ], + "score": 1.0, + "content": "8", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 505, + 149 + ], + "lines": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 82, + 505, + 95 + ], + "score": 1.0, + "content": "Hyperparameters. We examine the games with different combinations of hyperparameters, since", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "spans": [ + { + "bbox": [ + 106, + 94, + 505, + 106 + ], + "score": 1.0, + "content": "they are easy to parallize on each training nodes without the need to connect to an production envi-", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "spans": [ + { + "bbox": [ + 105, + 104, + 505, + 117 + ], + "score": 1.0, + "content": "ronment emulator. With NNI, the learning rates for value and policy network are set to log-uniform", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 113, + 505, + 129 + ], + "spans": [ + { + "bbox": [ + 104, + 113, + 117, + 129 + ], + "score": 1.0, + "content": "in", + "type": "text" + }, + { + "bbox": [ + 117, + 115, + 198, + 127 + ], + "score": 0.58, + "content": "[ 3 \\times 1 0 ^ { - 4 } , 3 \\times 1 0 ^ { - 3 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 198, + 113, + 215, + 129 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 216, + 115, + 296, + 127 + ], + "score": 0.89, + "content": "[ 1 \\times 1 0 ^ { - 4 } , 1 \\times \\mathsf { \\bar { 1 } 0 ^ { - 3 } } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 297, + 113, + 318, + 129 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 318, + 117, + 326, + 125 + ], + "score": 0.8, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 327, + 113, + 505, + 129 + ], + "score": 1.0, + "content": "in parameterized replay is set to log-uniform", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 126, + 506, + 139 + ], + "spans": [ + { + "bbox": [ + 105, + 126, + 271, + 139 + ], + "score": 1.0, + "content": "in [0.2, 1.0]. The action noise threshold", + "type": "text" + }, + { + "bbox": [ + 271, + 127, + 278, + 136 + ], + "score": 0.71, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 278, + 126, + 506, + 139 + ], + "score": 1.0, + "content": "in parameter space noise is set to log-uniform between", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 136, + 465, + 150 + ], + "spans": [ + { + "bbox": [ + 105, + 136, + 272, + 150 + ], + "score": 1.0, + "content": "[0.05, 0.8]. Tn soft parameter update, the", + "type": "text" + }, + { + "bbox": [ + 272, + 139, + 279, + 147 + ], + "score": 0.78, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 279, + 136, + 374, + 150 + ], + "score": 1.0, + "content": "is set to log-uniform in", + "type": "text" + }, + { + "bbox": [ + 374, + 137, + 460, + 149 + ], + "score": 0.91, + "content": "[ 1 \\times 1 0 ^ { - 3 } , 1 \\times 1 0 ^ { - 2 } ]", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 136, + 465, + 150 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 2.5, + "bbox_fs": [ + 104, + 82, + 506, + 150 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 154, + 505, + 199 + ], + "lines": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "spans": [ + { + "bbox": [ + 105, + 154, + 505, + 166 + ], + "score": 1.0, + "content": "In Platform-v0 and Goal-v0, we found the agent is not very sensitive to the range we set, such", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 165, + 505, + 177 + ], + "spans": [ + { + "bbox": [ + 105, + 165, + 118, + 177 + ], + "score": 1.0, + "content": "as", + "type": "text" + }, + { + "bbox": [ + 118, + 166, + 137, + 176 + ], + "score": 0.25, + "content": "\\alpha , \\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 138, + 165, + 156, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 156, + 167, + 163, + 175 + ], + "score": 0.77, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 163, + 165, + 188, + 177 + ], + "score": 1.0, + "content": ". The", + "type": "text" + }, + { + "bbox": [ + 188, + 166, + 195, + 175 + ], + "score": 0.76, + "content": "\\delta", + "type": "inline_equation" + }, + { + "bbox": [ + 195, + 165, + 252, + 177 + ], + "score": 1.0, + "content": "values of top", + "type": "text" + }, + { + "bbox": [ + 252, + 165, + 272, + 176 + ], + "score": 0.86, + "content": "20 \\%", + "type": "inline_equation" + }, + { + "bbox": [ + 272, + 165, + 425, + 177 + ], + "score": 1.0, + "content": "trials vary between [0.05, 0.3], while", + "type": "text" + }, + { + "bbox": [ + 426, + 167, + 434, + 175 + ], + "score": 0.78, + "content": "\\alpha", + "type": "inline_equation" + }, + { + "bbox": [ + 434, + 165, + 453, + 177 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 453, + 167, + 460, + 175 + ], + "score": 0.76, + "content": "\\tau", + "type": "inline_equation" + }, + { + "bbox": [ + 461, + 165, + 505, + 177 + ], + "score": 1.0, + "content": "values are", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 177, + 504, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 177, + 504, + 189 + ], + "score": 1.0, + "content": "evenly scattered in the defined range. In these settings, usually the agent prefers slightly larger value", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 187, + 261, + 200 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 261, + 200 + ], + "score": 1.0, + "content": "learning rate than policy learning rate.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7.5, + "bbox_fs": [ + 105, + 154, + 505, + 200 + ] + }, + { + "type": "table", + "bbox": [ + 212, + 211, + 397, + 257 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 212, + 211, + 397, + 257 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 212, + 211, + 397, + 257 + ], + "spans": [ + { + "bbox": [ + 212, + 211, + 397, + 257 + ], + "score": 0.972, + "html": "
Average Eval ReturnOurPA-DDPG
Platform-v00.95730.3113
Goal-v034.20-6.208
", + "type": "table", + "image_path": "43fb6a7fec969297f5ab873d3137dea0d0470694b6859f537189d55399ea27e8.jpg" + } + ] + } + ], + "index": 11, + "virtual_lines": [ + { + "bbox": [ + 212, + 211, + 397, + 226.33333333333334 + ], + "spans": [], + "index": 10 + }, + { + "bbox": [ + 212, + 226.33333333333334, + 397, + 241.66666666666669 + ], + "spans": [], + "index": 11 + }, + { + "bbox": [ + 212, + 241.66666666666669, + 397, + 257.0 + ], + "spans": [], + "index": 12 + } + ] + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 106, + 270, + 505, + 314 + ], + "lines": [ + { + "bbox": [ + 106, + 270, + 505, + 282 + ], + "spans": [ + { + "bbox": [ + 106, + 270, + 505, + 282 + ], + "score": 1.0, + "content": "Table 3: Average evaluation results (the average of all training rewards and final evaluation reward)", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "spans": [ + { + "bbox": [ + 106, + 281, + 505, + 293 + ], + "score": 1.0, + "content": "on benchmarks Platform-v0 and Goal-v0 with PA-DDPG (Hausknecht & Stone (2015)), MP-DQN", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "spans": [ + { + "bbox": [ + 105, + 292, + 506, + 305 + ], + "score": 1.0, + "content": "(Bester et al. (2019)) and P-DQN (Xiong et al. (2018)). We use reported number from the papers,", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 303, + 337, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 303, + 337, + 315 + ], + "score": 1.0, + "content": "while last two methods report another metric on Goal-v0.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 270, + 506, + 315 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 341, + 211, + 354 + ], + "lines": [ + { + "bbox": [ + 105, + 340, + 213, + 356 + ], + "spans": [ + { + "bbox": [ + 105, + 340, + 213, + 356 + ], + "score": 1.0, + "content": "5 RELATED WORK", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 107, + 367, + 505, + 577 + ], + "lines": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 367, + 505, + 381 + ], + "score": 1.0, + "content": "While the aforementioned algorithms and techniques work on discrete or continuous action spaces,", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 379, + 505, + 392 + ], + "score": 1.0, + "content": "it is not trivial to apply them on parameterized action space, since such discrete-continuous hybrid", + "type": "text" + } + ], + "index": 19 + }, + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 402 + ], + "score": 1.0, + "content": "action space is hard to parameterized by a single distribution. A related series of work is to combine", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "spans": [ + { + "bbox": [ + 106, + 401, + 506, + 413 + ], + "score": 1.0, + "content": "DDPG and DQN to optimize Q-value function on parameter actions. There are two classes of", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 505, + 424 + ], + "score": 1.0, + "content": "methods that belong to them: PA-DDPG (Hausknecht & Stone (2015)) based on DDPG and P-", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 505, + 435 + ], + "score": 1.0, + "content": "DQN (Xiong et al. (2018)) based on DQN. They are Q-Learning-based methods which select the", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "spans": [ + { + "bbox": [ + 106, + 434, + 505, + 447 + ], + "score": 1.0, + "content": "best action by maximizing the Q-value function on discrete action space or learning a deterministic", + "type": "text" + } + ], + "index": 24 + }, + { + "bbox": [ + 105, + 444, + 506, + 458 + ], + "spans": [ + { + "bbox": [ + 105, + 444, + 506, + 458 + ], + "score": 1.0, + "content": "policy outputting best continuous action. However, they use different strategies to combine discrete", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "spans": [ + { + "bbox": [ + 106, + 456, + 505, + 468 + ], + "score": 1.0, + "content": "and continuous actions. Bester et al. (2019) (MP-DQN) extends P-DQN to tackle the problem", + "type": "text" + } + ], + "index": 26 + }, + { + "bbox": [ + 104, + 465, + 506, + 480 + ], + "spans": [ + { + "bbox": [ + 104, + 465, + 356, + 480 + ], + "score": 1.0, + "content": "that Q-value is a function of the joint action-parameter vector", + "type": "text" + }, + { + "bbox": [ + 356, + 466, + 424, + 479 + ], + "score": 0.92, + "content": "Q ( s ^ { \\prime } , k ^ { \\prime } , { \\bf x } ^ { \\tilde { Q } } ( s ^ { \\prime } ) )", + "type": "inline_equation" + }, + { + "bbox": [ + 424, + 465, + 506, + 480 + ], + "score": 1.0, + "content": "in normal PAMDP,", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "spans": [ + { + "bbox": [ + 106, + 478, + 505, + 491 + ], + "score": 1.0, + "content": "which may results in fault gradients. However, such problem does not exist in our slightly modified", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 106, + 489, + 506, + 501 + ], + "spans": [ + { + "bbox": [ + 106, + 489, + 275, + 501 + ], + "score": 1.0, + "content": "setting, since the action-parameter space", + "type": "text" + }, + { + "bbox": [ + 275, + 489, + 285, + 498 + ], + "score": 0.81, + "content": "\\mathcal { X }", + "type": "inline_equation" + }, + { + "bbox": [ + 286, + 489, + 506, + 501 + ], + "score": 1.0, + "content": "in the match plan generation is inherently defined to", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 105, + 497, + 506, + 514 + ], + "spans": [ + { + "bbox": [ + 105, + 497, + 187, + 514 + ], + "score": 1.0, + "content": "be shared for each", + "type": "text" + }, + { + "bbox": [ + 187, + 500, + 224, + 511 + ], + "score": 0.92, + "content": "k \\in \\mathcal { A } _ { d }", + "type": "inline_equation" + }, + { + "bbox": [ + 224, + 497, + 506, + 514 + ], + "score": 1.0, + "content": ". Masson et al. (2016) purposes a method to iteratively optimizing", + "type": "text" + } + ], + "index": 30 + }, + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "spans": [ + { + "bbox": [ + 105, + 510, + 505, + 524 + ], + "score": 1.0, + "content": "discrete and continuous actions by alternating between them. Another perspective (Klimek et al.", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 105, + 520, + 505, + 535 + ], + "spans": [ + { + "bbox": [ + 105, + 520, + 505, + 535 + ], + "score": 1.0, + "content": "(2017); Wei et al. (2018); Fu et al. (2019)) for a parameterized action space is to regard it as a two-", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "spans": [ + { + "bbox": [ + 106, + 533, + 505, + 545 + ], + "score": 1.0, + "content": "hierarchy action space, where an agent selects discrete action first and continuous parameter later.", + "type": "text" + } + ], + "index": 33 + }, + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "spans": [ + { + "bbox": [ + 105, + 542, + 505, + 556 + ], + "score": 1.0, + "content": "However, we do not consider this direction in current scheme because we share the same parameters", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 105, + 553, + 505, + 568 + ], + "spans": [ + { + "bbox": [ + 105, + 553, + 505, + 568 + ], + "score": 1.0, + "content": "for all discrete action in not very large scale. Therefore, the hierarchical methods may not bring a", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 106, + 566, + 226, + 579 + ], + "spans": [ + { + "bbox": [ + 106, + 566, + 226, + 579 + ], + "score": 1.0, + "content": "significant performance gain.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 27, + "bbox_fs": [ + 104, + 367, + 506, + 579 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 595, + 196, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 594, + 198, + 611 + ], + "spans": [ + { + "bbox": [ + 105, + 594, + 198, + 611 + ], + "score": 1.0, + "content": "6 DISCUSSIONS", + "type": "text" + } + ], + "index": 37 + } + ], + "index": 37 + }, + { + "type": "text", + "bbox": [ + 107, + 621, + 505, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "spans": [ + { + "bbox": [ + 105, + 622, + 505, + 635 + ], + "score": 1.0, + "content": "In this paper, we present a parameterized action RL match plan generation method which extends the", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "spans": [ + { + "bbox": [ + 105, + 633, + 504, + 645 + ], + "score": 1.0, + "content": "plan generation to the general case without any predefined knowledge. Key to address the problem", + "type": "text" + } + ], + "index": 39 + }, + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "spans": [ + { + "bbox": [ + 105, + 644, + 506, + 657 + ], + "score": 1.0, + "content": "are normalized softmax values of discrete actions to enable gradients backpropagation, parameter", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 105, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "space noise on parameters of the policy for unifying the exploration direction in both discrete and", + "type": "text" + } + ], + "index": 41 + }, + { + "bbox": [ + 106, + 667, + 505, + 678 + ], + "spans": [ + { + "bbox": [ + 106, + 667, + 505, + 678 + ], + "score": 1.0, + "content": "continuous spaces, and recurrent deterministic policies with prioritized replay buffer to accelerate", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "score": 1.0, + "content": "and stabilize the training. Our algorithm can be applied to not only the match plan generation envi-", + "type": "text" + } + ], + "index": 43 + }, + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "spans": [ + { + "bbox": [ + 105, + 688, + 505, + 700 + ], + "score": 1.0, + "content": "ronment, but also other similar parameteried action environments. The experiment results demon-", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "spans": [ + { + "bbox": [ + 105, + 699, + 505, + 712 + ], + "score": 1.0, + "content": "strate our method outperforms the well-designed hand-crafted rules in Bing and serveral baseline", + "type": "text" + } + ], + "index": 45 + }, + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 506, + 723 + ], + "score": 1.0, + "content": "results in some existing PARL benchmarks. In this paper, we mainly discuss about offline training", + "type": "text" + } + ], + "index": 46 + }, + { + "bbox": [ + 105, + 721, + 459, + 734 + ], + "spans": [ + { + "bbox": [ + 105, + 721, + 459, + 734 + ], + "score": 1.0, + "content": "procedure. In the future, we plan to apply learned policy to the production environment.", + "type": "text" + } + ], + "index": 47 + } + ], + "index": 42.5, + "bbox_fs": [ + 105, + 622, + 506, + 734 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 503, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 220, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 220, + 123 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5 + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 503, + 153 + ], + "lines": [ + { + "bbox": [ + 107, + 130, + 504, + 142 + ], + "spans": [ + { + "bbox": [ + 107, + 130, + 504, + 142 + ], + "score": 1.0, + "content": "Craig J Bester, Steven D James, and George D Konidaris. Multi-pass q-networks for deep reinforce-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 141, + 474, + 152 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 474, + 152 + ], + "score": 1.0, + "content": "ment learning with parameterised action spaces. arXiv preprint arXiv:1905.04388, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 504, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 157, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 506, + 173 + ], + "score": 1.0, + "content": "Haotian Fu, Hongyao Tang, Jianye Hao, Zihan Lei, Yingfeng Chen, and Changjie Fan. Deep", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 116, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "multi-agent reinforcement learning with discrete-continuous hybrid action spaces. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 182, + 219, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 219, + 191 + ], + "score": 1.0, + "content": "arXiv:1903.04959, 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6 + }, + { + "type": "text", + "bbox": [ + 105, + 199, + 504, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 210, + 365, + 223 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 365, + 223 + ], + "score": 1.0, + "content": "actor-critic methods. arXiv preprint arXiv:1802.09477, 2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5 + }, + { + "type": "text", + "bbox": [ + 107, + 229, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 505, + 243 + ], + "score": 1.0, + "content": "Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 241, + 279, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 241, + 279, + 252 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1511.04143, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5 + }, + { + "type": "text", + "bbox": [ + 106, + 258, + 504, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "Nicolas Heess, Jonathan J Hunt, Timothy P Lillicrap, and David Silver. Memory-based control with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 270, + 388, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 388, + 282 + ], + "score": 1.0, + "content": "recurrent neural networks. arXiv preprint arXiv:1512.04455, 2015.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5 + }, + { + "type": "text", + "bbox": [ + 107, + 288, + 504, + 311 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 299, + 191, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 299, + 191, + 311 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5 + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 504, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 504, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 504, + 332 + ], + "score": 1.0, + "content": "Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent expe-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 329, + 347, + 341 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 347, + 341 + ], + "score": 1.0, + "content": "rience replay in distributed reinforcement learning. 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5 + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "Maciej Klimek, Henryk Michalewski, Piotr Mi, et al. Hierarchical reinforcement learning with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 359, + 386, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 359, + 386, + 371 + ], + "score": 1.0, + "content": "parameters. In Conference on Robot Learning, pp. 301–313, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5 + }, + { + "type": "text", + "bbox": [ + 108, + 376, + 503, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 400, + 254, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 400, + 254, + 410 + ], + "score": 1.0, + "content": "preprint arXiv:1509.02971, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21 + }, + { + "type": "text", + "bbox": [ + 104, + 417, + 503, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 504, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 504, + 431 + ], + "score": 1.0, + "content": "Warwick Masson, Pravesh Ranchod, and George Konidaris. Reinforcement learning with parame-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 429, + 428, + 442 + ], + "spans": [ + { + "bbox": [ + 116, + 429, + 428, + 442 + ], + "score": 1.0, + "content": "terized actions. In Thirtieth AAAI Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5 + }, + { + "type": "text", + "bbox": [ + 104, + 447, + 504, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 459, + 423, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 423, + 471 + ], + "score": 1.0, + "content": "projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5 + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "score": 1.0, + "content": "Azalia Mirhoseini, Hieu Pham, Quoc V Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean. Device placement optimization", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 499, + 506, + 513 + ], + "score": 1.0, + "content": "with reinforcement learning. In Proceedings of the 34th International Conference on Machine", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 510, + 342, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 342, + 523 + ], + "score": 1.0, + "content": "Learning-Volume 70, pp. 2430–2439. JMLR. org, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5 + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 551, + 420, + 563 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 420, + 563 + ], + "score": 1.0, + "content": "control through deep reinforcement learning. Nature, 518(7540):529, 2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32 + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 504, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "Matthias Plappert, Rein Houthooft, Prafulla Dhariwal, Szymon Sidor, Richard Y Chen, Xi Chen,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 580, + 504, + 593 + ], + "spans": [ + { + "bbox": [ + 116, + 580, + 504, + 593 + ], + "score": 1.0, + "content": "Tamim Asfour, Pieter Abbeel, and Marcin Andrychowicz. Parameter space noise for exploration.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 591, + 280, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 591, + 280, + 603 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1706.01905, 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35 + }, + { + "type": "text", + "bbox": [ + 105, + 609, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "score": 1.0, + "content": "Corby Rosset, Damien Jose, Gargi Ghosh, Bhaskar Mitra, and Saurabh Tiwary. Optimizing query", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 621, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 505, + 632 + ], + "score": 1.0, + "content": "evaluations using reinforcement learning for web search. In The 41st International ACM SIGIR", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 632, + 500, + 644 + ], + "spans": [ + { + "bbox": [ + 117, + 632, + 500, + 644 + ], + "score": 1.0, + "content": "Conference on Research & Development in Information Retrieval, pp. 1193–1196. ACM, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38 + }, + { + "type": "text", + "bbox": [ + 106, + 650, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 664 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 662, + 254, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 254, + 673 + ], + "score": 1.0, + "content": "preprint arXiv:1511.05952, 2015.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5 + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 690, + 310, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 310, + 703 + ], + "score": 1.0, + "content": "Deterministic policy gradient algorithms. 2014.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5 + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Ermo Wei, Drew Wicke, and Sean Luke. Hierarchical approaches for reinforcement learning in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 721, + 419, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 419, + 732 + ], + "score": 1.0, + "content": "parameterized action space. In 2018 AAAI Spring Symposium Series, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5 + } + ], + "page_idx": 8, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 302, + 751, + 309, + 759 + ], + "lines": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "spans": [ + { + "bbox": [ + 302, + 751, + 309, + 762 + ], + "score": 1.0, + "content": "9", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 107, + 81, + 176, + 94 + ], + "lines": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "spans": [ + { + "bbox": [ + 106, + 82, + 176, + 94 + ], + "score": 1.0, + "content": "REFERENCES", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "text", + "bbox": [ + 106, + 100, + 503, + 123 + ], + "lines": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "spans": [ + { + "bbox": [ + 105, + 99, + 505, + 113 + ], + "score": 1.0, + "content": "Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 111, + 220, + 123 + ], + "spans": [ + { + "bbox": [ + 115, + 111, + 220, + 123 + ], + "score": 1.0, + "content": "arXiv:1607.06450, 2016.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1.5, + "bbox_fs": [ + 105, + 99, + 505, + 123 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 129, + 503, + 153 + ], + "lines": [ + { + "bbox": [ + 107, + 130, + 504, + 142 + ], + "spans": [ + { + "bbox": [ + 107, + 130, + 504, + 142 + ], + "score": 1.0, + "content": "Craig J Bester, Steven D James, and George D Konidaris. Multi-pass q-networks for deep reinforce-", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 141, + 474, + 152 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 474, + 152 + ], + "score": 1.0, + "content": "ment learning with parameterised action spaces. arXiv preprint arXiv:1905.04388, 2019.", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5, + "bbox_fs": [ + 107, + 130, + 504, + 152 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 158, + 504, + 192 + ], + "lines": [ + { + "bbox": [ + 105, + 157, + 506, + 173 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 506, + 173 + ], + "score": 1.0, + "content": "Haotian Fu, Hongyao Tang, Jianye Hao, Zihan Lei, Yingfeng Chen, and Changjie Fan. Deep", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 171, + 505, + 183 + ], + "spans": [ + { + "bbox": [ + 116, + 171, + 505, + 183 + ], + "score": 1.0, + "content": "multi-agent reinforcement learning with discrete-continuous hybrid action spaces. arXiv preprint", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 182, + 219, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 219, + 191 + ], + "score": 1.0, + "content": "arXiv:1903.04959, 2019.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6, + "bbox_fs": [ + 105, + 157, + 506, + 191 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 199, + 504, + 222 + ], + "lines": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 199, + 505, + 213 + ], + "score": 1.0, + "content": "Scott Fujimoto, Herke van Hoof, and David Meger. Addressing function approximation error in", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 210, + 365, + 223 + ], + "spans": [ + { + "bbox": [ + 115, + 210, + 365, + 223 + ], + "score": 1.0, + "content": "actor-critic methods. arXiv preprint arXiv:1802.09477, 2018.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 8.5, + "bbox_fs": [ + 106, + 199, + 505, + 223 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 229, + 504, + 252 + ], + "lines": [ + { + "bbox": [ + 105, + 228, + 505, + 243 + ], + "spans": [ + { + "bbox": [ + 105, + 228, + 505, + 243 + ], + "score": 1.0, + "content": "Matthew Hausknecht and Peter Stone. Deep reinforcement learning in parameterized action space.", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 116, + 241, + 279, + 252 + ], + "spans": [ + { + "bbox": [ + 116, + 241, + 279, + 252 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1511.04143, 2015.", + "type": "text" + } + ], + "index": 11 + } + ], + "index": 10.5, + "bbox_fs": [ + 105, + 228, + 505, + 252 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 258, + 504, + 282 + ], + "lines": [ + { + "bbox": [ + 106, + 258, + 505, + 271 + ], + "spans": [ + { + "bbox": [ + 106, + 258, + 505, + 271 + ], + "score": 1.0, + "content": "Nicolas Heess, Jonathan J Hunt, Timothy P Lillicrap, and David Silver. Memory-based control with", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 115, + 270, + 388, + 282 + ], + "spans": [ + { + "bbox": [ + 115, + 270, + 388, + 282 + ], + "score": 1.0, + "content": "recurrent neural networks. arXiv preprint arXiv:1512.04455, 2015.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 12.5, + "bbox_fs": [ + 106, + 258, + 505, + 282 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 288, + 504, + 311 + ], + "lines": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "spans": [ + { + "bbox": [ + 105, + 287, + 505, + 302 + ], + "score": 1.0, + "content": "Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. ¨ Neural computation, 9(8):", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 116, + 299, + 191, + 311 + ], + "spans": [ + { + "bbox": [ + 116, + 299, + 191, + 311 + ], + "score": 1.0, + "content": "1735–1780, 1997.", + "type": "text" + } + ], + "index": 15 + } + ], + "index": 14.5, + "bbox_fs": [ + 105, + 287, + 505, + 311 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 317, + 504, + 341 + ], + "lines": [ + { + "bbox": [ + 106, + 316, + 504, + 332 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 504, + 332 + ], + "score": 1.0, + "content": "Steven Kapturowski, Georg Ostrovski, John Quan, Remi Munos, and Will Dabney. Recurrent expe-", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 115, + 329, + 347, + 341 + ], + "spans": [ + { + "bbox": [ + 115, + 329, + 347, + 341 + ], + "score": 1.0, + "content": "rience replay in distributed reinforcement learning. 2018.", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 16.5, + "bbox_fs": [ + 106, + 316, + 504, + 341 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 347, + 504, + 371 + ], + "lines": [ + { + "bbox": [ + 106, + 347, + 505, + 361 + ], + "spans": [ + { + "bbox": [ + 106, + 347, + 505, + 361 + ], + "score": 1.0, + "content": "Maciej Klimek, Henryk Michalewski, Piotr Mi, et al. Hierarchical reinforcement learning with", + "type": "text" + } + ], + "index": 18 + }, + { + "bbox": [ + 115, + 359, + 386, + 371 + ], + "spans": [ + { + "bbox": [ + 115, + 359, + 386, + 371 + ], + "score": 1.0, + "content": "parameters. In Conference on Robot Learning, pp. 301–313, 2017.", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 18.5, + "bbox_fs": [ + 106, + 347, + 505, + 371 + ] + }, + { + "type": "text", + "bbox": [ + 108, + 376, + 503, + 411 + ], + "lines": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 376, + 505, + 390 + ], + "score": 1.0, + "content": "Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa,", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 115, + 388, + 505, + 401 + ], + "spans": [ + { + "bbox": [ + 115, + 388, + 505, + 401 + ], + "score": 1.0, + "content": "David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning. arXiv", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 115, + 400, + 254, + 410 + ], + "spans": [ + { + "bbox": [ + 115, + 400, + 254, + 410 + ], + "score": 1.0, + "content": "preprint arXiv:1509.02971, 2015.", + "type": "text" + } + ], + "index": 22 + } + ], + "index": 21, + "bbox_fs": [ + 106, + 376, + 505, + 410 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 417, + 503, + 441 + ], + "lines": [ + { + "bbox": [ + 106, + 417, + 504, + 431 + ], + "spans": [ + { + "bbox": [ + 106, + 417, + 504, + 431 + ], + "score": 1.0, + "content": "Warwick Masson, Pravesh Ranchod, and George Konidaris. Reinforcement learning with parame-", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 116, + 429, + 428, + 442 + ], + "spans": [ + { + "bbox": [ + 116, + 429, + 428, + 442 + ], + "score": 1.0, + "content": "terized actions. In Thirtieth AAAI Conference on Artificial Intelligence, 2016.", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 23.5, + "bbox_fs": [ + 106, + 417, + 504, + 442 + ] + }, + { + "type": "text", + "bbox": [ + 104, + 447, + 504, + 471 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 505, + 460 + ], + "score": 1.0, + "content": "Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and", + "type": "text" + } + ], + "index": 25 + }, + { + "bbox": [ + 115, + 459, + 423, + 471 + ], + "spans": [ + { + "bbox": [ + 115, + 459, + 423, + 471 + ], + "score": 1.0, + "content": "projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018.", + "type": "text" + } + ], + "index": 26 + } + ], + "index": 25.5, + "bbox_fs": [ + 106, + 447, + 505, + 471 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 476, + 504, + 522 + ], + "lines": [ + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "spans": [ + { + "bbox": [ + 105, + 476, + 506, + 490 + ], + "score": 1.0, + "content": "Azalia Mirhoseini, Hieu Pham, Quoc V Le, Benoit Steiner, Rasmus Larsen, Yuefeng Zhou, Naveen", + "type": "text" + } + ], + "index": 27 + }, + { + "bbox": [ + 115, + 488, + 505, + 501 + ], + "spans": [ + { + "bbox": [ + 115, + 488, + 505, + 501 + ], + "score": 1.0, + "content": "Kumar, Mohammad Norouzi, Samy Bengio, and Jeff Dean. Device placement optimization", + "type": "text" + } + ], + "index": 28 + }, + { + "bbox": [ + 115, + 499, + 506, + 513 + ], + "spans": [ + { + "bbox": [ + 115, + 499, + 506, + 513 + ], + "score": 1.0, + "content": "with reinforcement learning. In Proceedings of the 34th International Conference on Machine", + "type": "text" + } + ], + "index": 29 + }, + { + "bbox": [ + 115, + 510, + 342, + 523 + ], + "spans": [ + { + "bbox": [ + 115, + 510, + 342, + 523 + ], + "score": 1.0, + "content": "Learning-Volume 70, pp. 2430–2439. JMLR. org, 2017.", + "type": "text" + } + ], + "index": 30 + } + ], + "index": 28.5, + "bbox_fs": [ + 105, + 476, + 506, + 523 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 528, + 504, + 563 + ], + "lines": [ + { + "bbox": [ + 106, + 528, + 505, + 542 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 505, + 542 + ], + "score": 1.0, + "content": "Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Belle-", + "type": "text" + } + ], + "index": 31 + }, + { + "bbox": [ + 115, + 540, + 505, + 551 + ], + "spans": [ + { + "bbox": [ + 115, + 540, + 505, + 551 + ], + "score": 1.0, + "content": "mare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level", + "type": "text" + } + ], + "index": 32 + }, + { + "bbox": [ + 115, + 551, + 420, + 563 + ], + "spans": [ + { + "bbox": [ + 115, + 551, + 420, + 563 + ], + "score": 1.0, + "content": "control through deep reinforcement learning. Nature, 518(7540):529, 2015.", + "type": "text" + } + ], + "index": 33 + } + ], + "index": 32, + "bbox_fs": [ + 106, + 528, + 505, + 563 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 568, + 504, + 603 + ], + "lines": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "spans": [ + { + "bbox": [ + 105, + 568, + 506, + 582 + ], + "score": 1.0, + "content": "Matthias Plappert, Rein Houthooft, Prafulla Dhariwal, Szymon Sidor, Richard Y Chen, Xi Chen,", + "type": "text" + } + ], + "index": 34 + }, + { + "bbox": [ + 116, + 580, + 504, + 593 + ], + "spans": [ + { + "bbox": [ + 116, + 580, + 504, + 593 + ], + "score": 1.0, + "content": "Tamim Asfour, Pieter Abbeel, and Marcin Andrychowicz. Parameter space noise for exploration.", + "type": "text" + } + ], + "index": 35 + }, + { + "bbox": [ + 116, + 591, + 280, + 603 + ], + "spans": [ + { + "bbox": [ + 116, + 591, + 280, + 603 + ], + "score": 1.0, + "content": "arXiv preprint arXiv:1706.01905, 2017.", + "type": "text" + } + ], + "index": 36 + } + ], + "index": 35, + "bbox_fs": [ + 105, + 568, + 506, + 603 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 609, + 504, + 644 + ], + "lines": [ + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 608, + 506, + 624 + ], + "score": 1.0, + "content": "Corby Rosset, Damien Jose, Gargi Ghosh, Bhaskar Mitra, and Saurabh Tiwary. Optimizing query", + "type": "text" + } + ], + "index": 37 + }, + { + "bbox": [ + 115, + 621, + 505, + 632 + ], + "spans": [ + { + "bbox": [ + 115, + 621, + 505, + 632 + ], + "score": 1.0, + "content": "evaluations using reinforcement learning for web search. In The 41st International ACM SIGIR", + "type": "text" + } + ], + "index": 38 + }, + { + "bbox": [ + 117, + 632, + 500, + 644 + ], + "spans": [ + { + "bbox": [ + 117, + 632, + 500, + 644 + ], + "score": 1.0, + "content": "Conference on Research & Development in Information Retrieval, pp. 1193–1196. ACM, 2018.", + "type": "text" + } + ], + "index": 39 + } + ], + "index": 38, + "bbox_fs": [ + 105, + 608, + 506, + 644 + ] + }, + { + "type": "text", + "bbox": [ + 106, + 650, + 504, + 673 + ], + "lines": [ + { + "bbox": [ + 106, + 649, + 505, + 664 + ], + "spans": [ + { + "bbox": [ + 106, + 649, + 505, + 664 + ], + "score": 1.0, + "content": "Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Prioritized experience replay. arXiv", + "type": "text" + } + ], + "index": 40 + }, + { + "bbox": [ + 115, + 662, + 254, + 673 + ], + "spans": [ + { + "bbox": [ + 115, + 662, + 254, + 673 + ], + "score": 1.0, + "content": "preprint arXiv:1511.05952, 2015.", + "type": "text" + } + ], + "index": 41 + } + ], + "index": 40.5, + "bbox_fs": [ + 106, + 649, + 505, + 673 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 680, + 503, + 703 + ], + "lines": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 680, + 505, + 693 + ], + "score": 1.0, + "content": "David Silver, Guy Lever, Nicolas Heess, Thomas Degris, Daan Wierstra, and Martin Riedmiller.", + "type": "text" + } + ], + "index": 42 + }, + { + "bbox": [ + 116, + 690, + 310, + 703 + ], + "spans": [ + { + "bbox": [ + 116, + 690, + 310, + 703 + ], + "score": 1.0, + "content": "Deterministic policy gradient algorithms. 2014.", + "type": "text" + } + ], + "index": 43 + } + ], + "index": 42.5, + "bbox_fs": [ + 106, + 680, + 505, + 703 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 709, + 503, + 732 + ], + "lines": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "spans": [ + { + "bbox": [ + 105, + 708, + 505, + 722 + ], + "score": 1.0, + "content": "Ermo Wei, Drew Wicke, and Sean Luke. Hierarchical approaches for reinforcement learning in", + "type": "text" + } + ], + "index": 44 + }, + { + "bbox": [ + 115, + 721, + 419, + 732 + ], + "spans": [ + { + "bbox": [ + 115, + 721, + 419, + 732 + ], + "score": 1.0, + "content": "parameterized action space. In 2018 AAAI Spring Symposium Series, 2018.", + "type": "text" + } + ], + "index": 45 + } + ], + "index": 44.5, + "bbox_fs": [ + 105, + 708, + 505, + 732 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "score": 1.0, + "content": "Ian H Witten, Ian H Witten, Alistair Moffat, Timothy C Bell, Timothy C Bell, and Timothy C Bell.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "Managing gigabytes: compressing and indexing documents and images. Morgan Kaufmann,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 142, + 116 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 142, + 116 + ], + "score": 1.0, + "content": "1999.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 504, + 156 + ], + "lines": [ + { + "bbox": [ + 104, + 122, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 122, + 506, + 138 + ], + "score": 1.0, + "content": "Jiechao Xiong, Qing Wang, Zhuoran Yang, Peng Sun, Lei Han, Yang Zheng, Haobo Fu, Tong", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "score": 1.0, + "content": "Zhang, Ji Liu, and Han Liu. Parametrized deep q-networks learning: Reinforcement learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 463, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 463, + 158 + ], + "score": 1.0, + "content": "with discrete-continuous hybrid action space. arXiv preprint arXiv:1810.06394, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4 + }, + { + "type": "text", + "bbox": [ + 105, + 164, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "score": 1.0, + "content": "Justin Zobel and Alistair Moffat. Inverted files for text search engines. ACM computing surveys", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 212, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 212, + 187 + ], + "score": 1.0, + "content": "(CSUR), 38(2):6, 2006.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5 + } + ], + "page_idx": 9, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "10", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 26, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "text", + "bbox": [ + 107, + 82, + 504, + 115 + ], + "lines": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "spans": [ + { + "bbox": [ + 105, + 81, + 504, + 95 + ], + "score": 1.0, + "content": "Ian H Witten, Ian H Witten, Alistair Moffat, Timothy C Bell, Timothy C Bell, and Timothy C Bell.", + "type": "text" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "spans": [ + { + "bbox": [ + 115, + 93, + 505, + 107 + ], + "score": 1.0, + "content": "Managing gigabytes: compressing and indexing documents and images. Morgan Kaufmann,", + "type": "text" + } + ], + "index": 1 + }, + { + "bbox": [ + 115, + 104, + 142, + 116 + ], + "spans": [ + { + "bbox": [ + 115, + 104, + 142, + 116 + ], + "score": 1.0, + "content": "1999.", + "type": "text" + } + ], + "index": 2 + } + ], + "index": 1, + "bbox_fs": [ + 105, + 81, + 505, + 116 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 123, + 504, + 156 + ], + "lines": [ + { + "bbox": [ + 104, + 122, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 104, + 122, + 506, + 138 + ], + "score": 1.0, + "content": "Jiechao Xiong, Qing Wang, Zhuoran Yang, Peng Sun, Lei Han, Yang Zheng, Haobo Fu, Tong", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 115, + 134, + 506, + 148 + ], + "score": 1.0, + "content": "Zhang, Ji Liu, and Han Liu. Parametrized deep q-networks learning: Reinforcement learning", + "type": "text" + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 145, + 463, + 158 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 463, + 158 + ], + "score": 1.0, + "content": "with discrete-continuous hybrid action space. arXiv preprint arXiv:1810.06394, 2018.", + "type": "text" + } + ], + "index": 5 + } + ], + "index": 4, + "bbox_fs": [ + 104, + 122, + 506, + 158 + ] + }, + { + "type": "text", + "bbox": [ + 105, + 164, + 504, + 186 + ], + "lines": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 162, + 505, + 178 + ], + "score": 1.0, + "content": "Justin Zobel and Alistair Moffat. Inverted files for text search engines. ACM computing surveys", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 116, + 175, + 212, + 187 + ], + "spans": [ + { + "bbox": [ + 116, + 175, + 212, + 187 + ], + "score": 1.0, + "content": "(CSUR), 38(2):6, 2006.", + "type": "text" + } + ], + "index": 7 + } + ], + "index": 6.5, + "bbox_fs": [ + 105, + 162, + 505, + 187 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 183, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 184, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 184, + 96 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 193, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 194, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 194, + 119 + ], + "score": 1.0, + "content": "A.1 PSEUDOCODE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 128, + 504, + 150 + ], + "lines": [ + { + "bbox": [ + 107, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "Algorithm pseudocode for the algorithm is provided in Algorithm 1 which includes all aforemen-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 165, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 165, + 150 + ], + "score": 1.0, + "content": "tioned details.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5 + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 159, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 161, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 161, + 176 + ], + "score": 1.0, + "content": "Algorithm 1", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 106, + 174, + 495, + 439 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 174, + 495, + 439 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 174, + 495, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 495, + 439 + ], + "score": 0.967, + "html": "
Input: Empty replay buffer D, init parameter noise std δparam, action noise threshold δ Initialize policy parameters 0,value parameters Φ ←0Q for each episode do Perturb policy parameters θ ← θ +(O,Oparam) and target policy parameters θtarget Observe state St, output a disturbed action embedding A =p(st；0） Compute executing action at by taking max over k
Execute parameterized action at in the environment server Observe next state St+1,reward rt, done signal dt denoting if St+1 is terminal t,rt, St+1,dt) to the buffer D If St+1 is terminal, reset to an initial state so for each update if update-condition do
Sample a minibatch from prioritized replay buffer D with specific priorities Compute and transform target actions with disturbed target policy network
Update parameter noise std Oparam using empirical distance d(μ, μ)
Compute targets for TD-error with min Q-value in the twin Q-networks Update Q-network parameters using gradient descent
if policy update frequency then
Update policy network parameters θ with Update target parameters with polyak averaging end if
", + "type": "table", + "image_path": "6805768dc1374f48ea940804c4f043f9541f672a807ee8463e41885f8a35b386.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 106, + 174, + 495, + 262.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 262.3333333333333, + 495, + 350.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 350.66666666666663, + 495, + 438.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 108, + 462, + 284, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 285, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 285, + 475 + ], + "score": 1.0, + "content": "A.2 FURTHER EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 293, + 496 + ], + "score": 1.0, + "content": "Hyperparameter search. We use Microsoft", + "type": "text" + }, + { + "bbox": [ + 293, + 483, + 316, + 495 + ], + "score": 0.79, + "content": "\\mathrm { N N I } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "and OpenPAI3 to search hyperparameters. The", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "final metric to report to NNI is set to the sum of average training reward and final evaluation reward", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 268, + 519 + ], + "score": 1.0, + "content": "(repeated 1,000 times) final_metric", + "type": "text" + }, + { + "bbox": [ + 269, + 507, + 277, + 515 + ], + "score": 0.72, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "(avg_reward+eval_reward)/2. The intermediate", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "metric is set to evaluation reward per 1,000 episodes. We also use the early stop assessor. Each GPU", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 528, + 402, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 402, + 540 + ], + "score": 1.0, + "content": "server node connects to a production environment emulator with ethernet.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11 + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "Model architecture. Both policy and value networks use two fully connected layers with 512", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "hidden units and a LSTM layer (Hochreiter & Schmidhuber (1997)). Each layer also uses a layer", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "normalization (LayerNorm, Ba et al. (2016)) as suggested by Plappert et al. (2017) in consideration", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "of stability for noise applied on parameters, and follows a ReLU activation. Both output heads of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 596, + 401, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 401, + 610 + ], + "score": 1.0, + "content": "the policy network has a hidden layer with sof tmax or tanh activations.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16 + }, + { + "type": "title", + "bbox": [ + 108, + 622, + 292, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 294, + 635 + ], + "score": 1.0, + "content": "A.3 FURTHER ENVIRONMENTAL DETAILS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 700 + ], + "lines": [ + { + "bbox": [ + 107, + 645, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 107, + 645, + 504, + 655 + ], + "score": 1.0, + "content": "Accumulated values. Note that, for accumulated values in received states, rewards and outputted", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "action-parameters, we use the difference (delta values) from the last step. For states and rewards, it", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 115, + 680 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 666, + 177, + 678 + ], + "score": 0.84, + "content": "s _ { t } = s _ { t } ^ { \\prime } - s _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 665, + 181, + 680 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 182, + 666, + 243, + 678 + ], + "score": 0.91, + "content": "r _ { t } = r _ { t } ^ { \\prime } - r _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 665, + 300, + 680 + ], + "score": 1.0, + "content": ", where where", + "type": "text" + }, + { + "bbox": [ + 301, + 666, + 309, + 676 + ], + "score": 0.86, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 665, + 327, + 680 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 666, + 336, + 676 + ], + "score": 0.85, + "content": "r ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "denote raw state and reward. For actions,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 292, + 690 + ], + "score": 1.0, + "content": "the agent outputs raw action-parameter output", + "type": "text" + }, + { + "bbox": [ + 292, + 677, + 301, + 687 + ], + "score": 0.85, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ", and the emulator converts it to accumulated value", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 686, + 174, + 703 + ], + "spans": [ + { + "bbox": [ + 107, + 688, + 169, + 700 + ], + "score": 0.89, + "content": "a _ { t } = a _ { t } ^ { \\prime } + a _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 686, + 174, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22 + } + ], + "page_idx": 10, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 119, + 710, + 241, + 732 + ], + "lines": [ + { + "bbox": [ + 119, + 708, + 241, + 722 + ], + "spans": [ + { + "bbox": [ + 119, + 708, + 241, + 722 + ], + "score": 1.0, + "content": "2https://github.com/microsoft/nni", + "type": "text" + } + ] + }, + { + "bbox": [ + 119, + 721, + 240, + 732 + ], + "spans": [ + { + "bbox": [ + 119, + 721, + 240, + 732 + ], + "score": 1.0, + "content": "3https://github.com/microsoft/pai", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "spans": [ + { + "bbox": [ + 107, + 26, + 308, + 38 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 310, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 765 + ], + "score": 1.0, + "content": "", + "type": "text", + "height": 15, + "width": 13 + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "title", + "bbox": [ + 108, + 81, + 183, + 93 + ], + "lines": [ + { + "bbox": [ + 106, + 79, + 184, + 96 + ], + "spans": [ + { + "bbox": [ + 106, + 79, + 184, + 96 + ], + "score": 1.0, + "content": "A APPENDIX", + "type": "text" + } + ], + "index": 0 + } + ], + "index": 0 + }, + { + "type": "title", + "bbox": [ + 108, + 106, + 193, + 118 + ], + "lines": [ + { + "bbox": [ + 106, + 106, + 194, + 119 + ], + "spans": [ + { + "bbox": [ + 106, + 106, + 194, + 119 + ], + "score": 1.0, + "content": "A.1 PSEUDOCODE", + "type": "text" + } + ], + "index": 1 + } + ], + "index": 1 + }, + { + "type": "text", + "bbox": [ + 108, + 128, + 504, + 150 + ], + "lines": [ + { + "bbox": [ + 107, + 128, + 505, + 140 + ], + "spans": [ + { + "bbox": [ + 107, + 128, + 505, + 140 + ], + "score": 1.0, + "content": "Algorithm pseudocode for the algorithm is provided in Algorithm 1 which includes all aforemen-", + "type": "text" + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 138, + 165, + 150 + ], + "spans": [ + { + "bbox": [ + 106, + 138, + 165, + 150 + ], + "score": 1.0, + "content": "tioned details.", + "type": "text" + } + ], + "index": 3 + } + ], + "index": 2.5, + "bbox_fs": [ + 106, + 128, + 505, + 150 + ] + }, + { + "type": "title", + "bbox": [ + 107, + 162, + 159, + 173 + ], + "lines": [ + { + "bbox": [ + 106, + 161, + 161, + 176 + ], + "spans": [ + { + "bbox": [ + 106, + 161, + 161, + 176 + ], + "score": 1.0, + "content": "Algorithm 1", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 106, + 174, + 495, + 439 + ], + "blocks": [ + { + "type": "table_body", + "bbox": [ + 106, + 174, + 495, + 439 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 106, + 174, + 495, + 439 + ], + "spans": [ + { + "bbox": [ + 106, + 174, + 495, + 439 + ], + "score": 0.967, + "html": "
Input: Empty replay buffer D, init parameter noise std δparam, action noise threshold δ Initialize policy parameters 0,value parameters Φ ←0Q for each episode do Perturb policy parameters θ ← θ +(O,Oparam) and target policy parameters θtarget Observe state St, output a disturbed action embedding A =p(st；0） Compute executing action at by taking max over k
Execute parameterized action at in the environment server Observe next state St+1,reward rt, done signal dt denoting if St+1 is terminal t,rt, St+1,dt) to the buffer D If St+1 is terminal, reset to an initial state so for each update if update-condition do
Sample a minibatch from prioritized replay buffer D with specific priorities Compute and transform target actions with disturbed target policy network
Update parameter noise std Oparam using empirical distance d(μ, μ)
Compute targets for TD-error with min Q-value in the twin Q-networks Update Q-network parameters using gradient descent
if policy update frequency then
Update policy network parameters θ with Update target parameters with polyak averaging end if
", + "type": "table", + "image_path": "6805768dc1374f48ea940804c4f043f9541f672a807ee8463e41885f8a35b386.jpg" + } + ] + } + ], + "index": 6, + "virtual_lines": [ + { + "bbox": [ + 106, + 174, + 495, + 262.3333333333333 + ], + "spans": [], + "index": 5 + }, + { + "bbox": [ + 106, + 262.3333333333333, + 495, + 350.66666666666663 + ], + "spans": [], + "index": 6 + }, + { + "bbox": [ + 106, + 350.66666666666663, + 495, + 438.99999999999994 + ], + "spans": [], + "index": 7 + } + ] + } + ], + "index": 6 + }, + { + "type": "title", + "bbox": [ + 108, + 462, + 284, + 474 + ], + "lines": [ + { + "bbox": [ + 106, + 462, + 285, + 475 + ], + "spans": [ + { + "bbox": [ + 106, + 462, + 285, + 475 + ], + "score": 1.0, + "content": "A.2 FURTHER EXPERIMENTAL DETAILS", + "type": "text" + } + ], + "index": 8 + } + ], + "index": 8 + }, + { + "type": "text", + "bbox": [ + 106, + 483, + 505, + 539 + ], + "lines": [ + { + "bbox": [ + 105, + 483, + 505, + 496 + ], + "spans": [ + { + "bbox": [ + 105, + 483, + 293, + 496 + ], + "score": 1.0, + "content": "Hyperparameter search. We use Microsoft", + "type": "text" + }, + { + "bbox": [ + 293, + 483, + 316, + 495 + ], + "score": 0.79, + "content": "\\mathrm { N N I } ^ { 2 }", + "type": "inline_equation" + }, + { + "bbox": [ + 317, + 483, + 505, + 496 + ], + "score": 1.0, + "content": "and OpenPAI3 to search hyperparameters. The", + "type": "text" + } + ], + "index": 9 + }, + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "spans": [ + { + "bbox": [ + 106, + 495, + 505, + 507 + ], + "score": 1.0, + "content": "final metric to report to NNI is set to the sum of average training reward and final evaluation reward", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 505, + 505, + 519 + ], + "spans": [ + { + "bbox": [ + 106, + 505, + 268, + 519 + ], + "score": 1.0, + "content": "(repeated 1,000 times) final_metric", + "type": "text" + }, + { + "bbox": [ + 269, + 507, + 277, + 515 + ], + "score": 0.72, + "content": "=", + "type": "inline_equation" + }, + { + "bbox": [ + 277, + 505, + 505, + 519 + ], + "score": 1.0, + "content": "(avg_reward+eval_reward)/2. The intermediate", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "spans": [ + { + "bbox": [ + 106, + 517, + 505, + 529 + ], + "score": 1.0, + "content": "metric is set to evaluation reward per 1,000 episodes. We also use the early stop assessor. Each GPU", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 106, + 528, + 402, + 540 + ], + "spans": [ + { + "bbox": [ + 106, + 528, + 402, + 540 + ], + "score": 1.0, + "content": "server node connects to a production environment emulator with ethernet.", + "type": "text" + } + ], + "index": 13 + } + ], + "index": 11, + "bbox_fs": [ + 105, + 483, + 505, + 540 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 552, + 505, + 608 + ], + "lines": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "spans": [ + { + "bbox": [ + 105, + 552, + 505, + 565 + ], + "score": 1.0, + "content": "Model architecture. Both policy and value networks use two fully connected layers with 512", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "spans": [ + { + "bbox": [ + 105, + 563, + 505, + 577 + ], + "score": 1.0, + "content": "hidden units and a LSTM layer (Hochreiter & Schmidhuber (1997)). Each layer also uses a layer", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "spans": [ + { + "bbox": [ + 105, + 574, + 505, + 587 + ], + "score": 1.0, + "content": "normalization (LayerNorm, Ba et al. (2016)) as suggested by Plappert et al. (2017) in consideration", + "type": "text" + } + ], + "index": 16 + }, + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "spans": [ + { + "bbox": [ + 106, + 586, + 505, + 598 + ], + "score": 1.0, + "content": "of stability for noise applied on parameters, and follows a ReLU activation. Both output heads of", + "type": "text" + } + ], + "index": 17 + }, + { + "bbox": [ + 106, + 596, + 401, + 610 + ], + "spans": [ + { + "bbox": [ + 106, + 596, + 401, + 610 + ], + "score": 1.0, + "content": "the policy network has a hidden layer with sof tmax or tanh activations.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 16, + "bbox_fs": [ + 105, + 552, + 505, + 610 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 622, + 292, + 634 + ], + "lines": [ + { + "bbox": [ + 106, + 622, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 106, + 622, + 294, + 635 + ], + "score": 1.0, + "content": "A.3 FURTHER ENVIRONMENTAL DETAILS", + "type": "text" + } + ], + "index": 19 + } + ], + "index": 19 + }, + { + "type": "text", + "bbox": [ + 107, + 644, + 505, + 700 + ], + "lines": [ + { + "bbox": [ + 107, + 645, + 504, + 655 + ], + "spans": [ + { + "bbox": [ + 107, + 645, + 504, + 655 + ], + "score": 1.0, + "content": "Accumulated values. Note that, for accumulated values in received states, rewards and outputted", + "type": "text" + } + ], + "index": 20 + }, + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "spans": [ + { + "bbox": [ + 106, + 655, + 505, + 667 + ], + "score": 1.0, + "content": "action-parameters, we use the difference (delta values) from the last step. For states and rewards, it", + "type": "text" + } + ], + "index": 21 + }, + { + "bbox": [ + 105, + 665, + 506, + 680 + ], + "spans": [ + { + "bbox": [ + 105, + 665, + 115, + 680 + ], + "score": 1.0, + "content": "is", + "type": "text" + }, + { + "bbox": [ + 116, + 666, + 177, + 678 + ], + "score": 0.84, + "content": "s _ { t } = s _ { t } ^ { \\prime } - s _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 178, + 665, + 181, + 680 + ], + "score": 1.0, + "content": ",", + "type": "text" + }, + { + "bbox": [ + 182, + 666, + 243, + 678 + ], + "score": 0.91, + "content": "r _ { t } = r _ { t } ^ { \\prime } - r _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 243, + 665, + 300, + 680 + ], + "score": 1.0, + "content": ", where where", + "type": "text" + }, + { + "bbox": [ + 301, + 666, + 309, + 676 + ], + "score": 0.86, + "content": "s ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 310, + 665, + 327, + 680 + ], + "score": 1.0, + "content": "and", + "type": "text" + }, + { + "bbox": [ + 328, + 666, + 336, + 676 + ], + "score": 0.85, + "content": "r ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 337, + 665, + 506, + 680 + ], + "score": 1.0, + "content": "denote raw state and reward. For actions,", + "type": "text" + } + ], + "index": 22 + }, + { + "bbox": [ + 105, + 677, + 505, + 690 + ], + "spans": [ + { + "bbox": [ + 105, + 677, + 292, + 690 + ], + "score": 1.0, + "content": "the agent outputs raw action-parameter output", + "type": "text" + }, + { + "bbox": [ + 292, + 677, + 301, + 687 + ], + "score": 0.85, + "content": "a ^ { \\prime }", + "type": "inline_equation" + }, + { + "bbox": [ + 301, + 677, + 505, + 690 + ], + "score": 1.0, + "content": ", and the emulator converts it to accumulated value", + "type": "text" + } + ], + "index": 23 + }, + { + "bbox": [ + 107, + 686, + 174, + 703 + ], + "spans": [ + { + "bbox": [ + 107, + 688, + 169, + 700 + ], + "score": 0.89, + "content": "a _ { t } = a _ { t } ^ { \\prime } + a _ { t - 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 170, + 686, + 174, + 703 + ], + "score": 1.0, + "content": ".", + "type": "text" + } + ], + "index": 24 + } + ], + "index": 22, + "bbox_fs": [ + 105, + 645, + 506, + 703 + ] + } + ] + }, + { + "preproc_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 84, + 477, + 236 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 134, + 84, + 477, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 84, + 477, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 84, + 477, + 236 + ], + "score": 0.969, + "type": "image", + "image_path": "6d04d136c67be4f2400a72fd7e2f02c247f2d8e0e0954f2f1782df4d09b152ff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 134, + 84, + 477, + 134.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 134, + 134.66666666666666, + 477, + 185.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 134, + 185.33333333333331, + 477, + 235.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 249, + 504, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "score": 1.0, + "content": "Figure 6: Best evaluation learning curves on both environments during tuning. The agent achieves maximum", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 258, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 287, + 270 + ], + "score": 1.0, + "content": "possible reward on both environments (50 and 1).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 290, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "Inverted index in Bing. In the inverted index system in Bing, there are two key steps: (i) the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 301, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 314 + ], + "score": 1.0, + "content": "system rewrites the query to enlarge candidate set with specified match plan (or search plan), (ii) it", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "returns candidates with top ranking scores. We just refer to the math plan part regarding to gener-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "ation, but not user-input content. In the match plan generation, the goal is to generate the optimal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 332, + 249, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 249, + 348 + ], + "score": 1.0, + "content": "search plan (policy) for each query.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7 + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 371 + ], + "score": 1.0, + "content": "Closed-loop and open-loop control. The production match plan is generated online in open-loop", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "(feedforward) without taking runtime system signals into consideration for the latency and imple-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "mentation consideration. However, RL is closed-loop which takes feedback from system signals", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "to make decisions. This may increase the online overhead, but it can be tackled by converting the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "reflective policy to a shooting-style action sequence by predicting with learned transition dynamics.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 422, + 425 + ], + "score": 1.0, + "content": "In the implementation, the environment emulator receives an action sequence", + "type": "text" + }, + { + "bbox": [ + 423, + 414, + 459, + 424 + ], + "score": 0.9, + "content": "a _ { 0 } , . . . , a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "and return", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 424, + 255, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 126, + 435 + ], + "score": 0.88, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 424, + 255, + 436 + ], + "score": 1.0, + "content": "to simulate the open-loop style.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13 + }, + { + "type": "title", + "bbox": [ + 108, + 448, + 290, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 291, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 291, + 461 + ], + "score": 1.0, + "content": "A.4 MORE TRAINING FOR BENCHMARKS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 441, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 443, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 443, + 483 + ], + "score": 1.0, + "content": "We provide the best evaluation learning curves on both environments during tuning.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18 + } + ], + "page_idx": 11, + "page_size": [ + 612, + 792 + ], + "discarded_blocks": [ + { + "type": "discarded", + "bbox": [ + 107, + 27, + 308, + 37 + ], + "lines": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "spans": [ + { + "bbox": [ + 106, + 25, + 309, + 39 + ], + "score": 1.0, + "content": "Under review as a conference paper at ICLR 2020", + "type": "text" + } + ] + } + ] + }, + { + "type": "discarded", + "bbox": [ + 300, + 751, + 311, + 760 + ], + "lines": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 313, + 764 + ], + "score": 1.0, + "content": "12", + "type": "text" + } + ] + } + ] + } + ], + "para_blocks": [ + { + "type": "image", + "bbox": [ + 134, + 84, + 477, + 236 + ], + "blocks": [ + { + "type": "image_body", + "bbox": [ + 134, + 84, + 477, + 236 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 134, + 84, + 477, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 84, + 477, + 236 + ], + "score": 0.969, + "type": "image", + "image_path": "6d04d136c67be4f2400a72fd7e2f02c247f2d8e0e0954f2f1782df4d09b152ff.jpg" + } + ] + } + ], + "index": 1, + "virtual_lines": [ + { + "bbox": [ + 134, + 84, + 477, + 134.66666666666666 + ], + "spans": [], + "index": 0 + }, + { + "bbox": [ + 134, + 134.66666666666666, + 477, + 185.33333333333331 + ], + "spans": [], + "index": 1 + }, + { + "bbox": [ + 134, + 185.33333333333331, + 477, + 235.99999999999997 + ], + "spans": [], + "index": 2 + } + ] + }, + { + "type": "image_caption", + "bbox": [ + 107, + 249, + 504, + 270 + ], + "group_id": 0, + "lines": [ + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "spans": [ + { + "bbox": [ + 105, + 248, + 505, + 262 + ], + "score": 1.0, + "content": "Figure 6: Best evaluation learning curves on both environments during tuning. The agent achieves maximum", + "type": "text" + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 258, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 105, + 258, + 287, + 270 + ], + "score": 1.0, + "content": "possible reward on both environments (50 and 1).", + "type": "text" + } + ], + "index": 4 + } + ], + "index": 3.5 + } + ], + "index": 2.25 + }, + { + "type": "text", + "bbox": [ + 107, + 290, + 505, + 345 + ], + "lines": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "spans": [ + { + "bbox": [ + 105, + 290, + 506, + 303 + ], + "score": 1.0, + "content": "Inverted index in Bing. In the inverted index system in Bing, there are two key steps: (i) the", + "type": "text" + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 301, + 506, + 314 + ], + "spans": [ + { + "bbox": [ + 105, + 301, + 506, + 314 + ], + "score": 1.0, + "content": "system rewrites the query to enlarge candidate set with specified match plan (or search plan), (ii) it", + "type": "text" + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 105, + 311, + 505, + 325 + ], + "score": 1.0, + "content": "returns candidates with top ranking scores. We just refer to the math plan part regarding to gener-", + "type": "text" + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "spans": [ + { + "bbox": [ + 105, + 323, + 505, + 336 + ], + "score": 1.0, + "content": "ation, but not user-input content. In the match plan generation, the goal is to generate the optimal", + "type": "text" + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 332, + 249, + 348 + ], + "spans": [ + { + "bbox": [ + 105, + 332, + 249, + 348 + ], + "score": 1.0, + "content": "search plan (policy) for each query.", + "type": "text" + } + ], + "index": 9 + } + ], + "index": 7, + "bbox_fs": [ + 105, + 290, + 506, + 348 + ] + }, + { + "type": "text", + "bbox": [ + 107, + 357, + 505, + 435 + ], + "lines": [ + { + "bbox": [ + 106, + 357, + 505, + 371 + ], + "spans": [ + { + "bbox": [ + 106, + 357, + 505, + 371 + ], + "score": 1.0, + "content": "Closed-loop and open-loop control. The production match plan is generated online in open-loop", + "type": "text" + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 368, + 505, + 381 + ], + "spans": [ + { + "bbox": [ + 106, + 368, + 505, + 381 + ], + "score": 1.0, + "content": "(feedforward) without taking runtime system signals into consideration for the latency and imple-", + "type": "text" + } + ], + "index": 11 + }, + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "spans": [ + { + "bbox": [ + 105, + 380, + 505, + 392 + ], + "score": 1.0, + "content": "mentation consideration. However, RL is closed-loop which takes feedback from system signals", + "type": "text" + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "spans": [ + { + "bbox": [ + 105, + 390, + 505, + 403 + ], + "score": 1.0, + "content": "to make decisions. This may increase the online overhead, but it can be tackled by converting the", + "type": "text" + } + ], + "index": 13 + }, + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 105, + 401, + 505, + 415 + ], + "score": 1.0, + "content": "reflective policy to a shooting-style action sequence by predicting with learned transition dynamics.", + "type": "text" + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 412, + 505, + 425 + ], + "spans": [ + { + "bbox": [ + 105, + 412, + 422, + 425 + ], + "score": 1.0, + "content": "In the implementation, the environment emulator receives an action sequence", + "type": "text" + }, + { + "bbox": [ + 423, + 414, + 459, + 424 + ], + "score": 0.9, + "content": "a _ { 0 } , . . . , a _ { t }", + "type": "inline_equation" + }, + { + "bbox": [ + 460, + 412, + 505, + 425 + ], + "score": 1.0, + "content": "and return", + "type": "text" + } + ], + "index": 15 + }, + { + "bbox": [ + 106, + 424, + 255, + 436 + ], + "spans": [ + { + "bbox": [ + 106, + 425, + 126, + 435 + ], + "score": 0.88, + "content": "s _ { t + 1 }", + "type": "inline_equation" + }, + { + "bbox": [ + 126, + 424, + 255, + 436 + ], + "score": 1.0, + "content": "to simulate the open-loop style.", + "type": "text" + } + ], + "index": 16 + } + ], + "index": 13, + "bbox_fs": [ + 105, + 357, + 505, + 436 + ] + }, + { + "type": "title", + "bbox": [ + 108, + 448, + 290, + 459 + ], + "lines": [ + { + "bbox": [ + 106, + 447, + 291, + 461 + ], + "spans": [ + { + "bbox": [ + 106, + 447, + 291, + 461 + ], + "score": 1.0, + "content": "A.4 MORE TRAINING FOR BENCHMARKS", + "type": "text" + } + ], + "index": 17 + } + ], + "index": 17 + }, + { + "type": "text", + "bbox": [ + 106, + 469, + 441, + 480 + ], + "lines": [ + { + "bbox": [ + 105, + 467, + 443, + 483 + ], + "spans": [ + { + "bbox": [ + 105, + 467, + 443, + 483 + ], + "score": 1.0, + "content": "We provide the best evaluation learning curves on both environments during tuning.", + "type": "text" + } + ], + "index": 18 + } + ], + "index": 18, + "bbox_fs": [ + 105, + 467, + 443, + 483 + ] + } + ] + } + ], + "_backend": "pipeline", + "_version_name": "2.2.2" +} \ No newline at end of file diff --git a/parse/train/rkgZaT4tDr/rkgZaT4tDr_model.json b/parse/train/rkgZaT4tDr/rkgZaT4tDr_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fbc98920937a3daa9a6c8f064d06a1de57ac7544 --- /dev/null +++ b/parse/train/rkgZaT4tDr/rkgZaT4tDr_model.json @@ -0,0 +1,14945 @@ +[ + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 398, + 590, + 1302, + 590, + 1302, + 956, + 398, + 956 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 299, + 1457, + 1402, + 1457, + 1402, + 1640, + 299, + 1640 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1089, + 1403, + 1089, + 1403, + 1272, + 298, + 1272 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 299, + 1656, + 1401, + 1656, + 1401, + 1810, + 299, + 1810 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 299, + 1288, + 1402, + 1288, + 1402, + 1441, + 299, + 1441 + ], + "score": 0.974 + }, + { + "category_id": 1, + "poly": [ + 299, + 1824, + 1401, + 1824, + 1401, + 1978, + 299, + 1978 + ], + "score": 0.971 + }, + { + "category_id": 0, + "poly": [ + 301, + 219, + 1399, + 219, + 1399, + 324, + 301, + 324 + ], + "score": 0.965 + }, + { + "category_id": 1, + "poly": [ + 313, + 378, + 680, + 378, + 680, + 438, + 313, + 438 + ], + "score": 0.934 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.899 + }, + { + "category_id": 0, + "poly": [ + 302, + 1018, + 572, + 1018, + 572, + 1053, + 302, + 1053 + ], + "score": 0.895 + }, + { + "category_id": 2, + "poly": [ + 336, + 2006, + 567, + 2006, + 567, + 2033, + 336, + 2033 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 773, + 519, + 927, + 519, + 927, + 553, + 773, + 553 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 857, + 2088, + 857, + 2112, + 841, + 2112 + ], + "score": 0.713 + }, + { + "category_id": 13, + "poly": [ + 830, + 870, + 849, + 870, + 849, + 893, + 830, + 893 + ], + "score": 0.44, + "latex": "n" + }, + { + "category_id": 13, + "poly": [ + 414, + 1317, + 482, + 1317, + 482, + 1351, + 414, + 1351 + ], + "score": 0.41, + "latex": "\\mathrm { B i n g ^ { 1 } }" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 218.0, + 1403.0, + 218.0, + 1403.0, + 273.0, + 294.0, + 273.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 275.0, + 1404.0, + 275.0, + 1404.0, + 328.0, + 295.0, + 328.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1015.0, + 579.0, + 1015.0, + 579.0, + 1062.0, + 293.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 332.0, + 2000.0, + 572.0, + 2000.0, + 572.0, + 2039.0, + 332.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 518.0, + 932.0, + 518.0, + 932.0, + 556.0, + 770.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 860.0, + 2087.0, + 860.0, + 2117.0, + 840.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 590.0, + 1305.0, + 590.0, + 1305.0, + 624.0, + 394.0, + 624.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 620.0, + 1306.0, + 620.0, + 1306.0, + 655.0, + 394.0, + 655.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 649.0, + 1304.0, + 649.0, + 1304.0, + 688.0, + 393.0, + 688.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 681.0, + 1305.0, + 681.0, + 1305.0, + 715.0, + 393.0, + 715.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 713.0, + 1306.0, + 713.0, + 1306.0, + 747.0, + 394.0, + 747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 743.0, + 1305.0, + 743.0, + 1305.0, + 778.0, + 393.0, + 778.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 773.0, + 1305.0, + 773.0, + 1305.0, + 808.0, + 393.0, + 808.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 395.0, + 805.0, + 1305.0, + 805.0, + 1305.0, + 837.0, + 395.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 834.0, + 1306.0, + 834.0, + 1306.0, + 867.0, + 394.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 866.0, + 829.0, + 866.0, + 829.0, + 898.0, + 394.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 866.0, + 1305.0, + 866.0, + 1305.0, + 898.0, + 850.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 393.0, + 896.0, + 1305.0, + 896.0, + 1305.0, + 928.0, + 393.0, + 928.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 394.0, + 926.0, + 1299.0, + 926.0, + 1299.0, + 959.0, + 394.0, + 959.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1456.0, + 1404.0, + 1456.0, + 1404.0, + 1492.0, + 294.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1485.0, + 1406.0, + 1485.0, + 1406.0, + 1523.0, + 294.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1514.0, + 1406.0, + 1514.0, + 1406.0, + 1556.0, + 293.0, + 1556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1550.0, + 1404.0, + 1550.0, + 1404.0, + 1581.0, + 294.0, + 1581.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1576.0, + 1407.0, + 1576.0, + 1407.0, + 1615.0, + 293.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1607.0, + 899.0, + 1607.0, + 899.0, + 1645.0, + 294.0, + 1645.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1084.0, + 1405.0, + 1084.0, + 1405.0, + 1129.0, + 292.0, + 1129.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1117.0, + 1405.0, + 1117.0, + 1405.0, + 1155.0, + 293.0, + 1155.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1151.0, + 1403.0, + 1151.0, + 1403.0, + 1183.0, + 296.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1179.0, + 1405.0, + 1179.0, + 1405.0, + 1218.0, + 292.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1211.0, + 1405.0, + 1211.0, + 1405.0, + 1246.0, + 294.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1241.0, + 909.0, + 1241.0, + 909.0, + 1277.0, + 294.0, + 1277.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1654.0, + 1403.0, + 1654.0, + 1403.0, + 1693.0, + 295.0, + 1693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1685.0, + 1405.0, + 1685.0, + 1405.0, + 1722.0, + 295.0, + 1722.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1717.0, + 1405.0, + 1717.0, + 1405.0, + 1753.0, + 293.0, + 1753.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1748.0, + 1405.0, + 1748.0, + 1405.0, + 1781.0, + 295.0, + 1781.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1778.0, + 1353.0, + 1778.0, + 1353.0, + 1814.0, + 295.0, + 1814.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1285.0, + 1406.0, + 1285.0, + 1406.0, + 1325.0, + 294.0, + 1325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1317.0, + 413.0, + 1317.0, + 413.0, + 1353.0, + 295.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 483.0, + 1317.0, + 1404.0, + 1317.0, + 1404.0, + 1353.0, + 483.0, + 1353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1347.0, + 1404.0, + 1347.0, + 1404.0, + 1387.0, + 294.0, + 1387.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1379.0, + 1406.0, + 1379.0, + 1406.0, + 1416.0, + 295.0, + 1416.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1409.0, + 814.0, + 1409.0, + 814.0, + 1445.0, + 295.0, + 1445.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1823.0, + 1405.0, + 1823.0, + 1405.0, + 1863.0, + 294.0, + 1863.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1853.0, + 1406.0, + 1853.0, + 1406.0, + 1892.0, + 293.0, + 1892.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1886.0, + 1406.0, + 1886.0, + 1406.0, + 1923.0, + 293.0, + 1923.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1913.0, + 1405.0, + 1913.0, + 1405.0, + 1957.0, + 292.0, + 1957.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1945.0, + 1405.0, + 1945.0, + 1405.0, + 1983.0, + 293.0, + 1983.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.0, + 379.0, + 560.0, + 379.0, + 560.0, + 411.0, + 315.0, + 411.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 312.0, + 407.0, + 681.0, + 407.0, + 681.0, + 442.0, + 312.0, + 442.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 0, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1333, + 1404, + 1333, + 1404, + 1580, + 297, + 1580 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 297, + 660, + 1403, + 660, + 1403, + 938, + 297, + 938 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 297, + 952, + 1404, + 952, + 1404, + 1166, + 297, + 1166 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1881, + 1403, + 1881, + 1403, + 2035, + 298, + 2035 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 298, + 1594, + 1405, + 1594, + 1405, + 1686, + 298, + 1686 + ], + "score": 0.973 + }, + { + "category_id": 8, + "poly": [ + 675, + 1695, + 1024, + 1695, + 1024, + 1766, + 675, + 1766 + ], + "score": 0.962 + }, + { + "category_id": 3, + "poly": [ + 528, + 241, + 1172, + 241, + 1172, + 407, + 528, + 407 + ], + "score": 0.96 + }, + { + "category_id": 1, + "poly": [ + 297, + 1772, + 1402, + 1772, + 1402, + 1835, + 297, + 1835 + ], + "score": 0.953 + }, + { + "category_id": 1, + "poly": [ + 294, + 583, + 1399, + 583, + 1399, + 647, + 294, + 647 + ], + "score": 0.951 + }, + { + "category_id": 4, + "poly": [ + 296, + 451, + 1399, + 451, + 1399, + 537, + 296, + 537 + ], + "score": 0.947 + }, + { + "category_id": 8, + "poly": [ + 617, + 1840, + 1087, + 1840, + 1087, + 1877, + 617, + 1877 + ], + "score": 0.942 + }, + { + "category_id": 0, + "poly": [ + 301, + 1276, + 708, + 1276, + 708, + 1309, + 301, + 1309 + ], + "score": 0.9 + }, + { + "category_id": 0, + "poly": [ + 301, + 1208, + 557, + 1208, + 557, + 1245, + 301, + 1245 + ], + "score": 0.885 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.731 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 104, + 297, + 104 + ], + "score": 0.63 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 104, + 297, + 104 + ], + "score": 0.266 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.186 + }, + { + "category_id": 14, + "poly": [ + 674, + 1692, + 1022, + 1692, + 1022, + 1768, + 674, + 1768 + ], + "score": 0.93, + "latex": "\\mathcal { A } = \\bigcup _ { k \\in \\mathcal { A } _ { d } } \\left\\{ ( k , x _ { k } ) | x _ { k } \\in \\mathcal { X } _ { k } \\right\\} ," + }, + { + "category_id": 13, + "poly": [ + 602, + 1772, + 774, + 1772, + 774, + 1805, + 602, + 1805 + ], + "score": 0.93, + "latex": "a \\in \\mathcal { A } _ { d } = [ K ]" + }, + { + "category_id": 14, + "poly": [ + 612, + 1839, + 1083, + 1839, + 1083, + 1877, + 612, + 1877 + ], + "score": 0.92, + "latex": "\\mathcal { A } = \\{ ( k , x ) | k \\in \\mathcal { A } _ { d } , x \\in \\mathcal { X } \\} = \\mathcal { A } _ { d } \\times \\mathcal { X } ," + }, + { + "category_id": 13, + "poly": [ + 372, + 1395, + 494, + 1395, + 494, + 1429, + 372, + 1429 + ], + "score": 0.92, + "latex": "\\gamma \\in \\ [ 0 , 1 )" + }, + { + "category_id": 13, + "poly": [ + 968, + 1425, + 1177, + 1425, + 1177, + 1458, + 968, + 1458 + ], + "score": 0.92, + "latex": "( S , \\mathcal { A } , T , R , \\Omega , \\dot { \\mathcal { O } } )" + }, + { + "category_id": 13, + "poly": [ + 1213, + 1487, + 1401, + 1487, + 1401, + 1516, + 1213, + 1516 + ], + "score": 0.91, + "latex": "R : S \\times \\mathcal { A } \\mathbb { R }" + }, + { + "category_id": 13, + "poly": [ + 851, + 1457, + 990, + 1457, + 990, + 1488, + 851, + 1488 + ], + "score": 0.9, + "latex": "( S , { \\mathcal { A } } , T , R )" + }, + { + "category_id": 13, + "poly": [ + 896, + 1487, + 1154, + 1487, + 1154, + 1518, + 896, + 1518 + ], + "score": 0.9, + "latex": "\\mathcal { T } : \\mathcal { S } \\times \\mathcal { S } \\times \\mathcal { A } \\mathbb { R } _ { + }" + }, + { + "category_id": 13, + "poly": [ + 753, + 1883, + 791, + 1883, + 791, + 1912, + 753, + 1912 + ], + "score": 0.88, + "latex": "\\mathbf { \\mathcal { A } } _ { d }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1806, + 333, + 1806, + 333, + 1834, + 298, + 1834 + ], + "score": 0.86, + "latex": "\\mathcal { X } _ { a }" + }, + { + "category_id": 13, + "poly": [ + 1085, + 1459, + 1108, + 1459, + 1108, + 1484, + 1085, + 1484 + ], + "score": 0.81, + "latex": "s" + }, + { + "category_id": 13, + "poly": [ + 1227, + 1549, + 1252, + 1549, + 1252, + 1575, + 1227, + 1575 + ], + "score": 0.79, + "latex": "\\mathcal { O }" + }, + { + "category_id": 13, + "poly": [ + 1164, + 1883, + 1191, + 1883, + 1191, + 1909, + 1164, + 1909 + ], + "score": 0.78, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1331, + 1458, + 1357, + 1458, + 1357, + 1484, + 1331, + 1484 + ], + "score": 0.77, + "latex": "\\mathcal { A }" + }, + { + "category_id": 13, + "poly": [ + 996, + 1519, + 1019, + 1519, + 1019, + 1546, + 996, + 1546 + ], + "score": 0.77, + "latex": "\\Omega" + }, + { + "category_id": 13, + "poly": [ + 615, + 1488, + 639, + 1488, + 639, + 1514, + 615, + 1514 + ], + "score": 0.76, + "latex": "T" + }, + { + "category_id": 13, + "poly": [ + 1254, + 882, + 1272, + 882, + 1272, + 903, + 1254, + 903 + ], + "score": 0.49, + "latex": "n" + }, + { + "category_id": 15, + "poly": [ + 588.0, + 241.0, + 764.0, + 241.0, + 764.0, + 274.0, + 588.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 247.0, + 812.0, + 247.0, + 812.0, + 272.0, + 770.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 820.0, + 247.0, + 861.0, + 247.0, + 861.0, + 272.0, + 820.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 868.0, + 246.0, + 910.0, + 246.0, + 910.0, + 271.0, + 868.0, + 271.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 952.0, + 251.0, + 979.0, + 251.0, + 979.0, + 266.0, + 952.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1010.0, + 245.0, + 1061.0, + 245.0, + 1061.0, + 272.0, + 1010.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1065.0, + 246.0, + 1112.0, + 246.0, + 1112.0, + 272.0, + 1065.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1120.0, + 247.0, + 1162.0, + 247.0, + 1162.0, + 272.0, + 1120.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 261.0, + 584.0, + 261.0, + 584.0, + 290.0, + 530.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 609.0, + 271.0, + 699.0, + 271.0, + 699.0, + 306.0, + 609.0, + 306.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 721.0, + 275.0, + 762.0, + 275.0, + 762.0, + 301.0, + 721.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 277.0, + 811.0, + 277.0, + 811.0, + 299.0, + 771.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 275.0, + 862.0, + 275.0, + 862.0, + 300.0, + 821.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 869.0, + 275.0, + 910.0, + 275.0, + 910.0, + 300.0, + 869.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 280.0, + 980.0, + 280.0, + 980.0, + 295.0, + 951.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 275.0, + 1061.0, + 275.0, + 1061.0, + 299.0, + 1011.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 275.0, + 1112.0, + 275.0, + 1112.0, + 299.0, + 1064.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1118.0, + 276.0, + 1165.0, + 276.0, + 1165.0, + 299.0, + 1118.0, + 299.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 529.0, + 316.0, + 620.0, + 316.0, + 620.0, + 343.0, + 529.0, + 343.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 329.0, + 827.0, + 329.0, + 827.0, + 361.0, + 740.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1000.0, + 346.0, + 1018.0, + 346.0, + 1018.0, + 353.0, + 1000.0, + 353.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1054.0, + 332.0, + 1092.0, + 332.0, + 1092.0, + 357.0, + 1054.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 740.0, + 351.0, + 781.0, + 351.0, + 781.0, + 377.0, + 740.0, + 377.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 353.0, + 897.0, + 353.0, + 897.0, + 374.0, + 857.0, + 374.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 530.0, + 380.0, + 685.0, + 380.0, + 685.0, + 403.0, + 530.0, + 403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 375.0, + 762.0, + 375.0, + 762.0, + 400.0, + 719.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 770.0, + 376.0, + 811.0, + 376.0, + 811.0, + 399.0, + 770.0, + 399.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 819.0, + 380.0, + 842.0, + 380.0, + 842.0, + 395.0, + 819.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 844.0, + 372.0, + 897.0, + 372.0, + 897.0, + 400.0, + 844.0, + 400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 859.0, + 334.5, + 944.0, + 334.5, + 944.0, + 358.5, + 859.0, + 358.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 449.0, + 1406.0, + 449.0, + 1406.0, + 485.0, + 295.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 478.0, + 1404.0, + 478.0, + 1404.0, + 513.0, + 294.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 506.0, + 914.0, + 506.0, + 914.0, + 539.0, + 292.0, + 539.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1276.0, + 712.0, + 1276.0, + 712.0, + 1312.0, + 295.0, + 1312.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1204.0, + 563.0, + 1204.0, + 563.0, + 1254.0, + 291.0, + 1254.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2120.0, + 838.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1331.0, + 1405.0, + 1331.0, + 1405.0, + 1370.0, + 294.0, + 1370.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1366.0, + 1406.0, + 1366.0, + 1406.0, + 1397.0, + 292.0, + 1397.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1395.0, + 371.0, + 1395.0, + 371.0, + 1429.0, + 294.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1395.0, + 1404.0, + 1395.0, + 1404.0, + 1429.0, + 495.0, + 1429.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1420.0, + 967.0, + 1420.0, + 967.0, + 1463.0, + 292.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1178.0, + 1420.0, + 1405.0, + 1420.0, + 1405.0, + 1463.0, + 1178.0, + 1463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 850.0, + 1457.0, + 850.0, + 1490.0, + 294.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 991.0, + 1457.0, + 1084.0, + 1457.0, + 1084.0, + 1490.0, + 991.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1109.0, + 1457.0, + 1330.0, + 1457.0, + 1330.0, + 1490.0, + 1109.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1358.0, + 1457.0, + 1405.0, + 1457.0, + 1405.0, + 1490.0, + 1358.0, + 1490.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1487.0, + 614.0, + 1487.0, + 614.0, + 1519.0, + 292.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 640.0, + 1487.0, + 895.0, + 1487.0, + 895.0, + 1519.0, + 640.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 1487.0, + 1212.0, + 1487.0, + 1212.0, + 1519.0, + 1155.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1487.0, + 1406.0, + 1487.0, + 1406.0, + 1519.0, + 1402.0, + 1519.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1518.0, + 995.0, + 1518.0, + 995.0, + 1552.0, + 294.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1020.0, + 1518.0, + 1406.0, + 1518.0, + 1406.0, + 1552.0, + 1020.0, + 1552.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1548.0, + 1226.0, + 1548.0, + 1226.0, + 1582.0, + 294.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1253.0, + 1548.0, + 1262.0, + 1548.0, + 1262.0, + 1582.0, + 1253.0, + 1582.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 661.0, + 1405.0, + 661.0, + 1405.0, + 698.0, + 293.0, + 698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 691.0, + 1405.0, + 691.0, + 1405.0, + 727.0, + 292.0, + 727.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 719.0, + 1403.0, + 719.0, + 1403.0, + 762.0, + 292.0, + 762.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 749.0, + 1404.0, + 749.0, + 1404.0, + 792.0, + 292.0, + 792.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 783.0, + 1403.0, + 783.0, + 1403.0, + 820.0, + 296.0, + 820.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 816.0, + 1404.0, + 816.0, + 1404.0, + 849.0, + 296.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 846.0, + 1404.0, + 846.0, + 1404.0, + 879.0, + 296.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 872.0, + 1253.0, + 872.0, + 1253.0, + 912.0, + 292.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1273.0, + 872.0, + 1407.0, + 872.0, + 1407.0, + 912.0, + 1273.0, + 912.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 904.0, + 1260.0, + 904.0, + 1260.0, + 942.0, + 294.0, + 942.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 952.0, + 1405.0, + 952.0, + 1405.0, + 987.0, + 296.0, + 987.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 984.0, + 1406.0, + 984.0, + 1406.0, + 1019.0, + 295.0, + 1019.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1015.0, + 1405.0, + 1015.0, + 1405.0, + 1047.0, + 292.0, + 1047.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1044.0, + 1406.0, + 1044.0, + 1406.0, + 1079.0, + 294.0, + 1079.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1074.0, + 1405.0, + 1074.0, + 1405.0, + 1108.0, + 295.0, + 1108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1103.0, + 1404.0, + 1103.0, + 1404.0, + 1140.0, + 292.0, + 1140.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1138.0, + 555.0, + 1138.0, + 555.0, + 1168.0, + 295.0, + 1168.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1881.0, + 752.0, + 1881.0, + 752.0, + 1918.0, + 296.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 792.0, + 1881.0, + 1163.0, + 1881.0, + 1163.0, + 1918.0, + 792.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1192.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1918.0, + 1192.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1945.0, + 293.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1941.0, + 1408.0, + 1941.0, + 1408.0, + 1980.0, + 292.0, + 1980.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1971.0, + 1405.0, + 1971.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 2000.0, + 1262.0, + 2000.0, + 1262.0, + 2040.0, + 293.0, + 2040.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 1405.0, + 1593.0, + 1405.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1626.0, + 1405.0, + 1626.0, + 1405.0, + 1659.0, + 295.0, + 1659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1657.0, + 1175.0, + 1657.0, + 1175.0, + 1690.0, + 295.0, + 1690.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1769.0, + 601.0, + 1769.0, + 601.0, + 1809.0, + 295.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 775.0, + 1769.0, + 1404.0, + 1769.0, + 1404.0, + 1809.0, + 775.0, + 1809.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 334.0, + 1805.0, + 1033.0, + 1805.0, + 1033.0, + 1837.0, + 334.0, + 1837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 583.0, + 1404.0, + 583.0, + 1404.0, + 619.0, + 295.0, + 619.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 615.0, + 1071.0, + 615.0, + 1071.0, + 651.0, + 295.0, + 651.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 1, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1424, + 1405, + 1424, + 1405, + 1672, + 297, + 1672 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 927, + 1403, + 927, + 1403, + 1143, + 298, + 1143 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 414, + 1404, + 414, + 1404, + 599, + 297, + 599 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1158, + 1403, + 1158, + 1403, + 1311, + 298, + 1311 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 299, + 631, + 1403, + 631, + 1403, + 756, + 299, + 756 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 771, + 1404, + 771, + 1404, + 893, + 298, + 893 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 297, + 1804, + 1403, + 1804, + 1403, + 1934, + 297, + 1934 + ], + "score": 0.975 + }, + { + "category_id": 1, + "poly": [ + 300, + 286, + 1400, + 286, + 1400, + 380, + 300, + 380 + ], + "score": 0.964 + }, + { + "category_id": 8, + "poly": [ + 578, + 1950, + 1116, + 1950, + 1116, + 2045, + 578, + 2045 + ], + "score": 0.96 + }, + { + "category_id": 8, + "poly": [ + 525, + 1707, + 1170, + 1707, + 1170, + 1785, + 525, + 1785 + ], + "score": 0.952 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.908 + }, + { + "category_id": 0, + "poly": [ + 299, + 1355, + 481, + 1355, + 481, + 1391, + 299, + 1391 + ], + "score": 0.902 + }, + { + "category_id": 9, + "poly": [ + 1366, + 1730, + 1400, + 1730, + 1400, + 1761, + 1366, + 1761 + ], + "score": 0.892 + }, + { + "category_id": 9, + "poly": [ + 1365, + 1981, + 1400, + 1981, + 1400, + 2011, + 1365, + 2011 + ], + "score": 0.878 + }, + { + "category_id": 0, + "poly": [ + 297, + 229, + 917, + 229, + 917, + 262, + 297, + 262 + ], + "score": 0.771 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.662 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 859, + 2088, + 859, + 2112, + 841, + 2112 + ], + "score": 0.581 + }, + { + "category_id": 1, + "poly": [ + 297, + 229, + 917, + 229, + 917, + 262, + 297, + 262 + ], + "score": 0.118 + }, + { + "category_id": 14, + "poly": [ + 578, + 1948, + 1119, + 1948, + 1119, + 2044, + 578, + 2044 + ], + "score": 0.94, + "latex": "L _ { x } \\left( \\theta _ { x } \\right) = \\underset { s \\sim D } { \\mathbb { E } } \\left[ - \\sum _ { k = 1 } ^ { K } Q \\left( s , k , x _ { k } \\left( s ; \\theta _ { x } \\right) ; \\theta _ { Q } \\right) \\right]" + }, + { + "category_id": 14, + "poly": [ + 526, + 1706, + 1172, + 1706, + 1172, + 1785, + 526, + 1785 + ], + "score": 0.93, + "latex": "\\begin{array} { r } { Q \\left( s , k , x \\right) = \\underset { r , s ^ { \\prime } } { \\mathbb { E } } \\left[ r + \\gamma \\underset { k ^ { \\prime } } { \\operatorname* { m a x } } \\ \\underset { x ^ { \\prime } \\in \\mathcal { X } } { \\operatorname* { s u p } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x ^ { \\prime } \\right) \\vert s , k , x \\right] . } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 297, + 1548, + 480, + 1548, + 480, + 1582, + 297, + 1582 + ], + "score": 0.92, + "latex": "k \\in \\mathcal { A } _ { d } = [ K ]" + }, + { + "category_id": 13, + "poly": [ + 298, + 1837, + 556, + 1837, + 556, + 1871, + 298, + 1871 + ], + "score": 0.9, + "latex": "x _ { k } ( s ; \\theta _ { x } ) : { \\cal { S } } { \\mathcal { X } } _ { k }" + }, + { + "category_id": 13, + "poly": [ + 819, + 664, + 857, + 664, + 857, + 694, + 819, + 694 + ], + "score": 0.89, + "latex": "\\mathbf { \\mathcal { A } } _ { d }" + }, + { + "category_id": 13, + "poly": [ + 795, + 567, + 874, + 567, + 874, + 599, + 795, + 599 + ], + "score": 0.89, + "latex": "[ - 1 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 1037, + 1837, + 1401, + 1837, + 1401, + 1876, + 1037, + 1876 + ], + "score": 0.88, + "latex": "\\begin{array} { r } { \\operatorname* { m a x } _ { k ^ { \\prime } } \\operatorname* { s u p } _ { x _ { k ^ { \\prime } } \\in \\mathcal { X } _ { k ^ { \\prime } } } Q \\left( s ^ { \\prime } , k ^ { \\prime } , x _ { k ^ { \\prime } } \\right) } \\end{array}" + }, + { + "category_id": 13, + "poly": [ + 559, + 724, + 638, + 724, + 638, + 758, + 559, + 758 + ], + "score": 0.86, + "latex": "[ - 1 , 1 ]" + }, + { + "category_id": 13, + "poly": [ + 863, + 637, + 891, + 637, + 891, + 663, + 863, + 663 + ], + "score": 0.86, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1208, + 1612, + 1227, + 1612, + 1227, + 1637, + 1208, + 1637 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 748, + 1839, + 768, + 1839, + 768, + 1865, + 748, + 1865 + ], + "score": 0.82, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 864, + 1580, + 894, + 1580, + 894, + 1606, + 864, + 1606 + ], + "score": 0.81, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 1314, + 1549, + 1343, + 1549, + 1343, + 1576, + 1314, + 1576 + ], + "score": 0.8, + "latex": "K" + }, + { + "category_id": 13, + "poly": [ + 298, + 696, + 324, + 696, + 324, + 721, + 298, + 721 + ], + "score": 0.8, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1116, + 1874, + 1135, + 1874, + 1135, + 1899, + 1116, + 1899 + ], + "score": 0.78, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 503, + 1646, + 521, + 1646, + 521, + 1667, + 503, + 1667 + ], + "score": 0.78, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 836, + 1878, + 855, + 1878, + 855, + 1899, + 836, + 1899 + ], + "score": 0.76, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 1036, + 1841, + 1111, + 1841, + 1111, + 1870, + 1036, + 1870 + ], + "score": 0.27, + "latex": "\\mathrm { m a x } _ { k ^ { \\prime } }" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1350.0, + 486.0, + 1350.0, + 486.0, + 1400.0, + 292.0, + 1400.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 921.0, + 227.0, + 921.0, + 265.0, + 292.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2085.0, + 862.0, + 2085.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1426.0, + 1405.0, + 1426.0, + 1405.0, + 1460.0, + 294.0, + 1460.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1454.0, + 1407.0, + 1454.0, + 1407.0, + 1492.0, + 292.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1486.0, + 1403.0, + 1486.0, + 1403.0, + 1523.0, + 292.0, + 1523.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1518.0, + 1405.0, + 1518.0, + 1405.0, + 1550.0, + 292.0, + 1550.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1546.0, + 296.0, + 1546.0, + 296.0, + 1585.0, + 291.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1546.0, + 1313.0, + 1546.0, + 1313.0, + 1585.0, + 481.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1344.0, + 1546.0, + 1406.0, + 1546.0, + 1406.0, + 1585.0, + 1344.0, + 1585.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1579.0, + 863.0, + 1579.0, + 863.0, + 1613.0, + 295.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 895.0, + 1579.0, + 1405.0, + 1579.0, + 1405.0, + 1613.0, + 895.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1607.0, + 1207.0, + 1607.0, + 1207.0, + 1646.0, + 292.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1228.0, + 1607.0, + 1407.0, + 1607.0, + 1407.0, + 1646.0, + 1228.0, + 1646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1637.0, + 502.0, + 1637.0, + 502.0, + 1676.0, + 294.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 522.0, + 1637.0, + 659.0, + 1637.0, + 659.0, + 1676.0, + 522.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 925.0, + 1404.0, + 925.0, + 1404.0, + 965.0, + 293.0, + 965.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 961.0, + 1403.0, + 961.0, + 1403.0, + 992.0, + 296.0, + 992.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 988.0, + 1405.0, + 988.0, + 1405.0, + 1023.0, + 293.0, + 1023.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1021.0, + 1404.0, + 1021.0, + 1404.0, + 1052.0, + 296.0, + 1052.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1052.0, + 1404.0, + 1052.0, + 1404.0, + 1083.0, + 296.0, + 1083.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1082.0, + 1404.0, + 1082.0, + 1404.0, + 1113.0, + 296.0, + 1113.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1109.0, + 1325.0, + 1109.0, + 1325.0, + 1149.0, + 293.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 414.0, + 1405.0, + 414.0, + 1405.0, + 449.0, + 295.0, + 449.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 444.0, + 1407.0, + 444.0, + 1407.0, + 481.0, + 292.0, + 481.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 476.0, + 1406.0, + 476.0, + 1406.0, + 510.0, + 294.0, + 510.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 505.0, + 1406.0, + 505.0, + 1406.0, + 541.0, + 294.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 1404.0, + 536.0, + 1404.0, + 573.0, + 294.0, + 573.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 566.0, + 794.0, + 566.0, + 794.0, + 602.0, + 295.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 875.0, + 566.0, + 1235.0, + 566.0, + 1235.0, + 602.0, + 875.0, + 602.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1157.0, + 1403.0, + 1157.0, + 1403.0, + 1194.0, + 296.0, + 1194.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1190.0, + 1404.0, + 1190.0, + 1404.0, + 1223.0, + 296.0, + 1223.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1220.0, + 1405.0, + 1220.0, + 1405.0, + 1253.0, + 293.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1245.0, + 1404.0, + 1245.0, + 1404.0, + 1290.0, + 292.0, + 1290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1279.0, + 691.0, + 1279.0, + 691.0, + 1315.0, + 292.0, + 1315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 631.0, + 862.0, + 631.0, + 862.0, + 667.0, + 294.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 892.0, + 631.0, + 1403.0, + 631.0, + 1403.0, + 667.0, + 892.0, + 667.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 662.0, + 818.0, + 662.0, + 818.0, + 699.0, + 293.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 662.0, + 1405.0, + 662.0, + 1405.0, + 699.0, + 858.0, + 699.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 691.0, + 297.0, + 691.0, + 297.0, + 729.0, + 292.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 325.0, + 691.0, + 1405.0, + 691.0, + 1405.0, + 729.0, + 325.0, + 729.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 724.0, + 558.0, + 724.0, + 558.0, + 759.0, + 294.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 639.0, + 724.0, + 1250.0, + 724.0, + 1250.0, + 759.0, + 639.0, + 759.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 768.0, + 1409.0, + 768.0, + 1409.0, + 807.0, + 292.0, + 807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 801.0, + 1408.0, + 801.0, + 1408.0, + 837.0, + 293.0, + 837.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 832.0, + 1408.0, + 832.0, + 1408.0, + 868.0, + 293.0, + 868.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 862.0, + 967.0, + 862.0, + 967.0, + 896.0, + 295.0, + 896.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1803.0, + 1405.0, + 1803.0, + 1405.0, + 1843.0, + 292.0, + 1843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 288.0, + 1831.0, + 297.0, + 1831.0, + 297.0, + 1881.0, + 288.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 557.0, + 1831.0, + 747.0, + 1831.0, + 747.0, + 1881.0, + 557.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 769.0, + 1831.0, + 1035.0, + 1831.0, + 1035.0, + 1881.0, + 769.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1402.0, + 1831.0, + 1408.0, + 1831.0, + 1408.0, + 1881.0, + 1402.0, + 1881.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1870.0, + 835.0, + 1870.0, + 835.0, + 1905.0, + 294.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 856.0, + 1870.0, + 1115.0, + 1870.0, + 1115.0, + 1905.0, + 856.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1136.0, + 1870.0, + 1405.0, + 1870.0, + 1405.0, + 1905.0, + 1136.0, + 1905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1903.0, + 1216.0, + 1903.0, + 1216.0, + 1936.0, + 295.0, + 1936.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 286.0, + 1404.0, + 286.0, + 1404.0, + 322.0, + 294.0, + 322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 317.0, + 1405.0, + 317.0, + 1405.0, + 352.0, + 295.0, + 352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 349.0, + 782.0, + 349.0, + 782.0, + 383.0, + 296.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 921.0, + 227.0, + 921.0, + 265.0, + 292.0, + 265.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 2, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 297, + 1348, + 1405, + 1348, + 1405, + 1626, + 297, + 1626 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 1850, + 1403, + 1850, + 1403, + 2034, + 297, + 2034 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 297, + 948, + 1403, + 948, + 1403, + 1134, + 297, + 1134 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 1148, + 1404, + 1148, + 1404, + 1333, + 297, + 1333 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1680, + 1403, + 1680, + 1403, + 1836, + 298, + 1836 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 703, + 1404, + 703, + 1404, + 894, + 297, + 894 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 503, + 1404, + 503, + 1404, + 690, + 298, + 690 + ], + "score": 0.978 + }, + { + "category_id": 3, + "poly": [ + 486, + 219, + 1199, + 219, + 1199, + 362, + 486, + 362 + ], + "score": 0.965 + }, + { + "category_id": 4, + "poly": [ + 654, + 398, + 1044, + 398, + 1044, + 429, + 654, + 429 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 297, + 75, + 858, + 75, + 858, + 105, + 297, + 105 + ], + "score": 0.91 + }, + { + "category_id": 2, + "poly": [ + 841, + 2089, + 858, + 2089, + 858, + 2111, + 841, + 2111 + ], + "score": 0.766 + }, + { + "category_id": 13, + "poly": [ + 639, + 1444, + 916, + 1444, + 916, + 1501, + 639, + 1501 + ], + "score": 0.94, + "latex": "\\mathbb { E } _ { s } \\left[ \\left( \\pi _ { x } ( s ) _ { i } - \\widetilde { \\pi } _ { x } ( s ) _ { i } \\right) ^ { 2 } \\right]" + }, + { + "category_id": 13, + "poly": [ + 917, + 735, + 979, + 735, + 979, + 774, + 917, + 774 + ], + "score": 0.93, + "latex": "a _ { t } ^ { s o f t }" + }, + { + "category_id": 13, + "poly": [ + 1111, + 1496, + 1394, + 1496, + 1394, + 1537, + 1111, + 1537 + ], + "score": 0.93, + "latex": "D _ { \\mathrm { K L } } ( \\pi _ { k } ^ { s o f t } ( s ) \\Vert \\widetilde { \\pi } _ { k } ^ { s o f t } ( s ) )" + }, + { + "category_id": 13, + "poly": [ + 441, + 1384, + 691, + 1384, + 691, + 1418, + 441, + 1418 + ], + "score": 0.92, + "latex": "d ( \\pi , \\widetilde { \\pi } ) = D _ { \\mathrm { K L } } ( \\pi \\| \\widetilde { \\pi } )" + }, + { + "category_id": 13, + "poly": [ + 864, + 1180, + 1028, + 1180, + 1028, + 1210, + 864, + 1210 + ], + "score": 0.92, + "latex": "\\mathcal { A } = \\mathcal { A } _ { d } \\times \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 803, + 735, + 865, + 735, + 865, + 773, + 803, + 773 + ], + "score": 0.92, + "latex": "k _ { t } ^ { s o f t }" + }, + { + "category_id": 13, + "poly": [ + 1128, + 505, + 1240, + 505, + 1240, + 538, + 1128, + 538 + ], + "score": 0.92, + "latex": "\\nabla L _ { x } \\left( \\theta _ { x } \\right)" + }, + { + "category_id": 13, + "poly": [ + 1155, + 1385, + 1351, + 1385, + 1351, + 1419, + 1155, + 1419 + ], + "score": 0.92, + "latex": "\\mu ( s ; \\theta _ { \\mu } ) , \\widetilde { \\mu } ( s ; \\theta _ { \\mu } )" + }, + { + "category_id": 13, + "poly": [ + 920, + 1348, + 1250, + 1348, + 1250, + 1388, + 920, + 1388 + ], + "score": 0.92, + "latex": "( \\pi _ { k } ^ { s o f t } ( s ) , \\pi _ { x } ( s ) ) = \\mu ( s ; \\theta _ { \\mu } ) ." + }, + { + "category_id": 13, + "poly": [ + 1142, + 702, + 1344, + 702, + 1344, + 737, + 1142, + 737 + ], + "score": 0.92, + "latex": "\\mu : \\mathcal { S } \\to \\mathbb { R } ^ { K } \\times \\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 1238, + 1741, + 1400, + 1741, + 1400, + 1776, + 1238, + 1776 + ], + "score": 0.91, + "latex": "( s , k ^ { s o f t } , x , r )" + }, + { + "category_id": 13, + "poly": [ + 985, + 567, + 1048, + 567, + 1048, + 600, + 985, + 600 + ], + "score": 0.91, + "latex": "\\nabla L _ { Q }" + }, + { + "category_id": 13, + "poly": [ + 311, + 735, + 717, + 735, + 717, + 774, + 311, + 774 + ], + "score": 0.91, + "latex": "\\mathrm { ~ ` ~ } _ { \\mu } \\mathrm { ~ a s : ~ } k _ { t } ^ { s o f t } , x _ { t } = a _ { t } ^ { s o f t } = \\mu ( s _ { t } ; \\theta _ { \\mu } )" + }, + { + "category_id": 13, + "poly": [ + 296, + 1453, + 364, + 1453, + 364, + 1489, + 296, + 1489 + ], + "score": 0.91, + "latex": "\\widetilde { \\pi } _ { x } ( s )" + }, + { + "category_id": 13, + "poly": [ + 764, + 1012, + 839, + 1012, + 839, + 1040, + 764, + 1040 + ], + "score": 0.89, + "latex": "\\epsilon < 1" + }, + { + "category_id": 13, + "poly": [ + 297, + 739, + 328, + 739, + 328, + 774, + 297, + 774 + ], + "score": 0.83, + "latex": "\\theta _ { \\mu }" + }, + { + "category_id": 13, + "poly": [ + 1142, + 1856, + 1169, + 1856, + 1169, + 1883, + 1142, + 1883 + ], + "score": 0.83, + "latex": "p _ { i }" + }, + { + "category_id": 13, + "poly": [ + 734, + 1242, + 753, + 1242, + 753, + 1268, + 734, + 1268 + ], + "score": 0.81, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 1129, + 1540, + 1149, + 1540, + 1149, + 1561, + 1129, + 1561 + ], + "score": 0.78, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 1127, + 1246, + 1146, + 1246, + 1146, + 1269, + 1127, + 1269 + ], + "score": 0.75, + "latex": "x" + }, + { + "category_id": 13, + "poly": [ + 932, + 1565, + 950, + 1565, + 950, + 1591, + 932, + 1591 + ], + "score": 0.73, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 489, + 1246, + 504, + 1246, + 504, + 1268, + 489, + 1268 + ], + "score": 0.66, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 657, + 987, + 671, + 987, + 671, + 1008, + 657, + 1008 + ], + "score": 0.62, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 1150, + 802, + 1228, + 802, + 1228, + 833, + 1150, + 833 + ], + "score": 0.4, + "latex": "( t a n h )" + }, + { + "category_id": 15, + "poly": [ + 834.0, + 225.0, + 910.0, + 225.0, + 910.0, + 274.0, + 834.0, + 274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 225.0, + 1066.0, + 225.0, + 1066.0, + 272.0, + 1005.0, + 272.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1146.0, + 237.0, + 1187.0, + 237.0, + 1187.0, + 262.0, + 1146.0, + 262.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 502.0, + 281.0, + 541.0, + 281.0, + 541.0, + 305.0, + 502.0, + 305.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 268.0, + 637.0, + 268.0, + 637.0, + 290.0, + 550.0, + 290.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 661.0, + 272.0, + 743.0, + 272.0, + 743.0, + 315.0, + 661.0, + 315.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 782.0, + 281.0, + 827.0, + 281.0, + 827.0, + 304.0, + 782.0, + 304.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 874.0, + 282.0, + 1010.0, + 282.0, + 1010.0, + 302.0, + 874.0, + 302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 275.0, + 1191.0, + 275.0, + 1191.0, + 295.0, + 1042.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 835.0, + 311.0, + 910.0, + 311.0, + 910.0, + 357.0, + 835.0, + 357.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 959.0, + 323.0, + 1038.0, + 323.0, + 1038.0, + 348.0, + 959.0, + 348.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 651.0, + 393.0, + 1048.0, + 393.0, + 1048.0, + 434.0, + 651.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 2086.0, + 862.0, + 2086.0, + 862.0, + 2118.0, + 838.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 287.0, + 1337.0, + 919.0, + 1337.0, + 919.0, + 1403.0, + 287.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1251.0, + 1337.0, + 1413.0, + 1337.0, + 1413.0, + 1403.0, + 1251.0, + 1403.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1382.0, + 440.0, + 1382.0, + 440.0, + 1423.0, + 294.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 692.0, + 1382.0, + 1154.0, + 1382.0, + 1154.0, + 1423.0, + 692.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1352.0, + 1382.0, + 1405.0, + 1382.0, + 1405.0, + 1423.0, + 1352.0, + 1423.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1413.0, + 1405.0, + 1413.0, + 1405.0, + 1450.0, + 295.0, + 1450.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 290.0, + 1450.0, + 295.0, + 1450.0, + 295.0, + 1495.0, + 290.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 365.0, + 1450.0, + 638.0, + 1450.0, + 638.0, + 1495.0, + 365.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 917.0, + 1450.0, + 1406.0, + 1450.0, + 1406.0, + 1495.0, + 917.0, + 1495.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 281.0, + 1471.0, + 1110.0, + 1471.0, + 1110.0, + 1570.0, + 281.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1395.0, + 1471.0, + 1416.0, + 1471.0, + 1416.0, + 1570.0, + 1395.0, + 1570.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1562.0, + 931.0, + 1562.0, + 931.0, + 1599.0, + 295.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 951.0, + 1562.0, + 1405.0, + 1562.0, + 1405.0, + 1599.0, + 951.0, + 1599.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1593.0, + 1245.0, + 1593.0, + 1245.0, + 1630.0, + 295.0, + 1630.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1849.0, + 1141.0, + 1849.0, + 1141.0, + 1885.0, + 295.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1170.0, + 1849.0, + 1403.0, + 1849.0, + 1403.0, + 1885.0, + 1170.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1915.0, + 293.0, + 1915.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1909.0, + 1407.0, + 1909.0, + 1407.0, + 1950.0, + 291.0, + 1950.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1941.0, + 1405.0, + 1941.0, + 1405.0, + 1979.0, + 294.0, + 1979.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1967.0, + 1405.0, + 1967.0, + 1405.0, + 2012.0, + 292.0, + 2012.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 2003.0, + 1155.0, + 2003.0, + 1155.0, + 2037.0, + 292.0, + 2037.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 946.0, + 1405.0, + 946.0, + 1405.0, + 986.0, + 294.0, + 986.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 978.0, + 656.0, + 978.0, + 656.0, + 1017.0, + 292.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 672.0, + 978.0, + 1405.0, + 978.0, + 1405.0, + 1017.0, + 672.0, + 1017.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1009.0, + 763.0, + 1009.0, + 763.0, + 1046.0, + 292.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 1009.0, + 1405.0, + 1009.0, + 1405.0, + 1046.0, + 840.0, + 1046.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1041.0, + 1407.0, + 1041.0, + 1407.0, + 1077.0, + 293.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1072.0, + 1407.0, + 1072.0, + 1407.0, + 1107.0, + 293.0, + 1107.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1104.0, + 1003.0, + 1104.0, + 1003.0, + 1135.0, + 295.0, + 1135.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1150.0, + 1405.0, + 1150.0, + 1405.0, + 1182.0, + 296.0, + 1182.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1179.0, + 863.0, + 1179.0, + 863.0, + 1215.0, + 294.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1029.0, + 1179.0, + 1404.0, + 1179.0, + 1404.0, + 1215.0, + 1029.0, + 1215.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1210.0, + 1405.0, + 1210.0, + 1405.0, + 1246.0, + 295.0, + 1246.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1240.0, + 488.0, + 1240.0, + 488.0, + 1276.0, + 295.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 505.0, + 1240.0, + 733.0, + 1240.0, + 733.0, + 1276.0, + 505.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 754.0, + 1240.0, + 1126.0, + 1240.0, + 1126.0, + 1276.0, + 754.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1147.0, + 1240.0, + 1404.0, + 1240.0, + 1404.0, + 1276.0, + 1147.0, + 1276.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1270.0, + 1406.0, + 1270.0, + 1406.0, + 1307.0, + 292.0, + 1307.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1302.0, + 1262.0, + 1302.0, + 1262.0, + 1337.0, + 294.0, + 1337.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1678.0, + 1405.0, + 1678.0, + 1405.0, + 1718.0, + 293.0, + 1718.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1711.0, + 1405.0, + 1711.0, + 1405.0, + 1747.0, + 293.0, + 1747.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 289.0, + 1734.0, + 1237.0, + 1734.0, + 1237.0, + 1783.0, + 289.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1401.0, + 1734.0, + 1408.0, + 1734.0, + 1408.0, + 1783.0, + 1401.0, + 1783.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1773.0, + 1407.0, + 1773.0, + 1407.0, + 1810.0, + 294.0, + 1810.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1805.0, + 1215.0, + 1805.0, + 1215.0, + 1838.0, + 296.0, + 1838.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 700.0, + 1141.0, + 700.0, + 1141.0, + 740.0, + 291.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1345.0, + 700.0, + 1406.0, + 700.0, + 1406.0, + 740.0, + 1345.0, + 740.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 285.0, + 723.0, + 296.0, + 723.0, + 296.0, + 787.0, + 285.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 718.0, + 723.0, + 802.0, + 723.0, + 802.0, + 787.0, + 718.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 866.0, + 723.0, + 916.0, + 723.0, + 916.0, + 787.0, + 866.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 980.0, + 723.0, + 1411.0, + 723.0, + 1411.0, + 787.0, + 980.0, + 787.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 766.0, + 1405.0, + 766.0, + 1405.0, + 806.0, + 294.0, + 806.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 798.0, + 1149.0, + 798.0, + 1149.0, + 836.0, + 292.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 798.0, + 1405.0, + 798.0, + 1405.0, + 836.0, + 1229.0, + 836.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 830.0, + 1405.0, + 830.0, + 1405.0, + 866.0, + 294.0, + 866.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 861.0, + 1404.0, + 861.0, + 1404.0, + 897.0, + 294.0, + 897.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 504.0, + 1127.0, + 504.0, + 1127.0, + 540.0, + 294.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1241.0, + 504.0, + 1404.0, + 504.0, + 1404.0, + 540.0, + 1241.0, + 540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 536.0, + 1408.0, + 536.0, + 1408.0, + 572.0, + 294.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 567.0, + 984.0, + 567.0, + 984.0, + 600.0, + 293.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 567.0, + 1405.0, + 567.0, + 1405.0, + 600.0, + 1049.0, + 600.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 597.0, + 1405.0, + 597.0, + 1405.0, + 632.0, + 294.0, + 632.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 626.0, + 1404.0, + 626.0, + 1404.0, + 663.0, + 292.0, + 663.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 657.0, + 1195.0, + 657.0, + 1195.0, + 696.0, + 293.0, + 696.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 3, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1403, + 1403, + 1403, + 1403, + 1618, + 298, + 1618 + ], + "score": 0.982 + }, + { + "category_id": 1, + "poly": [ + 298, + 1633, + 1404, + 1633, + 1404, + 1816, + 298, + 1816 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 1185, + 1403, + 1185, + 1403, + 1309, + 299, + 1309 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 297, + 228, + 1402, + 228, + 1402, + 354, + 297, + 354 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 791, + 1404, + 791, + 1404, + 894, + 299, + 894 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 537, + 1404, + 537, + 1404, + 667, + 298, + 667 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 300, + 1910, + 1402, + 1910, + 1402, + 2034, + 300, + 2034 + ], + "score": 0.975 + }, + { + "category_id": 8, + "poly": [ + 544, + 681, + 1153, + 681, + 1153, + 775, + 544, + 775 + ], + "score": 0.955 + }, + { + "category_id": 1, + "poly": [ + 293, + 367, + 1401, + 367, + 1401, + 433, + 293, + 433 + ], + "score": 0.955 + }, + { + "category_id": 8, + "poly": [ + 564, + 907, + 1133, + 907, + 1133, + 982, + 564, + 982 + ], + "score": 0.953 + }, + { + "category_id": 8, + "poly": [ + 433, + 445, + 1265, + 445, + 1265, + 525, + 433, + 525 + ], + "score": 0.944 + }, + { + "category_id": 1, + "poly": [ + 300, + 993, + 962, + 993, + 962, + 1026, + 300, + 1026 + ], + "score": 0.934 + }, + { + "category_id": 1, + "poly": [ + 298, + 1039, + 1217, + 1039, + 1217, + 1073, + 298, + 1073 + ], + "score": 0.923 + }, + { + "category_id": 0, + "poly": [ + 300, + 1116, + 558, + 1116, + 558, + 1151, + 300, + 1151 + ], + "score": 0.911 + }, + { + "category_id": 0, + "poly": [ + 299, + 1345, + 651, + 1345, + 651, + 1377, + 299, + 1377 + ], + "score": 0.904 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.9 + }, + { + "category_id": 9, + "poly": [ + 1366, + 714, + 1400, + 714, + 1400, + 743, + 1366, + 743 + ], + "score": 0.894 + }, + { + "category_id": 9, + "poly": [ + 1366, + 463, + 1400, + 463, + 1400, + 494, + 1366, + 494 + ], + "score": 0.887 + }, + { + "category_id": 9, + "poly": [ + 1365, + 929, + 1400, + 929, + 1400, + 959, + 1365, + 959 + ], + "score": 0.88 + }, + { + "category_id": 2, + "poly": [ + 841, + 2088, + 858, + 2088, + 858, + 2112, + 841, + 2112 + ], + "score": 0.714 + }, + { + "category_id": 0, + "poly": [ + 313, + 1853, + 1041, + 1853, + 1041, + 1885, + 313, + 1885 + ], + "score": 0.71 + }, + { + "category_id": 0, + "poly": [ + 297, + 1852, + 1045, + 1852, + 1045, + 1885, + 297, + 1885 + ], + "score": 0.289 + }, + { + "category_id": 14, + "poly": [ + 564, + 906, + 1134, + 906, + 1134, + 983, + 564, + 983 + ], + "score": 0.93, + "latex": "\\nabla L _ { Q } \\left( \\theta _ { Q } \\right) = \\nabla \\mathbb { E } \\left[ \\frac { 1 } { 2 } \\left( y - Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) \\right) ^ { 2 } \\right] ," + }, + { + "category_id": 13, + "poly": [ + 373, + 789, + 746, + 789, + 746, + 829, + 373, + 829 + ], + "score": 0.93, + "latex": "a ^ { * } = ( \\operatorname* { m a x } k _ { t } ^ { s o f t } , x _ { t } ) = ( k _ { t } , x _ { t } )" + }, + { + "category_id": 14, + "poly": [ + 544, + 680, + 1153, + 680, + 1153, + 776, + 544, + 776 + ], + "score": 0.93, + "latex": "y = h \\left( \\sum _ { k = 0 } ^ { n - 1 } r _ { t + k } \\gamma ^ { k } + \\gamma ^ { n } h ^ { - 1 } \\left( Q \\left( s _ { t + n } , a ^ { \\ast } ; \\theta _ { Q } ^ { - } \\right) \\right) \\right) ," + }, + { + "category_id": 14, + "poly": [ + 432, + 442, + 1264, + 442, + 1264, + 527, + 432, + 527 + ], + "score": 0.92, + "latex": "\\nabla L _ { \\mu } \\left( \\theta _ { \\mu } \\right) = \\nabla _ { s \\sim D } \\left[ Q \\left( s , \\mu ( s ; \\theta _ { \\mu } ) ; \\theta _ { Q } \\right) \\right] = \\nabla \\frac { 1 } { \\left| D \\right| } \\sum _ { s \\in D } Q \\left( s , k ^ { s o f t } , x ; \\theta _ { Q } \\right) ," + }, + { + "category_id": 13, + "poly": [ + 407, + 629, + 821, + 629, + 821, + 670, + 407, + 670 + ], + "score": 0.92, + "latex": "h ( x ) = \\mathrm { { s i g n } } ( x ) ( \\sqrt { | x | } + 1 - 1 ) + \\epsilon x" + }, + { + "category_id": 13, + "poly": [ + 344, + 828, + 379, + 828, + 379, + 866, + 344, + 866 + ], + "score": 0.89, + "latex": "\\theta _ { Q } ^ { - }" + }, + { + "category_id": 13, + "poly": [ + 372, + 540, + 398, + 540, + 398, + 566, + 372, + 566 + ], + "score": 0.83, + "latex": "D" + }, + { + "category_id": 13, + "poly": [ + 989, + 801, + 1008, + 801, + 1008, + 828, + 989, + 828 + ], + "score": 0.8, + "latex": "y" + }, + { + "category_id": 13, + "poly": [ + 698, + 327, + 719, + 327, + 719, + 349, + 698, + 349 + ], + "score": 0.77, + "latex": "\\sigma" + }, + { + "category_id": 13, + "poly": [ + 916, + 232, + 939, + 232, + 939, + 262, + 916, + 262 + ], + "score": 0.56, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 298, + 570, + 323, + 570, + 323, + 597, + 298, + 597 + ], + "score": 0.32, + "latex": "\\&" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1116.0, + 561.0, + 1116.0, + 561.0, + 1156.0, + 293.0, + 1156.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1346.0, + 653.0, + 1346.0, + 653.0, + 1379.0, + 295.0, + 1379.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2085.0, + 861.0, + 2085.0, + 861.0, + 2120.0, + 839.0, + 2120.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 309.0, + 1854.0, + 1047.0, + 1854.0, + 1047.0, + 1890.0, + 309.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1853.0, + 1048.0, + 1853.0, + 1048.0, + 1888.0, + 295.0, + 1888.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1403.0, + 1403.0, + 1403.0, + 1403.0, + 1438.0, + 294.0, + 1438.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1432.0, + 1404.0, + 1432.0, + 1404.0, + 1472.0, + 292.0, + 1472.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1462.0, + 1405.0, + 1462.0, + 1405.0, + 1498.0, + 293.0, + 1498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1492.0, + 1408.0, + 1492.0, + 1408.0, + 1533.0, + 291.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1524.0, + 1404.0, + 1524.0, + 1404.0, + 1561.0, + 293.0, + 1561.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1552.0, + 1405.0, + 1552.0, + 1405.0, + 1593.0, + 293.0, + 1593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1586.0, + 1137.0, + 1586.0, + 1137.0, + 1622.0, + 293.0, + 1622.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1632.0, + 1406.0, + 1632.0, + 1406.0, + 1668.0, + 296.0, + 1668.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1663.0, + 1405.0, + 1663.0, + 1405.0, + 1698.0, + 294.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1694.0, + 1406.0, + 1694.0, + 1406.0, + 1730.0, + 293.0, + 1730.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1722.0, + 1404.0, + 1722.0, + 1404.0, + 1761.0, + 293.0, + 1761.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1754.0, + 1406.0, + 1754.0, + 1406.0, + 1793.0, + 293.0, + 1793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1783.0, + 1389.0, + 1783.0, + 1389.0, + 1824.0, + 292.0, + 1824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1184.0, + 1403.0, + 1184.0, + 1403.0, + 1221.0, + 292.0, + 1221.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1214.0, + 1405.0, + 1214.0, + 1405.0, + 1253.0, + 293.0, + 1253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1248.0, + 1404.0, + 1248.0, + 1404.0, + 1281.0, + 294.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1277.0, + 996.0, + 1277.0, + 996.0, + 1310.0, + 294.0, + 1310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 227.0, + 915.0, + 227.0, + 915.0, + 266.0, + 292.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 940.0, + 227.0, + 1404.0, + 227.0, + 1404.0, + 266.0, + 940.0, + 266.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 262.0, + 1407.0, + 262.0, + 1407.0, + 296.0, + 291.0, + 296.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 292.0, + 1404.0, + 292.0, + 1404.0, + 329.0, + 293.0, + 329.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 322.0, + 697.0, + 322.0, + 697.0, + 358.0, + 295.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 720.0, + 322.0, + 1301.0, + 322.0, + 1301.0, + 358.0, + 720.0, + 358.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 286.0, + 786.0, + 372.0, + 786.0, + 372.0, + 839.0, + 286.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 747.0, + 786.0, + 988.0, + 786.0, + 988.0, + 839.0, + 747.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1009.0, + 786.0, + 1405.0, + 786.0, + 1405.0, + 839.0, + 1009.0, + 839.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 827.0, + 343.0, + 827.0, + 343.0, + 867.0, + 295.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 380.0, + 827.0, + 1405.0, + 827.0, + 1405.0, + 867.0, + 380.0, + 867.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 860.0, + 1018.0, + 860.0, + 1018.0, + 898.0, + 293.0, + 898.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 539.0, + 371.0, + 539.0, + 371.0, + 572.0, + 296.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 399.0, + 539.0, + 1405.0, + 539.0, + 1405.0, + 572.0, + 399.0, + 572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 566.0, + 297.0, + 566.0, + 297.0, + 604.0, + 292.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 566.0, + 1408.0, + 566.0, + 1408.0, + 604.0, + 324.0, + 604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 601.0, + 1404.0, + 601.0, + 1404.0, + 634.0, + 296.0, + 634.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 633.0, + 406.0, + 633.0, + 406.0, + 670.0, + 294.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 822.0, + 633.0, + 1142.0, + 633.0, + 1142.0, + 670.0, + 822.0, + 670.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1945.0, + 296.0, + 1945.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1940.0, + 1406.0, + 1940.0, + 1406.0, + 1978.0, + 293.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1974.0, + 1403.0, + 1974.0, + 1403.0, + 2007.0, + 295.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2002.0, + 1406.0, + 2002.0, + 1406.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 369.0, + 1403.0, + 369.0, + 1403.0, + 401.0, + 297.0, + 401.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 401.0, + 1378.0, + 401.0, + 1378.0, + 434.0, + 296.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 986.0, + 965.0, + 986.0, + 965.0, + 1036.0, + 293.0, + 1036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1037.0, + 1220.0, + 1037.0, + 1220.0, + 1077.0, + 296.0, + 1077.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 4, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1457, + 1403, + 1457, + 1403, + 1671, + 298, + 1671 + ], + "score": 0.979 + }, + { + "category_id": 1, + "poly": [ + 298, + 1850, + 1403, + 1850, + 1403, + 2034, + 298, + 2034 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 1256, + 1404, + 1256, + 1404, + 1440, + 298, + 1440 + ], + "score": 0.977 + }, + { + "category_id": 1, + "poly": [ + 299, + 1712, + 1402, + 1712, + 1402, + 1835, + 299, + 1835 + ], + "score": 0.976 + }, + { + "category_id": 3, + "poly": [ + 307, + 228, + 931, + 228, + 931, + 698, + 307, + 698 + ], + "score": 0.95 + }, + { + "category_id": 3, + "poly": [ + 1025, + 243, + 1319, + 243, + 1319, + 720, + 1025, + 720 + ], + "score": 0.948 + }, + { + "category_id": 3, + "poly": [ + 313, + 844, + 1376, + 844, + 1376, + 1117, + 313, + 1117 + ], + "score": 0.948 + }, + { + "category_id": 4, + "poly": [ + 320, + 756, + 826, + 756, + 826, + 790, + 320, + 790 + ], + "score": 0.918 + }, + { + "category_id": 4, + "poly": [ + 1006, + 756, + 1353, + 756, + 1353, + 788, + 1006, + 788 + ], + "score": 0.917 + }, + { + "category_id": 4, + "poly": [ + 550, + 1163, + 1146, + 1163, + 1146, + 1195, + 550, + 1195 + ], + "score": 0.912 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.86 + }, + { + "category_id": 2, + "poly": [ + 840, + 2089, + 858, + 2089, + 858, + 2112, + 840, + 2112 + ], + "score": 0.782 + }, + { + "category_id": 2, + "poly": [ + 299, + 76, + 856, + 76, + 856, + 104, + 299, + 104 + ], + "score": 0.101 + }, + { + "category_id": 13, + "poly": [ + 893, + 1257, + 1004, + 1257, + 1004, + 1291, + 893, + 1291 + ], + "score": 0.92, + "latex": "( - 7 , - 1 ) )" + }, + { + "category_id": 13, + "poly": [ + 1190, + 1549, + 1261, + 1549, + 1261, + 1579, + 1190, + 1579 + ], + "score": 0.57, + "latex": "- 1 0 0 )" + }, + { + "category_id": 15, + "poly": [ + 888.0, + 272.0, + 918.0, + 272.0, + 918.0, + 301.0, + 888.0, + 301.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 291.0, + 339.0, + 291.0, + 339.0, + 310.0, + 320.0, + 310.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 887.0, + 335.0, + 920.0, + 335.0, + 920.0, + 368.0, + 887.0, + 368.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 374.0, + 338.0, + 374.0, + 338.0, + 391.0, + 320.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 406.0, + 908.0, + 406.0, + 908.0, + 428.0, + 890.0, + 428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 456.0, + 339.0, + 456.0, + 339.0, + 475.0, + 320.0, + 475.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 890.0, + 469.0, + 932.0, + 469.0, + 932.0, + 498.0, + 890.0, + 498.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 538.0, + 340.0, + 538.0, + 340.0, + 559.0, + 317.0, + 559.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 536.0, + 931.0, + 536.0, + 931.0, + 562.0, + 891.0, + 562.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 891.0, + 602.0, + 931.0, + 602.0, + 931.0, + 627.0, + 891.0, + 627.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 300.0, + 617.0, + 343.0, + 617.0, + 343.0, + 646.0, + 300.0, + 646.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 371.0, + 672.0, + 399.0, + 672.0, + 399.0, + 695.0, + 371.0, + 695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 475.0, + 671.0, + 491.0, + 671.0, + 491.0, + 692.0, + 475.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 576.0, + 672.0, + 590.0, + 672.0, + 590.0, + 693.0, + 576.0, + 693.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 676.0, + 673.0, + 691.0, + 673.0, + 691.0, + 692.0, + 676.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 776.0, + 672.0, + 791.0, + 672.0, + 791.0, + 692.0, + 776.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1047.0, + 253.0, + 1065.0, + 253.0, + 1065.0, + 270.0, + 1047.0, + 270.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 287.0, + 1065.0, + 287.0, + 1065.0, + 300.0, + 1053.0, + 300.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1028.0, + 295.0, + 1045.0, + 295.0, + 1045.0, + 361.0, + 1028.0, + 361.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 318.0, + 1064.0, + 318.0, + 1064.0, + 333.0, + 1043.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1043.0, + 349.0, + 1064.0, + 349.0, + 1064.0, + 365.0, + 1043.0, + 365.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1282.0, + 368.0, + 1305.0, + 368.0, + 1305.0, + 383.0, + 1282.0, + 383.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1040.0, + 379.0, + 1066.0, + 379.0, + 1066.0, + 398.0, + 1040.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 380.0, + 1308.0, + 380.0, + 1308.0, + 398.0, + 1280.0, + 398.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1070.0, + 405.0, + 1086.0, + 405.0, + 1086.0, + 417.0, + 1070.0, + 417.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1088.0, + 403.0, + 1318.0, + 403.0, + 1318.0, + 419.0, + 1088.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1167.0, + 414.0, + 1210.0, + 414.0, + 1210.0, + 434.0, + 1167.0, + 434.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1080.0, + 430.0, + 1278.0, + 430.0, + 1278.0, + 458.0, + 1080.0, + 458.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 525.0, + 1065.0, + 525.0, + 1065.0, + 544.0, + 1042.0, + 544.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1011.0, + 543.0, + 1070.0, + 543.0, + 1070.0, + 631.0, + 1011.0, + 631.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1042.0, + 551.0, + 1063.0, + 551.0, + 1063.0, + 571.0, + 1042.0, + 571.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1049.0, + 579.0, + 1062.0, + 579.0, + 1062.0, + 593.0, + 1049.0, + 593.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1102.0, + 582.0, + 1121.0, + 582.0, + 1121.0, + 589.0, + 1102.0, + 589.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1155.0, + 569.0, + 1300.0, + 569.0, + 1300.0, + 605.0, + 1155.0, + 605.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 629.0, + 1062.0, + 629.0, + 1062.0, + 648.0, + 1036.0, + 648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1230.0, + 627.0, + 1303.0, + 627.0, + 1303.0, + 643.0, + 1230.0, + 643.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 637.0, + 1294.0, + 637.0, + 1294.0, + 659.0, + 1229.0, + 659.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1064.0, + 662.0, + 1314.0, + 662.0, + 1314.0, + 678.0, + 1064.0, + 678.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1164.0, + 675.0, + 1205.0, + 675.0, + 1205.0, + 692.0, + 1164.0, + 692.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 690.0, + 1314.0, + 690.0, + 1314.0, + 717.0, + 1036.0, + 717.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 418.0, + 848.0, + 563.0, + 848.0, + 563.0, + 870.0, + 418.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 768.0, + 848.0, + 949.0, + 848.0, + 949.0, + 870.0, + 768.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1143.0, + 848.0, + 1308.0, + 848.0, + 1308.0, + 870.0, + 1143.0, + 870.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 858.0, + 357.0, + 858.0, + 357.0, + 879.0, + 322.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 691.0, + 858.0, + 726.0, + 858.0, + 726.0, + 879.0, + 691.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1059.0, + 858.0, + 1094.0, + 858.0, + 1094.0, + 879.0, + 1059.0, + 879.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 886.0, + 355.0, + 886.0, + 355.0, + 908.0, + 322.0, + 908.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 879.0, + 722.0, + 879.0, + 722.0, + 901.0, + 689.0, + 901.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1057.0, + 883.0, + 1089.0, + 883.0, + 1089.0, + 905.0, + 1057.0, + 905.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 299.0, + 916.0, + 365.0, + 916.0, + 365.0, + 997.0, + 299.0, + 997.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 665.0, + 901.0, + 733.0, + 901.0, + 733.0, + 1001.0, + 665.0, + 1001.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1036.0, + 908.0, + 1101.0, + 908.0, + 1101.0, + 998.0, + 1036.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 977.0, + 354.0, + 977.0, + 354.0, + 998.0, + 322.0, + 998.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 689.0, + 991.0, + 722.0, + 991.0, + 722.0, + 1013.0, + 689.0, + 1013.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 982.0, + 1090.0, + 982.0, + 1090.0, + 1008.0, + 1058.0, + 1008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1006.0, + 354.0, + 1006.0, + 354.0, + 1028.0, + 321.0, + 1028.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 688.0, + 1012.0, + 722.0, + 1012.0, + 722.0, + 1038.0, + 688.0, + 1038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1058.0, + 1010.0, + 1090.0, + 1010.0, + 1090.0, + 1032.0, + 1058.0, + 1032.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 440.0, + 1045.0, + 471.0, + 1045.0, + 471.0, + 1066.0, + 440.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 481.0, + 1046.0, + 500.0, + 1046.0, + 500.0, + 1064.0, + 481.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 514.0, + 1046.0, + 536.0, + 1046.0, + 536.0, + 1064.0, + 514.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 545.0, + 1043.0, + 642.0, + 1043.0, + 642.0, + 1067.0, + 545.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1045.0, + 840.0, + 1045.0, + 840.0, + 1066.0, + 809.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 851.0, + 1048.0, + 868.0, + 1048.0, + 868.0, + 1062.0, + 851.0, + 1062.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 1046.0, + 904.0, + 1046.0, + 904.0, + 1064.0, + 882.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 915.0, + 1045.0, + 1009.0, + 1045.0, + 1009.0, + 1066.0, + 915.0, + 1066.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1179.0, + 1047.0, + 1206.0, + 1047.0, + 1206.0, + 1064.0, + 1179.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1248.0, + 1046.0, + 1271.0, + 1046.0, + 1271.0, + 1064.0, + 1248.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1284.0, + 1043.0, + 1347.0, + 1043.0, + 1347.0, + 1067.0, + 1284.0, + 1067.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1350.0, + 1047.0, + 1375.0, + 1047.0, + 1375.0, + 1064.0, + 1350.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 456.0, + 1057.0, + 526.0, + 1057.0, + 526.0, + 1077.0, + 456.0, + 1077.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 809.0, + 1056.0, + 907.0, + 1056.0, + 907.0, + 1078.0, + 809.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1181.0, + 1056.0, + 1269.0, + 1056.0, + 1269.0, + 1078.0, + 1181.0, + 1078.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 383.0, + 1081.0, + 577.0, + 1081.0, + 577.0, + 1114.0, + 383.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 719.0, + 1082.0, + 980.0, + 1082.0, + 980.0, + 1116.0, + 719.0, + 1116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1100.0, + 1081.0, + 1333.0, + 1081.0, + 1333.0, + 1114.0, + 1100.0, + 1114.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 315.75, + 1032.5, + 440.75, + 1032.5, + 440.75, + 1074.5, + 315.75, + 1074.5 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 683.0, + 1033.0, + 808.0, + 1033.0, + 808.0, + 1075.0, + 683.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1053.0, + 1033.0, + 1176.0, + 1033.0, + 1176.0, + 1075.0, + 1053.0, + 1075.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1210.0, + 1046.0, + 1239.0, + 1046.0, + 1239.0, + 1064.0, + 1210.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 317.0, + 751.0, + 830.0, + 751.0, + 830.0, + 796.0, + 317.0, + 796.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1004.0, + 753.0, + 1354.0, + 753.0, + 1354.0, + 793.0, + 1004.0, + 793.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 550.0, + 1161.0, + 1150.0, + 1161.0, + 1150.0, + 1198.0, + 550.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 857.0, + 73.0, + 857.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1457.0, + 1405.0, + 1457.0, + 1405.0, + 1492.0, + 294.0, + 1492.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1490.0, + 1404.0, + 1490.0, + 1404.0, + 1521.0, + 296.0, + 1521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1517.0, + 1404.0, + 1517.0, + 1404.0, + 1553.0, + 292.0, + 1553.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1549.0, + 1189.0, + 1549.0, + 1189.0, + 1584.0, + 296.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1262.0, + 1549.0, + 1403.0, + 1549.0, + 1403.0, + 1584.0, + 1262.0, + 1584.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1577.0, + 1405.0, + 1577.0, + 1405.0, + 1613.0, + 293.0, + 1613.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1609.0, + 1405.0, + 1609.0, + 1405.0, + 1642.0, + 293.0, + 1642.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1639.0, + 1388.0, + 1639.0, + 1388.0, + 1676.0, + 293.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1849.0, + 1405.0, + 1849.0, + 1405.0, + 1885.0, + 296.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1917.0, + 294.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1912.0, + 1404.0, + 1912.0, + 1404.0, + 1947.0, + 296.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1945.0, + 1405.0, + 1945.0, + 1405.0, + 1976.0, + 296.0, + 1976.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1971.0, + 1406.0, + 1971.0, + 1406.0, + 2008.0, + 292.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1404.0, + 2003.0, + 1404.0, + 2038.0, + 294.0, + 2038.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1257.0, + 892.0, + 1257.0, + 892.0, + 1292.0, + 294.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1005.0, + 1257.0, + 1405.0, + 1257.0, + 1405.0, + 1292.0, + 1005.0, + 1292.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1288.0, + 1404.0, + 1288.0, + 1404.0, + 1321.0, + 293.0, + 1321.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1320.0, + 1402.0, + 1320.0, + 1402.0, + 1352.0, + 296.0, + 1352.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1351.0, + 1401.0, + 1351.0, + 1401.0, + 1382.0, + 296.0, + 1382.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1377.0, + 1406.0, + 1377.0, + 1406.0, + 1415.0, + 293.0, + 1415.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1408.0, + 1047.0, + 1408.0, + 1047.0, + 1446.0, + 295.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1711.0, + 1404.0, + 1711.0, + 1404.0, + 1746.0, + 294.0, + 1746.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1740.0, + 1406.0, + 1740.0, + 1406.0, + 1779.0, + 293.0, + 1779.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1775.0, + 1406.0, + 1775.0, + 1406.0, + 1807.0, + 295.0, + 1807.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1805.0, + 1080.0, + 1805.0, + 1080.0, + 1837.0, + 294.0, + 1837.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 5, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1028, + 1403, + 1028, + 1403, + 1334, + 298, + 1334 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 1350, + 1403, + 1350, + 1403, + 1565, + 298, + 1565 + ], + "score": 0.981 + }, + { + "category_id": 1, + "poly": [ + 298, + 1580, + 1403, + 1580, + 1403, + 1764, + 298, + 1764 + ], + "score": 0.98 + }, + { + "category_id": 1, + "poly": [ + 299, + 789, + 1403, + 789, + 1403, + 912, + 299, + 912 + ], + "score": 0.979 + }, + { + "category_id": 5, + "poly": [ + 320, + 491, + 1378, + 491, + 1378, + 652, + 320, + 652 + ], + "score": 0.978, + "html": "
OverallRanking Score (Quality)Seek Count (Efficiency)
MLP+Prioritized+14.54+8.56-5.98
RNN+Prioritized+32.63+14.44-18.19
RNN + Prioritized Sample+32.67+15.29-17.38
" + }, + { + "category_id": 1, + "poly": [ + 299, + 1880, + 1402, + 1880, + 1402, + 2033, + 299, + 2033 + ], + "score": 0.977 + }, + { + "category_id": 5, + "poly": [ + 397, + 224, + 1299, + 224, + 1299, + 353, + 397, + 353 + ], + "score": 0.975, + "html": "
OverallRanking Score (Quality)Seek Count (Efficiency)
OU Noise+1.00+1.20-0.20
Param Noise+1.69+0.82-0.87
" + }, + { + "category_id": 6, + "poly": [ + 299, + 387, + 1403, + 387, + 1403, + 451, + 299, + 451 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.909 + }, + { + "category_id": 0, + "poly": [ + 300, + 1816, + 663, + 1816, + 663, + 1848, + 300, + 1848 + ], + "score": 0.897 + }, + { + "category_id": 6, + "poly": [ + 486, + 686, + 1210, + 686, + 1210, + 720, + 486, + 720 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 300, + 966, + 580, + 966, + 580, + 997, + 300, + 997 + ], + "score": 0.891 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 858, + 2087, + 858, + 2111, + 842, + 2111 + ], + "score": 0.649 + }, + { + "category_id": 2, + "poly": [ + 842, + 2087, + 859, + 2087, + 859, + 2111, + 842, + 2111 + ], + "score": 0.272 + }, + { + "category_id": 13, + "poly": [ + 395, + 1182, + 488, + 1182, + 488, + 1217, + 395, + 1217 + ], + "score": 0.93, + "latex": "\\widetilde \\mu ( s ; \\theta _ { \\mu } )" + }, + { + "category_id": 13, + "poly": [ + 1190, + 1218, + 1205, + 1218, + 1205, + 1240, + 1190, + 1240 + ], + "score": 0.66, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 343, + 1127, + 357, + 1127, + 357, + 1150, + 343, + 1150 + ], + "score": 0.64, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 920, + 1127, + 935, + 1127, + 935, + 1150, + 920, + 1150 + ], + "score": 0.61, + "latex": "\\epsilon" + }, + { + "category_id": 13, + "poly": [ + 788, + 420, + 823, + 420, + 823, + 447, + 788, + 447 + ], + "score": 0.57, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 805, + 1158, + 820, + 1158, + 820, + 1179, + 805, + 1179 + ], + "score": 0.41, + "latex": "\\epsilon" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 386.0, + 1404.0, + 386.0, + 1404.0, + 424.0, + 294.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 415.0, + 787.0, + 415.0, + 787.0, + 455.0, + 294.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 824.0, + 415.0, + 1404.0, + 415.0, + 1404.0, + 455.0, + 824.0, + 455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1813.0, + 667.0, + 1813.0, + 667.0, + 1852.0, + 294.0, + 1852.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 488.0, + 685.0, + 1213.0, + 685.0, + 1213.0, + 725.0, + 488.0, + 725.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 964.0, + 585.0, + 964.0, + 585.0, + 1000.0, + 296.0, + 1000.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2086.0, + 860.0, + 2086.0, + 860.0, + 2118.0, + 839.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 840.0, + 2086.0, + 861.0, + 2086.0, + 861.0, + 2118.0, + 840.0, + 2118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1027.0, + 1404.0, + 1027.0, + 1404.0, + 1064.0, + 294.0, + 1064.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1060.0, + 1405.0, + 1060.0, + 1405.0, + 1096.0, + 294.0, + 1096.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1090.0, + 1404.0, + 1090.0, + 1404.0, + 1125.0, + 294.0, + 1125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1121.0, + 342.0, + 1121.0, + 342.0, + 1157.0, + 294.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 358.0, + 1121.0, + 919.0, + 1121.0, + 919.0, + 1157.0, + 358.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 936.0, + 1121.0, + 1404.0, + 1121.0, + 1404.0, + 1157.0, + 936.0, + 1157.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1151.0, + 804.0, + 1151.0, + 804.0, + 1187.0, + 293.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 821.0, + 1151.0, + 1403.0, + 1151.0, + 1403.0, + 1187.0, + 821.0, + 1187.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1182.0, + 394.0, + 1182.0, + 394.0, + 1218.0, + 294.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 489.0, + 1182.0, + 1405.0, + 1182.0, + 1405.0, + 1218.0, + 489.0, + 1218.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1211.0, + 1189.0, + 1211.0, + 1189.0, + 1248.0, + 292.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1206.0, + 1211.0, + 1407.0, + 1211.0, + 1407.0, + 1248.0, + 1206.0, + 1248.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1243.0, + 1405.0, + 1243.0, + 1405.0, + 1279.0, + 293.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1274.0, + 1407.0, + 1274.0, + 1407.0, + 1309.0, + 294.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1306.0, + 533.0, + 1306.0, + 533.0, + 1334.0, + 293.0, + 1334.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1349.0, + 1405.0, + 1349.0, + 1405.0, + 1384.0, + 294.0, + 1384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1382.0, + 1400.0, + 1382.0, + 1400.0, + 1413.0, + 297.0, + 1413.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1412.0, + 1405.0, + 1412.0, + 1405.0, + 1446.0, + 294.0, + 1446.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1443.0, + 1404.0, + 1443.0, + 1404.0, + 1476.0, + 294.0, + 1476.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1471.0, + 1405.0, + 1471.0, + 1405.0, + 1508.0, + 293.0, + 1508.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1503.0, + 1405.0, + 1503.0, + 1405.0, + 1540.0, + 293.0, + 1540.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1536.0, + 776.0, + 1536.0, + 776.0, + 1567.0, + 296.0, + 1567.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1579.0, + 1405.0, + 1579.0, + 1405.0, + 1615.0, + 296.0, + 1615.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1610.0, + 1405.0, + 1610.0, + 1405.0, + 1647.0, + 293.0, + 1647.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1642.0, + 1405.0, + 1642.0, + 1405.0, + 1677.0, + 294.0, + 1677.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1673.0, + 1404.0, + 1673.0, + 1404.0, + 1705.0, + 294.0, + 1705.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1701.0, + 1407.0, + 1701.0, + 1407.0, + 1739.0, + 292.0, + 1739.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1733.0, + 1078.0, + 1733.0, + 1078.0, + 1767.0, + 293.0, + 1767.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 791.0, + 1403.0, + 791.0, + 1403.0, + 824.0, + 295.0, + 824.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 819.0, + 1404.0, + 819.0, + 1404.0, + 856.0, + 293.0, + 856.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 853.0, + 1403.0, + 853.0, + 1403.0, + 885.0, + 295.0, + 885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 882.0, + 935.0, + 882.0, + 935.0, + 916.0, + 294.0, + 916.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 1879.0, + 1404.0, + 1879.0, + 1404.0, + 1917.0, + 297.0, + 1917.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1913.0, + 1404.0, + 1913.0, + 1404.0, + 1946.0, + 295.0, + 1946.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1942.0, + 1403.0, + 1942.0, + 1403.0, + 1975.0, + 295.0, + 1975.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1972.0, + 1404.0, + 1972.0, + 1404.0, + 2009.0, + 295.0, + 2009.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 2004.0, + 1206.0, + 2004.0, + 1206.0, + 2037.0, + 295.0, + 2037.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 6, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1021, + 1405, + 1021, + 1405, + 1604, + 298, + 1604 + ], + "score": 0.984 + }, + { + "category_id": 1, + "poly": [ + 298, + 1727, + 1404, + 1727, + 1404, + 2035, + 298, + 2035 + ], + "score": 0.983 + }, + { + "category_id": 1, + "poly": [ + 298, + 229, + 1403, + 229, + 1403, + 415, + 298, + 415 + ], + "score": 0.978 + }, + { + "category_id": 5, + "poly": [ + 591, + 588, + 1105, + 588, + 1105, + 716, + 591, + 716 + ], + "score": 0.972, + "html": "
Average Eval ReturnOurPA-DDPG
Platform-v00.95730.3113
Goal-v034.20-6.208
" + }, + { + "category_id": 1, + "poly": [ + 298, + 429, + 1404, + 429, + 1404, + 553, + 298, + 553 + ], + "score": 0.944 + }, + { + "category_id": 0, + "poly": [ + 300, + 948, + 587, + 948, + 587, + 984, + 300, + 984 + ], + "score": 0.913 + }, + { + "category_id": 0, + "poly": [ + 299, + 1654, + 545, + 1654, + 545, + 1689, + 299, + 1689 + ], + "score": 0.899 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 104, + 298, + 104 + ], + "score": 0.897 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 858, + 2088, + 858, + 2112, + 840, + 2112 + ], + "score": 0.794 + }, + { + "category_id": 1, + "poly": [ + 297, + 750, + 1405, + 750, + 1405, + 874, + 297, + 874 + ], + "score": 0.488 + }, + { + "category_id": 6, + "poly": [ + 297, + 750, + 1405, + 750, + 1405, + 874, + 297, + 874 + ], + "score": 0.478 + }, + { + "category_id": 13, + "poly": [ + 521, + 1389, + 623, + 1389, + 623, + 1420, + 521, + 1420 + ], + "score": 0.92, + "latex": "k \\in \\mathcal { A } _ { d }" + }, + { + "category_id": 13, + "poly": [ + 990, + 1295, + 1179, + 1295, + 1179, + 1331, + 990, + 1331 + ], + "score": 0.92, + "latex": "Q ( s ^ { \\prime } , k ^ { \\prime } , { \\bf x } ^ { \\tilde { Q } } ( s ^ { \\prime } ) )" + }, + { + "category_id": 13, + "poly": [ + 1041, + 381, + 1279, + 381, + 1279, + 416, + 1041, + 416 + ], + "score": 0.91, + "latex": "[ 1 \\times 1 0 ^ { - 3 } , 1 \\times 1 0 ^ { - 2 } ]" + }, + { + "category_id": 13, + "poly": [ + 600, + 320, + 824, + 320, + 824, + 354, + 600, + 354 + ], + "score": 0.89, + "latex": "[ 1 \\times 1 0 ^ { - 4 } , 1 \\times \\mathsf { \\bar { 1 } 0 ^ { - 3 } } ]" + }, + { + "category_id": 13, + "poly": [ + 701, + 461, + 756, + 461, + 756, + 490, + 701, + 490 + ], + "score": 0.86, + "latex": "20 \\%" + }, + { + "category_id": 13, + "poly": [ + 766, + 1360, + 794, + 1360, + 794, + 1386, + 766, + 1386 + ], + "score": 0.81, + "latex": "\\mathcal { X }" + }, + { + "category_id": 13, + "poly": [ + 886, + 326, + 908, + 326, + 908, + 349, + 886, + 349 + ], + "score": 0.8, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 1184, + 465, + 1207, + 465, + 1207, + 488, + 1184, + 488 + ], + "score": 0.78, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 757, + 388, + 776, + 388, + 776, + 410, + 757, + 410 + ], + "score": 0.78, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 435, + 466, + 454, + 466, + 454, + 488, + 435, + 488 + ], + "score": 0.77, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 524, + 462, + 542, + 462, + 542, + 488, + 524, + 488 + ], + "score": 0.76, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 1260, + 466, + 1280, + 466, + 1280, + 488, + 1260, + 488 + ], + "score": 0.76, + "latex": "\\tau" + }, + { + "category_id": 13, + "poly": [ + 755, + 354, + 773, + 354, + 773, + 380, + 755, + 380 + ], + "score": 0.71, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 326, + 320, + 551, + 320, + 551, + 354, + 326, + 354 + ], + "score": 0.58, + "latex": "[ 3 \\times 1 0 ^ { - 4 } , 3 \\times 1 0 ^ { - 3 } ]" + }, + { + "category_id": 13, + "poly": [ + 364, + 462, + 381, + 462, + 381, + 488, + 364, + 488 + ], + "score": 0.55, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 329, + 465, + 351, + 465, + 351, + 489, + 329, + 489 + ], + "score": 0.54, + "latex": "\\alpha" + }, + { + "category_id": 13, + "poly": [ + 329, + 462, + 383, + 462, + 383, + 490, + 329, + 490 + ], + "score": 0.25, + "latex": "\\alpha , \\delta" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 946.0, + 593.0, + 946.0, + 593.0, + 989.0, + 292.0, + 989.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1650.0, + 550.0, + 1650.0, + 550.0, + 1698.0, + 293.0, + 1698.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 858.0, + 72.0, + 858.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 836.0, + 2085.0, + 859.0, + 2085.0, + 859.0, + 2116.0, + 836.0, + 2116.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 752.0, + 1405.0, + 752.0, + 1405.0, + 785.0, + 295.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 783.0, + 1403.0, + 783.0, + 1403.0, + 816.0, + 295.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 844.0, + 938.0, + 844.0, + 938.0, + 876.0, + 296.0, + 876.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1021.0, + 1405.0, + 1021.0, + 1405.0, + 1060.0, + 295.0, + 1060.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1054.0, + 1405.0, + 1054.0, + 1405.0, + 1089.0, + 292.0, + 1089.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1085.0, + 1405.0, + 1085.0, + 1405.0, + 1118.0, + 293.0, + 1118.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1114.0, + 1406.0, + 1114.0, + 1406.0, + 1149.0, + 295.0, + 1149.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1146.0, + 1405.0, + 1146.0, + 1405.0, + 1178.0, + 293.0, + 1178.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1176.0, + 1405.0, + 1176.0, + 1405.0, + 1211.0, + 295.0, + 1211.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1207.0, + 1405.0, + 1207.0, + 1405.0, + 1242.0, + 295.0, + 1242.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1236.0, + 1406.0, + 1236.0, + 1406.0, + 1274.0, + 292.0, + 1274.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1267.0, + 1403.0, + 1267.0, + 1403.0, + 1302.0, + 295.0, + 1302.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 1294.0, + 989.0, + 1294.0, + 989.0, + 1335.0, + 291.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1180.0, + 1294.0, + 1406.0, + 1294.0, + 1406.0, + 1335.0, + 1180.0, + 1335.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1329.0, + 1405.0, + 1329.0, + 1405.0, + 1364.0, + 296.0, + 1364.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1359.0, + 765.0, + 1359.0, + 765.0, + 1394.0, + 295.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 795.0, + 1359.0, + 1406.0, + 1359.0, + 1406.0, + 1394.0, + 795.0, + 1394.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1383.0, + 520.0, + 1383.0, + 520.0, + 1428.0, + 292.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 624.0, + 1383.0, + 1406.0, + 1383.0, + 1406.0, + 1428.0, + 624.0, + 1428.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1418.0, + 1405.0, + 1418.0, + 1405.0, + 1456.0, + 293.0, + 1456.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1446.0, + 1405.0, + 1446.0, + 1405.0, + 1488.0, + 293.0, + 1488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1481.0, + 1405.0, + 1481.0, + 1405.0, + 1516.0, + 295.0, + 1516.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1508.0, + 1405.0, + 1508.0, + 1405.0, + 1546.0, + 292.0, + 1546.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1538.0, + 1405.0, + 1538.0, + 1405.0, + 1580.0, + 293.0, + 1580.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1574.0, + 628.0, + 1574.0, + 628.0, + 1609.0, + 296.0, + 1609.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1728.0, + 1405.0, + 1728.0, + 1405.0, + 1764.0, + 293.0, + 1764.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1759.0, + 1402.0, + 1759.0, + 1402.0, + 1794.0, + 294.0, + 1794.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1789.0, + 1406.0, + 1789.0, + 1406.0, + 1825.0, + 293.0, + 1825.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1854.0, + 293.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1853.0, + 1404.0, + 1853.0, + 1404.0, + 1885.0, + 296.0, + 1885.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1882.0, + 1405.0, + 1882.0, + 1405.0, + 1918.0, + 294.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1912.0, + 1405.0, + 1912.0, + 1405.0, + 1947.0, + 293.0, + 1947.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1942.0, + 1404.0, + 1942.0, + 1404.0, + 1978.0, + 294.0, + 1978.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1969.0, + 1406.0, + 1969.0, + 1406.0, + 2011.0, + 292.0, + 2011.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 2003.0, + 1277.0, + 2003.0, + 1277.0, + 2039.0, + 294.0, + 2039.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 229.0, + 1405.0, + 229.0, + 1405.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 263.0, + 1404.0, + 263.0, + 1404.0, + 295.0, + 296.0, + 295.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 290.0, + 1405.0, + 290.0, + 1405.0, + 326.0, + 293.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 316.0, + 325.0, + 316.0, + 325.0, + 359.0, + 291.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 552.0, + 316.0, + 599.0, + 316.0, + 599.0, + 359.0, + 552.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 825.0, + 316.0, + 885.0, + 316.0, + 885.0, + 359.0, + 825.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 316.0, + 1405.0, + 316.0, + 1405.0, + 359.0, + 909.0, + 359.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 350.0, + 754.0, + 350.0, + 754.0, + 388.0, + 292.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 774.0, + 350.0, + 1407.0, + 350.0, + 1407.0, + 388.0, + 774.0, + 388.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 380.0, + 756.0, + 380.0, + 756.0, + 419.0, + 292.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 777.0, + 380.0, + 1040.0, + 380.0, + 1040.0, + 419.0, + 777.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1280.0, + 380.0, + 1292.0, + 380.0, + 1292.0, + 419.0, + 1280.0, + 419.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 428.0, + 1405.0, + 428.0, + 1405.0, + 463.0, + 293.0, + 463.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 461.0, + 328.0, + 461.0, + 328.0, + 493.0, + 294.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 384.0, + 461.0, + 434.0, + 461.0, + 434.0, + 493.0, + 384.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 455.0, + 461.0, + 523.0, + 461.0, + 523.0, + 493.0, + 455.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 543.0, + 461.0, + 700.0, + 461.0, + 700.0, + 493.0, + 543.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 757.0, + 461.0, + 1183.0, + 461.0, + 1183.0, + 493.0, + 757.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1208.0, + 461.0, + 1259.0, + 461.0, + 1259.0, + 493.0, + 1208.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1281.0, + 461.0, + 1405.0, + 461.0, + 1405.0, + 493.0, + 1281.0, + 493.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 492.0, + 1402.0, + 492.0, + 1402.0, + 525.0, + 296.0, + 525.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 521.0, + 727.0, + 521.0, + 727.0, + 556.0, + 295.0, + 556.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 752.0, + 1405.0, + 752.0, + 1405.0, + 785.0, + 295.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 783.0, + 1403.0, + 783.0, + 1403.0, + 816.0, + 295.0, + 816.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 812.0, + 1406.0, + 812.0, + 1406.0, + 849.0, + 294.0, + 849.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 844.0, + 938.0, + 844.0, + 938.0, + 876.0, + 296.0, + 876.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 7, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 299, + 227, + 489, + 227, + 489, + 262, + 299, + 262 + ], + "score": 0.799 + }, + { + "category_id": 2, + "poly": [ + 840, + 2088, + 859, + 2088, + 859, + 2111, + 840, + 2111 + ], + "score": 0.774 + }, + { + "category_id": 1, + "poly": [ + 296, + 278, + 1399, + 278, + 1399, + 342, + 296, + 342 + ], + "score": 0.739 + }, + { + "category_id": 1, + "poly": [ + 299, + 637, + 1401, + 637, + 1401, + 702, + 299, + 702 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 293, + 554, + 1402, + 554, + 1402, + 619, + 293, + 619 + ], + "score": 0.697 + }, + { + "category_id": 1, + "poly": [ + 298, + 802, + 1401, + 802, + 1401, + 866, + 298, + 866 + ], + "score": 0.696 + }, + { + "category_id": 1, + "poly": [ + 296, + 718, + 1402, + 718, + 1402, + 785, + 296, + 785 + ], + "score": 0.694 + }, + { + "category_id": 1, + "poly": [ + 295, + 359, + 1399, + 359, + 1399, + 425, + 295, + 425 + ], + "score": 0.69 + }, + { + "category_id": 1, + "poly": [ + 296, + 965, + 1402, + 965, + 1402, + 1033, + 296, + 1033 + ], + "score": 0.689 + }, + { + "category_id": 1, + "poly": [ + 302, + 1047, + 1399, + 1047, + 1399, + 1143, + 302, + 1143 + ], + "score": 0.675 + }, + { + "category_id": 1, + "poly": [ + 298, + 1971, + 1399, + 1971, + 1399, + 2035, + 298, + 2035 + ], + "score": 0.669 + }, + { + "category_id": 1, + "poly": [ + 291, + 1160, + 1399, + 1160, + 1399, + 1226, + 291, + 1226 + ], + "score": 0.649 + }, + { + "category_id": 1, + "poly": [ + 297, + 441, + 1400, + 441, + 1400, + 536, + 297, + 536 + ], + "score": 0.636 + }, + { + "category_id": 1, + "poly": [ + 296, + 1806, + 1400, + 1806, + 1400, + 1871, + 296, + 1871 + ], + "score": 0.632 + }, + { + "category_id": 1, + "poly": [ + 289, + 1242, + 1400, + 1242, + 1400, + 1309, + 289, + 1309 + ], + "score": 0.628 + }, + { + "category_id": 1, + "poly": [ + 298, + 1889, + 1398, + 1889, + 1398, + 1954, + 298, + 1954 + ], + "score": 0.621 + }, + { + "category_id": 1, + "poly": [ + 297, + 883, + 1400, + 883, + 1400, + 949, + 297, + 949 + ], + "score": 0.609 + }, + { + "category_id": 1, + "poly": [ + 299, + 1580, + 1402, + 1580, + 1402, + 1676, + 299, + 1676 + ], + "score": 0.597 + }, + { + "category_id": 1, + "poly": [ + 297, + 1468, + 1400, + 1468, + 1400, + 1564, + 297, + 1564 + ], + "score": 0.579 + }, + { + "category_id": 1, + "poly": [ + 297, + 1324, + 1402, + 1324, + 1402, + 1452, + 297, + 1452 + ], + "score": 0.564 + }, + { + "category_id": 1, + "poly": [ + 294, + 1693, + 1402, + 1693, + 1402, + 1790, + 294, + 1790 + ], + "score": 0.552 + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 230.0, + 490.0, + 230.0, + 490.0, + 263.0, + 297.0, + 263.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 839.0, + 2087.0, + 861.0, + 2087.0, + 861.0, + 2117.0, + 839.0, + 2117.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 275.0, + 1405.0, + 275.0, + 1405.0, + 316.0, + 292.0, + 316.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 309.0, + 612.0, + 309.0, + 612.0, + 342.0, + 321.0, + 342.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 635.0, + 1403.0, + 635.0, + 1403.0, + 675.0, + 294.0, + 675.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 671.0, + 776.0, + 671.0, + 776.0, + 700.0, + 324.0, + 700.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 554.0, + 1405.0, + 554.0, + 1405.0, + 592.0, + 295.0, + 592.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 585.0, + 1014.0, + 585.0, + 1014.0, + 620.0, + 321.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 799.0, + 1405.0, + 799.0, + 1405.0, + 840.0, + 293.0, + 840.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 833.0, + 532.0, + 833.0, + 532.0, + 865.0, + 324.0, + 865.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 719.0, + 1403.0, + 719.0, + 1403.0, + 755.0, + 295.0, + 755.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 752.0, + 1078.0, + 752.0, + 1078.0, + 785.0, + 321.0, + 785.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 362.0, + 1401.0, + 362.0, + 1401.0, + 395.0, + 298.0, + 395.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 392.0, + 1317.0, + 392.0, + 1317.0, + 424.0, + 321.0, + 424.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 966.0, + 1403.0, + 966.0, + 1403.0, + 1003.0, + 295.0, + 1003.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 998.0, + 1073.0, + 998.0, + 1073.0, + 1033.0, + 321.0, + 1033.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1045.0, + 1404.0, + 1045.0, + 1404.0, + 1086.0, + 296.0, + 1086.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1079.0, + 1404.0, + 1079.0, + 1404.0, + 1115.0, + 321.0, + 1115.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1112.0, + 707.0, + 1112.0, + 707.0, + 1141.0, + 320.0, + 1141.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1968.0, + 1405.0, + 1968.0, + 1405.0, + 2008.0, + 294.0, + 2008.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 2003.0, + 1164.0, + 2003.0, + 1164.0, + 2036.0, + 322.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1159.0, + 1402.0, + 1159.0, + 1402.0, + 1198.0, + 296.0, + 1198.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1192.0, + 1190.0, + 1192.0, + 1190.0, + 1228.0, + 323.0, + 1228.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 437.0, + 1406.0, + 437.0, + 1406.0, + 482.0, + 292.0, + 482.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 475.0, + 1404.0, + 475.0, + 1404.0, + 509.0, + 323.0, + 509.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 506.0, + 610.0, + 506.0, + 610.0, + 533.0, + 322.0, + 533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1805.0, + 1405.0, + 1805.0, + 1405.0, + 1845.0, + 295.0, + 1845.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 320.0, + 1839.0, + 706.0, + 1839.0, + 706.0, + 1872.0, + 320.0, + 1872.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1243.0, + 1405.0, + 1243.0, + 1405.0, + 1279.0, + 295.0, + 1279.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1276.0, + 1177.0, + 1276.0, + 1177.0, + 1309.0, + 321.0, + 1309.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1889.0, + 1403.0, + 1889.0, + 1403.0, + 1925.0, + 295.0, + 1925.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1919.0, + 862.0, + 1919.0, + 862.0, + 1955.0, + 323.0, + 1955.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 880.0, + 1402.0, + 880.0, + 1402.0, + 924.0, + 295.0, + 924.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 916.0, + 966.0, + 916.0, + 966.0, + 948.0, + 322.0, + 948.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1579.0, + 1406.0, + 1579.0, + 1406.0, + 1618.0, + 293.0, + 1618.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1612.0, + 1402.0, + 1612.0, + 1402.0, + 1648.0, + 323.0, + 1648.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 1644.0, + 778.0, + 1644.0, + 778.0, + 1676.0, + 323.0, + 1676.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1468.0, + 1405.0, + 1468.0, + 1405.0, + 1506.0, + 296.0, + 1506.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1501.0, + 1405.0, + 1501.0, + 1405.0, + 1533.0, + 322.0, + 1533.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1532.0, + 1169.0, + 1532.0, + 1169.0, + 1564.0, + 322.0, + 1564.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1323.0, + 1406.0, + 1323.0, + 1406.0, + 1363.0, + 292.0, + 1363.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1356.0, + 1404.0, + 1356.0, + 1404.0, + 1393.0, + 322.0, + 1393.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1388.0, + 1406.0, + 1388.0, + 1406.0, + 1425.0, + 322.0, + 1425.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 1418.0, + 951.0, + 1418.0, + 951.0, + 1455.0, + 322.0, + 1455.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1689.0, + 1407.0, + 1689.0, + 1407.0, + 1735.0, + 293.0, + 1735.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 1727.0, + 1404.0, + 1727.0, + 1404.0, + 1757.0, + 321.0, + 1757.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 326.0, + 1756.0, + 1390.0, + 1756.0, + 1390.0, + 1791.0, + 326.0, + 1791.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 8, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 228, + 1400, + 228, + 1400, + 322, + 298, + 322 + ], + "score": 0.92 + }, + { + "category_id": 1, + "poly": [ + 298, + 343, + 1401, + 343, + 1401, + 436, + 298, + 436 + ], + "score": 0.906 + }, + { + "category_id": 1, + "poly": [ + 292, + 456, + 1402, + 456, + 1402, + 518, + 292, + 518 + ], + "score": 0.876 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.831 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.52 + }, + { + "category_id": 2, + "poly": [ + 298, + 74, + 858, + 74, + 858, + 105, + 298, + 105 + ], + "score": 0.214 + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2084.0, + 871.0, + 2084.0, + 871.0, + 2123.0, + 832.0, + 2123.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 72.0, + 857.0, + 72.0, + 857.0, + 108.0, + 297.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 226.0, + 1402.0, + 226.0, + 1402.0, + 265.0, + 294.0, + 265.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 260.0, + 1405.0, + 260.0, + 1405.0, + 298.0, + 321.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 322.0, + 290.0, + 397.0, + 290.0, + 397.0, + 324.0, + 322.0, + 324.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 291.0, + 339.0, + 1407.0, + 339.0, + 1407.0, + 384.0, + 291.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 321.0, + 373.0, + 1406.0, + 373.0, + 1406.0, + 412.0, + 321.0, + 412.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 324.0, + 405.0, + 1288.0, + 405.0, + 1288.0, + 439.0, + 324.0, + 439.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 450.0, + 1405.0, + 450.0, + 1405.0, + 496.0, + 293.0, + 496.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 323.0, + 487.0, + 590.0, + 487.0, + 590.0, + 520.0, + 323.0, + 520.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 9, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 1536, + 1405, + 1536, + 1405, + 1690, + 298, + 1690 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 297, + 1343, + 1403, + 1343, + 1403, + 1499, + 297, + 1499 + ], + "score": 0.976 + }, + { + "category_id": 1, + "poly": [ + 298, + 1790, + 1404, + 1790, + 1404, + 1945, + 298, + 1945 + ], + "score": 0.971 + }, + { + "category_id": 5, + "poly": [ + 297, + 486, + 1375, + 486, + 1375, + 1220, + 297, + 1220 + ], + "score": 0.967, + "html": "
Input: Empty replay buffer D, init parameter noise std δparam, action noise threshold δ Initialize policy parameters 0,value parameters Φ ←0Q for each episode do Perturb policy parameters θ ← θ +(O,Oparam) and target policy parameters θtarget Observe state St, output a disturbed action embedding A =p(st；0） Compute executing action at by taking max over k
Execute parameterized action at in the environment server Observe next state St+1,reward rt, done signal dt denoting if St+1 is terminal t,rt, St+1,dt) to the buffer D If St+1 is terminal, reset to an initial state so for each update if update-condition do
Sample a minibatch from prioritized replay buffer D with specific priorities Compute and transform target actions with disturbed target policy network
Update parameter noise std Oparam using empirical distance d(μ, μ)
Compute targets for TD-error with min Q-value in the twin Q-networks Update Q-network parameters using gradient descent
if policy update frequency then
Update policy network parameters θ with Update target parameters with polyak averaging end if
" + }, + { + "category_id": 2, + "poly": [ + 333, + 1973, + 670, + 1973, + 670, + 2034, + 333, + 2034 + ], + "score": 0.946 + }, + { + "category_id": 1, + "poly": [ + 302, + 357, + 1401, + 357, + 1401, + 418, + 302, + 418 + ], + "score": 0.939 + }, + { + "category_id": 0, + "poly": [ + 302, + 1286, + 789, + 1286, + 789, + 1318, + 302, + 1318 + ], + "score": 0.906 + }, + { + "category_id": 2, + "poly": [ + 298, + 75, + 858, + 75, + 858, + 105, + 298, + 105 + ], + "score": 0.905 + }, + { + "category_id": 0, + "poly": [ + 300, + 227, + 509, + 227, + 509, + 261, + 300, + 261 + ], + "score": 0.891 + }, + { + "category_id": 0, + "poly": [ + 301, + 1730, + 813, + 1730, + 813, + 1763, + 301, + 1763 + ], + "score": 0.886 + }, + { + "category_id": 0, + "poly": [ + 300, + 297, + 538, + 297, + 538, + 330, + 300, + 330 + ], + "score": 0.853 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 863, + 2088, + 863, + 2113, + 836, + 2113 + ], + "score": 0.851 + }, + { + "category_id": 0, + "poly": [ + 298, + 452, + 444, + 452, + 444, + 483, + 298, + 483 + ], + "score": 0.804 + }, + { + "category_id": 13, + "poly": [ + 971, + 653, + 1169, + 653, + 1169, + 692, + 971, + 692 + ], + "score": 0.92, + "latex": "k _ { t } ^ { s o f t } = \\tilde { \\mu } _ { \\theta } ( s _ { t } ; \\tilde { \\theta } )" + }, + { + "category_id": 13, + "poly": [ + 659, + 616, + 906, + 616, + 906, + 654, + 659, + 654 + ], + "score": 0.92, + "latex": "\\tilde { \\theta } \\gets \\theta + \\mathcal { N } ( 0 , \\sigma _ { \\mathrm { p a r a m } } )" + }, + { + "category_id": 13, + "poly": [ + 1066, + 939, + 1150, + 939, + 1150, + 972, + 1066, + 972 + ], + "score": 0.92, + "latex": "d ( \\mu , \\tilde { \\mu } )" + }, + { + "category_id": 13, + "poly": [ + 506, + 1851, + 676, + 1851, + 676, + 1885, + 506, + 1885 + ], + "score": 0.91, + "latex": "r _ { t } = r _ { t } ^ { \\prime } - r _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 942, + 782, + 1192, + 782, + 1192, + 819, + 942, + 819 + ], + "score": 0.91, + "latex": "( s _ { t } , a _ { t } ^ { s o f t } , r _ { t } , s _ { t + 1 } , d _ { t } )" + }, + { + "category_id": 13, + "poly": [ + 392, + 819, + 445, + 819, + 445, + 849, + 392, + 849 + ], + "score": 0.9, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 577, + 754, + 630, + 754, + 630, + 784, + 577, + 784 + ], + "score": 0.9, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 298, + 1912, + 472, + 1912, + 472, + 1946, + 298, + 1946 + ], + "score": 0.89, + "latex": "a _ { t } = a _ { t } ^ { \\prime } + a _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 717, + 942, + 786, + 942, + 786, + 972, + 717, + 972 + ], + "score": 0.89, + "latex": "\\sigma _ { \\mathrm { p a r a m } }" + }, + { + "category_id": 13, + "poly": [ + 894, + 751, + 922, + 751, + 922, + 781, + 894, + 781 + ], + "score": 0.87, + "latex": "d _ { t }" + }, + { + "category_id": 13, + "poly": [ + 1053, + 754, + 1106, + 754, + 1106, + 782, + 1053, + 782 + ], + "score": 0.87, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 13, + "poly": [ + 837, + 1852, + 861, + 1852, + 861, + 1879, + 837, + 1879 + ], + "score": 0.86, + "latex": "s ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 700, + 725, + 728, + 725, + 728, + 750, + 700, + 750 + ], + "score": 0.86, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 813, + 1882, + 837, + 1882, + 837, + 1909, + 813, + 1909 + ], + "score": 0.85, + "latex": "a ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 912, + 1852, + 936, + 1852, + 936, + 1879, + 912, + 1879 + ], + "score": 0.85, + "latex": "r ^ { \\prime }" + }, + { + "category_id": 13, + "poly": [ + 750, + 1000, + 770, + 1000, + 770, + 1030, + 750, + 1030 + ], + "score": 0.85, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 663, + 694, + 691, + 694, + 691, + 720, + 663, + 720 + ], + "score": 0.84, + "latex": "a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 522, + 664, + 548, + 664, + 548, + 689, + 522, + 689 + ], + "score": 0.84, + "latex": "s _ { t }" + }, + { + "category_id": 13, + "poly": [ + 724, + 754, + 749, + 754, + 749, + 781, + 724, + 781 + ], + "score": 0.84, + "latex": "r _ { t }" + }, + { + "category_id": 13, + "poly": [ + 323, + 1851, + 494, + 1851, + 494, + 1885, + 323, + 1885 + ], + "score": 0.84, + "latex": "s _ { t } = s _ { t } ^ { \\prime } - s _ { t - 1 }" + }, + { + "category_id": 13, + "poly": [ + 639, + 491, + 664, + 491, + 664, + 516, + 639, + 516 + ], + "score": 0.82, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 831, + 823, + 859, + 823, + 859, + 847, + 831, + 847 + ], + "score": 0.81, + "latex": "s _ { 0 }" + }, + { + "category_id": 13, + "poly": [ + 967, + 879, + 993, + 879, + 993, + 905, + 967, + 905 + ], + "score": 0.79, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 1338, + 787, + 1364, + 787, + 1364, + 813, + 1338, + 813 + ], + "score": 0.79, + "latex": "\\mathcal { D }" + }, + { + "category_id": 13, + "poly": [ + 816, + 1343, + 880, + 1343, + 880, + 1375, + 816, + 1375 + ], + "score": 0.79, + "latex": "\\mathrm { N N I } ^ { 2 }" + }, + { + "category_id": 13, + "poly": [ + 856, + 520, + 877, + 520, + 877, + 551, + 856, + 551 + ], + "score": 0.77, + "latex": "\\phi" + }, + { + "category_id": 13, + "poly": [ + 846, + 971, + 869, + 971, + 869, + 1000, + 846, + 1000 + ], + "score": 0.77, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 634, + 520, + 651, + 520, + 651, + 546, + 634, + 546 + ], + "score": 0.76, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 1066, + 972, + 1089, + 972, + 1089, + 1000, + 1066, + 1000 + ], + "score": 0.75, + "latex": "Q" + }, + { + "category_id": 13, + "poly": [ + 1077, + 548, + 1213, + 548, + 1213, + 591, + 1077, + 591 + ], + "score": 0.75, + "latex": "\\theta _ { Q } ^ { \\mathrm { t a r g e t } } \\theta _ { Q }" + }, + { + "category_id": 13, + "poly": [ + 838, + 1061, + 855, + 1061, + 855, + 1087, + 838, + 1087 + ], + "score": 0.74, + "latex": "\\theta" + }, + { + "category_id": 13, + "poly": [ + 915, + 691, + 934, + 691, + 934, + 717, + 915, + 717 + ], + "score": 0.73, + "latex": "k" + }, + { + "category_id": 13, + "poly": [ + 941, + 495, + 1026, + 495, + 1026, + 522, + 941, + 522 + ], + "score": 0.73, + "latex": "\\sigma _ { p a r a m }" + }, + { + "category_id": 13, + "poly": [ + 748, + 1411, + 770, + 1411, + 770, + 1433, + 748, + 1433 + ], + "score": 0.72, + "latex": "=" + }, + { + "category_id": 13, + "poly": [ + 1289, + 491, + 1304, + 491, + 1304, + 515, + 1289, + 515 + ], + "score": 0.64, + "latex": "\\delta" + }, + { + "category_id": 13, + "poly": [ + 933, + 550, + 1064, + 550, + 1064, + 588, + 933, + 588 + ], + "score": 0.58, + "latex": "\\theta _ { \\mu } ^ { \\mathrm { t a r g e t } } \\theta _ { \\mu }" + }, + { + "category_id": 13, + "poly": [ + 1231, + 616, + 1247, + 616, + 1247, + 647, + 1231, + 647 + ], + "score": 0.58, + "latex": "\\tilde { \\theta }" + }, + { + "category_id": 15, + "poly": [ + 331.0, + 1969.0, + 670.0, + 1969.0, + 670.0, + 2007.0, + 331.0, + 2007.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 333.0, + 2003.0, + 668.0, + 2003.0, + 668.0, + 2036.0, + 333.0, + 2036.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1284.0, + 794.0, + 1284.0, + 794.0, + 1322.0, + 296.0, + 1322.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 73.0, + 858.0, + 73.0, + 858.0, + 108.0, + 298.0, + 108.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 222.0, + 513.0, + 222.0, + 513.0, + 268.0, + 295.0, + 268.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1729.0, + 817.0, + 1729.0, + 817.0, + 1766.0, + 295.0, + 1766.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 297.0, + 539.0, + 297.0, + 539.0, + 333.0, + 296.0, + 333.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 868.0, + 2085.0, + 868.0, + 2125.0, + 832.0, + 2125.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 448.0, + 448.0, + 448.0, + 448.0, + 491.0, + 295.0, + 491.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1534.0, + 1405.0, + 1534.0, + 1405.0, + 1572.0, + 293.0, + 1572.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1564.0, + 1405.0, + 1564.0, + 1405.0, + 1604.0, + 293.0, + 1604.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1596.0, + 1405.0, + 1596.0, + 1405.0, + 1633.0, + 293.0, + 1633.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1629.0, + 1405.0, + 1629.0, + 1405.0, + 1662.0, + 296.0, + 1662.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1658.0, + 1116.0, + 1658.0, + 1116.0, + 1695.0, + 295.0, + 1695.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1344.0, + 815.0, + 1344.0, + 815.0, + 1380.0, + 293.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 881.0, + 1344.0, + 1405.0, + 1344.0, + 1405.0, + 1380.0, + 881.0, + 1380.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1377.0, + 1404.0, + 1377.0, + 1404.0, + 1410.0, + 295.0, + 1410.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1405.0, + 747.0, + 1405.0, + 747.0, + 1442.0, + 295.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 771.0, + 1405.0, + 1405.0, + 1405.0, + 1405.0, + 1442.0, + 771.0, + 1442.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1437.0, + 1403.0, + 1437.0, + 1403.0, + 1470.0, + 295.0, + 1470.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1468.0, + 1118.0, + 1468.0, + 1118.0, + 1501.0, + 295.0, + 1501.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 1792.0, + 1402.0, + 1792.0, + 1402.0, + 1821.0, + 298.0, + 1821.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1821.0, + 1405.0, + 1821.0, + 1405.0, + 1854.0, + 296.0, + 1854.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1849.0, + 322.0, + 1849.0, + 322.0, + 1890.0, + 292.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 495.0, + 1849.0, + 505.0, + 1849.0, + 505.0, + 1890.0, + 495.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 677.0, + 1849.0, + 836.0, + 1849.0, + 836.0, + 1890.0, + 677.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 862.0, + 1849.0, + 911.0, + 1849.0, + 911.0, + 1890.0, + 862.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 937.0, + 1849.0, + 1408.0, + 1849.0, + 1408.0, + 1890.0, + 937.0, + 1890.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1881.0, + 812.0, + 1881.0, + 812.0, + 1918.0, + 293.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 838.0, + 1881.0, + 1405.0, + 1881.0, + 1405.0, + 1918.0, + 838.0, + 1918.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1908.0, + 297.0, + 1908.0, + 297.0, + 1953.0, + 293.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 473.0, + 1908.0, + 486.0, + 1908.0, + 486.0, + 1953.0, + 473.0, + 1953.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 298.0, + 356.0, + 1404.0, + 356.0, + 1404.0, + 391.0, + 298.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 297.0, + 386.0, + 459.0, + 386.0, + 459.0, + 419.0, + 297.0, + 419.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 10, + "width": 1700, + "height": 2200 + } + }, + { + "layout_dets": [ + { + "category_id": 1, + "poly": [ + 298, + 994, + 1404, + 994, + 1404, + 1209, + 298, + 1209 + ], + "score": 0.978 + }, + { + "category_id": 1, + "poly": [ + 298, + 806, + 1403, + 806, + 1403, + 960, + 298, + 960 + ], + "score": 0.975 + }, + { + "category_id": 3, + "poly": [ + 374, + 236, + 1326, + 236, + 1326, + 656, + 374, + 656 + ], + "score": 0.969 + }, + { + "category_id": 4, + "poly": [ + 298, + 693, + 1401, + 693, + 1401, + 750, + 298, + 750 + ], + "score": 0.937 + }, + { + "category_id": 2, + "poly": [ + 298, + 76, + 858, + 76, + 858, + 104, + 298, + 104 + ], + "score": 0.893 + }, + { + "category_id": 0, + "poly": [ + 302, + 1245, + 807, + 1245, + 807, + 1277, + 302, + 1277 + ], + "score": 0.882 + }, + { + "category_id": 2, + "poly": [ + 836, + 2088, + 865, + 2088, + 865, + 2112, + 836, + 2112 + ], + "score": 0.84 + }, + { + "category_id": 1, + "poly": [ + 297, + 1303, + 1227, + 1303, + 1227, + 1336, + 297, + 1336 + ], + "score": 0.746 + }, + { + "category_id": 4, + "poly": [ + 299, + 693, + 1399, + 693, + 1399, + 750, + 299, + 750 + ], + "score": 0.151 + }, + { + "category_id": 13, + "poly": [ + 1175, + 1152, + 1277, + 1152, + 1277, + 1178, + 1175, + 1178 + ], + "score": 0.9, + "latex": "a _ { 0 } , . . . , a _ { t }" + }, + { + "category_id": 13, + "poly": [ + 297, + 1182, + 351, + 1182, + 351, + 1211, + 297, + 1211 + ], + "score": 0.88, + "latex": "s _ { t + 1 }" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 241.0, + 391.0, + 241.0, + 391.0, + 253.0, + 374.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 415.0, + 242.0, + 439.0, + 242.0, + 439.0, + 250.0, + 415.0, + 250.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 240.0, + 875.0, + 240.0, + 875.0, + 253.0, + 858.0, + 253.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 376.0, + 286.0, + 390.0, + 286.0, + 390.0, + 298.0, + 376.0, + 298.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 964.0, + 317.0, + 974.0, + 317.0, + 974.0, + 326.0, + 964.0, + 326.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1304.0, + 317.0, + 1312.0, + 317.0, + 1312.0, + 325.0, + 1304.0, + 325.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 331.0, + 388.0, + 331.0, + 388.0, + 341.0, + 377.0, + 341.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 857.0, + 359.0, + 873.0, + 359.0, + 873.0, + 372.0, + 857.0, + 372.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 377.0, + 376.0, + 388.0, + 376.0, + 388.0, + 384.0, + 377.0, + 384.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 909.0, + 383.0, + 918.0, + 383.0, + 918.0, + 391.0, + 909.0, + 391.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 421.0, + 388.0, + 421.0, + 388.0, + 430.0, + 378.0, + 430.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 858.0, + 420.0, + 872.0, + 420.0, + 872.0, + 432.0, + 858.0, + 432.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 823.0, + 460.0, + 849.0, + 460.0, + 849.0, + 480.0, + 823.0, + 480.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 855.0, + 477.0, + 874.0, + 477.0, + 874.0, + 494.0, + 855.0, + 494.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 882.0, + 475.0, + 899.0, + 475.0, + 899.0, + 488.0, + 882.0, + 488.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1229.0, + 475.0, + 1238.0, + 475.0, + 1238.0, + 485.0, + 1229.0, + 485.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 789.0, + 503.0, + 800.0, + 503.0, + 800.0, + 513.0, + 789.0, + 513.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 374.0, + 509.0, + 389.0, + 509.0, + 389.0, + 521.0, + 374.0, + 521.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1112.0, + 526.0, + 1122.0, + 526.0, + 1122.0, + 537.0, + 1112.0, + 537.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 734.0, + 533.0, + 743.0, + 533.0, + 743.0, + 541.0, + 734.0, + 541.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 373.0, + 553.0, + 389.0, + 553.0, + 389.0, + 565.0, + 373.0, + 565.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 378.0, + 600.0, + 386.0, + 600.0, + 386.0, + 607.0, + 378.0, + 607.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 496.0, + 603.0, + 828.0, + 603.0, + 828.0, + 617.0, + 496.0, + 617.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 864.0, + 601.0, + 1316.0, + 601.0, + 1316.0, + 620.0, + 864.0, + 620.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1317.0, + 600.0, + 1328.0, + 600.0, + 1328.0, + 610.0, + 1317.0, + 610.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 532.0, + 622.0, + 672.0, + 622.0, + 672.0, + 657.0, + 532.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 998.0, + 622.0, + 1179.0, + 622.0, + 1179.0, + 657.0, + 998.0, + 657.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 850.0, + 531.0, + 893.0, + 531.0, + 893.0, + 554.0, + 850.0, + 554.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 690.0, + 1403.0, + 690.0, + 1403.0, + 728.0, + 293.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 719.0, + 798.0, + 719.0, + 798.0, + 752.0, + 293.0, + 752.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 70.0, + 859.0, + 70.0, + 859.0, + 109.0, + 296.0, + 109.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 296.0, + 1243.0, + 811.0, + 1243.0, + 811.0, + 1281.0, + 296.0, + 1281.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 832.0, + 2085.0, + 870.0, + 2085.0, + 870.0, + 2124.0, + 832.0, + 2124.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 690.0, + 1403.0, + 690.0, + 1403.0, + 728.0, + 294.0, + 728.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 720.0, + 796.0, + 720.0, + 796.0, + 751.0, + 294.0, + 751.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 992.0, + 1405.0, + 992.0, + 1405.0, + 1031.0, + 295.0, + 1031.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 295.0, + 1024.0, + 1403.0, + 1024.0, + 1403.0, + 1061.0, + 295.0, + 1061.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 293.0, + 1056.0, + 1404.0, + 1056.0, + 1404.0, + 1090.0, + 293.0, + 1090.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1084.0, + 1404.0, + 1084.0, + 1404.0, + 1121.0, + 292.0, + 1121.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1115.0, + 1404.0, + 1115.0, + 1404.0, + 1153.0, + 292.0, + 1153.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1146.0, + 1174.0, + 1146.0, + 1174.0, + 1183.0, + 292.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 1278.0, + 1146.0, + 1404.0, + 1146.0, + 1404.0, + 1183.0, + 1278.0, + 1183.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 1179.0, + 296.0, + 1179.0, + 296.0, + 1213.0, + 292.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 352.0, + 1179.0, + 709.0, + 1179.0, + 709.0, + 1213.0, + 352.0, + 1213.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 806.0, + 1407.0, + 806.0, + 1407.0, + 843.0, + 294.0, + 843.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 837.0, + 1407.0, + 837.0, + 1407.0, + 874.0, + 294.0, + 874.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 292.0, + 866.0, + 1405.0, + 866.0, + 1405.0, + 904.0, + 292.0, + 904.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 898.0, + 1405.0, + 898.0, + 1405.0, + 935.0, + 294.0, + 935.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 924.0, + 694.0, + 924.0, + 694.0, + 968.0, + 294.0, + 968.0 + ], + "score": 1.0, + "text": "" + }, + { + "category_id": 15, + "poly": [ + 294.0, + 1298.0, + 1232.0, + 1298.0, + 1232.0, + 1342.0, + 294.0, + 1342.0 + ], + "score": 1.0, + "text": "" + } + ], + "page_info": { + "page_no": 11, + "width": 1700, + "height": 2200 + } + } +] \ No newline at end of file diff --git a/vlm/train/0WASBV4xkhy/0.png b/vlm/train/0WASBV4xkhy/0.png new file mode 100644 index 0000000000000000000000000000000000000000..71baf121b2c889e298812ee5e364fe0ed3997874 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3fba467d766f3a5138d5f39648ec3372efc9c81d2bbd0ec2123046d03bdcdca6 +size 419106 diff --git a/vlm/train/0WASBV4xkhy/1.png b/vlm/train/0WASBV4xkhy/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a13f51f8538e4694adfacfa77b699faa21be65f7 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6f123d56e016ee0c6affde2265ad4d2e8857e93dbca9790739d5ca45bcb1352 +size 561601 diff --git a/vlm/train/0WASBV4xkhy/10.png b/vlm/train/0WASBV4xkhy/10.png new file mode 100644 index 0000000000000000000000000000000000000000..94f995d41958a52cadfd012372ec86e694cf8bf5 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:078ca82e7512d3ca7513e47cd1bb0667b082ae2a686ed79c20d96989b918bf75 +size 539647 diff --git a/vlm/train/0WASBV4xkhy/11.png b/vlm/train/0WASBV4xkhy/11.png new file mode 100644 index 0000000000000000000000000000000000000000..220ef61e6809d68807fc151f55e303a2ecdfa70a --- /dev/null +++ b/vlm/train/0WASBV4xkhy/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:452e2643bc2350c5413f15d78fcb20b172e7f9bb7cf734e91b54cb8211715a81 +size 554819 diff --git a/vlm/train/0WASBV4xkhy/12.png b/vlm/train/0WASBV4xkhy/12.png new file mode 100644 index 0000000000000000000000000000000000000000..1c1663f1d5337f77cf24cdfe9c0e1da24614dfad --- /dev/null +++ b/vlm/train/0WASBV4xkhy/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cff77dc5c99eb26469d6728d8b17582d19eb98d75bf40fe4578b1f442678cdd3 +size 458621 diff --git a/vlm/train/0WASBV4xkhy/2.png b/vlm/train/0WASBV4xkhy/2.png new file mode 100644 index 0000000000000000000000000000000000000000..b502ee28f5eb24726bcbcd88046eae078be56bea --- /dev/null +++ b/vlm/train/0WASBV4xkhy/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:708d8c2609eeb74acd77de0264dadc157d621abb656776ba9d81290e60f34dc5 +size 490846 diff --git a/vlm/train/0WASBV4xkhy/3.png b/vlm/train/0WASBV4xkhy/3.png new file mode 100644 index 0000000000000000000000000000000000000000..a95070fcfd8131ef6bedafa479b85ba50d8610c5 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69039e0d8915403f5866607f7cc66580f7bb68617bd140fdee5f65667999fc3a +size 606234 diff --git a/vlm/train/0WASBV4xkhy/4.png b/vlm/train/0WASBV4xkhy/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f7d5eb764760cc1a9423e5d872c2c1ac270431aa --- /dev/null +++ b/vlm/train/0WASBV4xkhy/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8bcbf0bc188dcc499ee5e4495c04e1dcfb420c8b06c35c3cee7daafdde5a8de +size 519021 diff --git a/vlm/train/0WASBV4xkhy/5.png b/vlm/train/0WASBV4xkhy/5.png new file mode 100644 index 0000000000000000000000000000000000000000..1674123fd2cf8fc8ae38702573af105051447e94 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c02bd519303d7986b0eff579d58c6a43684fd45a43b9d09d677ec0a8cfcb3aeb +size 521788 diff --git a/vlm/train/0WASBV4xkhy/6.png b/vlm/train/0WASBV4xkhy/6.png new file mode 100644 index 0000000000000000000000000000000000000000..4d2bb18ef0797cbef686b5a8400d765bfa480e48 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0aa38b1b8907a84e21057a9c8b7c5ba54ee3f48abb89a059839591124651cfb +size 522475 diff --git a/vlm/train/0WASBV4xkhy/7.png b/vlm/train/0WASBV4xkhy/7.png new file mode 100644 index 0000000000000000000000000000000000000000..b1ecb7f997db4a32ce01048e496cd1f036338538 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aad56f333ae823f3218157e44ba3e0e23b05706d685cf84c8af4dcf49984169c +size 504560 diff --git a/vlm/train/0WASBV4xkhy/8.png b/vlm/train/0WASBV4xkhy/8.png new file mode 100644 index 0000000000000000000000000000000000000000..7ca140658ea5367d6d95e7254e95e8e137e8c7e0 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d332bcf5a41d899bb926272201ae5f1c07bc40ca33f6313d99318ffadbaa0fda +size 610299 diff --git a/vlm/train/0WASBV4xkhy/9.png b/vlm/train/0WASBV4xkhy/9.png new file mode 100644 index 0000000000000000000000000000000000000000..9ae68236ccc70d00d536b4fa44c8ee67672520c3 --- /dev/null +++ b/vlm/train/0WASBV4xkhy/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2c2bfa1b0877d7c9b4705701616979e92712eef1c360afc7af958b762e1122e +size 494381 diff --git a/vlm/train/27acGyyI1BY/0.png b/vlm/train/27acGyyI1BY/0.png new file mode 100644 index 0000000000000000000000000000000000000000..53669b223e53508e4a53d408fc36e9d22ba905ca --- /dev/null +++ b/vlm/train/27acGyyI1BY/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca090cfbec2b9d3fe21463b18cf4b332940daa7f20128c968f55751f3e0f60de +size 516021 diff --git a/vlm/train/27acGyyI1BY/1.png b/vlm/train/27acGyyI1BY/1.png new file mode 100644 index 0000000000000000000000000000000000000000..5b314af467a930af3efb2d5bb0435f331eae4206 --- /dev/null +++ b/vlm/train/27acGyyI1BY/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8895732c6039ab5f5e7729f552a770796863c4ac10ad054687e1b7d47652beae +size 597025 diff --git a/vlm/train/27acGyyI1BY/10.png b/vlm/train/27acGyyI1BY/10.png new file mode 100644 index 0000000000000000000000000000000000000000..7dc7692e7aa985d4a15e13b9c75b03ec797e4e99 --- /dev/null +++ b/vlm/train/27acGyyI1BY/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61db4f895c47390e65b0e01118fc6f40e4620fa38d1ec27428a3157d8cb4659c +size 246535 diff --git a/vlm/train/27acGyyI1BY/11.png b/vlm/train/27acGyyI1BY/11.png new file mode 100644 index 0000000000000000000000000000000000000000..49139621bdd02fb2b47452bd11dc00d62efade5c --- /dev/null +++ b/vlm/train/27acGyyI1BY/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03812857d015359fdbe1961600276531aa6bae481dfdcd41321f641ac5f9e79d +size 407417 diff --git a/vlm/train/27acGyyI1BY/12.png b/vlm/train/27acGyyI1BY/12.png new file mode 100644 index 0000000000000000000000000000000000000000..e8c2bc771b53ab97358ef6f409c4a074c8383ca2 --- /dev/null +++ b/vlm/train/27acGyyI1BY/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f349f848ee467d2a958e4ec1f7c54e8fb5a814ef7830282189f27d3d689d27b4 +size 418953 diff --git a/vlm/train/27acGyyI1BY/13.png b/vlm/train/27acGyyI1BY/13.png new file mode 100644 index 0000000000000000000000000000000000000000..5bd14413762f4ab8ee220c2f46d12d5b389b7ebf --- /dev/null +++ b/vlm/train/27acGyyI1BY/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ee8ac0f6abf7156b0f846e6c2c9330ec01193c9a50f9b0882c48e855ef46817 +size 485708 diff --git a/vlm/train/27acGyyI1BY/14.png b/vlm/train/27acGyyI1BY/14.png new file mode 100644 index 0000000000000000000000000000000000000000..6730af6fc912c075649a879b7d8abcaa9b617601 --- /dev/null +++ b/vlm/train/27acGyyI1BY/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78df9613285b0bde44dfacf3abcffcfe36dfcd4264b56fb1f3db8bfd106b1518 +size 462748 diff --git a/vlm/train/27acGyyI1BY/15.png b/vlm/train/27acGyyI1BY/15.png new file mode 100644 index 0000000000000000000000000000000000000000..013e098a2e6882f2f7d56f5a2d77f8d367de97dd --- /dev/null +++ b/vlm/train/27acGyyI1BY/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2686986f99faf2ebb6e7bf2052d91a157ba962cd2ab7c319a399a360e7c84c8e +size 497364 diff --git a/vlm/train/27acGyyI1BY/16.png b/vlm/train/27acGyyI1BY/16.png new file mode 100644 index 0000000000000000000000000000000000000000..65f3645daa8e367754b10c39c686b2795705d2cd --- /dev/null +++ b/vlm/train/27acGyyI1BY/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac9a6a0d3ca34b8f3eb597906ab32e9cdd974c546623b9b041b4b9e6245aa49d +size 643020 diff --git a/vlm/train/27acGyyI1BY/17.png b/vlm/train/27acGyyI1BY/17.png new file mode 100644 index 0000000000000000000000000000000000000000..93ac9d889612bdb1042d013a01ba89d96b379d9f --- /dev/null +++ b/vlm/train/27acGyyI1BY/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0766d88a8c9350776abc60b29165a4279668ab6efacbef593b506211c708e09d +size 441019 diff --git a/vlm/train/27acGyyI1BY/18.png b/vlm/train/27acGyyI1BY/18.png new file mode 100644 index 0000000000000000000000000000000000000000..f18a2e14693f86b2e5935654fd48d4bbb19e0bf7 --- /dev/null +++ b/vlm/train/27acGyyI1BY/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e310206b2ab63d6d116538a521d04309d1a01226769bec18f5e84b276e4fd13 +size 451119 diff --git a/vlm/train/27acGyyI1BY/2.png b/vlm/train/27acGyyI1BY/2.png new file mode 100644 index 0000000000000000000000000000000000000000..f8c593f1bf52aca04d8d2cf64e77c917b4d9f922 --- /dev/null +++ b/vlm/train/27acGyyI1BY/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c371f92e24cd16e4b2923e153da5e024c73a131e2ae6ba1c473d3c1e30ea666 +size 572856 diff --git a/vlm/train/27acGyyI1BY/3.png b/vlm/train/27acGyyI1BY/3.png new file mode 100644 index 0000000000000000000000000000000000000000..17a43f4ec6f3cd42b042a11c3f2bf91e4156374e --- /dev/null +++ b/vlm/train/27acGyyI1BY/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffc1fc9843a4cf9394253663f09eacfbf896aa8528f1681939d913a245fe0190 +size 521410 diff --git a/vlm/train/27acGyyI1BY/4.png b/vlm/train/27acGyyI1BY/4.png new file mode 100644 index 0000000000000000000000000000000000000000..39b14e340cd68be3a72a5d45d043ea5f4e6f06db --- /dev/null +++ b/vlm/train/27acGyyI1BY/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:309b79e50ef9797aa46647fac10d22d0c1dbc37483d28f63b84dd1d4b0317bba +size 561426 diff --git a/vlm/train/27acGyyI1BY/5.png b/vlm/train/27acGyyI1BY/5.png new file mode 100644 index 0000000000000000000000000000000000000000..954998add61944243eed8187336b8f4905fee51d --- /dev/null +++ b/vlm/train/27acGyyI1BY/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7422d22ab4612aedce5667036d78273b88c9d1b62dcfd9a783dbcb186fd6f0d +size 540594 diff --git a/vlm/train/27acGyyI1BY/6.png b/vlm/train/27acGyyI1BY/6.png new file mode 100644 index 0000000000000000000000000000000000000000..6a915a1e294ef3fcf99304ac3aa9fc7a07ef0439 --- /dev/null +++ b/vlm/train/27acGyyI1BY/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:322068f1bf0fcab9aa8ccedcb2d631ca3062889e54c68bee53f8b25432112b42 +size 515757 diff --git a/vlm/train/27acGyyI1BY/7.png b/vlm/train/27acGyyI1BY/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e7cac8a1abf661ca2ff80b8d728c6ba5993ce851 --- /dev/null +++ b/vlm/train/27acGyyI1BY/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a09424d33239f91960f5b55d9afd7f70a087a603bc08bb98c6c1141831dfc742 +size 671169 diff --git a/vlm/train/27acGyyI1BY/8.png b/vlm/train/27acGyyI1BY/8.png new file mode 100644 index 0000000000000000000000000000000000000000..4f5ca3fc0f45e6289f74776fa3e4ab1f81d8e7a6 --- /dev/null +++ b/vlm/train/27acGyyI1BY/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:870ff4d220e37c65d62a14324750ef50d61da91d19c6b0e37757a04f6969a167 +size 576681 diff --git a/vlm/train/27acGyyI1BY/9.png b/vlm/train/27acGyyI1BY/9.png new file mode 100644 index 0000000000000000000000000000000000000000..228ee796a75e059fff979d32aebe29d907046503 --- /dev/null +++ b/vlm/train/27acGyyI1BY/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f64a99b1c95800b38abac792bd283f07308ad3a12cb30b3ba1abc7a4af26ce19 +size 488892 diff --git a/vlm/train/6vaActvpcp3/0.png b/vlm/train/6vaActvpcp3/0.png new file mode 100644 index 0000000000000000000000000000000000000000..b2bd1a8827d264d01fdf1047eebf4a5460cef488 --- /dev/null +++ b/vlm/train/6vaActvpcp3/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22fc40ffc52b5351cb97ccba18934072dcee8da9859aff1bbbf2660ee527c9fa +size 447884 diff --git a/vlm/train/6vaActvpcp3/1.png b/vlm/train/6vaActvpcp3/1.png new file mode 100644 index 0000000000000000000000000000000000000000..585a2d29616366566a92f3a270996569d7a2acfe --- /dev/null +++ b/vlm/train/6vaActvpcp3/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a442dee963d01e943a2dacabf395cc653b850167b52129ac4736afe801d30ae8 +size 509736 diff --git a/vlm/train/6vaActvpcp3/10.png b/vlm/train/6vaActvpcp3/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a88d24eb8b81da5ada4a3be3b446691ca76b2b96 --- /dev/null +++ b/vlm/train/6vaActvpcp3/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abe67e87cadce0899bfe442e813080da2ee63331c2383327c4c8424156a4797b +size 601517 diff --git a/vlm/train/6vaActvpcp3/11.png b/vlm/train/6vaActvpcp3/11.png new file mode 100644 index 0000000000000000000000000000000000000000..da24b51fb730c7eb5da99f898e3a93ba16a29594 --- /dev/null +++ b/vlm/train/6vaActvpcp3/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ce1d7333c2830609a4b990d98b9177552e284734ae881fd2e215a76e593da18 +size 625306 diff --git a/vlm/train/6vaActvpcp3/12.png b/vlm/train/6vaActvpcp3/12.png new file mode 100644 index 0000000000000000000000000000000000000000..f84249474a7943a5e60ec8cbd8ae358d3ba0e9ef --- /dev/null +++ b/vlm/train/6vaActvpcp3/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f62f1b3dddcb3bc9901b5a4f9d746fc54fdd69f929422d7e548b2a1398bd02e +size 175256 diff --git a/vlm/train/6vaActvpcp3/2.png b/vlm/train/6vaActvpcp3/2.png new file mode 100644 index 0000000000000000000000000000000000000000..42d6369d8ad49eb31b757053a8373d643d38af9d --- /dev/null +++ b/vlm/train/6vaActvpcp3/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:136f7891c57233ba6834aa02d1c82e7465bb3d1ccb91caac5c55853d0b7a45f3 +size 463673 diff --git a/vlm/train/6vaActvpcp3/3.png b/vlm/train/6vaActvpcp3/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e5f69a844e90334cb9048da45241ba4bd043aae5 --- /dev/null +++ b/vlm/train/6vaActvpcp3/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a792d4c937f7ba2b0b7f69d16fedf4a41b504071b8e95bc9a7f7189ed082fbaa +size 439081 diff --git a/vlm/train/6vaActvpcp3/4.png b/vlm/train/6vaActvpcp3/4.png new file mode 100644 index 0000000000000000000000000000000000000000..d9b4e47c6ef5203f77970f1838a85297d1bac1c9 --- /dev/null +++ b/vlm/train/6vaActvpcp3/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3a59a0c9c1fec171b7973ef3c70c8dc401bb5dbd1bf4244dbbc7ed7f582bc4a +size 535649 diff --git a/vlm/train/6vaActvpcp3/5.png b/vlm/train/6vaActvpcp3/5.png new file mode 100644 index 0000000000000000000000000000000000000000..9759c2d19d051ae4a98b492b4e59d99a2f39b20e --- /dev/null +++ b/vlm/train/6vaActvpcp3/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54764dcb21df1b7d1ed9de9b8f8827434eb75f5d6425a1de57406884da2daeaf +size 431378 diff --git a/vlm/train/6vaActvpcp3/6.png b/vlm/train/6vaActvpcp3/6.png new file mode 100644 index 0000000000000000000000000000000000000000..fef16aa1cf6f09d9fed6ab4a11943f9297f6030d --- /dev/null +++ b/vlm/train/6vaActvpcp3/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a5f78ba5ae72b14d3507c2ea1776aa0a945f7c2d8eb09e84b819ebc9e7321ff +size 547350 diff --git a/vlm/train/6vaActvpcp3/7.png b/vlm/train/6vaActvpcp3/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f562f74d2e959566fd92b48f64d6a0d0c3277b6e --- /dev/null +++ b/vlm/train/6vaActvpcp3/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73c16b37a321f1cf0bafc33c3b3c3412fff628519948c5a9bcc37ff520f5829e +size 488205 diff --git a/vlm/train/6vaActvpcp3/8.png b/vlm/train/6vaActvpcp3/8.png new file mode 100644 index 0000000000000000000000000000000000000000..fa321dad1a1972e6811d67f45f46a5b0eb8451e0 --- /dev/null +++ b/vlm/train/6vaActvpcp3/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02d69bfaa54bd4854db1e9760f29c105b0d805cd7608d79dcecc0d9842ab85a2 +size 446689 diff --git a/vlm/train/6vaActvpcp3/9.png b/vlm/train/6vaActvpcp3/9.png new file mode 100644 index 0000000000000000000000000000000000000000..c63df4a5ebf83c77e66f450a3ca217edd81b20fd --- /dev/null +++ b/vlm/train/6vaActvpcp3/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39dfb159989956430079f9c3f8d1e3aaf922d7cfd0d9e6eb4364ac9ca82ec8e3 +size 476486 diff --git a/vlm/train/B1ckMDqlg/0.png b/vlm/train/B1ckMDqlg/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6f73389a01e3f7569b9177a3db746cf7009c61fc --- /dev/null +++ b/vlm/train/B1ckMDqlg/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1945e1ece5bd0e1000ae7280d650669a657fc74df5f4107ac6ea44f2627c2427 +size 453938 diff --git a/vlm/train/B1ckMDqlg/1.png b/vlm/train/B1ckMDqlg/1.png new file mode 100644 index 0000000000000000000000000000000000000000..8157eaa8c40e65a97b95441356a8f43e201abda4 --- /dev/null +++ b/vlm/train/B1ckMDqlg/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1410d535a87d72262e560bc5e04625502e6eb83e39d66fa1d23fb78ecbdd4a93 +size 433306 diff --git a/vlm/train/B1ckMDqlg/10.png b/vlm/train/B1ckMDqlg/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d39f7a4c84d4df3bbc08ae0fc2e4160256db9013 --- /dev/null +++ b/vlm/train/B1ckMDqlg/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd0a881bc87a59236c482ded4ea37f087ba30e44e0d928a8b0061ab447355ff2 +size 511642 diff --git a/vlm/train/B1ckMDqlg/11.png b/vlm/train/B1ckMDqlg/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5abf43bad16bb980cf2d82abc4f5632573b6f0b3 --- /dev/null +++ b/vlm/train/B1ckMDqlg/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:27550250b032fc50b03cd9c387f91b71ba5d606bc8d95523fa742be588bac4c3 +size 217245 diff --git a/vlm/train/B1ckMDqlg/12.png b/vlm/train/B1ckMDqlg/12.png new file mode 100644 index 0000000000000000000000000000000000000000..d1764504e23f8d9d135a907fa6710794d5c5f0d6 --- /dev/null +++ b/vlm/train/B1ckMDqlg/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffe3bc0f8a0a902abc127fd4676e76a60601d7dc05a74615b44f7ccb7de37d8e +size 404692 diff --git a/vlm/train/B1ckMDqlg/13.png b/vlm/train/B1ckMDqlg/13.png new file mode 100644 index 0000000000000000000000000000000000000000..ae5d2d43fc9bd10aff62580ff40dab62c79eefed --- /dev/null +++ b/vlm/train/B1ckMDqlg/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1797f4647b4ec502d8122a95e04e0b6fcfc308d64a0f083d376d9ad8724bd3de +size 487646 diff --git a/vlm/train/B1ckMDqlg/14.png b/vlm/train/B1ckMDqlg/14.png new file mode 100644 index 0000000000000000000000000000000000000000..b25ace7355d827ebeabfccacbbe9f15a068bc095 --- /dev/null +++ b/vlm/train/B1ckMDqlg/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38b91002be4dbe043dd7a39e09e6c60d880f2fc7b2a590cf847ea8ef1550d218 +size 491298 diff --git a/vlm/train/B1ckMDqlg/15.png b/vlm/train/B1ckMDqlg/15.png new file mode 100644 index 0000000000000000000000000000000000000000..8bd6b11f9bbbc2d8d3b315f55fcbf430a37d2edd --- /dev/null +++ b/vlm/train/B1ckMDqlg/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58a8a7c9c389835b1f757c81575b71cc572b8c4cf2ee65b043efe14d081a164f +size 549803 diff --git a/vlm/train/B1ckMDqlg/16.png b/vlm/train/B1ckMDqlg/16.png new file mode 100644 index 0000000000000000000000000000000000000000..2973e8d709c5622cfbb3927019653bf76d7dd079 --- /dev/null +++ b/vlm/train/B1ckMDqlg/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:766462801e17a74c11eb9d835055a0ffb93a645061960a552109922c84335ef3 +size 597378 diff --git a/vlm/train/B1ckMDqlg/17.png b/vlm/train/B1ckMDqlg/17.png new file mode 100644 index 0000000000000000000000000000000000000000..6d81141b2c0c76346ec048d5cda25e740db8c48d --- /dev/null +++ b/vlm/train/B1ckMDqlg/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ba698ed5a145821120f16f3a0ec97f2eb33f2a6a19d5a13ffbf86438a77b3d8 +size 444199 diff --git a/vlm/train/B1ckMDqlg/18.png b/vlm/train/B1ckMDqlg/18.png new file mode 100644 index 0000000000000000000000000000000000000000..705ea2e436d3ba0d58295baf6f084f6eb30c08c7 --- /dev/null +++ b/vlm/train/B1ckMDqlg/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6da71a9062952553a42f313f08391f82b2682c391635d100e048b073be440c91 +size 362055 diff --git a/vlm/train/B1ckMDqlg/2.png b/vlm/train/B1ckMDqlg/2.png new file mode 100644 index 0000000000000000000000000000000000000000..6848b94639dbca1a308d2fccac036d3d9b294a82 --- /dev/null +++ b/vlm/train/B1ckMDqlg/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c86206d1bccc6e9a3aa317b799aedb0538053409dddc301e2b1c8a35087538c1 +size 547206 diff --git a/vlm/train/B1ckMDqlg/3.png b/vlm/train/B1ckMDqlg/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e478e70897f1ba6f9828022e9e9a1ac319629ab7 --- /dev/null +++ b/vlm/train/B1ckMDqlg/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:970969f45c034954c65cf003ca81a1bc635959c3be44b442a665ed449950b2d9 +size 503261 diff --git a/vlm/train/B1ckMDqlg/4.png b/vlm/train/B1ckMDqlg/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f73f79dcc10c5436aa1a84326eefd19ab48332e0 --- /dev/null +++ b/vlm/train/B1ckMDqlg/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ccfd488d55f6ea29ec9e21551e58d69571808520326e0fd18824de3d875fc17 +size 519176 diff --git a/vlm/train/B1ckMDqlg/5.png b/vlm/train/B1ckMDqlg/5.png new file mode 100644 index 0000000000000000000000000000000000000000..03a79ddb1a44ca4302e1ecea1214348998423992 --- /dev/null +++ b/vlm/train/B1ckMDqlg/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4acff78d2d04f02ad6922e1ddcd633d0d8439414cb586dafc7a0fd16c4404f55 +size 493950 diff --git a/vlm/train/B1ckMDqlg/6.png b/vlm/train/B1ckMDqlg/6.png new file mode 100644 index 0000000000000000000000000000000000000000..9f79c4ff5e36311475389b15f80c3e157d192b96 --- /dev/null +++ b/vlm/train/B1ckMDqlg/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:abca44350978328e299dbc0f5fb8dc26bb388703916864979073b1f9fd326f93 +size 485819 diff --git a/vlm/train/B1ckMDqlg/7.png b/vlm/train/B1ckMDqlg/7.png new file mode 100644 index 0000000000000000000000000000000000000000..10aec768424995131df2e91f5053019e56cd7c75 --- /dev/null +++ b/vlm/train/B1ckMDqlg/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9415ad7ec75df2024929969f98c6f3367bd71d1d4749a7430a10c8a5782cc3b +size 511736 diff --git a/vlm/train/B1ckMDqlg/8.png b/vlm/train/B1ckMDqlg/8.png new file mode 100644 index 0000000000000000000000000000000000000000..54e8a83c245d7b227fe34522548b38554781e6fd --- /dev/null +++ b/vlm/train/B1ckMDqlg/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1583b2fa248ca06d662c3bc380864fe69e706c257a5c3d70fae330d35c35d53b +size 537282 diff --git a/vlm/train/B1ckMDqlg/9.png b/vlm/train/B1ckMDqlg/9.png new file mode 100644 index 0000000000000000000000000000000000000000..22332ef9faf0db86eb23fc615751fa672363b265 --- /dev/null +++ b/vlm/train/B1ckMDqlg/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26dc863bc9c8ffa4d0f4d0f5eb641353c8ada3153b0a8caf9efebecb665e9a4d +size 545564 diff --git a/vlm/train/B1l6qiR5F7/0.png b/vlm/train/B1l6qiR5F7/0.png new file mode 100644 index 0000000000000000000000000000000000000000..90401ae2c3db1e4f8ffc46d0216852c6805767a4 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:339ade146f9c576a51f9cf532c6d28969f2fbb4a4cf699a0fac613877c6ef69f +size 429769 diff --git a/vlm/train/B1l6qiR5F7/1.png b/vlm/train/B1l6qiR5F7/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a76d5fd38b4b8a6f2f9c9b4b52ef509a4d23ceb7 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c8be7b86c7aac389372d36361aa84eaea372f29af041979710c0e06f316faec +size 581427 diff --git a/vlm/train/B1l6qiR5F7/10.png b/vlm/train/B1l6qiR5F7/10.png new file mode 100644 index 0000000000000000000000000000000000000000..ad86131e85d5be3616db037d7c464618d683aec7 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14a2f85a60b9d04a9a48aba2d1864391c1db718396a0b0bbc3eafb4208a443c3 +size 559916 diff --git a/vlm/train/B1l6qiR5F7/11.png b/vlm/train/B1l6qiR5F7/11.png new file mode 100644 index 0000000000000000000000000000000000000000..2b22151801797adc9088f1e3fc627ba43ba004c7 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9944be3df7060b04b51e20a9a5b9416a8569541b6df0c180f9804145c8de126b +size 504065 diff --git a/vlm/train/B1l6qiR5F7/12.png b/vlm/train/B1l6qiR5F7/12.png new file mode 100644 index 0000000000000000000000000000000000000000..ae8831f469e1acecfd2a98d7a4cf2b8bd0d87a72 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9e5a5404bdbf4947021d386a715bad97ed4a60093ecb51670d0ee6859dd13d9 +size 399718 diff --git a/vlm/train/B1l6qiR5F7/13.png b/vlm/train/B1l6qiR5F7/13.png new file mode 100644 index 0000000000000000000000000000000000000000..ff7961e94b40b508689134b48a3425bf04214ac5 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8dabcf78864ec684f114843722e315e700aa37c150395449958a4257f3dd6570 +size 201030 diff --git a/vlm/train/B1l6qiR5F7/2.png b/vlm/train/B1l6qiR5F7/2.png new file mode 100644 index 0000000000000000000000000000000000000000..07b1bf407d453cdde0d3708577397caefd034436 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d340c782d1b6591981424547a68309740e5c67a392ef8d6789c5b30e5ee07924 +size 618640 diff --git a/vlm/train/B1l6qiR5F7/3.png b/vlm/train/B1l6qiR5F7/3.png new file mode 100644 index 0000000000000000000000000000000000000000..5d52869d9f2d5acd918178c5f11e0d1d3542ebac --- /dev/null +++ b/vlm/train/B1l6qiR5F7/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21abda44e865aab6bf8c3da9a7591803435d6449c65acbc0dac7507647659fc7 +size 436207 diff --git a/vlm/train/B1l6qiR5F7/4.png b/vlm/train/B1l6qiR5F7/4.png new file mode 100644 index 0000000000000000000000000000000000000000..007e62c1b098ca6c61a76f4f5861a90aa8e06643 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47e13f1b2310f3d2932f53e35bb7a638c33b60627a80d1c289106035cbfb39cb +size 453018 diff --git a/vlm/train/B1l6qiR5F7/5.png b/vlm/train/B1l6qiR5F7/5.png new file mode 100644 index 0000000000000000000000000000000000000000..da35bdeceae2813231f85bf0ba2a8db1011fc268 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1206631d7fdfcf12788fccbb0c3de085423517f2a86c41fe6d107643d6a3600f +size 503175 diff --git a/vlm/train/B1l6qiR5F7/6.png b/vlm/train/B1l6qiR5F7/6.png new file mode 100644 index 0000000000000000000000000000000000000000..506cfde86eaf4dbdba7b9708f53eb41faa2ff5d6 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e782e8b530286d37b72f842ef03225533e06af6dc52b5c0e420a57eabb15e44 +size 546970 diff --git a/vlm/train/B1l6qiR5F7/7.png b/vlm/train/B1l6qiR5F7/7.png new file mode 100644 index 0000000000000000000000000000000000000000..5f4742f53639c131d8d5cafe4ee59c2862a1bd17 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e743abd336682b31ce48da7a7a3f042efaaf3336fd875cb327fd504d51e8d323 +size 569692 diff --git a/vlm/train/B1l6qiR5F7/8.png b/vlm/train/B1l6qiR5F7/8.png new file mode 100644 index 0000000000000000000000000000000000000000..266a3578cf2893d67918e7bde52f586c459a4191 --- /dev/null +++ b/vlm/train/B1l6qiR5F7/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be6c1c7c2ecbc11058d0495f6abd4a0abcc6c16b16fdf55fe4ab1e07938c75ca +size 374770 diff --git a/vlm/train/B1l6qiR5F7/9.png b/vlm/train/B1l6qiR5F7/9.png new file mode 100644 index 0000000000000000000000000000000000000000..a74d447a21966c799f1f9190bbc503013441f2fd --- /dev/null +++ b/vlm/train/B1l6qiR5F7/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f934dc3b24d4b7a97b1f1e0fc462373ad00817e4c37c2e6b32b007492db14e3c +size 531800 diff --git a/vlm/train/BJOFETxR-/0.png b/vlm/train/BJOFETxR-/0.png new file mode 100644 index 0000000000000000000000000000000000000000..89a6ebaeeab45d52a090351a3a4b98477d371b31 --- /dev/null +++ b/vlm/train/BJOFETxR-/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:574fdb9cb433723dc096217fb5ca572da32035f427a41335a1fec549b52c96fe +size 468904 diff --git a/vlm/train/BJOFETxR-/1.png b/vlm/train/BJOFETxR-/1.png new file mode 100644 index 0000000000000000000000000000000000000000..2179987d6b35b3a151c0be83664099ffee35443b --- /dev/null +++ b/vlm/train/BJOFETxR-/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe9ef6b0dad4816bbd2b5f71a605187f2d30d18d2d22783570f6bc9a1cd9b19b +size 561191 diff --git a/vlm/train/BJOFETxR-/10.png b/vlm/train/BJOFETxR-/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d1cec6114f84da8a31476c3f9a2e6db41470eea1 --- /dev/null +++ b/vlm/train/BJOFETxR-/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f17473e27c2619639041a46fee584cda7101ff3889c2a3083ccd20d58c6b48e +size 362732 diff --git a/vlm/train/BJOFETxR-/11.png b/vlm/train/BJOFETxR-/11.png new file mode 100644 index 0000000000000000000000000000000000000000..4ded198e88a431471b5d55d8bab9f16a43403e52 --- /dev/null +++ b/vlm/train/BJOFETxR-/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b19318c10d5bfcb79454b1804563b463214ec0e5d33e8e50a198ad46c60512e0 +size 197192 diff --git a/vlm/train/BJOFETxR-/12.png b/vlm/train/BJOFETxR-/12.png new file mode 100644 index 0000000000000000000000000000000000000000..d4c8c58e2954a3769a04369e635a9a4d9aa32452 --- /dev/null +++ b/vlm/train/BJOFETxR-/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b65a4940948aadc91309d2e081e67a54706a11b980f1c795a267efd5735e543 +size 191212 diff --git a/vlm/train/BJOFETxR-/13.png b/vlm/train/BJOFETxR-/13.png new file mode 100644 index 0000000000000000000000000000000000000000..93b79ccd5e004324c9bc0f6d91d9d9f223204793 --- /dev/null +++ b/vlm/train/BJOFETxR-/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c77bb11addd6d9c98396e2a0c78b47f35b83e5b777af0d32b6b82f6dfeb11886 +size 210091 diff --git a/vlm/train/BJOFETxR-/14.png b/vlm/train/BJOFETxR-/14.png new file mode 100644 index 0000000000000000000000000000000000000000..27a351bcc4cb7375b0ddd305f261612628fb7802 --- /dev/null +++ b/vlm/train/BJOFETxR-/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f2ccd505c57571ac07344d93525777fa5d1d5b2866b1780d6d3a8e432599cb3 +size 224209 diff --git a/vlm/train/BJOFETxR-/15.png b/vlm/train/BJOFETxR-/15.png new file mode 100644 index 0000000000000000000000000000000000000000..9add213b04ab988d713dc4cb5e018a74bdceeb58 --- /dev/null +++ b/vlm/train/BJOFETxR-/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1396f878883c45f5dcb6c32c11ad2438e2bc821e7ec9c8c44a95d7994560a305 +size 425865 diff --git a/vlm/train/BJOFETxR-/16.png b/vlm/train/BJOFETxR-/16.png new file mode 100644 index 0000000000000000000000000000000000000000..50d8a0b97c45cddfcbb860327eaaf6b7f06b7d74 --- /dev/null +++ b/vlm/train/BJOFETxR-/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a12ff2076aeaa2b50e164eb8169fff40812293ac181ed42aa7639a39da4bfc7 +size 84008 diff --git a/vlm/train/BJOFETxR-/2.png b/vlm/train/BJOFETxR-/2.png new file mode 100644 index 0000000000000000000000000000000000000000..4bcd373300122808d8333061e9da9391eb3d6b43 --- /dev/null +++ b/vlm/train/BJOFETxR-/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab6f4f587da693e0de41e3f4510964ce295e7feb5e2c762478564c3d82c8a619 +size 553033 diff --git a/vlm/train/BJOFETxR-/3.png b/vlm/train/BJOFETxR-/3.png new file mode 100644 index 0000000000000000000000000000000000000000..9af3c232939900d765d7a47f2a77a1456bb3e24e --- /dev/null +++ b/vlm/train/BJOFETxR-/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:346e2440cf3ddd62b915d9f9fe8edd8b502ee50cfcdb65a23fa9684f954895ef +size 572629 diff --git a/vlm/train/BJOFETxR-/4.png b/vlm/train/BJOFETxR-/4.png new file mode 100644 index 0000000000000000000000000000000000000000..41cb42e0e477cfa4e4993934d2ed00e6a6254e3d --- /dev/null +++ b/vlm/train/BJOFETxR-/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1932aacdf08a972448732f713974a694f1df0c52913c2c4d7c44badaaa63663 +size 618244 diff --git a/vlm/train/BJOFETxR-/5.png b/vlm/train/BJOFETxR-/5.png new file mode 100644 index 0000000000000000000000000000000000000000..d651be4a64805fa09735c6e5e8ecab0b54b62e10 --- /dev/null +++ b/vlm/train/BJOFETxR-/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1edc0a8e93a5b399eb75134d7f477f114659596afdea22bab80c373686d32731 +size 575892 diff --git a/vlm/train/BJOFETxR-/6.png b/vlm/train/BJOFETxR-/6.png new file mode 100644 index 0000000000000000000000000000000000000000..85c47a3b0374fa56a046c6870d1e433010e832c5 --- /dev/null +++ b/vlm/train/BJOFETxR-/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f769989cbd5f048c9dda3281303f624704ce77173a6290ff6e717639048c6048 +size 490488 diff --git a/vlm/train/BJOFETxR-/7.png b/vlm/train/BJOFETxR-/7.png new file mode 100644 index 0000000000000000000000000000000000000000..065d021bde59c373998d57a7b40be31fb89450ea --- /dev/null +++ b/vlm/train/BJOFETxR-/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1f98571508d88660ecff30db84add8b86c8ec07efc76aa266fdb48f6b57229a +size 561934 diff --git a/vlm/train/BJOFETxR-/8.png b/vlm/train/BJOFETxR-/8.png new file mode 100644 index 0000000000000000000000000000000000000000..748e4871dacac0cd77da55200ce3c145d22c107c --- /dev/null +++ b/vlm/train/BJOFETxR-/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a559adcce204c1ce5e0489a675d7d01d2029dd4192566f82c1ac2fdd1e46834 +size 537590 diff --git a/vlm/train/BJOFETxR-/9.png b/vlm/train/BJOFETxR-/9.png new file mode 100644 index 0000000000000000000000000000000000000000..18fb5c9abaa2d487976362b73f05cc6395be67e3 --- /dev/null +++ b/vlm/train/BJOFETxR-/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6107f84953bcb090f5143b4871f808ce323794f372bbd9f8170f7cbe521903fc +size 571648 diff --git a/vlm/train/BJg4NgBKvH/0.png b/vlm/train/BJg4NgBKvH/0.png new file mode 100644 index 0000000000000000000000000000000000000000..fbcd3fe50d9bf1c6d60f52dbec5dea482a87e45b --- /dev/null +++ b/vlm/train/BJg4NgBKvH/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e36cd46d0b66bae562c0fb718bf5be1464efe442e0374ff441466355a6abfc8 +size 494385 diff --git a/vlm/train/BJg4NgBKvH/1.png b/vlm/train/BJg4NgBKvH/1.png new file mode 100644 index 0000000000000000000000000000000000000000..984786504631b4955fa978e0ffd087b845885639 --- /dev/null +++ b/vlm/train/BJg4NgBKvH/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e15e4bdbbe3077b812ea944d31b74a93ffa0deb846dc6f6661e07d64ad37f854 +size 452344 diff --git a/vlm/train/BJg4NgBKvH/10.png b/vlm/train/BJg4NgBKvH/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a81c3df5c6c0dde141a3f5bad69ccb28cc47ac4e --- /dev/null +++ b/vlm/train/BJg4NgBKvH/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ece5f25bc90bcbd73982163878cd5b31323a3728783c29e29c6299daa7d86b8d +size 59050 diff --git a/vlm/train/BJg4NgBKvH/2.png b/vlm/train/BJg4NgBKvH/2.png new file mode 100644 index 0000000000000000000000000000000000000000..27e69b62c87bc5adc8c29a0143ee32cdf4cb138b --- /dev/null +++ b/vlm/train/BJg4NgBKvH/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:022e83f170929334d2c93518c7e942da9b493b9e39e188aef91868b258c0f072 +size 597191 diff --git a/vlm/train/BJg4NgBKvH/3.png b/vlm/train/BJg4NgBKvH/3.png new file mode 100644 index 0000000000000000000000000000000000000000..cc87a98d83cbe7a2c01e5b0a9915cb50fb05a302 --- /dev/null +++ b/vlm/train/BJg4NgBKvH/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87e3cf83fc3e20043e9cd69fd0270a6d19794916b91bc216d2e4bcd177bdfe05 +size 504536 diff --git a/vlm/train/BJg4NgBKvH/4.png b/vlm/train/BJg4NgBKvH/4.png new file mode 100644 index 0000000000000000000000000000000000000000..d72736ee3e3c821c6d324ffb5f4c0d7b424f2475 --- /dev/null +++ b/vlm/train/BJg4NgBKvH/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a28199503a4b4c95dc4c916662940f151b68e9eedfc48ab3d6ddab1e47e34f64 +size 550382 diff --git a/vlm/train/BJg4NgBKvH/5.png b/vlm/train/BJg4NgBKvH/5.png new file mode 100644 index 0000000000000000000000000000000000000000..b5937fed769335798526d53e7f3cd93098c92a1b --- /dev/null +++ b/vlm/train/BJg4NgBKvH/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed326aeaa1f240417ad13ba940dd18524e190e114abe9812af380c9a61911fdc +size 546440 diff --git a/vlm/train/BJg4NgBKvH/6.png b/vlm/train/BJg4NgBKvH/6.png new file mode 100644 index 0000000000000000000000000000000000000000..7e43723dc769b60152bd5474ac8863faeb4d8679 --- /dev/null +++ b/vlm/train/BJg4NgBKvH/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25fd8983ecf7a6b9f53bc37b8dccde9fe7506749d4fc307774eb01e80978aae2 +size 491505 diff --git a/vlm/train/BJg4NgBKvH/7.png b/vlm/train/BJg4NgBKvH/7.png new file mode 100644 index 0000000000000000000000000000000000000000..90b6745057e200076d93e492f42154998da33719 --- /dev/null +++ b/vlm/train/BJg4NgBKvH/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c85d8e93917b522d1afd11b0476a9f976e153a2ceff3396c44af110fac5ddfe +size 423606 diff --git a/vlm/train/BJg4NgBKvH/8.png b/vlm/train/BJg4NgBKvH/8.png new file mode 100644 index 0000000000000000000000000000000000000000..fdcd0ceb42cf0343c438e9ca7eed4a5ef0b1c6de --- /dev/null +++ b/vlm/train/BJg4NgBKvH/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43d2b0b9252b373779790753c4adedbd50f02e857ca9096c0a1334b095d0fdac +size 481180 diff --git a/vlm/train/BJg4NgBKvH/9.png b/vlm/train/BJg4NgBKvH/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0312cdde273aab5b969d34f3cc72bf36b58839de --- /dev/null +++ b/vlm/train/BJg4NgBKvH/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bddf888afc4d1218e48bc6a910a9c1de231cff98ef74ace4728e4981f3c22b7a +size 540183 diff --git a/vlm/train/BJk7Gf-CZ/10.png b/vlm/train/BJk7Gf-CZ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..76441635d5aeadd128b1a5ea863f3f432f4cbe7d --- /dev/null +++ b/vlm/train/BJk7Gf-CZ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21bfba829cb2dba7aae476319d3c9c42e287cab8cf1eb9bc4898b764c4e5cf60 +size 332354 diff --git a/vlm/train/BJk7Gf-CZ/3.png b/vlm/train/BJk7Gf-CZ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..8d167a6ffe7b8c6ec26f340e78f1a505fede466e --- /dev/null +++ b/vlm/train/BJk7Gf-CZ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bf3b1f29b5ce2ce5a705333d51c31d37f01125dd8e87d802c2dbbaa22f8f52b +size 443526 diff --git a/vlm/train/BJk7Gf-CZ/6.png b/vlm/train/BJk7Gf-CZ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..1dcb16bcc393e92ebe33ae44748702e8a1cece74 --- /dev/null +++ b/vlm/train/BJk7Gf-CZ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a873bccbdd91952e8e4690dd6980b199e287423ab2cfc4b964f1155efca4c5bc +size 449347 diff --git a/vlm/train/BJk7Gf-CZ/9.png b/vlm/train/BJk7Gf-CZ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0e61bd39e3c573a19e8225b3369352eee0bc07b1 --- /dev/null +++ b/vlm/train/BJk7Gf-CZ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20f09f74e46492e87e3ccbc6e9af64d69762dc1ccba5bc8d9688d87369c3b157 +size 370247 diff --git a/vlm/train/BJlZ5ySKPH/0.png b/vlm/train/BJlZ5ySKPH/0.png new file mode 100644 index 0000000000000000000000000000000000000000..34c18895bbdda2d1f9d07d1f1a18a0132ddbb9fa --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0e0fade781782fdeb0d1649811e228536ade3b44308e26bfc95a8ef19c9b6e82 +size 497681 diff --git a/vlm/train/BJlZ5ySKPH/1.png b/vlm/train/BJlZ5ySKPH/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9115bfe1a45550f6dec33bc83265f890c2538da6 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb4fa78873411338bbe1e2329f001da503e9fe5dff6bae728797c6e4c256fd0c +size 488918 diff --git a/vlm/train/BJlZ5ySKPH/10.png b/vlm/train/BJlZ5ySKPH/10.png new file mode 100644 index 0000000000000000000000000000000000000000..ad5284a9f3252b6af1206da932e07b77bbe571cd --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:887b89450a8c0f8753f1e97c779e2c1cd03491c868e0bfd8c080b284f9789500 +size 532988 diff --git a/vlm/train/BJlZ5ySKPH/11.png b/vlm/train/BJlZ5ySKPH/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5d7d8f2693ac58fe580bf7c5a20fdd49d3d002dc --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b540a0b08172ec5ad93e53cb7f6741d43267a246a4a8e5d726434f950902dfea +size 588203 diff --git a/vlm/train/BJlZ5ySKPH/12.png b/vlm/train/BJlZ5ySKPH/12.png new file mode 100644 index 0000000000000000000000000000000000000000..60950c9fba5ec553e2465f71026e7eccaa6fa8f5 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a76ee595aef3f51901f4ce3cb2f2e14a5ee2cef677f14484dd49a9159a6beff4 +size 491182 diff --git a/vlm/train/BJlZ5ySKPH/13.png b/vlm/train/BJlZ5ySKPH/13.png new file mode 100644 index 0000000000000000000000000000000000000000..f2a1554082ca6c46b6e5f413c7e311f0ed185a92 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23d842bbc5742ed955875929f364fe0a3eb2ad43a79e361373fbe2f869b707e0 +size 414551 diff --git a/vlm/train/BJlZ5ySKPH/14.png b/vlm/train/BJlZ5ySKPH/14.png new file mode 100644 index 0000000000000000000000000000000000000000..2a3f816a16fa50d234c6875d9bfc10af5f29bb99 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f99468cef8d08dea8062743559ddd30dc344c9f1e254e8cf42a7d6b19d2a3366 +size 2083139 diff --git a/vlm/train/BJlZ5ySKPH/15.png b/vlm/train/BJlZ5ySKPH/15.png new file mode 100644 index 0000000000000000000000000000000000000000..d19e498ae6a61f193f2adcc1147ef16c9d376d1c --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4340139342ead6c5f08aba95fd62be863ea6d340c78b63e82966bf11af8650d5 +size 1649458 diff --git a/vlm/train/BJlZ5ySKPH/16.png b/vlm/train/BJlZ5ySKPH/16.png new file mode 100644 index 0000000000000000000000000000000000000000..735d440558b7eb16dcdf36f21ff703313614ebfd --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b8d0acb7bb2ca25180af3397c238077e60c30e546c7e9b439349b119470bc7e +size 2036251 diff --git a/vlm/train/BJlZ5ySKPH/17.png b/vlm/train/BJlZ5ySKPH/17.png new file mode 100644 index 0000000000000000000000000000000000000000..a598dd4434b7b473ac402265cf5574c1bdb36d55 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf75543f39becc6d24f766e30c24d8ae1fa03b24f9d5d39004d59d7c8638ed0e +size 1733539 diff --git a/vlm/train/BJlZ5ySKPH/18.png b/vlm/train/BJlZ5ySKPH/18.png new file mode 100644 index 0000000000000000000000000000000000000000..ee1ea675629f299b5575bfca5effe124ffa3024d --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd23ef6eb505adc88e1b595f9d627ad2681aa1b79ad2a5d48c742f3a71246642 +size 2018048 diff --git a/vlm/train/BJlZ5ySKPH/2.png b/vlm/train/BJlZ5ySKPH/2.png new file mode 100644 index 0000000000000000000000000000000000000000..ca78560a3bf332e0d4dad7f45b78f854860e63c4 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44c04c218336fd5cfb9e14ef51677d854020d83c28177579d2213d6da2a8b1bf +size 501797 diff --git a/vlm/train/BJlZ5ySKPH/3.png b/vlm/train/BJlZ5ySKPH/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b3006be18fdea562c388bac6b10866478bbc7dbb --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dc7255b80cb7eb41ebb37262e86c7d31efa5459ae95ec03ba22ca9391b83785 +size 405101 diff --git a/vlm/train/BJlZ5ySKPH/4.png b/vlm/train/BJlZ5ySKPH/4.png new file mode 100644 index 0000000000000000000000000000000000000000..5ed72d4657b7f95b3442ace075eb7d42ac0c304c --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48a76059b96870b5f4512d644b10947a19ac0780a03cd9fc8643c22803ff70da +size 925330 diff --git a/vlm/train/BJlZ5ySKPH/5.png b/vlm/train/BJlZ5ySKPH/5.png new file mode 100644 index 0000000000000000000000000000000000000000..ed03add8106c746f6a963238ba9d5a8b63f5b7dd --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c4e15224fbcc4c5be7580ad063504ebc29e73bd3eb58863dafa7cdd285a1361 +size 1049134 diff --git a/vlm/train/BJlZ5ySKPH/6.png b/vlm/train/BJlZ5ySKPH/6.png new file mode 100644 index 0000000000000000000000000000000000000000..c1c05277e9ba6710d7b8b75f8a571a45dfff80e1 --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01d38b4e67ec457588efa3171205560868307267745a2771f775e8ce830c5502 +size 1088401 diff --git a/vlm/train/BJlZ5ySKPH/7.png b/vlm/train/BJlZ5ySKPH/7.png new file mode 100644 index 0000000000000000000000000000000000000000..0f81179cba144e9f98c4d782dd59d00b52a3a71a --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b0a3c5e867baef32064daf46c441854b38f6280eb7156b431bc4336bd21c555 +size 500697 diff --git a/vlm/train/BJlZ5ySKPH/8.png b/vlm/train/BJlZ5ySKPH/8.png new file mode 100644 index 0000000000000000000000000000000000000000..ab74016a77a1a55ab1a89f813f44dcff441e734d --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa48da31644bc021f5b58249554a1ba18f83eb9a5d2ae21d36719eaa67319164 +size 503867 diff --git a/vlm/train/BJlZ5ySKPH/9.png b/vlm/train/BJlZ5ySKPH/9.png new file mode 100644 index 0000000000000000000000000000000000000000..ab3802a3df1b14318fab75cce8ffa528bd7db94d --- /dev/null +++ b/vlm/train/BJlZ5ySKPH/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a856cae5c96a28f10a8495bc14bef56c3f8cc82176f92a8ad68a57a33d44c05 +size 544538 diff --git a/vlm/train/BkUHlMZ0b/0.png b/vlm/train/BkUHlMZ0b/0.png new file mode 100644 index 0000000000000000000000000000000000000000..979c5dccdc01b773a8ada2971848c6e2b22fe8ef --- /dev/null +++ b/vlm/train/BkUHlMZ0b/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99fd9d0736372961f70c3da91ee4ca9ece80844311d63877d67908af77f21ebc +size 508176 diff --git a/vlm/train/BkUHlMZ0b/1.png b/vlm/train/BkUHlMZ0b/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0821e18079897c56e7178e754683977de20b0fbb --- /dev/null +++ b/vlm/train/BkUHlMZ0b/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e94ea9acbc80a91408e13a8c02ffd536e124fc6beb1dc5a50f90219f9afbbf1 +size 665484 diff --git a/vlm/train/BkUHlMZ0b/10.png b/vlm/train/BkUHlMZ0b/10.png new file mode 100644 index 0000000000000000000000000000000000000000..5000e081605379ad2c93c741cec220fadc3dcb96 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffb8ff44a5ecb593dd7e1ecbbb1a0608b8077d75e69af8ead8067e9fb128e5b1 +size 409685 diff --git a/vlm/train/BkUHlMZ0b/11.png b/vlm/train/BkUHlMZ0b/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5280a3b14ce93a1fc1b603a4ff56528e8fe4ab55 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:20dfc9405528e7b01ae300ec7ea9961f157e6e3456201703b555c3cd8f147007 +size 500767 diff --git a/vlm/train/BkUHlMZ0b/12.png b/vlm/train/BkUHlMZ0b/12.png new file mode 100644 index 0000000000000000000000000000000000000000..ec858d596272bd03d0eaa8486ef34a36844c4442 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:134d06fbed718f27c55c3642e97866eaed19986e81614d79bb939139e302255f +size 575289 diff --git a/vlm/train/BkUHlMZ0b/13.png b/vlm/train/BkUHlMZ0b/13.png new file mode 100644 index 0000000000000000000000000000000000000000..5058b227b582830ca2fa4170f337393891861d77 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62aa468303152f238c36dd9ac6c6580e42ad6cd8ce73d05dc37d6b1e4e586b82 +size 117692 diff --git a/vlm/train/BkUHlMZ0b/14.png b/vlm/train/BkUHlMZ0b/14.png new file mode 100644 index 0000000000000000000000000000000000000000..98f004e75ec785305d5ac6b5bba6bf63398f65dc --- /dev/null +++ b/vlm/train/BkUHlMZ0b/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd2e4486d9f05039f69e89b1e03ea70b2c8b0c70343294f573e0d56b13fe0af6 +size 364935 diff --git a/vlm/train/BkUHlMZ0b/15.png b/vlm/train/BkUHlMZ0b/15.png new file mode 100644 index 0000000000000000000000000000000000000000..545871f764d9d898a57e0cdf176863a698803fd3 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2aa0a8d9fe789cbc434092b048f4dfb24e003ccc29f2f87199b3d3078ac57fe +size 390420 diff --git a/vlm/train/BkUHlMZ0b/16.png b/vlm/train/BkUHlMZ0b/16.png new file mode 100644 index 0000000000000000000000000000000000000000..bb8bdcbe6ebd5f6f86da526dd756ba3278c4cd0d --- /dev/null +++ b/vlm/train/BkUHlMZ0b/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f01a3489405fe58f5b1f1a6f7014206c4d09a6339c00547cc542a24810b2013 +size 375244 diff --git a/vlm/train/BkUHlMZ0b/17.png b/vlm/train/BkUHlMZ0b/17.png new file mode 100644 index 0000000000000000000000000000000000000000..a1ec88d70f169d2fa4f59bb2a948bb42cf4ce474 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95fae56d370769ba93057169285837f5584f814a47b26c4949ababedff85c1a3 +size 112290 diff --git a/vlm/train/BkUHlMZ0b/2.png b/vlm/train/BkUHlMZ0b/2.png new file mode 100644 index 0000000000000000000000000000000000000000..d0ca94929e2cde40b0a3e7299645cf98268db202 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73849b497776251f612fcd7bab470d0ea8ac92d35192bf9b49d3117e3d8e4d40 +size 620170 diff --git a/vlm/train/BkUHlMZ0b/3.png b/vlm/train/BkUHlMZ0b/3.png new file mode 100644 index 0000000000000000000000000000000000000000..8894c1be3ed42667bdb762abc9fce360bbb9c300 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a14ace879f1ce1f5cb899033c53607a54746b1deb8500f634607d6f7b3049457 +size 602744 diff --git a/vlm/train/BkUHlMZ0b/4.png b/vlm/train/BkUHlMZ0b/4.png new file mode 100644 index 0000000000000000000000000000000000000000..eb5f3926fbabeb1911712f2e4e9ce95c30a57e1e --- /dev/null +++ b/vlm/train/BkUHlMZ0b/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3675f68b2836deb1a713d668953aa70040cb84131be93a65f4818b6db6edbd4b +size 578258 diff --git a/vlm/train/BkUHlMZ0b/5.png b/vlm/train/BkUHlMZ0b/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8b1ffbc8f4cac74d749157c39c9c4f8f12b2a281 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53e1ab9e7a92341bffe6a64bd8a12da468664d2721490b3b2ff97de998e7a3c2 +size 628814 diff --git a/vlm/train/BkUHlMZ0b/6.png b/vlm/train/BkUHlMZ0b/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ea838204a8a32155be62f89ab926eafd3fcd7a20 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59a40f509142f2839b50121ba8386899ee1fc62dd35b28fc06d0c390d2473f40 +size 503793 diff --git a/vlm/train/BkUHlMZ0b/7.png b/vlm/train/BkUHlMZ0b/7.png new file mode 100644 index 0000000000000000000000000000000000000000..65e4c790db116c6a14ce3c832ee2bd5beaf0642a --- /dev/null +++ b/vlm/train/BkUHlMZ0b/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a033881ad569bf7e265debc8892b9eea90dfd4295bee9b80a68972e543b47356 +size 528548 diff --git a/vlm/train/BkUHlMZ0b/8.png b/vlm/train/BkUHlMZ0b/8.png new file mode 100644 index 0000000000000000000000000000000000000000..c5ccc31aaa36d0ff45391888329c6ab97971f9e5 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e4621611a298f2118bcf05953b1c906d51087c73ba3a14b8b3f827ccc204307 +size 652562 diff --git a/vlm/train/BkUHlMZ0b/9.png b/vlm/train/BkUHlMZ0b/9.png new file mode 100644 index 0000000000000000000000000000000000000000..6b85e0b5f76bf2a91323649a656bc9e1f5e49dd4 --- /dev/null +++ b/vlm/train/BkUHlMZ0b/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea83795f93bf1e3560a21fa5f3520335333e284c70a887130f9e716fce7c8253 +size 494130 diff --git a/vlm/train/Bkg2viA5FQ/0.png b/vlm/train/Bkg2viA5FQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..a5c74de7b1286231e31d1ca20d66b00b576112c9 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a32945dcef823bb7d33c0cc3ae78d5667b18fc8985771257a183da0e29981b26 +size 490575 diff --git a/vlm/train/Bkg2viA5FQ/1.png b/vlm/train/Bkg2viA5FQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..8355db4e5dab86fe0f523ae8066e1b620096a933 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c63c5981144f8c2ed9a5a743471eabee7666690fecc964c15575be3c5ab0970 +size 535456 diff --git a/vlm/train/Bkg2viA5FQ/10.png b/vlm/train/Bkg2viA5FQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..8f5e3ebc602b771b19b9f351514b00a0f4bfc6a0 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5734fedd25b4fbb89c9e2e4cc9e74a4c7ffbda918184fdc883448c1a66104692 +size 510364 diff --git a/vlm/train/Bkg2viA5FQ/11.png b/vlm/train/Bkg2viA5FQ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..7f5eef3d588a80f664b819ced9ebd8daaaa5e822 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88bd8838513cfa6d5cd3c945f58061479fd785f48104c0ed826f6135426b495e +size 243634 diff --git a/vlm/train/Bkg2viA5FQ/12.png b/vlm/train/Bkg2viA5FQ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..1f49d84f52b7ffba62d5a9a4f1b26e60b0907f51 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26b9434dcdf03ea6ea14d667df9d2a0d1d89946bc63b7a5b960f13da1f7959c2 +size 279879 diff --git a/vlm/train/Bkg2viA5FQ/13.png b/vlm/train/Bkg2viA5FQ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..676009f6ddf885b95eaed68f10ea748b37c78631 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da2309ef510b8b85ef120199b69f6303eb43532b1257da9a5f24d6a8a4462a14 +size 276641 diff --git a/vlm/train/Bkg2viA5FQ/14.png b/vlm/train/Bkg2viA5FQ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..4fb8de9c9452aa7874821f8fe90e8b362b546c20 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:135474ce6843df3b3cbbf974f55ff2748ad26936da63ee09d9072fa554967351 +size 270663 diff --git a/vlm/train/Bkg2viA5FQ/15.png b/vlm/train/Bkg2viA5FQ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..00818f0c8368479940fc4e16038a54f7dac46497 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:941017a022a6662168d68af3d61977280c61a5ded61527e121077c28a18ed7f9 +size 358444 diff --git a/vlm/train/Bkg2viA5FQ/16.png b/vlm/train/Bkg2viA5FQ/16.png new file mode 100644 index 0000000000000000000000000000000000000000..e12b44ada4d86434e26b7e8d6a35cd5dadc60f8b --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e5a6d278ea808c1f3b29d83f68a30edc0aedb0bb924a7fce2aa117f057cdba7 +size 305549 diff --git a/vlm/train/Bkg2viA5FQ/17.png b/vlm/train/Bkg2viA5FQ/17.png new file mode 100644 index 0000000000000000000000000000000000000000..7668fbb24c6945e5efb9d2c11d8ff7c40e68d20b --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee87070645b8dbc3d01bf293ed07b4f8a534572adea8d0deb2f6e17ac9dc1044 +size 341089 diff --git a/vlm/train/Bkg2viA5FQ/18.png b/vlm/train/Bkg2viA5FQ/18.png new file mode 100644 index 0000000000000000000000000000000000000000..cc1295ab6c244a3ce9d53bcbe02255d0753e2224 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5097d6dac66f2fe333b227b9e33f24ed238f11f8faeeea7da78ab99bb2cf1ba +size 156027 diff --git a/vlm/train/Bkg2viA5FQ/19.png b/vlm/train/Bkg2viA5FQ/19.png new file mode 100644 index 0000000000000000000000000000000000000000..d12e8d724b13d4912a5fc45c283cbd0123e154f0 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efb24865ab4bb048bed2b7882101dc4b39815f3633df4f5a2c7e42cd053709be +size 353682 diff --git a/vlm/train/Bkg2viA5FQ/2.png b/vlm/train/Bkg2viA5FQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..f2c088a940982721c7882b632c0674e57d5e215e --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24feca4ba534b964e567a75594bc4e5dd527c7fd2bc315c47b69b55e432b17e0 +size 484115 diff --git a/vlm/train/Bkg2viA5FQ/20.png b/vlm/train/Bkg2viA5FQ/20.png new file mode 100644 index 0000000000000000000000000000000000000000..28f95488afb37627b03e8b29442636bccfd721b5 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:297acf50cd545d48b4b6c919e7e67470a6618893441473a87a4ca057d6ae6fdb +size 348802 diff --git a/vlm/train/Bkg2viA5FQ/21.png b/vlm/train/Bkg2viA5FQ/21.png new file mode 100644 index 0000000000000000000000000000000000000000..80200bc6461783eefc71d20502ec2d24701133b7 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90bb3f68f0aeda99bbdb7471e89304b990fac78f116fd1d92fac9a790b807bb3 +size 341907 diff --git a/vlm/train/Bkg2viA5FQ/22.png b/vlm/train/Bkg2viA5FQ/22.png new file mode 100644 index 0000000000000000000000000000000000000000..07cdd927c03e52237512874d226c691626865133 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:649d9b8d50ce901dae68a7b6ec6864149b57bbcd15c250c7da28ca018b0232cf +size 251799 diff --git a/vlm/train/Bkg2viA5FQ/23.png b/vlm/train/Bkg2viA5FQ/23.png new file mode 100644 index 0000000000000000000000000000000000000000..5fc97148292191284ab199ae48ddf166388c7774 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6931ff3eeaec795504a7a25c7a161ea3860ef7ca15b6a808ccc8c3c30944d2d +size 297643 diff --git a/vlm/train/Bkg2viA5FQ/24.png b/vlm/train/Bkg2viA5FQ/24.png new file mode 100644 index 0000000000000000000000000000000000000000..d8c8beb32ae45e4b413be4c6c53d8c2464c7d5b8 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58399fcb99c707bd1dbd3c9d44ddf4141ac0526c79d03db89e635a6527ad0196 +size 79236 diff --git a/vlm/train/Bkg2viA5FQ/25.png b/vlm/train/Bkg2viA5FQ/25.png new file mode 100644 index 0000000000000000000000000000000000000000..fccb94f0757fc3c4c1807238f1addf28ed4094e5 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e082d0545f53c21c542f2a76121867e764c1cf0b3a02bcbb55437ccf6beb89e7 +size 550365 diff --git a/vlm/train/Bkg2viA5FQ/26.png b/vlm/train/Bkg2viA5FQ/26.png new file mode 100644 index 0000000000000000000000000000000000000000..66723355539bd2ab4e794cf3fc993128f3937a73 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:267be9f1cbf846ee96236d7ab6b381902793e20d1f3c518fdd8819373e778526 +size 189175 diff --git a/vlm/train/Bkg2viA5FQ/27.png b/vlm/train/Bkg2viA5FQ/27.png new file mode 100644 index 0000000000000000000000000000000000000000..733601fbe07322d8659ee2524372059bf9cb3178 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b947222d8e5c2b67c87499be94c3142de3ec9cb2d5171b8b070b88508952623 +size 156070 diff --git a/vlm/train/Bkg2viA5FQ/28.png b/vlm/train/Bkg2viA5FQ/28.png new file mode 100644 index 0000000000000000000000000000000000000000..b09cb01d5ed286a1a1954670d92d6fb052fc0d9a --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51525bd74a1920fb500296dad7c965b86176885af391420dad4951c695653a00 +size 329630 diff --git a/vlm/train/Bkg2viA5FQ/29.png b/vlm/train/Bkg2viA5FQ/29.png new file mode 100644 index 0000000000000000000000000000000000000000..562bad5f788198fc9cbc7177e3c49ba3d6796eda --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b8d761d98b98056c103743c43007808ebfd88e06346ac0329115ffb32354133 +size 212645 diff --git a/vlm/train/Bkg2viA5FQ/3.png b/vlm/train/Bkg2viA5FQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e76a6af205fcdd9b22c81235152171d2dfc17603 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5d5c415d506f236a1d733b2dcbebb41a89e9dfb6a41e7f75d673ca611b506c8 +size 470345 diff --git a/vlm/train/Bkg2viA5FQ/30.png b/vlm/train/Bkg2viA5FQ/30.png new file mode 100644 index 0000000000000000000000000000000000000000..9fffe3f951bc86cebab3d2fe20862311f668052c --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:438f9ef957da7279a948cbec21eb7ba64d9976dc05aa7c45cfad8223ac3cbeb4 +size 394904 diff --git a/vlm/train/Bkg2viA5FQ/31.png b/vlm/train/Bkg2viA5FQ/31.png new file mode 100644 index 0000000000000000000000000000000000000000..dea4b29d1b998af1da184ac7f950ff7ceb15420c --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/31.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0d6b22a094707efa9faa84ea5cd62d48ab332dd15daad4f7476f848f7d110b9 +size 415620 diff --git a/vlm/train/Bkg2viA5FQ/32.png b/vlm/train/Bkg2viA5FQ/32.png new file mode 100644 index 0000000000000000000000000000000000000000..a34c5f59d1970297c7102a3eefa6d3e25977c212 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/32.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03503c86be9e0049c34d4e0c746affb24563541dd2c63082244082ee71d75fe1 +size 145685 diff --git a/vlm/train/Bkg2viA5FQ/33.png b/vlm/train/Bkg2viA5FQ/33.png new file mode 100644 index 0000000000000000000000000000000000000000..e2aa708dd1e1f3026b230abfbc96e9b77faf90e9 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/33.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72be460a7f4035426a082091133b0d1aeae2d86f7decd60381d8b66a63b6bce6 +size 256455 diff --git a/vlm/train/Bkg2viA5FQ/34.png b/vlm/train/Bkg2viA5FQ/34.png new file mode 100644 index 0000000000000000000000000000000000000000..efc9af9dde14b33729aa2b7d33bf6cf27164f9b1 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/34.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b055f508cd7bb398254c50b6f788e6223ddd022e3fafec1412e577559c4b864 +size 117834 diff --git a/vlm/train/Bkg2viA5FQ/35.png b/vlm/train/Bkg2viA5FQ/35.png new file mode 100644 index 0000000000000000000000000000000000000000..3358a0f7ba95c43aca0283f86495510f3d17743d --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/35.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0c10181a52fb3703bb4a83d718f481b3e8bf087fde4499a9834f72518aaa3ae +size 649035 diff --git a/vlm/train/Bkg2viA5FQ/36.png b/vlm/train/Bkg2viA5FQ/36.png new file mode 100644 index 0000000000000000000000000000000000000000..fba20e1b729b8263027d51463e70c6b6cf3bc197 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/36.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab7dfc370892dcff3fc29d7bd397474e26a7a91150e00bd8aee077b3777790bf +size 388371 diff --git a/vlm/train/Bkg2viA5FQ/37.png b/vlm/train/Bkg2viA5FQ/37.png new file mode 100644 index 0000000000000000000000000000000000000000..7f7f102df7b574140538d8655c8edd33c71219ee --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/37.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:707eff8693a048bfd1cc98a60009bb938286ec48e0f42e458e93553ed1a98bcd +size 402556 diff --git a/vlm/train/Bkg2viA5FQ/4.png b/vlm/train/Bkg2viA5FQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..586e2f00465d9e0001dda0e905485a17c872df49 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74d39d46798de76370dcff03ed6caf91195591a5115fc05ef512582497f9013c +size 497856 diff --git a/vlm/train/Bkg2viA5FQ/5.png b/vlm/train/Bkg2viA5FQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..965e1d08c74891407ed3a1143f7d7d540a4195db --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c1f94129bdb6c0c596e9204b739f0877c71346f16882e305adc8c04deed7d9e +size 549720 diff --git a/vlm/train/Bkg2viA5FQ/6.png b/vlm/train/Bkg2viA5FQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..bde470a6d592f59edaf65b9b5344fd6e48a52265 --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5ee99018be494eae9b4ef92a531ae62d71a34adca350e1e325042aff951c043 +size 566340 diff --git a/vlm/train/Bkg2viA5FQ/7.png b/vlm/train/Bkg2viA5FQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..7cabe5402c76cc73660b307d1a31877953e2bd0a --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9da9dbe1d5f98dcf01f821e459a63b37fab77b4aabfec62621b21e7e6dbfa2e2 +size 520526 diff --git a/vlm/train/Bkg2viA5FQ/8.png b/vlm/train/Bkg2viA5FQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8fdd4db90bcfaab3d729014d0947d33f4779073a --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5efc55fff32404b5304bfe4b708b7d11b4d76423b13b728a5bc593bfa5d990a7 +size 548577 diff --git a/vlm/train/Bkg2viA5FQ/9.png b/vlm/train/Bkg2viA5FQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..72259987f8f896428a8f1f493350890df15fc33a --- /dev/null +++ b/vlm/train/Bkg2viA5FQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77fa9921aa343564eb3a23b36313f2e1448be76478c98526a4c85b79aea6b02b +size 538579 diff --git a/vlm/train/BkgrBgSYDS/0.png b/vlm/train/BkgrBgSYDS/0.png new file mode 100644 index 0000000000000000000000000000000000000000..5ad63898d2daaf80d6a76ee08546de3e48642e28 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:daa6ba830d0f665757401df3a583cc6c40a238296941871e194118cf7e534818 +size 503375 diff --git a/vlm/train/BkgrBgSYDS/1.png b/vlm/train/BkgrBgSYDS/1.png new file mode 100644 index 0000000000000000000000000000000000000000..cc0e2c82f774389a175c0c2ed1c6a070f0680cf2 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6f44c73ac14cf3b1dc626b5a43cc7edd97b2835bf3b0ba2449e8a3d9c129ecf +size 609449 diff --git a/vlm/train/BkgrBgSYDS/10.png b/vlm/train/BkgrBgSYDS/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d6af4f4bd9b157ffcfb07c437fcad671db3ff47a --- /dev/null +++ b/vlm/train/BkgrBgSYDS/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:548e8e52df8f5af06922da9bb591a94d54d7515229803364e154fac146b4ec65 +size 556426 diff --git a/vlm/train/BkgrBgSYDS/11.png b/vlm/train/BkgrBgSYDS/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5e76edbd53fbad1f82ed9477846256f43948f02a --- /dev/null +++ b/vlm/train/BkgrBgSYDS/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da7fe4855ebcce9a29c04ce0a2bec60136b860c69ac222c29d0e20e4a716bc2e +size 588645 diff --git a/vlm/train/BkgrBgSYDS/12.png b/vlm/train/BkgrBgSYDS/12.png new file mode 100644 index 0000000000000000000000000000000000000000..73d1217e0c98f7285447b02da561135ec3cf060a --- /dev/null +++ b/vlm/train/BkgrBgSYDS/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cd0896d0f7bb4795f7fc0901e3bd301d81b544751c04ae1f3911fbe8dff1c70 +size 538631 diff --git a/vlm/train/BkgrBgSYDS/13.png b/vlm/train/BkgrBgSYDS/13.png new file mode 100644 index 0000000000000000000000000000000000000000..3efecec735f243fc29709ed5251fb84c7211a501 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8df38d997abf15a50bc659b40a608f56050ff73d58bee19b30adcff232b8738b +size 398014 diff --git a/vlm/train/BkgrBgSYDS/14.png b/vlm/train/BkgrBgSYDS/14.png new file mode 100644 index 0000000000000000000000000000000000000000..5d37f202c90df55a2a4fd2ac0abfee50d86b6ee6 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eed63ff4232f0807dbe88a8cfb2dd99d8d9138863e5ea8dd46c4c903057aa8da +size 595951 diff --git a/vlm/train/BkgrBgSYDS/15.png b/vlm/train/BkgrBgSYDS/15.png new file mode 100644 index 0000000000000000000000000000000000000000..3058b35e489a5d1bd1f8a195b6424600acc429f0 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:083e9aa92745b4f77074fffee996468748c459680d2445861920bc0ae18bcb30 +size 538896 diff --git a/vlm/train/BkgrBgSYDS/16.png b/vlm/train/BkgrBgSYDS/16.png new file mode 100644 index 0000000000000000000000000000000000000000..31b4554e999478fcb460acf5573e31577a9b8538 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:418770ffb6e63f7aea237cacd9e520a87a4a4e5aca5eef8e68529181ece1387d +size 560987 diff --git a/vlm/train/BkgrBgSYDS/17.png b/vlm/train/BkgrBgSYDS/17.png new file mode 100644 index 0000000000000000000000000000000000000000..94d512bc04ddb970cc8cafdcedbc9d7f7d89e8a0 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:286ff86e967c7cb9f24a009540b103694376624cf10169cb4a9a0a6139256e85 +size 493410 diff --git a/vlm/train/BkgrBgSYDS/18.png b/vlm/train/BkgrBgSYDS/18.png new file mode 100644 index 0000000000000000000000000000000000000000..6dfcd558a5812e08fc818fe57bd2fcad82fa7547 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b33ef11e3ab2733f8c901ae78e90c7de550227bfc297bc9e7f8deb46294b311d +size 405709 diff --git a/vlm/train/BkgrBgSYDS/19.png b/vlm/train/BkgrBgSYDS/19.png new file mode 100644 index 0000000000000000000000000000000000000000..acb1e63be39e6e130b678d861dc4b49244ff2cd8 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c77be7726612eb20bd26a190b4fb71996c541f75cd4b1bc2d7a2570d577213a +size 501603 diff --git a/vlm/train/BkgrBgSYDS/2.png b/vlm/train/BkgrBgSYDS/2.png new file mode 100644 index 0000000000000000000000000000000000000000..db8c25b03d9fdb0c395d73beacde43015f1e1311 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f58571603d7e7557b09a3729cd4fabd3d772201d1dc06999649fb023c8695412 +size 534991 diff --git a/vlm/train/BkgrBgSYDS/20.png b/vlm/train/BkgrBgSYDS/20.png new file mode 100644 index 0000000000000000000000000000000000000000..879b2ae523f5997bafa6dc35fe2e0be2bd00cd05 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9885fa2f0a55793c69a166c4c513cb3d530b06fb7b8dade2dc041199a3e9c924 +size 559335 diff --git a/vlm/train/BkgrBgSYDS/21.png b/vlm/train/BkgrBgSYDS/21.png new file mode 100644 index 0000000000000000000000000000000000000000..9173f6a821c7f4cb6481bc21b228a98b7eedbaed --- /dev/null +++ b/vlm/train/BkgrBgSYDS/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36b5e635c8de9eb34ecf5e190d324c488a6640184c1086f77a4dadbce9f7f54a +size 334197 diff --git a/vlm/train/BkgrBgSYDS/22.png b/vlm/train/BkgrBgSYDS/22.png new file mode 100644 index 0000000000000000000000000000000000000000..08d874afe0e161bb3116ee422a7d1cd18e43dea4 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e2cbf4cfab0a528e1cd380222151c708ece17b8321b2a3c83d663cd9d6373d2 +size 498376 diff --git a/vlm/train/BkgrBgSYDS/23.png b/vlm/train/BkgrBgSYDS/23.png new file mode 100644 index 0000000000000000000000000000000000000000..67f15a5129a7622c48da471c7cc8514d8833dc49 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b88937ac0c6a995f7937260c490889c83c1029e97ea361aefa27c600b014a510 +size 441558 diff --git a/vlm/train/BkgrBgSYDS/24.png b/vlm/train/BkgrBgSYDS/24.png new file mode 100644 index 0000000000000000000000000000000000000000..057bcc00e029cb44c1e455393b57da333ae3590d --- /dev/null +++ b/vlm/train/BkgrBgSYDS/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:529324d978aa614e10ccf531f73e53e99097504b38f82e0fe8e295befeb09745 +size 424514 diff --git a/vlm/train/BkgrBgSYDS/25.png b/vlm/train/BkgrBgSYDS/25.png new file mode 100644 index 0000000000000000000000000000000000000000..6e73bd14aba4aacb631b671092775206b0b36b72 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb6368d4849aa0808056fede42946f5df843232638dd4197b7033c0da5a35679 +size 437555 diff --git a/vlm/train/BkgrBgSYDS/26.png b/vlm/train/BkgrBgSYDS/26.png new file mode 100644 index 0000000000000000000000000000000000000000..af9e779a4c23ff8062867711f3a7aad2ef4371f0 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55252e0337ace12263b5e635331474ae7a6e5f4349fa4d92435cbb9643f5f68e +size 466126 diff --git a/vlm/train/BkgrBgSYDS/27.png b/vlm/train/BkgrBgSYDS/27.png new file mode 100644 index 0000000000000000000000000000000000000000..69ab8f63775b9a57c3581fa5cde654c5d51c0412 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8ce8d9225426eaf2bc18f86ebdeee42ad009e7ef2eaba659b0463ee3d488850 +size 390324 diff --git a/vlm/train/BkgrBgSYDS/28.png b/vlm/train/BkgrBgSYDS/28.png new file mode 100644 index 0000000000000000000000000000000000000000..88113de59422be8cdff4714f02846e34ecc191a4 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6edd3f656883809cd3ba971a12fab34f471866fedf2aacbeeb05479a0b7aab12 +size 382211 diff --git a/vlm/train/BkgrBgSYDS/29.png b/vlm/train/BkgrBgSYDS/29.png new file mode 100644 index 0000000000000000000000000000000000000000..36163264eb3c4e7ed3c6e9107f3eaab52bfc19da --- /dev/null +++ b/vlm/train/BkgrBgSYDS/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:637a7f356b034e731e9c7c9da9cb8ed574e7d0e9119a820260c11d577059377e +size 398528 diff --git a/vlm/train/BkgrBgSYDS/3.png b/vlm/train/BkgrBgSYDS/3.png new file mode 100644 index 0000000000000000000000000000000000000000..1ba727922220c08a881db4e625b5876cb67d2758 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4bd99a18806de772b8d0a550aa09371cfbc3f3915d045133d2f120394c4eb8ac +size 581467 diff --git a/vlm/train/BkgrBgSYDS/30.png b/vlm/train/BkgrBgSYDS/30.png new file mode 100644 index 0000000000000000000000000000000000000000..3644340ae623ac28d3df9d7842bdc3c58d5ad53a --- /dev/null +++ b/vlm/train/BkgrBgSYDS/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:747d02c0d8dd1d34f00006a5921460a0fc58c023f08bc55a6058d633ef502bd9 +size 368107 diff --git a/vlm/train/BkgrBgSYDS/31.png b/vlm/train/BkgrBgSYDS/31.png new file mode 100644 index 0000000000000000000000000000000000000000..af32e6d45dd6d60b9d1e0b051c3c416b1e83d54a --- /dev/null +++ b/vlm/train/BkgrBgSYDS/31.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80178d65db4887ee338342718e86072f7dadd88f87cc7f979e0b47a83c327b26 +size 288404 diff --git a/vlm/train/BkgrBgSYDS/32.png b/vlm/train/BkgrBgSYDS/32.png new file mode 100644 index 0000000000000000000000000000000000000000..72820ab19e746e5db9d06de232c19f4a17f671b8 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/32.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adcfdf9dbbdc4205743ae0a570e40e495069a51adfb5df8e0f47992470b756a4 +size 440994 diff --git a/vlm/train/BkgrBgSYDS/33.png b/vlm/train/BkgrBgSYDS/33.png new file mode 100644 index 0000000000000000000000000000000000000000..494c881910b8f788eb5b71ab194ce761d26deb43 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/33.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93d23fd3f58f88d8acca43f8e8e57e63991680d9cfd150a0d5c324c2e01463c7 +size 486699 diff --git a/vlm/train/BkgrBgSYDS/34.png b/vlm/train/BkgrBgSYDS/34.png new file mode 100644 index 0000000000000000000000000000000000000000..56115367ab30a95ac0241e42b53bc014d54e1d51 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/34.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b02fa28bf9b9c16d8fd472e035637ffdc835b17961595971b7c7016f954145c3 +size 517187 diff --git a/vlm/train/BkgrBgSYDS/35.png b/vlm/train/BkgrBgSYDS/35.png new file mode 100644 index 0000000000000000000000000000000000000000..2f9df80c32690f82c1b13e76a6ff516ae3afa617 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/35.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c34bd93027ae8a51ee4b966375065663781762d30303aea4cd8f78e3d4ac9ca7 +size 477627 diff --git a/vlm/train/BkgrBgSYDS/36.png b/vlm/train/BkgrBgSYDS/36.png new file mode 100644 index 0000000000000000000000000000000000000000..1e87c9114398fa74c566e00bbdd48a283d7b051b --- /dev/null +++ b/vlm/train/BkgrBgSYDS/36.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5306fa25506c088a84bfde320de2da8d6094237a992f862ce3fa7c6c9bd23d0f +size 371582 diff --git a/vlm/train/BkgrBgSYDS/37.png b/vlm/train/BkgrBgSYDS/37.png new file mode 100644 index 0000000000000000000000000000000000000000..f597e2c46d41e8f7dd8372f1b4ebea1e623601fa --- /dev/null +++ b/vlm/train/BkgrBgSYDS/37.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97f0962f3df4d25aaa32c40b71bc819ad63c8af0721704828b8c0c2d7416c397 +size 404895 diff --git a/vlm/train/BkgrBgSYDS/38.png b/vlm/train/BkgrBgSYDS/38.png new file mode 100644 index 0000000000000000000000000000000000000000..d825f3147334aff3de5cf4a81ff99e46a6de881e --- /dev/null +++ b/vlm/train/BkgrBgSYDS/38.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dee4a4fa9fc4b2fbf287bc496aa73c9761864b6520bc69e04b0958c6395400af +size 358333 diff --git a/vlm/train/BkgrBgSYDS/39.png b/vlm/train/BkgrBgSYDS/39.png new file mode 100644 index 0000000000000000000000000000000000000000..027deb863f1bd4b5c3077d7c99ad2622cebf4d1a --- /dev/null +++ b/vlm/train/BkgrBgSYDS/39.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd2229a81268ef7c798be01d1d9f518b1a64c4308a1ecd0d20b819849c999567 +size 362859 diff --git a/vlm/train/BkgrBgSYDS/4.png b/vlm/train/BkgrBgSYDS/4.png new file mode 100644 index 0000000000000000000000000000000000000000..62783ac075edfa6acf0aad615a574e8d040124e3 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:175a68c5442121c2e539f09919d456c05ed86943a5b94ff2582dbf146032135c +size 542542 diff --git a/vlm/train/BkgrBgSYDS/40.png b/vlm/train/BkgrBgSYDS/40.png new file mode 100644 index 0000000000000000000000000000000000000000..96672cda00df3aed53031aa337b6409a2d71faf9 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/40.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1630746c05509346e6f18772258102baaf42023bf73fd5deebc3a4937e571bc +size 360628 diff --git a/vlm/train/BkgrBgSYDS/41.png b/vlm/train/BkgrBgSYDS/41.png new file mode 100644 index 0000000000000000000000000000000000000000..b81b18bdd7ae3048f26decc2841bc8f4d665017b --- /dev/null +++ b/vlm/train/BkgrBgSYDS/41.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5664046e16cc0c0de30df9975ba7f7e8bf33aa271b7ac475ebb6bcdbafbe694 +size 553240 diff --git a/vlm/train/BkgrBgSYDS/42.png b/vlm/train/BkgrBgSYDS/42.png new file mode 100644 index 0000000000000000000000000000000000000000..10410300e1e27a9d2781acb93a9fbb73ed5f3c79 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/42.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fc3412dced9fe33b1b5c48540a22936ffeaafd01e3373849b0a48fddc6be0a3 +size 565616 diff --git a/vlm/train/BkgrBgSYDS/43.png b/vlm/train/BkgrBgSYDS/43.png new file mode 100644 index 0000000000000000000000000000000000000000..08a0fc18defd4a65b8e3480c4cf708c8a0fb2cc9 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/43.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:40872e9e61c483a7fc8b34a82ea774100877ba2e8f95ebf243baac3cf6cf95b2 +size 348925 diff --git a/vlm/train/BkgrBgSYDS/44.png b/vlm/train/BkgrBgSYDS/44.png new file mode 100644 index 0000000000000000000000000000000000000000..70d84e671672d1f98622ab05438d6333c2d97c74 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/44.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96fc0633ada8a91f9718ebaf497d33757af60192b460bc21ce455d6b42e258c9 +size 326299 diff --git a/vlm/train/BkgrBgSYDS/5.png b/vlm/train/BkgrBgSYDS/5.png new file mode 100644 index 0000000000000000000000000000000000000000..b17fd0ff73341c7f5d414f2494ecfab1108b678c --- /dev/null +++ b/vlm/train/BkgrBgSYDS/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c908907c39ad9094271bac11ecdd2482c82305192e29a7c6bf4f51f75e3309b2 +size 497937 diff --git a/vlm/train/BkgrBgSYDS/6.png b/vlm/train/BkgrBgSYDS/6.png new file mode 100644 index 0000000000000000000000000000000000000000..a001f880a33b57a40815426e45f9b8621aa9c65b --- /dev/null +++ b/vlm/train/BkgrBgSYDS/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1afe1610ea70edc856f51c825db2dd217b335b16cf8a60200771e067be7e327a +size 584095 diff --git a/vlm/train/BkgrBgSYDS/7.png b/vlm/train/BkgrBgSYDS/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f0d7539db33f2de55a6a4764b3be6cfbf6455156 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6e35f300fa9171ed0ef8d38ec3ecb8050f2058761cf8929d236b21adbf542cf +size 629867 diff --git a/vlm/train/BkgrBgSYDS/8.png b/vlm/train/BkgrBgSYDS/8.png new file mode 100644 index 0000000000000000000000000000000000000000..4ddc5496ce54c39f9ecd325f6ef153b6e133c3cc --- /dev/null +++ b/vlm/train/BkgrBgSYDS/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a03b5cb08dfedc9b1f52095258fde3f6f4a383ad071d1e9c7daad7fdeb3b3569 +size 546493 diff --git a/vlm/train/BkgrBgSYDS/9.png b/vlm/train/BkgrBgSYDS/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e8e16c57758a0e4743e17d71a2fa401326ef3814 --- /dev/null +++ b/vlm/train/BkgrBgSYDS/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:237e4d1e0e8d6c8d467b21c7a08a6898ccc6d6a881796735bb673deda3866314 +size 589171 diff --git a/vlm/train/BkrSv0lA-/0.png b/vlm/train/BkrSv0lA-/0.png new file mode 100644 index 0000000000000000000000000000000000000000..8f1364171c48b73aad724c910d9b6fb1b4e1d520 --- /dev/null +++ b/vlm/train/BkrSv0lA-/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b5f4e45ccede49c78ac5af158d25ff78673866b218da03aee16cd9288c76978 +size 519106 diff --git a/vlm/train/BkrSv0lA-/1.png b/vlm/train/BkrSv0lA-/1.png new file mode 100644 index 0000000000000000000000000000000000000000..5bd29a42a5f6561352d2855f4b09d9de0bc74e00 --- /dev/null +++ b/vlm/train/BkrSv0lA-/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f15533cc8f1821186f6faaa3402a7c49a310bba1e62c23b1b83da7569b6c168e +size 537473 diff --git a/vlm/train/BkrSv0lA-/10.png b/vlm/train/BkrSv0lA-/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a36947a1e27c2cf59ded30595edbc67ea43764f1 --- /dev/null +++ b/vlm/train/BkrSv0lA-/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb13455bbb9bdbc074d96804d0555f1c0b3e38b23260e2f16dfe015015870828 +size 304186 diff --git a/vlm/train/BkrSv0lA-/11.png b/vlm/train/BkrSv0lA-/11.png new file mode 100644 index 0000000000000000000000000000000000000000..9ddcc165d448e21df2152cac6c23484f79ce38d4 --- /dev/null +++ b/vlm/train/BkrSv0lA-/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49d5805759172cf9917ba53653068c390004fb506421097ab1cdfdfb6db44d52 +size 244991 diff --git a/vlm/train/BkrSv0lA-/12.png b/vlm/train/BkrSv0lA-/12.png new file mode 100644 index 0000000000000000000000000000000000000000..d15b6e296dd3ac83af29b19acd7d044c4da3b297 --- /dev/null +++ b/vlm/train/BkrSv0lA-/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b631cb0124e243b3d0d4b3972559f3430ce3d724e2226d7e1fc176eecc9cb31f +size 351229 diff --git a/vlm/train/BkrSv0lA-/13.png b/vlm/train/BkrSv0lA-/13.png new file mode 100644 index 0000000000000000000000000000000000000000..516c0fd49b2a95b23c55e01f4dbb9d32011b7b9c --- /dev/null +++ b/vlm/train/BkrSv0lA-/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3985a68ea7a21f9e87c123413792fe54d7b4a2a9a9745f027db5b6aaca0fb8eb +size 336589 diff --git a/vlm/train/BkrSv0lA-/14.png b/vlm/train/BkrSv0lA-/14.png new file mode 100644 index 0000000000000000000000000000000000000000..4108a58a0622768c801d0c766e071e634fa43384 --- /dev/null +++ b/vlm/train/BkrSv0lA-/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2770eb8e287618a2eefacb8112d0e38360ec1a0fdf2ed69edf3d3d47534b1437 +size 432317 diff --git a/vlm/train/BkrSv0lA-/15.png b/vlm/train/BkrSv0lA-/15.png new file mode 100644 index 0000000000000000000000000000000000000000..f121f1ebd9db42009acc9feedf41fa3772e8f65c --- /dev/null +++ b/vlm/train/BkrSv0lA-/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9870d3f9ef81a0bfa0c6b6f38095273a7ec9668cde83e44d8c1d053efdc7d216 +size 453333 diff --git a/vlm/train/BkrSv0lA-/2.png b/vlm/train/BkrSv0lA-/2.png new file mode 100644 index 0000000000000000000000000000000000000000..8deed2b16360b281f1c98a9fd6b9e67081d1cd44 --- /dev/null +++ b/vlm/train/BkrSv0lA-/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb1e167387506343922fa6719251f311cdd42d03f24eb6599bad9620237a950f +size 481001 diff --git a/vlm/train/BkrSv0lA-/3.png b/vlm/train/BkrSv0lA-/3.png new file mode 100644 index 0000000000000000000000000000000000000000..5739466820f815274d807ffdb5b4247855953ed5 --- /dev/null +++ b/vlm/train/BkrSv0lA-/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61b101b85aad8a9bd2f4ac39e0b35684b875b697bc79ba915d23a2d3376dfcd0 +size 568380 diff --git a/vlm/train/BkrSv0lA-/4.png b/vlm/train/BkrSv0lA-/4.png new file mode 100644 index 0000000000000000000000000000000000000000..e5a7b782125c24d8926d268039faae0092c921b7 --- /dev/null +++ b/vlm/train/BkrSv0lA-/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2cad8b1a11b96eaa86d3d667984d8ef9c547e44edc0142eff319684d5c14265 +size 399514 diff --git a/vlm/train/BkrSv0lA-/5.png b/vlm/train/BkrSv0lA-/5.png new file mode 100644 index 0000000000000000000000000000000000000000..f0b0308d75acc5fa9a36a42687d5e97af62750b0 --- /dev/null +++ b/vlm/train/BkrSv0lA-/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4d71f9ae8514fcc9348308791fd697bddeebb3edbd202f7c084350998b3d174 +size 530669 diff --git a/vlm/train/BkrSv0lA-/6.png b/vlm/train/BkrSv0lA-/6.png new file mode 100644 index 0000000000000000000000000000000000000000..8d448386e498e5568de4a21d0ad90164fa964a95 --- /dev/null +++ b/vlm/train/BkrSv0lA-/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:630a7810c8398005d07afa10b0ebd1d04a4687f9ea7e9c4139e116e9d21761d0 +size 458242 diff --git a/vlm/train/BkrSv0lA-/7.png b/vlm/train/BkrSv0lA-/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e4228ff2aee0c56ca54e3906d2f0286da0bde76b --- /dev/null +++ b/vlm/train/BkrSv0lA-/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c134fe6fbe6b31b7a72ea16c7a5958f6b88f70761b2a87a6579c75fafc2faf75 +size 532932 diff --git a/vlm/train/BkrSv0lA-/8.png b/vlm/train/BkrSv0lA-/8.png new file mode 100644 index 0000000000000000000000000000000000000000..7e96cf3008b61d8edd8321e6c1d59a6bd14af99a --- /dev/null +++ b/vlm/train/BkrSv0lA-/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d91c98454dd7f4d861c3074034c7552b0b89c1f38af48dfd8400ecc7a32bf116 +size 447865 diff --git a/vlm/train/BkrSv0lA-/9.png b/vlm/train/BkrSv0lA-/9.png new file mode 100644 index 0000000000000000000000000000000000000000..5ac8912a92b4093c5292da39825833da1bcf165b --- /dev/null +++ b/vlm/train/BkrSv0lA-/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04f7cf186891c528ae1311d5aa39ef580e5560565fbad781156483a623028feb +size 509036 diff --git a/vlm/train/Bkx0RjA9tX/0.png b/vlm/train/Bkx0RjA9tX/0.png new file mode 100644 index 0000000000000000000000000000000000000000..dd59b8ec66cf2f950798ebcef9685863d6b4cadb --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:208c1f22de427c5b60c35d02a5e9742fd244e5d250a45dffe952f48281e86f43 +size 471694 diff --git a/vlm/train/Bkx0RjA9tX/1.png b/vlm/train/Bkx0RjA9tX/1.png new file mode 100644 index 0000000000000000000000000000000000000000..fceb6044b483f1233e9e3204387437739fe419ce --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3beee52f969b65a496cbdc9ca736c729c6cb3db65bed0cebf3065c3013b44a55 +size 587145 diff --git a/vlm/train/Bkx0RjA9tX/10.png b/vlm/train/Bkx0RjA9tX/10.png new file mode 100644 index 0000000000000000000000000000000000000000..4058289a52870f0cfdbb9284ad1efb53b81f006e --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c22cf9d5200c7d8b845dcca0eb52508585ac837f6f24439d25a2d919989cbec +size 629446 diff --git a/vlm/train/Bkx0RjA9tX/11.png b/vlm/train/Bkx0RjA9tX/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a187f7eb129264802aa82f9ee86597fdc0c45254 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b97ea568368006e621eb76833899f22a43488b56045860312f4e10c40f81b0c7 +size 654267 diff --git a/vlm/train/Bkx0RjA9tX/12.png b/vlm/train/Bkx0RjA9tX/12.png new file mode 100644 index 0000000000000000000000000000000000000000..9450bdb371ec92f3d346bd94ecc2b98256dc7dc7 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ece4c9bc2bedf664983bd70d78ad4f330aca0d9b8658ff391330bfcea57183a5 +size 612086 diff --git a/vlm/train/Bkx0RjA9tX/13.png b/vlm/train/Bkx0RjA9tX/13.png new file mode 100644 index 0000000000000000000000000000000000000000..6ff60c634d2b5d1ea4d423ae0fe277731f7ea1de --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:331e73da9c8d6440a9602595a788b5fff8448dbb18a899a16cc1533640da3518 +size 381614 diff --git a/vlm/train/Bkx0RjA9tX/2.png b/vlm/train/Bkx0RjA9tX/2.png new file mode 100644 index 0000000000000000000000000000000000000000..6e0b29faa298b2e55becff6e70fa51bdf28443bf --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99da90fb47f6e146b835ae44a5b8cec26da1e8cf7842a644465eac33d0d1d9ec +size 574422 diff --git a/vlm/train/Bkx0RjA9tX/3.png b/vlm/train/Bkx0RjA9tX/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b12c0d0d6dc63d9fa381f3e599a513e5e873dd52 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c7ddf3fabd492a2eddf2aa46fa64c562766e6bdf95764d20d045b335f7e2bea +size 404405 diff --git a/vlm/train/Bkx0RjA9tX/4.png b/vlm/train/Bkx0RjA9tX/4.png new file mode 100644 index 0000000000000000000000000000000000000000..312e0f7090f38845a8f5910430cedd6e6450fbc6 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c7f435bd2f7df2f12e07ddc769b7bf446d599bc33a1f21aa844e89e7e1e154c +size 635450 diff --git a/vlm/train/Bkx0RjA9tX/5.png b/vlm/train/Bkx0RjA9tX/5.png new file mode 100644 index 0000000000000000000000000000000000000000..35702234656a7de9fd798bbec5fa307a312dcfd0 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:075552ce0dba95724517ee026267823175edf54f1aef533b531dbf32bba801e3 +size 451066 diff --git a/vlm/train/Bkx0RjA9tX/6.png b/vlm/train/Bkx0RjA9tX/6.png new file mode 100644 index 0000000000000000000000000000000000000000..30eb1aa61b4323a7a7a93c079e22568385a59981 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48d82d5667ec6dacea2490ff0a66b99bf642174d00c282b7d793ff31336588a4 +size 446884 diff --git a/vlm/train/Bkx0RjA9tX/7.png b/vlm/train/Bkx0RjA9tX/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c11b8fc0dd462b62349f515e8273314dbaa2b17c --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b10f4ad295fec6f9a2d8281bf84126b3d54fd29b47ad0adfc0a4f398501c50e1 +size 493311 diff --git a/vlm/train/Bkx0RjA9tX/8.png b/vlm/train/Bkx0RjA9tX/8.png new file mode 100644 index 0000000000000000000000000000000000000000..daf357875cb0807e26baf4508748a27000de83a0 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37c5ba35de6a07bbfc9f288593a461c50f51cc1f0624511a62c778105433c169 +size 498519 diff --git a/vlm/train/Bkx0RjA9tX/9.png b/vlm/train/Bkx0RjA9tX/9.png new file mode 100644 index 0000000000000000000000000000000000000000..8550b9d53b72e38377cc34d480586d71d99ecb85 --- /dev/null +++ b/vlm/train/Bkx0RjA9tX/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3abcdb92eb6ee1c2e4e422d7b66829208a0c31c82496e25b6a783e6fb9b01ace +size 605648 diff --git a/vlm/train/ByMHvs0cFQ/0.png b/vlm/train/ByMHvs0cFQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..cd8230e13355e158ff83d697af2b6a85713bce37 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05779e8d55e39a21a0cba4cac237d748746684f56bd8607ef892af5931fe9f9b +size 470994 diff --git a/vlm/train/ByMHvs0cFQ/1.png b/vlm/train/ByMHvs0cFQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..599d23846c609a695f04b16d608e8d2d2d9b1bd1 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59fe16732cb7396245b558d0f15df6270c4b8cc4b11724c7546459908737cd30 +size 623346 diff --git a/vlm/train/ByMHvs0cFQ/10.png b/vlm/train/ByMHvs0cFQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..ba5867ad2bdf13c23f7d8ce06b5f1719d9d63f61 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18a346d22f92f0e153e6511e0aec52d2e5bfe404d929f1728e9c1476a8c02c52 +size 504702 diff --git a/vlm/train/ByMHvs0cFQ/11.png b/vlm/train/ByMHvs0cFQ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a0bf7c01bcb0c51f0d5b76fbd486a5878314767c --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e2ddb1fa5298edf188a0f3008e4715b827461745bc627fa7f095dca1708a9a3 +size 501614 diff --git a/vlm/train/ByMHvs0cFQ/12.png b/vlm/train/ByMHvs0cFQ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..cf0a438a4f7658bd4971698b6156244c547c369e --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96d98ddea2e0c3659604a95cb82d7a08a667eb39d8ea91366f1a7cf876efb01d +size 373960 diff --git a/vlm/train/ByMHvs0cFQ/13.png b/vlm/train/ByMHvs0cFQ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..112941c151053f76bcc1ebef80ff605fe7bd4865 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1479d617d26e5d4d7549186832bfb2770e6800e896581ab709ec137499e22f56 +size 237585 diff --git a/vlm/train/ByMHvs0cFQ/14.png b/vlm/train/ByMHvs0cFQ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..22411de31400f6561ae5e08305bbf4c0b27c5d68 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc2b2e471d8c3c262cc2d04324c548250827a6c6bb99f60f7331d42cca1029b4 +size 330760 diff --git a/vlm/train/ByMHvs0cFQ/15.png b/vlm/train/ByMHvs0cFQ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..bd7479718850bc95ecce74d16c6a715a4735538d --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9854e018cb582923016046a960df6758a71bddf0e2f36ee0b0b392884367d7ba +size 307590 diff --git a/vlm/train/ByMHvs0cFQ/16.png b/vlm/train/ByMHvs0cFQ/16.png new file mode 100644 index 0000000000000000000000000000000000000000..7dfa84ab186b06c89052ab1dccae261b399b64d4 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75cbc3214ab1e967068d23f954bde5bd210b91a639952c22a6f1f5cb02a708ef +size 306331 diff --git a/vlm/train/ByMHvs0cFQ/17.png b/vlm/train/ByMHvs0cFQ/17.png new file mode 100644 index 0000000000000000000000000000000000000000..637c9480cb4f9fab3a3e31c7b7d3d55f560d4c81 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35206516697073a908a2a20712e342792ce5408ffa03692458e43617e848c5c7 +size 244260 diff --git a/vlm/train/ByMHvs0cFQ/18.png b/vlm/train/ByMHvs0cFQ/18.png new file mode 100644 index 0000000000000000000000000000000000000000..810439e989d3132c315661c1e46c8b494226a4ab --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82e5d42197b91398077c2ab889c2adebccc7ab634e3b4bc9a1b21e48961f7393 +size 75887 diff --git a/vlm/train/ByMHvs0cFQ/2.png b/vlm/train/ByMHvs0cFQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..23f409bfeede3222e3a87c907a03841dd4ffb7e5 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b06365648f303f9ec6f772036de928bef1de12a0ce940d64f7a49ca381059842 +size 584223 diff --git a/vlm/train/ByMHvs0cFQ/3.png b/vlm/train/ByMHvs0cFQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..12e0e64780ea901f48cbee4587e3abc2ffdedfdd --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d3b177d85cc9fb94ad50d1105c96649978ebd31758713532a56130eaea5f270 +size 483758 diff --git a/vlm/train/ByMHvs0cFQ/4.png b/vlm/train/ByMHvs0cFQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..7a1562a1e7a5ce00379a9f72d83015c5df651e20 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b71d02af092efc95ad9be3133241e0589eaf3593c9817bc50c47e15c6433ea3 +size 497555 diff --git a/vlm/train/ByMHvs0cFQ/5.png b/vlm/train/ByMHvs0cFQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..c2dd8b6860a46d49b322913bcb1353e7ffd27f6c --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9019b4c0e089f36edd956f4b896807cc80ebfab1c621f13ccc7f52860ddda50 +size 431029 diff --git a/vlm/train/ByMHvs0cFQ/6.png b/vlm/train/ByMHvs0cFQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..e9ba942be7c543d5fc3376717dcd8fe03daf78b5 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdebc2e5c34619ac986e45463561e54159ab40bffdcb2fa7a5445949e993dafb +size 585594 diff --git a/vlm/train/ByMHvs0cFQ/7.png b/vlm/train/ByMHvs0cFQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..a59b75a212d66c38b3137e23592f0de3dcbff192 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37fc6041c6263a1136a649121cc37e8c05b0d4856b158ab48d27dfc12988a1c1 +size 486193 diff --git a/vlm/train/ByMHvs0cFQ/8.png b/vlm/train/ByMHvs0cFQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..1a9c4e9b91a8757b505a7be69be72b344323af0c --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e8d44a7f6d24e50eac9c2e166e355a216cab43952debe39487f87f8b6b06dc1 +size 563395 diff --git a/vlm/train/ByMHvs0cFQ/9.png b/vlm/train/ByMHvs0cFQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..989e25035e6cfe35b020dfd6bcaa1dab49974b37 --- /dev/null +++ b/vlm/train/ByMHvs0cFQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fe4c5be2261365807af711470b5fced465a247aa338628d33bd3c8a7938b23d +size 621566 diff --git a/vlm/train/BydjJte0-/0.png b/vlm/train/BydjJte0-/0.png new file mode 100644 index 0000000000000000000000000000000000000000..e26d2babd7024914f9cd6b65e142b98b54c25d28 --- /dev/null +++ b/vlm/train/BydjJte0-/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:394cf55bd8c4297eddf4e0a9d954fbb34e0cb84474b85ff69df85fc6951ebd23 +size 514294 diff --git a/vlm/train/BydjJte0-/1.png b/vlm/train/BydjJte0-/1.png new file mode 100644 index 0000000000000000000000000000000000000000..cf6ca5ec7c9f9a81c7129a455fd9ad0827e04e8a --- /dev/null +++ b/vlm/train/BydjJte0-/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e2210be838b80759fe2d900e1a212cb8747c7eba04a4b80b005ef9abf304732 +size 570258 diff --git a/vlm/train/BydjJte0-/10.png b/vlm/train/BydjJte0-/10.png new file mode 100644 index 0000000000000000000000000000000000000000..5a189873b13daff84835776c62be92a2c224477e --- /dev/null +++ b/vlm/train/BydjJte0-/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:792fea67b47afd77b411b4acc9f3d36ded010ecd2e8dadb2872ab21350e339b4 +size 433099 diff --git a/vlm/train/BydjJte0-/11.png b/vlm/train/BydjJte0-/11.png new file mode 100644 index 0000000000000000000000000000000000000000..c396974d7ed7ce361f996f1a4457b007d852c358 --- /dev/null +++ b/vlm/train/BydjJte0-/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:01dc03b28be0a614daf9270a15f616575a48ab163e69bb1181b31081c1e5a086 +size 236545 diff --git a/vlm/train/BydjJte0-/12.png b/vlm/train/BydjJte0-/12.png new file mode 100644 index 0000000000000000000000000000000000000000..4931b3b8a638e6f0433f96fa66d2a1f7907167af --- /dev/null +++ b/vlm/train/BydjJte0-/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0b480503886b6cbc35e37b50e29e5c98011b96023f8880ba58ea6595e743f3e +size 423728 diff --git a/vlm/train/BydjJte0-/13.png b/vlm/train/BydjJte0-/13.png new file mode 100644 index 0000000000000000000000000000000000000000..89a9afb568411a6452fc2ced25caca2c72ec5947 --- /dev/null +++ b/vlm/train/BydjJte0-/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fcd89f41153a9f6f564fd18801093d2f82cb544c88fe65eb7a7701c47920d65f +size 376161 diff --git a/vlm/train/BydjJte0-/14.png b/vlm/train/BydjJte0-/14.png new file mode 100644 index 0000000000000000000000000000000000000000..cc89877536b291fd777d79c048fcd6903fce3617 --- /dev/null +++ b/vlm/train/BydjJte0-/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85bbde8b27b4a3e16b4673ce0bd34f6d8600cf1e33ea6a949146f5d2cc984f10 +size 473256 diff --git a/vlm/train/BydjJte0-/15.png b/vlm/train/BydjJte0-/15.png new file mode 100644 index 0000000000000000000000000000000000000000..5fec4efed48b0ac737f43ef1128cf7a766d2b63b --- /dev/null +++ b/vlm/train/BydjJte0-/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8af5bb60eb94e4226fcea98efeaba744ab036b3c24024d9ebdd951fcc8722fb6 +size 2452570 diff --git a/vlm/train/BydjJte0-/16.png b/vlm/train/BydjJte0-/16.png new file mode 100644 index 0000000000000000000000000000000000000000..7a92b794374ff3cda5e51ee00608022ae7a1c063 --- /dev/null +++ b/vlm/train/BydjJte0-/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae7e7fa7460c9f328292a1c63531e8be4f43b048f6a9634101b6175179f8c0b2 +size 509233 diff --git a/vlm/train/BydjJte0-/17.png b/vlm/train/BydjJte0-/17.png new file mode 100644 index 0000000000000000000000000000000000000000..de834a0d321a6841ed71ae1653883dcddaf3eaa1 --- /dev/null +++ b/vlm/train/BydjJte0-/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5513a2a46267e9ad0c61d58fd31f944dea56ee1279ce15a66210d727a6abdb07 +size 578169 diff --git a/vlm/train/BydjJte0-/18.png b/vlm/train/BydjJte0-/18.png new file mode 100644 index 0000000000000000000000000000000000000000..75cc325103f746992422808db68568ef8597e955 --- /dev/null +++ b/vlm/train/BydjJte0-/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c075105282d665f8fbddfc5cd592c1f9b9ce08a0c7687fcfdc24e915a6fe60c3 +size 115847 diff --git a/vlm/train/BydjJte0-/19.png b/vlm/train/BydjJte0-/19.png new file mode 100644 index 0000000000000000000000000000000000000000..ce8516825c24e9abf647671aaf8ef0fb706e9287 --- /dev/null +++ b/vlm/train/BydjJte0-/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b1599a22a91af71ba7c34caab7ab8c9517926c474219dfb4f954431fd468ef73 +size 115936 diff --git a/vlm/train/BydjJte0-/2.png b/vlm/train/BydjJte0-/2.png new file mode 100644 index 0000000000000000000000000000000000000000..a4c1bbe2dcba6fc5b795e223596d23a85ebb41b1 --- /dev/null +++ b/vlm/train/BydjJte0-/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da5bed77b770faaff165b55f6f1100ef3af6a7accfe00b0aae0f2a78cf1d8f65 +size 542037 diff --git a/vlm/train/BydjJte0-/3.png b/vlm/train/BydjJte0-/3.png new file mode 100644 index 0000000000000000000000000000000000000000..fb9869e2f8879eff5776133afae8aea0dae5e2b2 --- /dev/null +++ b/vlm/train/BydjJte0-/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:211cf4984ddffec7f8ac0e6f49953dca2c118fdc34a4264e788c3bbbcf792d68 +size 510568 diff --git a/vlm/train/BydjJte0-/4.png b/vlm/train/BydjJte0-/4.png new file mode 100644 index 0000000000000000000000000000000000000000..c717fbccc4409f2bd4b3efaf1207e1f3e7cd44b0 --- /dev/null +++ b/vlm/train/BydjJte0-/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dd06aa4362a33e6fb901d86909daea7e76209793f4b9ecf0092a8e9b4f048f9 +size 543837 diff --git a/vlm/train/BydjJte0-/5.png b/vlm/train/BydjJte0-/5.png new file mode 100644 index 0000000000000000000000000000000000000000..3cd9cf2020d95ee4dfbcc54e3e94d1ffac9a9ce2 --- /dev/null +++ b/vlm/train/BydjJte0-/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77f8bf1a359bf10a91d3e70ddb0bd86559b86892c87ce24c7d08a757b3395a76 +size 509455 diff --git a/vlm/train/BydjJte0-/6.png b/vlm/train/BydjJte0-/6.png new file mode 100644 index 0000000000000000000000000000000000000000..1bfe533672e378236aeed90bbe81c02bec718e61 --- /dev/null +++ b/vlm/train/BydjJte0-/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7edfa57583aaa58245cab085a905dd6abdf4484e7f611df958c1f0f4a9a846a5 +size 560597 diff --git a/vlm/train/BydjJte0-/7.png b/vlm/train/BydjJte0-/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c0b4fbf24ff6d12563608b001689e34f3f1de0ad --- /dev/null +++ b/vlm/train/BydjJte0-/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93be778abd34b8e000172c7a0964baa61c4659452c138315dd900352581f25c6 +size 534283 diff --git a/vlm/train/BydjJte0-/8.png b/vlm/train/BydjJte0-/8.png new file mode 100644 index 0000000000000000000000000000000000000000..807e4282763429960498b83b4ab4594d685358ba --- /dev/null +++ b/vlm/train/BydjJte0-/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:879afebef4a320eb16bc79a4b6b564eef55566241f2f7fa6503d71126b94955a +size 504335 diff --git a/vlm/train/BydjJte0-/9.png b/vlm/train/BydjJte0-/9.png new file mode 100644 index 0000000000000000000000000000000000000000..97f9402a2b91dbd7c979089fc4fde64061dc0277 --- /dev/null +++ b/vlm/train/BydjJte0-/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2a2418865430f8cbbfff8693b60e8b6109d14903a39417ae38001cc81434085 +size 587138 diff --git a/vlm/train/Byey7n05FQ/10.png b/vlm/train/Byey7n05FQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a9ce64fa69bd846f2ef9f5e2495cd364b75e8659 --- /dev/null +++ b/vlm/train/Byey7n05FQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5830aeaebbd3a9116a6aec15eb7a86c3e87f84bdf82b148b73354c7ed80885b7 +size 521577 diff --git a/vlm/train/Byey7n05FQ/14.png b/vlm/train/Byey7n05FQ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..1edbc7501cf8f1a39d2f459e5d62006528c8c3eb --- /dev/null +++ b/vlm/train/Byey7n05FQ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c1e1f4a13522f7c294941b1bde9b9ac51274ad4e8b6242a6541b2f1e6e2ef20 +size 318279 diff --git a/vlm/train/Byey7n05FQ/2.png b/vlm/train/Byey7n05FQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..49ee97eaf4d644df24610ed46d4884d297e7ecda --- /dev/null +++ b/vlm/train/Byey7n05FQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73c2ba982cf45f4a34c41d52356d5aef3afc555d41a5dd8013cdcea8e07a01c3 +size 522102 diff --git a/vlm/train/Byey7n05FQ/3.png b/vlm/train/Byey7n05FQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b7eb48264b8f7810680f00436da93f077e281e4e --- /dev/null +++ b/vlm/train/Byey7n05FQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e13064a4624669a9e1b3b9e7460ad8a2abca583ad06dad5260e26a620c887546 +size 530933 diff --git a/vlm/train/Byey7n05FQ/6.png b/vlm/train/Byey7n05FQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..17581edb7a76f48715d2041e7089f5aa3e77e5a6 --- /dev/null +++ b/vlm/train/Byey7n05FQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7d9ec4b36355498875ef064f83ec4a50f3b260497ff541951d967071dc8816a +size 615738 diff --git a/vlm/train/Byey7n05FQ/9.png b/vlm/train/Byey7n05FQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..edff04f8c2c244da4750c74ccbd4bb62461585b5 --- /dev/null +++ b/vlm/train/Byey7n05FQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd9f83d7fe5885170e697f77955e43f1ea5cae1838fd0bcde5c2937d51770fc4 +size 495237 diff --git a/vlm/train/D2TE6VTJG9/0.png b/vlm/train/D2TE6VTJG9/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6dc897e034a478768c48786c8c908fecdb431b6e --- /dev/null +++ b/vlm/train/D2TE6VTJG9/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92e3285a9639f33fcbf7b4196f6fa1a40ae86e27b3024ac925bf583ab7816fec +size 496120 diff --git a/vlm/train/D2TE6VTJG9/1.png b/vlm/train/D2TE6VTJG9/1.png new file mode 100644 index 0000000000000000000000000000000000000000..c92fcfe3562a9d3f498e4d6400aa7f85369f90e8 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ff49de804444baabc919315a3c80b5e87d4c83389ea140c7908018c63506a0f +size 604575 diff --git a/vlm/train/D2TE6VTJG9/10.png b/vlm/train/D2TE6VTJG9/10.png new file mode 100644 index 0000000000000000000000000000000000000000..7021f3770993840326c5c33a47e2fda451902bd3 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:888e05a09f6207a7326c6b5ac6626dc7da46b2b4664783dbd1f2c427782e0148 +size 549513 diff --git a/vlm/train/D2TE6VTJG9/11.png b/vlm/train/D2TE6VTJG9/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ae9c60dc9acbea2bf739652239eb5e569c735f2f --- /dev/null +++ b/vlm/train/D2TE6VTJG9/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f97a5b30c7d6a52ef7388dd167c495ed4f4e0322535765e64920008704366d01 +size 91235 diff --git a/vlm/train/D2TE6VTJG9/12.png b/vlm/train/D2TE6VTJG9/12.png new file mode 100644 index 0000000000000000000000000000000000000000..172a4bdc12538999b407b8c7e9e2cec7b007b958 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:220e7555f10d9545e21c6219ce2a033779bb27f948e35d7968f929925a14fd81 +size 393441 diff --git a/vlm/train/D2TE6VTJG9/13.png b/vlm/train/D2TE6VTJG9/13.png new file mode 100644 index 0000000000000000000000000000000000000000..e1275db545a810fc3b8b8652242dd2d0ba4142e1 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37e7896fc0d1d26dc39d9963dc6d388b478138535c3e1a4f6fe3233fba3a2b05 +size 311855 diff --git a/vlm/train/D2TE6VTJG9/14.png b/vlm/train/D2TE6VTJG9/14.png new file mode 100644 index 0000000000000000000000000000000000000000..408a572cf73417e4eeeb8efec1be9911d7cde76c --- /dev/null +++ b/vlm/train/D2TE6VTJG9/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1434ca451044f8ad56076fd4f6441858bcb55171a14b168f2ee93889bd26f762 +size 317859 diff --git a/vlm/train/D2TE6VTJG9/15.png b/vlm/train/D2TE6VTJG9/15.png new file mode 100644 index 0000000000000000000000000000000000000000..5ff1e22bc1d9a46aa1ef5ea99304fef5c9d1f76a --- /dev/null +++ b/vlm/train/D2TE6VTJG9/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0dee01f344328f31065ca6a60f7263f2bea28a59e323af5b1d6871cd47747191 +size 328687 diff --git a/vlm/train/D2TE6VTJG9/16.png b/vlm/train/D2TE6VTJG9/16.png new file mode 100644 index 0000000000000000000000000000000000000000..f2b2589b7b7d95ec2b88c7faa71f964ce42ac647 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a3afe3c97fd915e5a3e28da57c6d0b21d4c62268bf8ffea897d435f7b236e51 +size 314289 diff --git a/vlm/train/D2TE6VTJG9/17.png b/vlm/train/D2TE6VTJG9/17.png new file mode 100644 index 0000000000000000000000000000000000000000..4754f0df8ef02a402c03774b23ca8c9c4cc01c97 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:43e8c4ebb4e70e52715c337ac26cbf63bb4509d966ee15fa36ce9b3e79481463 +size 362241 diff --git a/vlm/train/D2TE6VTJG9/18.png b/vlm/train/D2TE6VTJG9/18.png new file mode 100644 index 0000000000000000000000000000000000000000..7ff18a313165bac527fcb8d1f974245dc926a7cf --- /dev/null +++ b/vlm/train/D2TE6VTJG9/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9163861943e11181df3d32ff7bf46f2a100beaf34fde5481a872f86e39f71b6c +size 422309 diff --git a/vlm/train/D2TE6VTJG9/19.png b/vlm/train/D2TE6VTJG9/19.png new file mode 100644 index 0000000000000000000000000000000000000000..f084843a18b61ee168f73695f542800728da7f21 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd628cec256b73ee6513e1166df21a3bebe2117edec71299280d377c47d807b1 +size 367543 diff --git a/vlm/train/D2TE6VTJG9/2.png b/vlm/train/D2TE6VTJG9/2.png new file mode 100644 index 0000000000000000000000000000000000000000..ab9e4d17c4156a3da3ab2eb44b879db6de34a29e --- /dev/null +++ b/vlm/train/D2TE6VTJG9/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f07066e1a3e922a26266d0a023bea74bd7a6fa438e04918f31ee98f459d8627 +size 526460 diff --git a/vlm/train/D2TE6VTJG9/20.png b/vlm/train/D2TE6VTJG9/20.png new file mode 100644 index 0000000000000000000000000000000000000000..2f6ecfefa600ae9c9d77f8a446d71349fe7522bb --- /dev/null +++ b/vlm/train/D2TE6VTJG9/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d4c5527f5db992d2f2da31c3820a91b0b1abb45d1a73a952c143af3d7e35cf3 +size 462075 diff --git a/vlm/train/D2TE6VTJG9/21.png b/vlm/train/D2TE6VTJG9/21.png new file mode 100644 index 0000000000000000000000000000000000000000..46dd493c450a5f2269927ed4277aef4fd0678676 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6d489b97bd407e4eeaf223904e29da34f7212f39153c1774beba1ebd7e96390 +size 412615 diff --git a/vlm/train/D2TE6VTJG9/22.png b/vlm/train/D2TE6VTJG9/22.png new file mode 100644 index 0000000000000000000000000000000000000000..e3af7ff11b2080ccd36b6b90e3a08ec32866607c --- /dev/null +++ b/vlm/train/D2TE6VTJG9/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0eeb2e57a0ab17d2fa14cc86ec8d54da4d5dec705cf219af57c94d28d15218f +size 322796 diff --git a/vlm/train/D2TE6VTJG9/23.png b/vlm/train/D2TE6VTJG9/23.png new file mode 100644 index 0000000000000000000000000000000000000000..19b8fd1efae78f8f24ec9fb02025a068f3c56f47 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec71d5bb7077f2463e7787d3c349fb61990cf5447111c26270f01d61f837a715 +size 374551 diff --git a/vlm/train/D2TE6VTJG9/24.png b/vlm/train/D2TE6VTJG9/24.png new file mode 100644 index 0000000000000000000000000000000000000000..1503dea6a2a8f3c544b597b0417f4e53efe48f79 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22799a39f96ba0d608b4ae325cb289dafc79e20d6a377f89315d44c57e9f402b +size 350754 diff --git a/vlm/train/D2TE6VTJG9/25.png b/vlm/train/D2TE6VTJG9/25.png new file mode 100644 index 0000000000000000000000000000000000000000..2d138d32d885f28fbf6c41d71c59ecd7addecee0 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5458f337f420ed8daa6517fd921475e5947dc21f7c50feb0f25c309d69017f89 +size 290506 diff --git a/vlm/train/D2TE6VTJG9/26.png b/vlm/train/D2TE6VTJG9/26.png new file mode 100644 index 0000000000000000000000000000000000000000..c8dbe53b506c3b2c405513ace5c8da396387855d --- /dev/null +++ b/vlm/train/D2TE6VTJG9/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6f3c54a4b13add380bfd699385c049f90f008de4306c54a3cb4d124e42ac793 +size 310606 diff --git a/vlm/train/D2TE6VTJG9/27.png b/vlm/train/D2TE6VTJG9/27.png new file mode 100644 index 0000000000000000000000000000000000000000..f461cd0419fc03c543eb4510a9c33c6e9ff8d2d3 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13acf87b530357dc0740529247029a0bc1ca3106a3e2d30be426b6be1188d35c +size 44956 diff --git a/vlm/train/D2TE6VTJG9/28.png b/vlm/train/D2TE6VTJG9/28.png new file mode 100644 index 0000000000000000000000000000000000000000..ad46bf94904a5fbeb20b93bb9e9ae6c687be3305 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a67e87e1b2b74bed763c5dfcddb48605b3f24b206bcca9cc282b0ac55dcac3aa +size 587243 diff --git a/vlm/train/D2TE6VTJG9/29.png b/vlm/train/D2TE6VTJG9/29.png new file mode 100644 index 0000000000000000000000000000000000000000..5a4e9c9a254e421f96eaaa1625160b30e95a4585 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c0211d1db799914b3abddee0e997407e9b323bbdffbc692f867113cc06249da +size 326738 diff --git a/vlm/train/D2TE6VTJG9/3.png b/vlm/train/D2TE6VTJG9/3.png new file mode 100644 index 0000000000000000000000000000000000000000..94990196420ba36f049dbca303914e09bfe71b9b --- /dev/null +++ b/vlm/train/D2TE6VTJG9/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56e67ab8124292aff08fb48091ca6d6340dbc3650d5601983eae39bfe976b141 +size 562241 diff --git a/vlm/train/D2TE6VTJG9/4.png b/vlm/train/D2TE6VTJG9/4.png new file mode 100644 index 0000000000000000000000000000000000000000..77e9cc0e548530cb6758589c7359e3200bcb13eb --- /dev/null +++ b/vlm/train/D2TE6VTJG9/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff7c9a14f813faaf54bfa858b009df2f91a658d2f80458ad04b52e74f1e99f91 +size 542797 diff --git a/vlm/train/D2TE6VTJG9/5.png b/vlm/train/D2TE6VTJG9/5.png new file mode 100644 index 0000000000000000000000000000000000000000..1f899434af92eccc24e82ab31b2f1fe578ec2dc2 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0b9a2e95482bf0134ca4299081d9ce4b25aa5c130ce93f8c13f934ff170d7c5 +size 535933 diff --git a/vlm/train/D2TE6VTJG9/6.png b/vlm/train/D2TE6VTJG9/6.png new file mode 100644 index 0000000000000000000000000000000000000000..31f2487be14a4189aef595a8e7825eefca154150 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4750694fdabc09da2307be8572930007c545274ff3951216ce71bb65f761676f +size 503243 diff --git a/vlm/train/D2TE6VTJG9/7.png b/vlm/train/D2TE6VTJG9/7.png new file mode 100644 index 0000000000000000000000000000000000000000..5223b800729188f7c2d4313f5d1875413e1bfdd2 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d3f07a3ab3c704537defa3b274b0db666e4107a74a38953b9998b53878bb7ab +size 549004 diff --git a/vlm/train/D2TE6VTJG9/8.png b/vlm/train/D2TE6VTJG9/8.png new file mode 100644 index 0000000000000000000000000000000000000000..1381d40a8e370746b00a34c189f0393dfef3ebe8 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb8ff9464db5b5fac8834671d349a6d439d2802cd30d685a2be84eda270f7862 +size 536889 diff --git a/vlm/train/D2TE6VTJG9/9.png b/vlm/train/D2TE6VTJG9/9.png new file mode 100644 index 0000000000000000000000000000000000000000..3d490c5d63b744efb313e3f1e94bf57dc17641c7 --- /dev/null +++ b/vlm/train/D2TE6VTJG9/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77f5ba78cf12c376c2121d2dc57c8f1ffba074cfc622e1eb9dd2907f8624d620 +size 542528 diff --git a/vlm/train/GWRkOYr4jxQ/0.png b/vlm/train/GWRkOYr4jxQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..00dc6d8f3244d75f17a423342ad1959e37185dc3 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84b3abec50dc38f921f35bacdb4767edafe2a057659b87d76c6764528f3b51f2 +size 485727 diff --git a/vlm/train/GWRkOYr4jxQ/1.png b/vlm/train/GWRkOYr4jxQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7cc8d0aee4de8e96b7f6dfff91ce9abb0789f658 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c45df8c8f9fd714ec53e500b0baef90f2783a4db93a1613d5ff564843374d72 +size 597518 diff --git a/vlm/train/GWRkOYr4jxQ/10.png b/vlm/train/GWRkOYr4jxQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..71dcfc55a6607dd993e891d66031330045c5c6d4 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:414f86cb946eb585aa5dbd36e4c62c2374ca93041d2c0b1d73577bae38fc7d94 +size 579343 diff --git a/vlm/train/GWRkOYr4jxQ/11.png b/vlm/train/GWRkOYr4jxQ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..f2a3e57854ad4b67f1b65453a8324868c00070ee --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e24cde909ff790eb8b6a40d58d276fa64c7777b8087b432c605edacf7290c81 +size 606230 diff --git a/vlm/train/GWRkOYr4jxQ/12.png b/vlm/train/GWRkOYr4jxQ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..cf6f1db09b607fcad35c2bc08044491d0113dda7 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bcf409ca0d926706c4b70bd3907950af30c3042b5751971a9cd6e6a79fd5728 +size 353777 diff --git a/vlm/train/GWRkOYr4jxQ/2.png b/vlm/train/GWRkOYr4jxQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..847f7054ef353f61ec4c871ab874fd72fba6d16a --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da88657a523b103ab6cff34fd70b3c2230fec65643eac50dc3a9339560c04569 +size 444796 diff --git a/vlm/train/GWRkOYr4jxQ/3.png b/vlm/train/GWRkOYr4jxQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b5c1be3bc8aff392b1b8f51fcd8a757ca6ac1c1e --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b5e3605dd8b588b2af3fa3b5e96ba81ee46321ee0c7d9bd816e61fb49d02d0f2 +size 502078 diff --git a/vlm/train/GWRkOYr4jxQ/4.png b/vlm/train/GWRkOYr4jxQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..7cd11bb1d45c519ae27fb3c04e86e76fb27d4dc9 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:708926c519b994bcd90f13736f0af45a0f4dd7e167eef2932803fac6306b0875 +size 536653 diff --git a/vlm/train/GWRkOYr4jxQ/5.png b/vlm/train/GWRkOYr4jxQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..e4887205812a9f6c4bbd9fff341a252f76a0333e --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bce25ac99a3da4cd6a20e10bee830ab52fc96938adf4619b2883625a32a606e6 +size 538336 diff --git a/vlm/train/GWRkOYr4jxQ/6.png b/vlm/train/GWRkOYr4jxQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..cf3b9f09148e95b701617683ed97b5f6ae8fc568 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d901a67ba752cebff92de2cf7d6b3e4de59db5f8e74087902c81a858e6c51e35 +size 641137 diff --git a/vlm/train/GWRkOYr4jxQ/7.png b/vlm/train/GWRkOYr4jxQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8396cec3dbfc66441a96e9929fc5ec462de916ed --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b25b18511b69db427c187ad29f3cfabeb262506a852bbfff04aeab9b765f1726 +size 620652 diff --git a/vlm/train/GWRkOYr4jxQ/8.png b/vlm/train/GWRkOYr4jxQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..59c1fdef545d0c4e1c8f901eb853460d7ebd4b44 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82e1aec7355b14fe164a84eff775ca4126f043a921161297b6c4e3b46428073d +size 503612 diff --git a/vlm/train/GWRkOYr4jxQ/9.png b/vlm/train/GWRkOYr4jxQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..783551de6d7fefc2ef604d1af8df8c8c0c661ba1 --- /dev/null +++ b/vlm/train/GWRkOYr4jxQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bd3b24df6a430164ed99cc66b23ad9519c323e4d65225f48df7b18b5c9c8bab +size 524562 diff --git a/vlm/train/H1Heentlx/0.png b/vlm/train/H1Heentlx/0.png new file mode 100644 index 0000000000000000000000000000000000000000..bfb5b5d27d1a6b9f0dcae593775299274d416c66 --- /dev/null +++ b/vlm/train/H1Heentlx/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59afefdcb63814d97fd8d8e3e3cb76f3e1615cbad2c85e260fd58803195e139c +size 498921 diff --git a/vlm/train/H1Heentlx/1.png b/vlm/train/H1Heentlx/1.png new file mode 100644 index 0000000000000000000000000000000000000000..694a58c07b4ef0f4f8050991c489f868e5473ed3 --- /dev/null +++ b/vlm/train/H1Heentlx/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a0d137e4045f06035f729d87f70340c2b868162311eed54e78ec8590e230714 +size 485908 diff --git a/vlm/train/H1Heentlx/10.png b/vlm/train/H1Heentlx/10.png new file mode 100644 index 0000000000000000000000000000000000000000..9744f15e1b078444e3f957971caa0aa24620f690 --- /dev/null +++ b/vlm/train/H1Heentlx/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f7edbd2eca135624c14efec79177f170ab4c89d0ff41a68b23b65674addc986 +size 503045 diff --git a/vlm/train/H1Heentlx/11.png b/vlm/train/H1Heentlx/11.png new file mode 100644 index 0000000000000000000000000000000000000000..f5447953dcefc5c7b2027e5db08486815b7a0612 --- /dev/null +++ b/vlm/train/H1Heentlx/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff22ad01c8bb180d6e714752cd6f86b6e45bf72c52e72f86a7e4b15e9ae8daa4 +size 558418 diff --git a/vlm/train/H1Heentlx/12.png b/vlm/train/H1Heentlx/12.png new file mode 100644 index 0000000000000000000000000000000000000000..01f5f2d9b716b9a765f3f78f00921c7c3c177bd2 --- /dev/null +++ b/vlm/train/H1Heentlx/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85cd2c41dfeb386622ca85549bffdb17c06b58e2ed39c8a500dfc99172691673 +size 375491 diff --git a/vlm/train/H1Heentlx/2.png b/vlm/train/H1Heentlx/2.png new file mode 100644 index 0000000000000000000000000000000000000000..8ecf651ae69663e529a0645e612b3cb0b21fe2dd --- /dev/null +++ b/vlm/train/H1Heentlx/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c61a6939980a2cf65cc8e8116fdc3f6a7cbce4290375a6c10d4943b87ec5926 +size 497142 diff --git a/vlm/train/H1Heentlx/3.png b/vlm/train/H1Heentlx/3.png new file mode 100644 index 0000000000000000000000000000000000000000..cdede1e100fd0ded049ccc94b55c895cec556d4d --- /dev/null +++ b/vlm/train/H1Heentlx/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89b0f2238f8b5f685b11d6d0748a71f258e4c624f3a4671a83ddec849e7a4b4a +size 450394 diff --git a/vlm/train/H1Heentlx/4.png b/vlm/train/H1Heentlx/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ff2c5c19cfdff8f3bc5bba0d6f56303b65aeb556 --- /dev/null +++ b/vlm/train/H1Heentlx/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4137229b806494e8df8c58c67a7cd80cc6e2f6676ad60f1dc832be6a92ed2d9a +size 451444 diff --git a/vlm/train/H1Heentlx/5.png b/vlm/train/H1Heentlx/5.png new file mode 100644 index 0000000000000000000000000000000000000000..c288a61259a6e50990ea5468a239e527de2ba401 --- /dev/null +++ b/vlm/train/H1Heentlx/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8960f820cfbbcaba1c375ecafd38659b15df411ad8c754ea35e1b8e1e60402e +size 522279 diff --git a/vlm/train/H1Heentlx/6.png b/vlm/train/H1Heentlx/6.png new file mode 100644 index 0000000000000000000000000000000000000000..62f3230273e585daa3cde2fa18cc9d01b14aa323 --- /dev/null +++ b/vlm/train/H1Heentlx/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7de970b7b74a83795250c041e558cf200526ab3bfafc9e667384a88cb7e1b7b +size 987916 diff --git a/vlm/train/H1Heentlx/7.png b/vlm/train/H1Heentlx/7.png new file mode 100644 index 0000000000000000000000000000000000000000..7fc9ab84e02a7792b8d17f35878c809aca62627a --- /dev/null +++ b/vlm/train/H1Heentlx/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4e95bc37ce27b22a3405df7ebcf5a8e821fe16807b0c1d4319bd6ce751d8e04 +size 1285407 diff --git a/vlm/train/H1Heentlx/8.png b/vlm/train/H1Heentlx/8.png new file mode 100644 index 0000000000000000000000000000000000000000..9a70952bae0650a8ce4136e3c117f5d29b0ac052 --- /dev/null +++ b/vlm/train/H1Heentlx/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47e0dc534368304ff6bed5559e9e9bfc649d5f6ee76fb882fdb9fd061f082508 +size 547588 diff --git a/vlm/train/H1Heentlx/9.png b/vlm/train/H1Heentlx/9.png new file mode 100644 index 0000000000000000000000000000000000000000..f85cead8491771cf9452df4daf5c701b9dbe1ac0 --- /dev/null +++ b/vlm/train/H1Heentlx/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21ee8b322893a205d06e58092e832e5ae376506c06585effa22edc0dd4a9d4dc +size 537413 diff --git a/vlm/train/H1WgVz-AZ/0.png b/vlm/train/H1WgVz-AZ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..baf41be70a8bed97931aab0af6a25ca10ba65e3f --- /dev/null +++ b/vlm/train/H1WgVz-AZ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6722ab856a9d18c0b3020804bd07ebdb5c770b4823e3d321ed2d213aeb8995fb +size 492296 diff --git a/vlm/train/H1WgVz-AZ/1.png b/vlm/train/H1WgVz-AZ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9e3153d93573634801316e0547513fa48b708c7a --- /dev/null +++ b/vlm/train/H1WgVz-AZ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaf248bb483b5a4122a62cf09562e0cc12768bffa1c2b825340537662611644a +size 507045 diff --git a/vlm/train/H1WgVz-AZ/10.png b/vlm/train/H1WgVz-AZ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..8450ba8cd00eb6fd1bcbf5927c33ddf6dffb5f15 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cff0f0aca1ce7db78d1d1bfc567b43b49f7c979eb878f3720365127712f4b082 +size 530866 diff --git a/vlm/train/H1WgVz-AZ/11.png b/vlm/train/H1WgVz-AZ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..6b2152540c94cf3ded3a046ecc5cea3ab36d36ed --- /dev/null +++ b/vlm/train/H1WgVz-AZ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8f209c9a4dd1eaaec0d5731057a10bb81e7017a3f497a75bbc1513c72cc355b +size 459254 diff --git a/vlm/train/H1WgVz-AZ/12.png b/vlm/train/H1WgVz-AZ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..38bde023358aefe3dba22c5259a5909949940734 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a8ca9695de3cfc06310410419c8f4df412d5f97b6a798163d48c44e4d80cb9a +size 452215 diff --git a/vlm/train/H1WgVz-AZ/13.png b/vlm/train/H1WgVz-AZ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..b73a729242d1f5aea09237ab793f3a46b3631f87 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33567e57d4a7a1df1def6689a0f3e55771df6562fbf5e5540039603c5d63a4dd +size 530010 diff --git a/vlm/train/H1WgVz-AZ/14.png b/vlm/train/H1WgVz-AZ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..ae364866e773dd40f7807a4a0e9cc876147c47e3 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:204b53fbb4563355475634a1107398f11ab8b647f2fc12d9fa90f02f410f8997 +size 339350 diff --git a/vlm/train/H1WgVz-AZ/2.png b/vlm/train/H1WgVz-AZ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..207082fa43d168e24e58ad2779c25945abd50844 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d290d013880054bbf1c401a7547972bd8f6a8be9e443db31239c6f2c2fa5658 +size 510426 diff --git a/vlm/train/H1WgVz-AZ/3.png b/vlm/train/H1WgVz-AZ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e4c02457b1750f216ce5fe118913aa09932bc9c8 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b3f8e3349740d24c28b6093a3a02da275712a6d8fcf6d8f35b65605b92e6fa9 +size 557343 diff --git a/vlm/train/H1WgVz-AZ/4.png b/vlm/train/H1WgVz-AZ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..35ca51f81df9347e395118da0b3dfa49485977e2 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c3063ce0761f5dba4b4178e1ccfef946a9b49354ff862f88097fcc08507118 +size 520523 diff --git a/vlm/train/H1WgVz-AZ/5.png b/vlm/train/H1WgVz-AZ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..ffd40ad492722922142901199d9c94c7763aa385 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d530fe227cce3cc0aa4c0f222d266bde7f7bb0e472623aa732446ddd6c938fe +size 550128 diff --git a/vlm/train/H1WgVz-AZ/6.png b/vlm/train/H1WgVz-AZ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..e1b119f5a26e256100d4add23473a94ac37e86b7 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c08862ebf09f19f5bae04ca128eff912029e8398b467c74f913960c7f073c52 +size 502160 diff --git a/vlm/train/H1WgVz-AZ/7.png b/vlm/train/H1WgVz-AZ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..cef789149caefb172701c91be1a663f04e149f6c --- /dev/null +++ b/vlm/train/H1WgVz-AZ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35ba6b94bb485c0119af909cc867513914c6da78564ed5106187c1020348dc32 +size 499146 diff --git a/vlm/train/H1WgVz-AZ/8.png b/vlm/train/H1WgVz-AZ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..3049b267202ce19e30da272ef73decd0f88c2423 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8125f435194848899ad673588fcb3be75ed1c6bb84c39c85b25e242e259ea3b +size 521160 diff --git a/vlm/train/H1WgVz-AZ/9.png b/vlm/train/H1WgVz-AZ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0f44cdb23764b6f2fab017f7921d9296d30a85a2 --- /dev/null +++ b/vlm/train/H1WgVz-AZ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9c5155dcb6443a264c116be3d32d123991d3cce5c7f03da8c32fd7f58fabde7 +size 452120 diff --git a/vlm/train/H1eH4n09KX/0.png b/vlm/train/H1eH4n09KX/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d45ee6fdf03590e457416f65bc82929cd2d53afd --- /dev/null +++ b/vlm/train/H1eH4n09KX/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:917cf43c54d78f95bfc5e3ef0e20462c06728530db046d83ecb0cf2cb5620d63 +size 520217 diff --git a/vlm/train/H1eH4n09KX/1.png b/vlm/train/H1eH4n09KX/1.png new file mode 100644 index 0000000000000000000000000000000000000000..4fa435b69d5272b3e3c55a1a8abea89ebb099031 --- /dev/null +++ b/vlm/train/H1eH4n09KX/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5c45cc78f1719946d8a942b9a33b0c311c0ecac296a952d243a7dc4acaa4a6b +size 693242 diff --git a/vlm/train/H1eH4n09KX/10.png b/vlm/train/H1eH4n09KX/10.png new file mode 100644 index 0000000000000000000000000000000000000000..fae0eff7e0632fa367d16bbdecf06d3389623cbc --- /dev/null +++ b/vlm/train/H1eH4n09KX/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ee4f960914a04e37a12b13e1d4c8aadbf59ea96d7e2229626ad1dc50092514f +size 469322 diff --git a/vlm/train/H1eH4n09KX/11.png b/vlm/train/H1eH4n09KX/11.png new file mode 100644 index 0000000000000000000000000000000000000000..599ae8deec94f0eaeb02a935a6700affedd1850a --- /dev/null +++ b/vlm/train/H1eH4n09KX/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8760e4d9f195ba7b0e2b3c3635b625a7fa1c31786fa966921f679f2ac077eb2f +size 253002 diff --git a/vlm/train/H1eH4n09KX/2.png b/vlm/train/H1eH4n09KX/2.png new file mode 100644 index 0000000000000000000000000000000000000000..52e77e3d04e4a33dd5a2a69d1ab01576f710996b --- /dev/null +++ b/vlm/train/H1eH4n09KX/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6547e1e072b8b36ef18b1cb1aa6bf73d6519d6bc422cde978fa6bf1981b4e1c1 +size 586572 diff --git a/vlm/train/H1eH4n09KX/3.png b/vlm/train/H1eH4n09KX/3.png new file mode 100644 index 0000000000000000000000000000000000000000..a2db41573d3402ed818d11a1a7810f12dcdbdfdb --- /dev/null +++ b/vlm/train/H1eH4n09KX/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad3f7e007fad959945045c8e6fa67d7d1e206f32d0249d9caef671d5dc46700e +size 480557 diff --git a/vlm/train/H1eH4n09KX/4.png b/vlm/train/H1eH4n09KX/4.png new file mode 100644 index 0000000000000000000000000000000000000000..60712e50f0b62ea4b930a7e6a7c297a46a61cacb --- /dev/null +++ b/vlm/train/H1eH4n09KX/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7ca940e77064d87c8a421034f2072e5dc19007700eb032d868128ded4bc35a9 +size 451099 diff --git a/vlm/train/H1eH4n09KX/5.png b/vlm/train/H1eH4n09KX/5.png new file mode 100644 index 0000000000000000000000000000000000000000..56a8c6381f5b2f82368b02584fd0242db117fe95 --- /dev/null +++ b/vlm/train/H1eH4n09KX/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f309cfae11ebfc54c79a057a39720f9887306ac222e60738099d47109433ce2b +size 581524 diff --git a/vlm/train/H1eH4n09KX/6.png b/vlm/train/H1eH4n09KX/6.png new file mode 100644 index 0000000000000000000000000000000000000000..494df5f5b4ae610e5cd39e63a5ddb6cb7dcc4286 --- /dev/null +++ b/vlm/train/H1eH4n09KX/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33b7344ecd4e78d5bf7ab47f8efb85d6bcb35a2aecbad900c1f13bc254e052c5 +size 541615 diff --git a/vlm/train/H1eH4n09KX/7.png b/vlm/train/H1eH4n09KX/7.png new file mode 100644 index 0000000000000000000000000000000000000000..d5f51b3660aff3277622f5bb827a4da5c5ea4e18 --- /dev/null +++ b/vlm/train/H1eH4n09KX/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77e7976d6adbd1ed364c13b12ef4fbf51fa9d426af8399b5aca47637e1fb35e2 +size 749750 diff --git a/vlm/train/H1eH4n09KX/8.png b/vlm/train/H1eH4n09KX/8.png new file mode 100644 index 0000000000000000000000000000000000000000..13ba59f6d950f4140c5d947aeb26bb91024c8f3f --- /dev/null +++ b/vlm/train/H1eH4n09KX/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7771a92068331412f8de274d4f6773b25c360851c8ca944e8f19d3898e6e0dc +size 463242 diff --git a/vlm/train/H1eH4n09KX/9.png b/vlm/train/H1eH4n09KX/9.png new file mode 100644 index 0000000000000000000000000000000000000000..dc6bcff4a78dd2d05fcc23ef836961aedc057bf3 --- /dev/null +++ b/vlm/train/H1eH4n09KX/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cd010469505603b2181667aca90e47ec337039347d4c43dc77f0b79f6710b83 +size 519662 diff --git a/vlm/train/H1ersoRqtm/0.png b/vlm/train/H1ersoRqtm/0.png new file mode 100644 index 0000000000000000000000000000000000000000..fcc74d1242d8e5fdad2598798729afe9e803e1d5 --- /dev/null +++ b/vlm/train/H1ersoRqtm/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c2054853704befff5ad03512f9c4cc65416122c729bb123311ff8f5b1c5cb48 +size 466487 diff --git a/vlm/train/H1ersoRqtm/1.png b/vlm/train/H1ersoRqtm/1.png new file mode 100644 index 0000000000000000000000000000000000000000..c8052c1ac78070d4575f363cfc30e899ff9d2d5e --- /dev/null +++ b/vlm/train/H1ersoRqtm/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a87681ed2c630807ebf4c5b6f8d3903b1222d454c190a27e09cc5bcfac4edd +size 504202 diff --git a/vlm/train/H1ersoRqtm/10.png b/vlm/train/H1ersoRqtm/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a818593ee239e8e0ec4ecce0e45201602f68128b --- /dev/null +++ b/vlm/train/H1ersoRqtm/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d89fce84b65abcfe8dda7403fd0874fe293eaf2adaa244be1594fa1324b99e4 +size 574025 diff --git a/vlm/train/H1ersoRqtm/11.png b/vlm/train/H1ersoRqtm/11.png new file mode 100644 index 0000000000000000000000000000000000000000..f8f0deff692f5bde82c87f85aa59b81de5ad0555 --- /dev/null +++ b/vlm/train/H1ersoRqtm/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe74e149549f0ce02b057a2af79aba6cdd3630f65bc0882e219733317c59261c +size 217377 diff --git a/vlm/train/H1ersoRqtm/12.png b/vlm/train/H1ersoRqtm/12.png new file mode 100644 index 0000000000000000000000000000000000000000..eed0afb722e79ab6aba17cf8f6e9ccdddf87036b --- /dev/null +++ b/vlm/train/H1ersoRqtm/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35c3c876918ba4835dd6144f47ea2ef3c68d144540a8e272a61d654b1b4ad290 +size 344160 diff --git a/vlm/train/H1ersoRqtm/13.png b/vlm/train/H1ersoRqtm/13.png new file mode 100644 index 0000000000000000000000000000000000000000..85566b4b5f2882e2ad23ba0fdcf824a5f2d5ef36 --- /dev/null +++ b/vlm/train/H1ersoRqtm/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dd60b758bd01effd39a31ef51d1000f223ef586dc5843b41b9a6aa76233f24b +size 315995 diff --git a/vlm/train/H1ersoRqtm/14.png b/vlm/train/H1ersoRqtm/14.png new file mode 100644 index 0000000000000000000000000000000000000000..8e0400aa31908738d536521b37d8ff6e027d2782 --- /dev/null +++ b/vlm/train/H1ersoRqtm/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:468fc3b78569255f0182d1496f89cfcc708c50db476550b2420936e850d5cdb3 +size 214753 diff --git a/vlm/train/H1ersoRqtm/15.png b/vlm/train/H1ersoRqtm/15.png new file mode 100644 index 0000000000000000000000000000000000000000..01da1ca94f81dc366cc5ad590fe4d00347dbc541 --- /dev/null +++ b/vlm/train/H1ersoRqtm/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5cd9d3b05898d66c0df38f787722484586e2d43e69a7fa8e1a1121d7c290aaa +size 502349 diff --git a/vlm/train/H1ersoRqtm/16.png b/vlm/train/H1ersoRqtm/16.png new file mode 100644 index 0000000000000000000000000000000000000000..31c0f299bbdc47d7c5cfaa7d7e5254a625a1e755 --- /dev/null +++ b/vlm/train/H1ersoRqtm/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67a012f4a6523ff415535fd9f42ea8284c6472c7d30594d8668d31e826475794 +size 432844 diff --git a/vlm/train/H1ersoRqtm/17.png b/vlm/train/H1ersoRqtm/17.png new file mode 100644 index 0000000000000000000000000000000000000000..06535799f082a09b013fb30a38fd94f9b6efbcfe --- /dev/null +++ b/vlm/train/H1ersoRqtm/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebfe3f224df7b16a95051bb83e1d0b96566852c5ba8c3f4ceccfdcda88ff90c8 +size 99750 diff --git a/vlm/train/H1ersoRqtm/2.png b/vlm/train/H1ersoRqtm/2.png new file mode 100644 index 0000000000000000000000000000000000000000..0664b509ce3435387d335f9c11785f2812ed851b --- /dev/null +++ b/vlm/train/H1ersoRqtm/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:646731ac9195f46ce511d8e8af88a11de4b1bc33875fcf56fa57d0ee4f2124aa +size 620394 diff --git a/vlm/train/H1ersoRqtm/3.png b/vlm/train/H1ersoRqtm/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3b2e330b830128a7d958157a04807a54eedf44d5 --- /dev/null +++ b/vlm/train/H1ersoRqtm/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bdd0968e115fb0735d8f203fe23be75f76b178979754440875454b58a0f33dc6 +size 567069 diff --git a/vlm/train/H1ersoRqtm/4.png b/vlm/train/H1ersoRqtm/4.png new file mode 100644 index 0000000000000000000000000000000000000000..155c46e9ac4a93cca8e39accc7e16fc923a7a027 --- /dev/null +++ b/vlm/train/H1ersoRqtm/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35af1e73294ee246bd2ab387295ca97ec8478b904d7c38dccd1ceb21b7612b0c +size 525771 diff --git a/vlm/train/H1ersoRqtm/5.png b/vlm/train/H1ersoRqtm/5.png new file mode 100644 index 0000000000000000000000000000000000000000..90ffc9dead659d107577f6ce99d2654f761f29a5 --- /dev/null +++ b/vlm/train/H1ersoRqtm/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecec88f2e151966bbebcd568a2e1b34bab0de9d69036ad49bf078fbaf84d92ea +size 459578 diff --git a/vlm/train/H1ersoRqtm/6.png b/vlm/train/H1ersoRqtm/6.png new file mode 100644 index 0000000000000000000000000000000000000000..66f9e8a1c5f510ccb59c55f041792544e7551c26 --- /dev/null +++ b/vlm/train/H1ersoRqtm/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:637d6fdaa2fc633d753c0007c9439101ba657e6d63270d90fcbcf8d1f141f0ec +size 493631 diff --git a/vlm/train/H1ersoRqtm/7.png b/vlm/train/H1ersoRqtm/7.png new file mode 100644 index 0000000000000000000000000000000000000000..53ec4cfa1a62233b2178fb936a7ac1b09c4eec9d --- /dev/null +++ b/vlm/train/H1ersoRqtm/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64690059aecf100a5638076f80fb275883ac13f3282203d2baaf1025fde01d78 +size 542201 diff --git a/vlm/train/H1ersoRqtm/8.png b/vlm/train/H1ersoRqtm/8.png new file mode 100644 index 0000000000000000000000000000000000000000..3fdde6e62e5e7e6da7250ac9dcdcce2d7cd515dc --- /dev/null +++ b/vlm/train/H1ersoRqtm/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72b69919bad41d5d0fb0ac101337938d500b3ab648a176513ec0404ac005e81e +size 516222 diff --git a/vlm/train/H1ersoRqtm/9.png b/vlm/train/H1ersoRqtm/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0467f213cfd5a6d19af65b6d0cb1f07ca51d4a78 --- /dev/null +++ b/vlm/train/H1ersoRqtm/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a2fcb2e1068559a3cf0ccc24dc354579413f08e1ae59027f6c39db2f1dffe10 +size 535736 diff --git a/vlm/train/H1lS8oA5YQ/0.png b/vlm/train/H1lS8oA5YQ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..1d6454804fad0701ec73d388854317968759b8b0 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5581f203f4f27b572c2767db08451d23a4a4de8398c944a35e0f30ece5d3e95 +size 479278 diff --git a/vlm/train/H1lS8oA5YQ/1.png b/vlm/train/H1lS8oA5YQ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..03e0f3e86966cec70685c5aa8fa3e555c0fc3287 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08db45457695260d9f560dad540c5095ebbcfe11e6c615af414a554fe7072d66 +size 555599 diff --git a/vlm/train/H1lS8oA5YQ/10.png b/vlm/train/H1lS8oA5YQ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..16780cf56d20280163436eecd3f7892e48232a7c --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b558a9d7416d24397b5a002545aad6281df0c3b906c5402d5cde0f9d86a3845 +size 473701 diff --git a/vlm/train/H1lS8oA5YQ/11.png b/vlm/train/H1lS8oA5YQ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a516e50ed318556517ccca2a034fb86e195660d8 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d23ca193a8f6191d61735edd0e4cb1979e5e759d5be56c5ef97aac562d7b195 +size 456561 diff --git a/vlm/train/H1lS8oA5YQ/12.png b/vlm/train/H1lS8oA5YQ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b893944f09d64642daa3fea95e7e362a88e5de97 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a1fc2ef013d8b69aed05e9f7d53671465f8469688ab71050a23db92252bda68 +size 524978 diff --git a/vlm/train/H1lS8oA5YQ/13.png b/vlm/train/H1lS8oA5YQ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..5fef48af5084f94130cace0445724b51c61ee3f4 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:48f4a20c52228381ccfaadc2c29d6158ffc75404f75b8d40f00e023710c35ec2 +size 249325 diff --git a/vlm/train/H1lS8oA5YQ/14.png b/vlm/train/H1lS8oA5YQ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..1b0ca19b40a207244da3d182af867bae60f387c8 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a254e5478eaa8495f787541525239375db8a614ed039595d554afc9bcfe9dce +size 302443 diff --git a/vlm/train/H1lS8oA5YQ/15.png b/vlm/train/H1lS8oA5YQ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..03fef107e070f92b859a6cd1d2d9064057a3c2dc --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:431bfbfd6abad7a33264851e57f7bdd9eba06e14337f6a24312ac5a3a04ce5c2 +size 506943 diff --git a/vlm/train/H1lS8oA5YQ/16.png b/vlm/train/H1lS8oA5YQ/16.png new file mode 100644 index 0000000000000000000000000000000000000000..cdd9bd4e6a66695f2b625dec101f5e69a6b6681e --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dc529563b68122399e7f1af40dfd8f5cae5ebf14c0941a6354f23241f1e9aaa +size 479751 diff --git a/vlm/train/H1lS8oA5YQ/17.png b/vlm/train/H1lS8oA5YQ/17.png new file mode 100644 index 0000000000000000000000000000000000000000..9ff9e0cdc2bf8cf2c96aed9e97147092c727a04c --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df3923d65a8308a8d5ab6003fed9fd8935303230ed1c0d40ecfde57a13826a57 +size 141112 diff --git a/vlm/train/H1lS8oA5YQ/18.png b/vlm/train/H1lS8oA5YQ/18.png new file mode 100644 index 0000000000000000000000000000000000000000..d74ce4036a8463450f70e47c7126ca04e0cfce22 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38ab5ccfa7c49cad955d6eb913d7e1ef94394535f277bc78bbd937b57092f05d +size 1541422 diff --git a/vlm/train/H1lS8oA5YQ/19.png b/vlm/train/H1lS8oA5YQ/19.png new file mode 100644 index 0000000000000000000000000000000000000000..8ade79d241c65962ae63ed68cae26311b3ef61a6 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9bdce1cb6e63b47889fb30625e257ba80066a4ddc0673d61f5741b5e2e5daff +size 1540819 diff --git a/vlm/train/H1lS8oA5YQ/2.png b/vlm/train/H1lS8oA5YQ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..91efd83896fc17be1c6f7eca23cca313764bc5fe --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fc9136b42cf48009213b18a4d1ac416d6c8b762228c3ab56c62fff53282eed0 +size 502473 diff --git a/vlm/train/H1lS8oA5YQ/20.png b/vlm/train/H1lS8oA5YQ/20.png new file mode 100644 index 0000000000000000000000000000000000000000..9b10bdc6093cc9a24acc9f38e8ba344ab6f89ca5 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b651c8d535de52cdd6554c75d65de61f8d76fe5a8d47340bea3428d52490c94 +size 1502078 diff --git a/vlm/train/H1lS8oA5YQ/21.png b/vlm/train/H1lS8oA5YQ/21.png new file mode 100644 index 0000000000000000000000000000000000000000..045fbd6fffb065f32ec8d4644b5a190cc78acaa2 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:356e45bcc7d8fbcc749a61877d1548b347a32fc42aeccf463674c12a0704ad62 +size 1401343 diff --git a/vlm/train/H1lS8oA5YQ/22.png b/vlm/train/H1lS8oA5YQ/22.png new file mode 100644 index 0000000000000000000000000000000000000000..3b77a141ec3c65c16afe1f78d5a56da7c6bc6231 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0875b55b70cdf5c2184c920f4f176d724dbf2592c6285d0cf684a9dbe65af398 +size 1571572 diff --git a/vlm/train/H1lS8oA5YQ/3.png b/vlm/train/H1lS8oA5YQ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..fdfeec838a0dc7e5396825e55c940f7065c243e0 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15094e60fcc027799d7e911b0d3bae21394df78994fb88b445e190fd5365d628 +size 630768 diff --git a/vlm/train/H1lS8oA5YQ/4.png b/vlm/train/H1lS8oA5YQ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..903bf7648aa8cd0f644b717226f60d35455d0e5c --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f607a546f12b7e7c2610b4b84d855ebe948ef0d758c40086c5a229b190ee9b8 +size 511761 diff --git a/vlm/train/H1lS8oA5YQ/5.png b/vlm/train/H1lS8oA5YQ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7a3dfd34db123e28cfda9b7f36384581ceeed89f --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f0b9fd4b58aec398a67287f9f958573b0d8076ae67a28062c26ee607947aa20 +size 560212 diff --git a/vlm/train/H1lS8oA5YQ/6.png b/vlm/train/H1lS8oA5YQ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..c1c2415b0d969ff001da2e9191dd4f26483a44b4 --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2d6bedf0480c279edc633018bbc34b3c74e6e3da926c8980ffdc8642177a4b3 +size 593095 diff --git a/vlm/train/H1lS8oA5YQ/7.png b/vlm/train/H1lS8oA5YQ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..901867c4df76852ea0828fd92d0d82db6f070cda --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6eef1276a490e951131656c1bedfb06c732fdf854b2c61c3dcf8aa3e9637af0f +size 654250 diff --git a/vlm/train/H1lS8oA5YQ/8.png b/vlm/train/H1lS8oA5YQ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..6a40051489cac8267b0735a2d3d43c0c1fb5dbed --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70d4746eae0ec184e49a4af6903c802398e738be32df773582b478dfe2604c45 +size 541556 diff --git a/vlm/train/H1lS8oA5YQ/9.png b/vlm/train/H1lS8oA5YQ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..fb4b4d22feb233faeac0398204b93016af358fed --- /dev/null +++ b/vlm/train/H1lS8oA5YQ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4c880b8336f261e5b952aa722f6050553a58e86d5405ce8beeb131aca264742 +size 261714 diff --git a/vlm/train/H1loF2NFwr/0.png b/vlm/train/H1loF2NFwr/0.png new file mode 100644 index 0000000000000000000000000000000000000000..e4d65fbc66448d4862b1c938fa8ed17b7df2e82c --- /dev/null +++ b/vlm/train/H1loF2NFwr/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f89ade6c95f3c0d463027eb6cc4920354098a812d86708fdd36dd416d70cfd0 +size 476003 diff --git a/vlm/train/H1loF2NFwr/1.png b/vlm/train/H1loF2NFwr/1.png new file mode 100644 index 0000000000000000000000000000000000000000..465c3300c71ea970383c71a2af94818525a66c22 --- /dev/null +++ b/vlm/train/H1loF2NFwr/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a3d9916237a6aabe5a5ed4438fe8ec94e9854007601e423a5a29b48389f4864 +size 634896 diff --git a/vlm/train/H1loF2NFwr/10.png b/vlm/train/H1loF2NFwr/10.png new file mode 100644 index 0000000000000000000000000000000000000000..143b179910acfd0d1c43b09f6b3b4ad5588a6bd6 --- /dev/null +++ b/vlm/train/H1loF2NFwr/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc3eed1a1a9627ecc71b9b34c4c088569083f4d859e789327ca1bcd0c8942874 +size 541248 diff --git a/vlm/train/H1loF2NFwr/11.png b/vlm/train/H1loF2NFwr/11.png new file mode 100644 index 0000000000000000000000000000000000000000..fe6d7423a4d8346eccd23bc8262778469e09fef0 --- /dev/null +++ b/vlm/train/H1loF2NFwr/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4aae1032d05c867ed0e45344cec36a20bd4ce6bb5a53cd503a8516babc093cea +size 477669 diff --git a/vlm/train/H1loF2NFwr/12.png b/vlm/train/H1loF2NFwr/12.png new file mode 100644 index 0000000000000000000000000000000000000000..18b1744d8438eb5c7d4d7e626c27fd0da77999b1 --- /dev/null +++ b/vlm/train/H1loF2NFwr/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a8984179d05aa5df70d7bb8d5dd8aece60bd9bd0661f0ed37367adb843bf477 +size 507321 diff --git a/vlm/train/H1loF2NFwr/13.png b/vlm/train/H1loF2NFwr/13.png new file mode 100644 index 0000000000000000000000000000000000000000..1dc2b98956c46722ea909a7c2c75592b5a8d07fb --- /dev/null +++ b/vlm/train/H1loF2NFwr/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3be2a6a73b740f0a6ca1a9baf47ae85450f5a24e7b79ba91936d623bcbc67823 +size 476563 diff --git a/vlm/train/H1loF2NFwr/14.png b/vlm/train/H1loF2NFwr/14.png new file mode 100644 index 0000000000000000000000000000000000000000..001945c5ea4708319f6828e97746589e68e76928 --- /dev/null +++ b/vlm/train/H1loF2NFwr/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e09b557181210a429b17324348ca4adf07208d79474cb7081158285bbc11df5 +size 611254 diff --git a/vlm/train/H1loF2NFwr/15.png b/vlm/train/H1loF2NFwr/15.png new file mode 100644 index 0000000000000000000000000000000000000000..6740400660be20e526a04860e59b68e6369e843f --- /dev/null +++ b/vlm/train/H1loF2NFwr/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:254f52466dd890040d23663ea7897cecc3eab41e1f3e698fd948011fa7b261dd +size 397195 diff --git a/vlm/train/H1loF2NFwr/2.png b/vlm/train/H1loF2NFwr/2.png new file mode 100644 index 0000000000000000000000000000000000000000..98d405e4d860be5f99d4794f83422e9a0be54907 --- /dev/null +++ b/vlm/train/H1loF2NFwr/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab543755fe4063553242d7025636fe0a318e056d0bb2dcf598dfbdd6ee712bab +size 582850 diff --git a/vlm/train/H1loF2NFwr/3.png b/vlm/train/H1loF2NFwr/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0ad606ae8c54730b2dcea5b83f6bed07803f6ff4 --- /dev/null +++ b/vlm/train/H1loF2NFwr/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a128f13c2af7014fd58dfb64705ba394b8971bf362c0098d691848430bb51d74 +size 527155 diff --git a/vlm/train/H1loF2NFwr/4.png b/vlm/train/H1loF2NFwr/4.png new file mode 100644 index 0000000000000000000000000000000000000000..9fe4db6bbcc954ca41f5f2584409abfd3e20811f --- /dev/null +++ b/vlm/train/H1loF2NFwr/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fc84e76a3ccb7a8a7c7305f7ed5243df99537e5accafa331fa18236585fe5e8 +size 585009 diff --git a/vlm/train/H1loF2NFwr/5.png b/vlm/train/H1loF2NFwr/5.png new file mode 100644 index 0000000000000000000000000000000000000000..c1ddd77537bfbccdf349402f81bc15ff3982b995 --- /dev/null +++ b/vlm/train/H1loF2NFwr/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:950e8e8c1e0d70cc810349289fc2c3a475f037ec69d63a766931b1c54085a3fa +size 604028 diff --git a/vlm/train/H1loF2NFwr/6.png b/vlm/train/H1loF2NFwr/6.png new file mode 100644 index 0000000000000000000000000000000000000000..f18cfd417d0e9e85fa453a893adaa735aaeadce1 --- /dev/null +++ b/vlm/train/H1loF2NFwr/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a124013b52f36622d6381df435fd29ccaed9d0af3d356fb98cd8437f5e39e4f9 +size 569903 diff --git a/vlm/train/H1loF2NFwr/7.png b/vlm/train/H1loF2NFwr/7.png new file mode 100644 index 0000000000000000000000000000000000000000..6f7b4d38f9e14035fd351a4836cdc87d6012e261 --- /dev/null +++ b/vlm/train/H1loF2NFwr/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8da731e176059a7bb3561c0a8c9b43e0f06e84b88cf79fa8c7038711aa16a5b3 +size 561971 diff --git a/vlm/train/H1loF2NFwr/8.png b/vlm/train/H1loF2NFwr/8.png new file mode 100644 index 0000000000000000000000000000000000000000..ffa6f686b7db5942bb356271360fb08abfcea118 --- /dev/null +++ b/vlm/train/H1loF2NFwr/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68c31f4319b82a6ff03a322ef313562ca9d52e518e5e4d12e5770d5b0c333d48 +size 509375 diff --git a/vlm/train/H1loF2NFwr/9.png b/vlm/train/H1loF2NFwr/9.png new file mode 100644 index 0000000000000000000000000000000000000000..fb739e323b852ef27d7c8745cd845648938abd19 --- /dev/null +++ b/vlm/train/H1loF2NFwr/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14b655459db7232c9a2bc8e84b2ba6779b7144884accaa09df335f55446a5575 +size 563263 diff --git a/vlm/train/H1q-TM-AW/0.png b/vlm/train/H1q-TM-AW/0.png new file mode 100644 index 0000000000000000000000000000000000000000..4c9d96af52656982e02d29deb2fad17895f89345 --- /dev/null +++ b/vlm/train/H1q-TM-AW/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecaaa00a3d2d142b7694bfb8c43f84cce055bd91aab3780c6f3a80cb52304749 +size 487693 diff --git a/vlm/train/H1q-TM-AW/1.png b/vlm/train/H1q-TM-AW/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a143540bca79b211e9577696b9baebeb80a2a68c --- /dev/null +++ b/vlm/train/H1q-TM-AW/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7aafcac4a5b4aa988ab71f148a9a670a71e0229fc387cb7e8f3e9e59c88872f4 +size 600574 diff --git a/vlm/train/H1q-TM-AW/11.png b/vlm/train/H1q-TM-AW/11.png new file mode 100644 index 0000000000000000000000000000000000000000..c7850007f0c4c36647edcf4a71b48977b0589a49 --- /dev/null +++ b/vlm/train/H1q-TM-AW/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74a15d39dc4946af5a1d9c619590be4976fb847537dc9b62660aa231b05efc3d +size 539298 diff --git a/vlm/train/H1q-TM-AW/12.png b/vlm/train/H1q-TM-AW/12.png new file mode 100644 index 0000000000000000000000000000000000000000..60847ff944f3047dbb7e6f04549fe782a742f477 --- /dev/null +++ b/vlm/train/H1q-TM-AW/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af092e9b5b57f63fd1ce09eb5df34cabdc01166da939d31fce62eadeef831f1d +size 126844 diff --git a/vlm/train/H1q-TM-AW/13.png b/vlm/train/H1q-TM-AW/13.png new file mode 100644 index 0000000000000000000000000000000000000000..a4cc3e37e543c4165a89deaca1c7bdfd86b1ceb7 --- /dev/null +++ b/vlm/train/H1q-TM-AW/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46736e10d0c55b9ca536a03dcdc774e7e0133813ab8ff48bda4329e35ac8fb1a +size 213010 diff --git a/vlm/train/H1q-TM-AW/14.png b/vlm/train/H1q-TM-AW/14.png new file mode 100644 index 0000000000000000000000000000000000000000..fe27623f7e422ae4ca40a3a9b3e2bf48cc2b500b --- /dev/null +++ b/vlm/train/H1q-TM-AW/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6a28fdc8b254fcac8cc086ce05746c347e3c6ced4ec1e0b82570679ca6dd1a48 +size 472392 diff --git a/vlm/train/H1q-TM-AW/15.png b/vlm/train/H1q-TM-AW/15.png new file mode 100644 index 0000000000000000000000000000000000000000..9d11734abeb4036296af039869850d2e971ab186 --- /dev/null +++ b/vlm/train/H1q-TM-AW/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c59eac1ecfb267f4339acc00b92df0ac8ac4a6c8d63a63bc3768c56327a22200 +size 455478 diff --git a/vlm/train/H1q-TM-AW/16.png b/vlm/train/H1q-TM-AW/16.png new file mode 100644 index 0000000000000000000000000000000000000000..97e4639d88573107597a13fcff11d9dab3cd1f3e --- /dev/null +++ b/vlm/train/H1q-TM-AW/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:66670c0f27dbb31ce82d076550bd96da66ac3bbc3d3c1e34f23ad67f44d4f906 +size 414884 diff --git a/vlm/train/H1q-TM-AW/17.png b/vlm/train/H1q-TM-AW/17.png new file mode 100644 index 0000000000000000000000000000000000000000..9484e58b0ab2a56a1afa7df745bf1263f46c84e3 --- /dev/null +++ b/vlm/train/H1q-TM-AW/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53f914a25394a45ec74291b603436052ee3ac2c4b403e60e4a21f6bf7b83fa8c +size 409812 diff --git a/vlm/train/H1q-TM-AW/18.png b/vlm/train/H1q-TM-AW/18.png new file mode 100644 index 0000000000000000000000000000000000000000..ea0aec0c9ff0cec186e8ed5b464575aab3ab47b9 --- /dev/null +++ b/vlm/train/H1q-TM-AW/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ce4f5e9e4860a629ffc86e2657ca6ccd573e0cb7db99ecc63fd185808552cb7 +size 172100 diff --git a/vlm/train/H1q-TM-AW/2.png b/vlm/train/H1q-TM-AW/2.png new file mode 100644 index 0000000000000000000000000000000000000000..3f0f2c042e4b27c4abb6f5b1f4870c39e8f6a37b --- /dev/null +++ b/vlm/train/H1q-TM-AW/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:53c80f1b95636c37b516d2053eff28d7d606f232035cad871bb18c7e1440b232 +size 542231 diff --git a/vlm/train/H1q-TM-AW/3.png b/vlm/train/H1q-TM-AW/3.png new file mode 100644 index 0000000000000000000000000000000000000000..52908bd8243a6789e4a027fe7e36d068cffc5547 --- /dev/null +++ b/vlm/train/H1q-TM-AW/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ad2d8a0546519483ad72736c3ab9e48a4c14f649e78b5bedf7090b493227e91 +size 438298 diff --git a/vlm/train/H1q-TM-AW/4.png b/vlm/train/H1q-TM-AW/4.png new file mode 100644 index 0000000000000000000000000000000000000000..2ad12157320bdc078ef2e66e11d23245be2cd96e --- /dev/null +++ b/vlm/train/H1q-TM-AW/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a7c65b80e8cffd8bce2a6f8654cc452f68cd625c93cb6eb58496956da35318a +size 384555 diff --git a/vlm/train/H1q-TM-AW/5.png b/vlm/train/H1q-TM-AW/5.png new file mode 100644 index 0000000000000000000000000000000000000000..bb7f9dd87494b6f857d938227f42d8c771c5aa02 --- /dev/null +++ b/vlm/train/H1q-TM-AW/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba8e3a317626ef02350d6b1380c7200d611d3eddafcd6f654f6a90029576ec58 +size 553284 diff --git a/vlm/train/H1q-TM-AW/7.png b/vlm/train/H1q-TM-AW/7.png new file mode 100644 index 0000000000000000000000000000000000000000..0884b2b55b9e67f96efe3cdb05cf19b247e11872 --- /dev/null +++ b/vlm/train/H1q-TM-AW/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fccc1fb15a66c129b269231b816cd97261207e2ebd5d4134a8f23873c3810ec +size 515305 diff --git a/vlm/train/H1q-TM-AW/8.png b/vlm/train/H1q-TM-AW/8.png new file mode 100644 index 0000000000000000000000000000000000000000..36d1047ac0f37bff9a60fc7be63de4d96917238b --- /dev/null +++ b/vlm/train/H1q-TM-AW/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d964890abe4c65b2656c7487287a0d031ce91d7279f92fb0d64950a3111c34d +size 410155 diff --git a/vlm/train/H1q-TM-AW/9.png b/vlm/train/H1q-TM-AW/9.png new file mode 100644 index 0000000000000000000000000000000000000000..19eae63b4f4be22c4a0448da4f171442e2120e87 --- /dev/null +++ b/vlm/train/H1q-TM-AW/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42300da31e24956e52c3e751426dfbf7d686a77517f08f9cd6ace600767b0259 +size 624208 diff --git a/vlm/train/HkeSdCEtDS/0.png b/vlm/train/HkeSdCEtDS/0.png new file mode 100644 index 0000000000000000000000000000000000000000..33b703112de24f90161d677af43a3dc6f9ccc6c3 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c93e74ef193e0aaf2fa48179eee7372f64cf1a5e57efa36de4ae29010e25b74 +size 500164 diff --git a/vlm/train/HkeSdCEtDS/1.png b/vlm/train/HkeSdCEtDS/1.png new file mode 100644 index 0000000000000000000000000000000000000000..9f77a7be00d2e4ca47b3d5b2076f344aae2c78d9 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:211d6139599bf1a2dee1380ef10434c39c5a4f6b7191b7afe254e3e99e81afa9 +size 576535 diff --git a/vlm/train/HkeSdCEtDS/10.png b/vlm/train/HkeSdCEtDS/10.png new file mode 100644 index 0000000000000000000000000000000000000000..fc5c796cd41f5c8731112a54d3aede1acfefebfc --- /dev/null +++ b/vlm/train/HkeSdCEtDS/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e6a5a8fd86e165ca1ec0514bbbad423278a1e49960664b8d9a6a80a6a750367a +size 255539 diff --git a/vlm/train/HkeSdCEtDS/11.png b/vlm/train/HkeSdCEtDS/11.png new file mode 100644 index 0000000000000000000000000000000000000000..1ee33da847b9f270448151f9a5cf39e226e4edf1 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:683955d4db7725ced063f5c8cc63ecdc70dffbee4934dab8573a46c5e561e676 +size 315394 diff --git a/vlm/train/HkeSdCEtDS/12.png b/vlm/train/HkeSdCEtDS/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b78517cd3ddafdc43a7c7f92944e1c0ac953faaf --- /dev/null +++ b/vlm/train/HkeSdCEtDS/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc4a153cec1ad8a475f69e7c1ec2fbd645c7599933d33d4110a8d4ceb9a64ee0 +size 310953 diff --git a/vlm/train/HkeSdCEtDS/13.png b/vlm/train/HkeSdCEtDS/13.png new file mode 100644 index 0000000000000000000000000000000000000000..0c6171e4bc8ce9be32b5bdc363e832d9f22ece8e --- /dev/null +++ b/vlm/train/HkeSdCEtDS/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc499283397f3d372843a97ff5b7bf689589c83f1c55cb44342ce3399c5f385d +size 368489 diff --git a/vlm/train/HkeSdCEtDS/14.png b/vlm/train/HkeSdCEtDS/14.png new file mode 100644 index 0000000000000000000000000000000000000000..d1b632e9a872d6febd66ff60df05f65202922ce2 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59fe25a2b1430d0b3c9c0f7971995d4a9899779e7686f73102e23386dd77c49d +size 321896 diff --git a/vlm/train/HkeSdCEtDS/15.png b/vlm/train/HkeSdCEtDS/15.png new file mode 100644 index 0000000000000000000000000000000000000000..a5d7f696d457bc607b1d8961186c5940f0d612a1 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7feacf583794e057fc44db98abadbb337e92a93e4d966e2ebd26865a0215bc5e +size 409839 diff --git a/vlm/train/HkeSdCEtDS/16.png b/vlm/train/HkeSdCEtDS/16.png new file mode 100644 index 0000000000000000000000000000000000000000..444ab4b0ea9fb531bb94c4c1cc217b137b0daf70 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11695ac340ea3a4293f14dc8c0acb4312cc82b9cb9368d48b0bd3141c2fa0a15 +size 195410 diff --git a/vlm/train/HkeSdCEtDS/2.png b/vlm/train/HkeSdCEtDS/2.png new file mode 100644 index 0000000000000000000000000000000000000000..28da3ef85d8563b1b55c8a4cb1e413c68d5a3064 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f72268d0466729c75c38cae96c8a2e1cfb61f49d769ee6ad8cd9dce02e40b06b +size 414766 diff --git a/vlm/train/HkeSdCEtDS/3.png b/vlm/train/HkeSdCEtDS/3.png new file mode 100644 index 0000000000000000000000000000000000000000..93ff00761f2e1bbbcfc0577094a34b13dc005103 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:504ca52920b84f25168db3e6ade304ea56f3336ceb3d12ea580781bdbeaa5c20 +size 528402 diff --git a/vlm/train/HkeSdCEtDS/4.png b/vlm/train/HkeSdCEtDS/4.png new file mode 100644 index 0000000000000000000000000000000000000000..43607f11232f7bbf724c8d7df73a5010ab4d9233 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a0daafb2857c845d591227796f4910853384cfc84b33d7dca3744f5e69e2a541 +size 518616 diff --git a/vlm/train/HkeSdCEtDS/5.png b/vlm/train/HkeSdCEtDS/5.png new file mode 100644 index 0000000000000000000000000000000000000000..92d9ad5e1cdde796e26f5e01b04eec350f15fe5a --- /dev/null +++ b/vlm/train/HkeSdCEtDS/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9dc91506db4480a45792753856d3cabb4f97d7e46ffc71e93386fb56a7ecdf8d +size 504249 diff --git a/vlm/train/HkeSdCEtDS/6.png b/vlm/train/HkeSdCEtDS/6.png new file mode 100644 index 0000000000000000000000000000000000000000..0aceb8f9352b737a3d851b8452fd9c2909784ef3 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0c6884c8aca779d44a24b1c3b8a57a60d675932be03f15b5a2aff91741fc7e71 +size 469775 diff --git a/vlm/train/HkeSdCEtDS/7.png b/vlm/train/HkeSdCEtDS/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8981143ff3994814bc9ce096ebdb3aa111d7a8de --- /dev/null +++ b/vlm/train/HkeSdCEtDS/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae81b7d3993e8e8ee5a492ff282d134cb6f070f7e9600ba82ed7ab4b056a4482 +size 544024 diff --git a/vlm/train/HkeSdCEtDS/8.png b/vlm/train/HkeSdCEtDS/8.png new file mode 100644 index 0000000000000000000000000000000000000000..9f6ce7ff745147c0cac4c4e4f0f9123b058a5461 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f225c396182d32258b951128a08610e775389fd7fcf01c779258c8c9b224f94a +size 611819 diff --git a/vlm/train/HkeSdCEtDS/9.png b/vlm/train/HkeSdCEtDS/9.png new file mode 100644 index 0000000000000000000000000000000000000000..73ff5b0e01b2790305c2e151cece194ba3a96806 --- /dev/null +++ b/vlm/train/HkeSdCEtDS/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65ac38ca05add0c88b5a85200db1f165a84adcd343fb5bc043a12313a453f893 +size 635772 diff --git a/vlm/train/HklBjCEKvH/0.png b/vlm/train/HklBjCEKvH/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c7fd4fa06bd190eebfe67e1825bc869db43ed3be --- /dev/null +++ b/vlm/train/HklBjCEKvH/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f84e4a30ceb3915240a28bc23d917a2511dd5194f879076f65a56a9f702b7175 +size 494058 diff --git a/vlm/train/HklBjCEKvH/1.png b/vlm/train/HklBjCEKvH/1.png new file mode 100644 index 0000000000000000000000000000000000000000..d488a1236fdbb1704f8c900e227ad66eaf74624b --- /dev/null +++ b/vlm/train/HklBjCEKvH/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11233e9f977dce5aecb431390cf4ef3c3a1001f45e1841f9ea66b4c08f4408f3 +size 503954 diff --git a/vlm/train/HklBjCEKvH/10.png b/vlm/train/HklBjCEKvH/10.png new file mode 100644 index 0000000000000000000000000000000000000000..a308a0c6cdcc1b443ea7fcf93f67d825b5076eb4 --- /dev/null +++ b/vlm/train/HklBjCEKvH/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7686a7b4baa8e7f8d1aa812cd5f5355bd8239ba56531d34ac3d0c96cfb320bec +size 208643 diff --git a/vlm/train/HklBjCEKvH/11.png b/vlm/train/HklBjCEKvH/11.png new file mode 100644 index 0000000000000000000000000000000000000000..3db61e05c816fe9d1529d215c00d1df156098f4a --- /dev/null +++ b/vlm/train/HklBjCEKvH/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3df06ba5ba995356fbc8f6d9719531a0fa806d890783818e5f8f260d06689461 +size 429645 diff --git a/vlm/train/HklBjCEKvH/12.png b/vlm/train/HklBjCEKvH/12.png new file mode 100644 index 0000000000000000000000000000000000000000..dcba678e0a13499e099ef889bad430186a2e59bf --- /dev/null +++ b/vlm/train/HklBjCEKvH/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0365642a334c1593952334dfea9a294a66f83a8be77ae71cf99f2352a981b4dd +size 389067 diff --git a/vlm/train/HklBjCEKvH/2.png b/vlm/train/HklBjCEKvH/2.png new file mode 100644 index 0000000000000000000000000000000000000000..b63677c7edcf9141cde339e1b5c16a1f334ba34c --- /dev/null +++ b/vlm/train/HklBjCEKvH/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a021bc652bcb1657beb43a423856e08f0163b350a86ceb3db9482bec036d4f6 +size 558001 diff --git a/vlm/train/HklBjCEKvH/3.png b/vlm/train/HklBjCEKvH/3.png new file mode 100644 index 0000000000000000000000000000000000000000..dc51321b03a753afc8f17a36339dd16e8c181f30 --- /dev/null +++ b/vlm/train/HklBjCEKvH/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:613b832f3d89cd4d6bf20c06da09844fc2a977ee953a43b7898f7660d610cc82 +size 461921 diff --git a/vlm/train/HklBjCEKvH/4.png b/vlm/train/HklBjCEKvH/4.png new file mode 100644 index 0000000000000000000000000000000000000000..528fb9f9ce189aa6ca4ac7340d2bbe7dac2881cd --- /dev/null +++ b/vlm/train/HklBjCEKvH/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72476ef0c745d5f060b2c136b773466413a27d570c2abe2741d8d4181f6ee6ab +size 423699 diff --git a/vlm/train/HklBjCEKvH/5.png b/vlm/train/HklBjCEKvH/5.png new file mode 100644 index 0000000000000000000000000000000000000000..47d4133f086ea279796360d58d2752c01da75faf --- /dev/null +++ b/vlm/train/HklBjCEKvH/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a0b5075ca1c5a5df59178e8497d1385b88e24dcf35b0f48a937d0ecaec95bdd +size 439947 diff --git a/vlm/train/HklBjCEKvH/6.png b/vlm/train/HklBjCEKvH/6.png new file mode 100644 index 0000000000000000000000000000000000000000..a8c39be2b06cdd233377d37913d825b64cc5e65f --- /dev/null +++ b/vlm/train/HklBjCEKvH/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33594db2ad280458182a745b5d2ec89708e2a4c8af23d9b4956c04097009c1f2 +size 523235 diff --git a/vlm/train/HklBjCEKvH/7.png b/vlm/train/HklBjCEKvH/7.png new file mode 100644 index 0000000000000000000000000000000000000000..d829ad84cd71e99d424ccdd6b023728da15ec45d --- /dev/null +++ b/vlm/train/HklBjCEKvH/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e402216d91d9a5e6810251e0504995edef48c1fd1b7f73c8ae6a3aec912af42a +size 437800 diff --git a/vlm/train/HklBjCEKvH/8.png b/vlm/train/HklBjCEKvH/8.png new file mode 100644 index 0000000000000000000000000000000000000000..1af08c235b30e8c5b6fe96e8dcb06044e5ea0a4f --- /dev/null +++ b/vlm/train/HklBjCEKvH/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:396bc8ff4766b8e18e1bf2567f008a7db561b4b81d5d61aa4e66a4ede4ec6610 +size 492119 diff --git a/vlm/train/HklBjCEKvH/9.png b/vlm/train/HklBjCEKvH/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d1b38d7f7597d7aaced06d8ebfd0e69f0f79f167 --- /dev/null +++ b/vlm/train/HklBjCEKvH/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24f350e8897e31a93905505773779e433abdd86319bd8d76ddc17f799da32cb7 +size 515285 diff --git a/vlm/train/HkxJpnVtPr/0.png b/vlm/train/HkxJpnVtPr/0.png new file mode 100644 index 0000000000000000000000000000000000000000..20a94c6a54383f29f2abb33c58ec54da1b8448d4 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5e7be0e304f1db0b1daa8d7abd6e87e542256147d17f57c56376ca76f714d28b +size 507628 diff --git a/vlm/train/HkxJpnVtPr/1.png b/vlm/train/HkxJpnVtPr/1.png new file mode 100644 index 0000000000000000000000000000000000000000..f519976a353789dcbaf8430317d1b3ff9bf1a23b --- /dev/null +++ b/vlm/train/HkxJpnVtPr/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2988f37997eb86674254056a385d627f18603392d24c3aca57d0f4065c63d4d5 +size 633981 diff --git a/vlm/train/HkxJpnVtPr/10.png b/vlm/train/HkxJpnVtPr/10.png new file mode 100644 index 0000000000000000000000000000000000000000..12a1b759783dfdd877c7e03f0178718e65263eac --- /dev/null +++ b/vlm/train/HkxJpnVtPr/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc4d117f05320adc592b12f6d5b4e34cbd0a0ccfe18ffb737c77128860b0b46f +size 552502 diff --git a/vlm/train/HkxJpnVtPr/11.png b/vlm/train/HkxJpnVtPr/11.png new file mode 100644 index 0000000000000000000000000000000000000000..657daff3a0d6eef470705e39b77807de375f66e3 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa31dacd161439687c8864718ce3edbce1452250536d6e87c9fdb9697404d3d0 +size 308666 diff --git a/vlm/train/HkxJpnVtPr/12.png b/vlm/train/HkxJpnVtPr/12.png new file mode 100644 index 0000000000000000000000000000000000000000..edfd708c3299aa40c809a11726d620a1ea7fb910 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9cf2217a3d774bfc38e66e829b5566fbf7d88ed06ac600dcec6abdd6c552d37 +size 307970 diff --git a/vlm/train/HkxJpnVtPr/13.png b/vlm/train/HkxJpnVtPr/13.png new file mode 100644 index 0000000000000000000000000000000000000000..5a7abe1ed89dceb6ba1be50186dac796b037f474 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:602ac043a90ed5b47db294949be48069a0dbbcfc564b9bc8f6f50505de8997b4 +size 340264 diff --git a/vlm/train/HkxJpnVtPr/14.png b/vlm/train/HkxJpnVtPr/14.png new file mode 100644 index 0000000000000000000000000000000000000000..7681bc4867434f844a3cd1a9bcfa903daa71b21e --- /dev/null +++ b/vlm/train/HkxJpnVtPr/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:06b2f8c9f67e347251ff12776864933e00d6a57c80838fc6a3e271ccbff6a2fa +size 338976 diff --git a/vlm/train/HkxJpnVtPr/15.png b/vlm/train/HkxJpnVtPr/15.png new file mode 100644 index 0000000000000000000000000000000000000000..845de3ce93116cfe0c31a4b40989a2dfd2037cf9 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4aaeed683764054697e6d04a4a1a992b4ade9ff8884b051613c4894faa19ed7 +size 363855 diff --git a/vlm/train/HkxJpnVtPr/16.png b/vlm/train/HkxJpnVtPr/16.png new file mode 100644 index 0000000000000000000000000000000000000000..0ac20cb32930d7c938a7bd6fcf666854577cb4dd --- /dev/null +++ b/vlm/train/HkxJpnVtPr/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ff0b987a4d6cbb9a96479a2aa6c9f22d16d76243fe1c7bbe877c63d20bd2ec4 +size 374879 diff --git a/vlm/train/HkxJpnVtPr/17.png b/vlm/train/HkxJpnVtPr/17.png new file mode 100644 index 0000000000000000000000000000000000000000..14123f892fa190b5601f7997c7744a99b27842c8 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d835edfbce2305380155c88a508d91301e986a24552af539379805ffc46fb551 +size 531748 diff --git a/vlm/train/HkxJpnVtPr/18.png b/vlm/train/HkxJpnVtPr/18.png new file mode 100644 index 0000000000000000000000000000000000000000..57f914b6515eb8b186d8afa2c30d07c8d8d054df --- /dev/null +++ b/vlm/train/HkxJpnVtPr/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c41d3dd5602b8b7bafac008bf316ef21858e5b73a93e207b3a10b8df09a87903 +size 504727 diff --git a/vlm/train/HkxJpnVtPr/19.png b/vlm/train/HkxJpnVtPr/19.png new file mode 100644 index 0000000000000000000000000000000000000000..5788c6f25e304bbd552c1d378d1a78548646c7ed --- /dev/null +++ b/vlm/train/HkxJpnVtPr/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e70c2d3eadc59f17139462162fb2829c131456ddf426cb9b7f346b7ffedf6739 +size 504759 diff --git a/vlm/train/HkxJpnVtPr/2.png b/vlm/train/HkxJpnVtPr/2.png new file mode 100644 index 0000000000000000000000000000000000000000..e9585965f860dee29859e2c32b484cca7a81fb04 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36ca7b13b2ffaaa102594a2544db00bebc6bcbb704b5cd2c9490b88203f95ca8 +size 418685 diff --git a/vlm/train/HkxJpnVtPr/20.png b/vlm/train/HkxJpnVtPr/20.png new file mode 100644 index 0000000000000000000000000000000000000000..3dc879c6705d63d49d2fbc67228ee764bda2abb3 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec955b231f1a27300c7ccaf1388819178289ea608e19c54187e25f7d6953b4ce +size 486156 diff --git a/vlm/train/HkxJpnVtPr/21.png b/vlm/train/HkxJpnVtPr/21.png new file mode 100644 index 0000000000000000000000000000000000000000..d141f8f9c6704d455f96de75d555d40c4a1071ed --- /dev/null +++ b/vlm/train/HkxJpnVtPr/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d97f6c6341c4f429e8c4c9ec15a00134e7e97cea63e7da6dd7bf4d3c9b43cc1 +size 235238 diff --git a/vlm/train/HkxJpnVtPr/3.png b/vlm/train/HkxJpnVtPr/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b573fd7749c5fb0fc30cf2328854e71efec7f2d4 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efc5033b7b6d1c98214b7193bad349de00ca9af4f4b5f0260ba57aa638cf31a4 +size 524683 diff --git a/vlm/train/HkxJpnVtPr/4.png b/vlm/train/HkxJpnVtPr/4.png new file mode 100644 index 0000000000000000000000000000000000000000..d67a5ad0e40ff1414449c00e02877eaa15a05437 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:318f3868576156d858382d0ffe5820d9fa8b679ef003aada1837e9994cd9660f +size 528269 diff --git a/vlm/train/HkxJpnVtPr/5.png b/vlm/train/HkxJpnVtPr/5.png new file mode 100644 index 0000000000000000000000000000000000000000..a24dca8ba8ba6eb454884528e3ab3e235c3e0418 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28061cc5bfe1a128d05d122f773e1c826766e0010e9109889aa3a7440a317631 +size 476928 diff --git a/vlm/train/HkxJpnVtPr/6.png b/vlm/train/HkxJpnVtPr/6.png new file mode 100644 index 0000000000000000000000000000000000000000..03b2ff8f22ce87ab5089410447ae7a5f952d8342 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af023b8cba7701c90c7b8fb562a1352eef1bed827992d952d3bd46e82073e982 +size 589179 diff --git a/vlm/train/HkxJpnVtPr/7.png b/vlm/train/HkxJpnVtPr/7.png new file mode 100644 index 0000000000000000000000000000000000000000..b0f5de7d8bd6e958f3cddcccec030bf283e48979 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fcded031d22e6f37aab9d42e1b13a1534c4966e72ff8b9109d611f668a622da +size 507083 diff --git a/vlm/train/HkxJpnVtPr/8.png b/vlm/train/HkxJpnVtPr/8.png new file mode 100644 index 0000000000000000000000000000000000000000..33040b57011889096d1224506aafe825179cdd39 --- /dev/null +++ b/vlm/train/HkxJpnVtPr/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6e951a16ed732045b208982a23af05bcaffcfcb231a3db9a62bb3499c23aa85 +size 667223 diff --git a/vlm/train/HkxJpnVtPr/9.png b/vlm/train/HkxJpnVtPr/9.png new file mode 100644 index 0000000000000000000000000000000000000000..6cd5d2aaf589397e25760f25c487580cd7eefc7c --- /dev/null +++ b/vlm/train/HkxJpnVtPr/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b949d722f724336afb8248089bebcd0bc1a2378add7ebb0f3fa726d68d78935 +size 659250 diff --git a/vlm/train/JXAyJeYqUkZ/0.png b/vlm/train/JXAyJeYqUkZ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..4bcf9143bdb0b0baf9fe186c9bb781cfe00f52d4 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36ab2fbaa4ef692d492b7ca7b9fb39973730d5332818c97d16c619a6b88b5126 +size 477652 diff --git a/vlm/train/JXAyJeYqUkZ/1.png b/vlm/train/JXAyJeYqUkZ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0bb23c5ad0d7097e00851b7bb5a64d90d9fcdec1 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f0665212e1b463f42490c5c1cca9046b201d18eab981156a6eb60ebc9cbf99d +size 510444 diff --git a/vlm/train/JXAyJeYqUkZ/10.png b/vlm/train/JXAyJeYqUkZ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..4c698574c11e0e655b289017e04b1e19e5880dfe --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e41d8eb49c670d34b0a2779c9b55568a64491ac4b1556672303541b4af92cee0 +size 571871 diff --git a/vlm/train/JXAyJeYqUkZ/11.png b/vlm/train/JXAyJeYqUkZ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..8d6e33e67ebfc32fc13f3a32aa30f6824bbf1dcd --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8465a1022681078c875ee268360a52c8e7a338bcb77a213318351873fd8fa42 +size 610500 diff --git a/vlm/train/JXAyJeYqUkZ/12.png b/vlm/train/JXAyJeYqUkZ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..9ae72a001c912c761c2268fd2471775c8b719ff2 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4d3d923d295171e223246d5cefec86115fcf93156c022a610d0ecb4326d88c9 +size 424486 diff --git a/vlm/train/JXAyJeYqUkZ/2.png b/vlm/train/JXAyJeYqUkZ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c3f5c335528b7ab7cc091c7a2ff682a1a55626fd --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1df33e5e63c4e19d74294000527f79b6156258825f8f3fdd5ff217010277af6 +size 497326 diff --git a/vlm/train/JXAyJeYqUkZ/3.png b/vlm/train/JXAyJeYqUkZ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..9a452cbc657e984cb19dd5bafabb8f4705f012e8 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5f5aa8071ba0c2a7594eab44e2488f149cd2ae4b0f3453e1f341c1932158b33 +size 561090 diff --git a/vlm/train/JXAyJeYqUkZ/4.png b/vlm/train/JXAyJeYqUkZ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..2a3866044c0bd798d8ea6ab12604881a09ac3b57 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecafe1da74ead835c54a4da477a5391d1977b184108761301c86a4c8412b511e +size 495767 diff --git a/vlm/train/JXAyJeYqUkZ/5.png b/vlm/train/JXAyJeYqUkZ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8b0c0bb9498533e799132ee2f1efb00743c21734 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47ec1ccb246c2ee41943d668a9f898814da448fac015fa334cc62e73f234a6fd +size 561719 diff --git a/vlm/train/JXAyJeYqUkZ/6.png b/vlm/train/JXAyJeYqUkZ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..e39ef5a7faf54a0e703b7fc58eb16c50e607c8c5 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:86c9b4d22ed93f99b943c7f670857c1b5a62ab4f9ab778e556f9aa5fa3e1ff8d +size 500720 diff --git a/vlm/train/JXAyJeYqUkZ/7.png b/vlm/train/JXAyJeYqUkZ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..4ba3f7f20e129a2b7eb0272b46eef8ad03930391 --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b50e6879e71c7108a39b6fe100506f7408052b657c4e78bfd7d5c531b1f9d941 +size 462438 diff --git a/vlm/train/JXAyJeYqUkZ/8.png b/vlm/train/JXAyJeYqUkZ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..f039c234d6c8df37165738d6feac4e236d5ad1df --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6adf1be77db11b807b4a50db45f681dd3d8c4d5915285f5905a27d1c2dfe685 +size 482561 diff --git a/vlm/train/JXAyJeYqUkZ/9.png b/vlm/train/JXAyJeYqUkZ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..45c7d311b3919903e85c6f3034f147758657034b --- /dev/null +++ b/vlm/train/JXAyJeYqUkZ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f46dbf04bf0831d46c93faf178561cd777245739886d6e28557d549912d55c3 +size 332882 diff --git a/vlm/train/Kloou2uk_Rz/0.png b/vlm/train/Kloou2uk_Rz/0.png new file mode 100644 index 0000000000000000000000000000000000000000..63952c8497099183c781314be0d9599719980278 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1de2592d5326098a5c05574607e6daf68ef54f420394d0b7e62a661cd8e7916f +size 430187 diff --git a/vlm/train/Kloou2uk_Rz/1.png b/vlm/train/Kloou2uk_Rz/1.png new file mode 100644 index 0000000000000000000000000000000000000000..0233cecd263507fd0c4c97f36f52eeb96c6ad919 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:99b3b9459eede77bee722084af6ce7de3bd254b623a2433eef8972bf21676290 +size 676778 diff --git a/vlm/train/Kloou2uk_Rz/10.png b/vlm/train/Kloou2uk_Rz/10.png new file mode 100644 index 0000000000000000000000000000000000000000..54e1070743b4ad32176c1100730197474ed1f012 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ad4ca048cf833d5530271f3440a3cb8b971e4dd92e648e59299f0b94b4bc76d +size 595949 diff --git a/vlm/train/Kloou2uk_Rz/11.png b/vlm/train/Kloou2uk_Rz/11.png new file mode 100644 index 0000000000000000000000000000000000000000..50cbc6d8a47eb8aa4aa4282c0982fa5370e4d9bc --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a97af169f71e8798119afac69fe55fb445ee93f740ea67c4bbc1e3c8a7864c1 +size 531763 diff --git a/vlm/train/Kloou2uk_Rz/2.png b/vlm/train/Kloou2uk_Rz/2.png new file mode 100644 index 0000000000000000000000000000000000000000..1b7ec68211cdfa0dc3890c874ed45e83a8eb1a3d --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d92e3988b947004342026c43c7c96cd74610df221c9fa81524e70536cbe23028 +size 633096 diff --git a/vlm/train/Kloou2uk_Rz/3.png b/vlm/train/Kloou2uk_Rz/3.png new file mode 100644 index 0000000000000000000000000000000000000000..6b4c0499c7cb25f5cc8153213e920f20b9a01b15 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:478ce17a9a22effd1562bd74402a855cfa2d0a72cc9532cad791580a5a66a1a1 +size 550706 diff --git a/vlm/train/Kloou2uk_Rz/4.png b/vlm/train/Kloou2uk_Rz/4.png new file mode 100644 index 0000000000000000000000000000000000000000..767cd13ee2148f1628df6c0896d81c53266d1472 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ef5cbebb93a8d6271455df8b1c55e5344ad5a1b22d046d261bba1e83e0133106 +size 584914 diff --git a/vlm/train/Kloou2uk_Rz/5.png b/vlm/train/Kloou2uk_Rz/5.png new file mode 100644 index 0000000000000000000000000000000000000000..0be69d8714083efc7ac60b4bf666c32abdf299b0 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44bedc6bd28a859dd825b60051dd3f07808f7c73de0a00df05c02f3eb3ec0e33 +size 616176 diff --git a/vlm/train/Kloou2uk_Rz/6.png b/vlm/train/Kloou2uk_Rz/6.png new file mode 100644 index 0000000000000000000000000000000000000000..754581dee7c6a8009733689e3593d8cd44929ad4 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3205670496691e81a844308798f80cf4ff46e39e3ec05b2dfa405cd1e2982d85 +size 675498 diff --git a/vlm/train/Kloou2uk_Rz/7.png b/vlm/train/Kloou2uk_Rz/7.png new file mode 100644 index 0000000000000000000000000000000000000000..9371d02b6dba906d4f461a788757b8a3d94e1229 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72d639c7a3b8715d2ff7343349d4d6d17a5815a14f105b60678264e336e03fe6 +size 676608 diff --git a/vlm/train/Kloou2uk_Rz/8.png b/vlm/train/Kloou2uk_Rz/8.png new file mode 100644 index 0000000000000000000000000000000000000000..220fa1f2a617b00ba0a0301d5a14ae01ab44dd82 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bb01deae5e424c5c37b7288e3a74991396f816b41f35d0bd6877c150a8d0544 +size 636109 diff --git a/vlm/train/Kloou2uk_Rz/9.png b/vlm/train/Kloou2uk_Rz/9.png new file mode 100644 index 0000000000000000000000000000000000000000..a123e678025880ec272f6a13fd750319f80033f0 --- /dev/null +++ b/vlm/train/Kloou2uk_Rz/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6dc1cb10357311701a325259c7578914ae2cec7efd6512371f9d9f16707beab +size 530560 diff --git a/vlm/train/LLoe0U9ShkN/0.png b/vlm/train/LLoe0U9ShkN/0.png new file mode 100644 index 0000000000000000000000000000000000000000..950f2a8c0e7947271d7e362167e8d78f94295bd3 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9f1076b916f049af7a73f8717b8c6b0e9e0a0cde40b6a826ff7672ca9542487 +size 502325 diff --git a/vlm/train/LLoe0U9ShkN/1.png b/vlm/train/LLoe0U9ShkN/1.png new file mode 100644 index 0000000000000000000000000000000000000000..b1345ad60694d57e618c54dbb70642a84becc2ea --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:19b8fffa269f511a0dfd1d4adb5f76eec49b7e0acd3d7837274fd1197998d213 +size 479097 diff --git a/vlm/train/LLoe0U9ShkN/10.png b/vlm/train/LLoe0U9ShkN/10.png new file mode 100644 index 0000000000000000000000000000000000000000..7d0d428cf392b448fe23f0749109cee14d9ebcfc --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1afc0ade0b29f6480c33df521517a3c43cbe887c77e407aac71e1ad827ec761b +size 499840 diff --git a/vlm/train/LLoe0U9ShkN/11.png b/vlm/train/LLoe0U9ShkN/11.png new file mode 100644 index 0000000000000000000000000000000000000000..daac98c2638670e89afe2cec9f0e4b1ae1c17772 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64feb888e35b063951db1d26bac5e73e2c20bc4e79184cb53631ea95cdbd06c1 +size 404251 diff --git a/vlm/train/LLoe0U9ShkN/12.png b/vlm/train/LLoe0U9ShkN/12.png new file mode 100644 index 0000000000000000000000000000000000000000..c6c65b7c04eb3bfc4e29533466dbc34ddc087578 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:204b7f5c0a8ae703e3a7d019614e8cb070eb9f40b7b71b37721b1faf36351581 +size 406559 diff --git a/vlm/train/LLoe0U9ShkN/13.png b/vlm/train/LLoe0U9ShkN/13.png new file mode 100644 index 0000000000000000000000000000000000000000..2de702c91757e933b3d7e0bdfbc55b79888ff1da --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b7da153213226f723bd0517928b24145e6ef849b3ef5131e512a42544309c3b +size 259009 diff --git a/vlm/train/LLoe0U9ShkN/14.png b/vlm/train/LLoe0U9ShkN/14.png new file mode 100644 index 0000000000000000000000000000000000000000..ee3dd4f5c8c882d36db8fe60d8991da930a93585 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aa7f015b5e438484ff8f31e3eced5f592438df74157e6055acb9e79368b2490 +size 369925 diff --git a/vlm/train/LLoe0U9ShkN/15.png b/vlm/train/LLoe0U9ShkN/15.png new file mode 100644 index 0000000000000000000000000000000000000000..b304046544561a2a7c73df1c7cdebb71d09908e9 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88672ccb88036412a52cb88d9736683efb14542c15d339c2f37b85f12fd70262 +size 420479 diff --git a/vlm/train/LLoe0U9ShkN/16.png b/vlm/train/LLoe0U9ShkN/16.png new file mode 100644 index 0000000000000000000000000000000000000000..069deed94975d96dce339017872762b55330c294 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d1c55a330a3724cf62576eb21034201ec44d7a92d4c0a1191eaf55109fc713c2 +size 501153 diff --git a/vlm/train/LLoe0U9ShkN/17.png b/vlm/train/LLoe0U9ShkN/17.png new file mode 100644 index 0000000000000000000000000000000000000000..b04d6fc90fba5fca924cbad75d0a8baa636015d2 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c7a10243ce576a3824db9664539f8f58f795e665b19c8f436d7893dfae99d8a +size 270953 diff --git a/vlm/train/LLoe0U9ShkN/18.png b/vlm/train/LLoe0U9ShkN/18.png new file mode 100644 index 0000000000000000000000000000000000000000..ab515015284bfe2c0f846fa50239723349ce6a5d --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:513aee0ee0b6bcabb9fcd9ba4e922aeb0026395e0544b75d0311ad1abead751f +size 451138 diff --git a/vlm/train/LLoe0U9ShkN/19.png b/vlm/train/LLoe0U9ShkN/19.png new file mode 100644 index 0000000000000000000000000000000000000000..082d37e820958931dce4a2b0cf8953caae6197d8 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7a339c42663423d353aed053ae6ce6fdccff72eac9079083d978e4ff87cc93d +size 398157 diff --git a/vlm/train/LLoe0U9ShkN/2.png b/vlm/train/LLoe0U9ShkN/2.png new file mode 100644 index 0000000000000000000000000000000000000000..b8474f83753e6603856673968eebd3e759e3656a --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f6f549864dd5b5b203cbefa4145e453a056a64e2849040087c440f95db00058 +size 447955 diff --git a/vlm/train/LLoe0U9ShkN/20.png b/vlm/train/LLoe0U9ShkN/20.png new file mode 100644 index 0000000000000000000000000000000000000000..5423f8e669ee35b1b46e92dc966d06eac1109cdf --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1119ea68b5b276673e975684fec92db573b5876d51a29e4eb4756de9876185ef +size 436447 diff --git a/vlm/train/LLoe0U9ShkN/21.png b/vlm/train/LLoe0U9ShkN/21.png new file mode 100644 index 0000000000000000000000000000000000000000..38d9beb232eedb464d68e58c91a47d2519564e79 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc97b234710bdc601ba3281df96bc4cdc2dcddad005f1d7e96cd51ade78ba537 +size 535761 diff --git a/vlm/train/LLoe0U9ShkN/22.png b/vlm/train/LLoe0U9ShkN/22.png new file mode 100644 index 0000000000000000000000000000000000000000..bef75e0bd343d549d081670b71b203ae0dc49300 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85c4a20ec9e803f3c2d67efeca79cb53c6caf3376f9600bc5ae2237a818e95cc +size 484248 diff --git a/vlm/train/LLoe0U9ShkN/23.png b/vlm/train/LLoe0U9ShkN/23.png new file mode 100644 index 0000000000000000000000000000000000000000..1cc2c6f7c49bb0cd1b10b190bf5c0e346f2226de --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6806583f1c5bf855f532460fb7e0d5e74051c45ff7fb6086a625ac61f7164814 +size 463412 diff --git a/vlm/train/LLoe0U9ShkN/24.png b/vlm/train/LLoe0U9ShkN/24.png new file mode 100644 index 0000000000000000000000000000000000000000..61cf98eb80ac4bf80488572c48057928c8dca385 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcc27821fae130db256d1e027d240e30517f546a7a01df62c86bdc7fa324db0b +size 432642 diff --git a/vlm/train/LLoe0U9ShkN/25.png b/vlm/train/LLoe0U9ShkN/25.png new file mode 100644 index 0000000000000000000000000000000000000000..5bac6c5f7131252424050c6b1da50cd78867742b --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc8b0ed536d2ac3c9cd2803411608b65c755d094ced381b8c1bac3025c59349f +size 453641 diff --git a/vlm/train/LLoe0U9ShkN/26.png b/vlm/train/LLoe0U9ShkN/26.png new file mode 100644 index 0000000000000000000000000000000000000000..525c75cb5ca97a125b404ac3e007fca5963ba79b --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60ca0ffa322d697980460b47eaf9e7c01fa9815b98cb3271aef3076e2f82cbe0 +size 289450 diff --git a/vlm/train/LLoe0U9ShkN/27.png b/vlm/train/LLoe0U9ShkN/27.png new file mode 100644 index 0000000000000000000000000000000000000000..19dbd616d14eb2ff5c0250fa21ef2ac7fed9b482 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69d729736056a4a78042b5665f776b59c30a294400cdfbe6a82447b05e865dfd +size 266322 diff --git a/vlm/train/LLoe0U9ShkN/28.png b/vlm/train/LLoe0U9ShkN/28.png new file mode 100644 index 0000000000000000000000000000000000000000..616a4cafaec60b358308f12d62756d863b8cf2e4 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9668144d2360835e712d78322dd577cf75df9da3cd77072ef5f986a0eb0e15b6 +size 271709 diff --git a/vlm/train/LLoe0U9ShkN/29.png b/vlm/train/LLoe0U9ShkN/29.png new file mode 100644 index 0000000000000000000000000000000000000000..29cac54786c15b7a62a8374efa3e547f08815c03 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81918ddd31138c65d0ae68a085e1d8b635d5dc17965180428cbdabdee4ae46e5 +size 279480 diff --git a/vlm/train/LLoe0U9ShkN/3.png b/vlm/train/LLoe0U9ShkN/3.png new file mode 100644 index 0000000000000000000000000000000000000000..71bd3fef10619e736dce162fdf603c1b962f83a6 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:645645566a4fdf98cccf38247781de2c30c28b1db841f510efe4ef4d30d90e29 +size 470279 diff --git a/vlm/train/LLoe0U9ShkN/30.png b/vlm/train/LLoe0U9ShkN/30.png new file mode 100644 index 0000000000000000000000000000000000000000..d55ae6b6be91c7ad89b62dd4e8590134b787023e --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/30.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a806bf1cc90836cf761e79c9ae4f501872d020966368a77abbd9c924a3a481c2 +size 256722 diff --git a/vlm/train/LLoe0U9ShkN/31.png b/vlm/train/LLoe0U9ShkN/31.png new file mode 100644 index 0000000000000000000000000000000000000000..03c8eb83c4ac12312cb2659c698ee500c1382d77 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/31.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af416d211377ae39561054b3f00ef99bb0358664accef6e6ab02e0cec7c38009 +size 254782 diff --git a/vlm/train/LLoe0U9ShkN/4.png b/vlm/train/LLoe0U9ShkN/4.png new file mode 100644 index 0000000000000000000000000000000000000000..0c964dd31b7b721b0eb9a34b9c705474cbd59e2b --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d394f989f2bc1baea38b3c5fde9874cd6117fa32823f15ae6fc8716130e077c +size 474881 diff --git a/vlm/train/LLoe0U9ShkN/5.png b/vlm/train/LLoe0U9ShkN/5.png new file mode 100644 index 0000000000000000000000000000000000000000..0364738a4e873487516af2f9573454fbcb073fc3 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b614049a110c5827ef6b95b7c2bf0fbb7a2bf8fc367eaea7cb3ce54978807e4 +size 429894 diff --git a/vlm/train/LLoe0U9ShkN/6.png b/vlm/train/LLoe0U9ShkN/6.png new file mode 100644 index 0000000000000000000000000000000000000000..f3a6661940ad212aa7f1b4e601a6469734d6c167 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59ce3879b0e45759f7bcccecd1a083fec48f31f661b5c393155dc466f6adae20 +size 429528 diff --git a/vlm/train/LLoe0U9ShkN/7.png b/vlm/train/LLoe0U9ShkN/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8f177949b6d17ed027f1b7c7341227aaa7971c83 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e735c921c9e2e1c8b6453af59f7111a50731e23425a9fcb5dab947e7477ba6d6 +size 539282 diff --git a/vlm/train/LLoe0U9ShkN/8.png b/vlm/train/LLoe0U9ShkN/8.png new file mode 100644 index 0000000000000000000000000000000000000000..61af06df07b0896ebd3947d2bd28efa3545ac29c --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10999120d42330e99f6901bbe688267671b7abad051f967dd230db43ed2fda3d +size 537219 diff --git a/vlm/train/LLoe0U9ShkN/9.png b/vlm/train/LLoe0U9ShkN/9.png new file mode 100644 index 0000000000000000000000000000000000000000..ebe68293555b3d7bf090c020ceb28454f8f0cb83 --- /dev/null +++ b/vlm/train/LLoe0U9ShkN/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2da40d8c6cd5bc54007b6c43a66c5553406e0a2a6b6afdc17d5871a92d06f26 +size 536302 diff --git a/vlm/train/LyjH88yV7F/0.png b/vlm/train/LyjH88yV7F/0.png new file mode 100644 index 0000000000000000000000000000000000000000..fffd80b1b2a1df2ee81e5ddf97772c54ce4ec940 --- /dev/null +++ b/vlm/train/LyjH88yV7F/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7215ee8633ac884bbdd570a5b722b4c6a271fb720f40fbf3607308b1cd07fc9 +size 413432 diff --git a/vlm/train/LyjH88yV7F/1.png b/vlm/train/LyjH88yV7F/1.png new file mode 100644 index 0000000000000000000000000000000000000000..5112ddb6e534dc1d75b0c75a76a9f07437f5b769 --- /dev/null +++ b/vlm/train/LyjH88yV7F/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c1651a3781213a974af288fdbee8b83c3f38717f77ac5e6540f15aebf6caad5 +size 473283 diff --git a/vlm/train/LyjH88yV7F/10.png b/vlm/train/LyjH88yV7F/10.png new file mode 100644 index 0000000000000000000000000000000000000000..84c541aa68eec475ee802a64b8e697b5b93052b7 --- /dev/null +++ b/vlm/train/LyjH88yV7F/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9bdbee3bfbd031a55b80121287df8a4fb718eda05099381e91b700a13661d23 +size 404279 diff --git a/vlm/train/LyjH88yV7F/2.png b/vlm/train/LyjH88yV7F/2.png new file mode 100644 index 0000000000000000000000000000000000000000..ac8e925d86ad0dd88c3c4dd77b6521a06c3079f1 --- /dev/null +++ b/vlm/train/LyjH88yV7F/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2bf42b6bcf4152cbe2d443360f632d05f9aec6d3a33749c8e5084ab2b6090db +size 582841 diff --git a/vlm/train/LyjH88yV7F/3.png b/vlm/train/LyjH88yV7F/3.png new file mode 100644 index 0000000000000000000000000000000000000000..1dfe0146b0a14ca7b230c1f67ea8c09eed85d432 --- /dev/null +++ b/vlm/train/LyjH88yV7F/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8c0aa74114f9da4d1ad26ccc74e5355e0acad42828dce5ad12fba0e2d6c3202 +size 536645 diff --git a/vlm/train/LyjH88yV7F/4.png b/vlm/train/LyjH88yV7F/4.png new file mode 100644 index 0000000000000000000000000000000000000000..ec5f9075ccb3610368aeaf9d6e5d6a9c1e170d78 --- /dev/null +++ b/vlm/train/LyjH88yV7F/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2c1ab80e9e521184ea6f1b200d3dba3940de46061cca5a1bd04012cd9560529 +size 478176 diff --git a/vlm/train/LyjH88yV7F/5.png b/vlm/train/LyjH88yV7F/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7330e5ebe151b82592e20738f020a8c743d7fc5a --- /dev/null +++ b/vlm/train/LyjH88yV7F/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7aa98c54c416633b8dbb725c2d405c1d69f81843748af4a042ce5e0d0898f71 +size 473153 diff --git a/vlm/train/LyjH88yV7F/6.png b/vlm/train/LyjH88yV7F/6.png new file mode 100644 index 0000000000000000000000000000000000000000..62187c85a737c7e47068386f5ba5cb2d52aae181 --- /dev/null +++ b/vlm/train/LyjH88yV7F/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7c1e858b1ef3240b71c604934404b50544298cec253fd076bdaec646dbff991 +size 497887 diff --git a/vlm/train/LyjH88yV7F/7.png b/vlm/train/LyjH88yV7F/7.png new file mode 100644 index 0000000000000000000000000000000000000000..371819d650b45a33eabfee67b1b1c9a126a48ce2 --- /dev/null +++ b/vlm/train/LyjH88yV7F/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be0da925ace35eaaa79f965c212ccf281ed4c82ca6fa6bb770c2b99f2a538126 +size 492918 diff --git a/vlm/train/LyjH88yV7F/8.png b/vlm/train/LyjH88yV7F/8.png new file mode 100644 index 0000000000000000000000000000000000000000..d9789b0dd41008964c49a1c0563206d9ee426019 --- /dev/null +++ b/vlm/train/LyjH88yV7F/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:026b26ed3db1b8ab1fb0916723fe7901473133c51c00e71aa1bf537ada9edbab +size 386742 diff --git a/vlm/train/LyjH88yV7F/9.png b/vlm/train/LyjH88yV7F/9.png new file mode 100644 index 0000000000000000000000000000000000000000..60fb38772faff6c8f9ca0e098cf3da55864ed4f0 --- /dev/null +++ b/vlm/train/LyjH88yV7F/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82b543016c6b9c798cf07b70df1fb39cfbc61b22b70e4fcfccac71f5e4b56b08 +size 516785 diff --git a/vlm/train/N51zJ7F3mw/0.png b/vlm/train/N51zJ7F3mw/0.png new file mode 100644 index 0000000000000000000000000000000000000000..79d5f0ba558eb919a5e75544210ba1532fc02c36 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcb68caa6f0522d81c4411d75254ad762a107541cb8911325fd846c2bb285e27 +size 438397 diff --git a/vlm/train/N51zJ7F3mw/1.png b/vlm/train/N51zJ7F3mw/1.png new file mode 100644 index 0000000000000000000000000000000000000000..b512154b6686cdf555381165a60c440cf209e8cf --- /dev/null +++ b/vlm/train/N51zJ7F3mw/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:259e6c39aa123dc74048398bb3ad79ba6eac0fc2cfb2700641ebd3b326964af0 +size 592057 diff --git a/vlm/train/N51zJ7F3mw/10.png b/vlm/train/N51zJ7F3mw/10.png new file mode 100644 index 0000000000000000000000000000000000000000..0a24fdd7ac23895ebb19b1ec3521804e711139a5 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f1ec00f371b403305a7c170933b6fda7614c6515edfbf841c47a42ab9daf7c1 +size 499413 diff --git a/vlm/train/N51zJ7F3mw/11.png b/vlm/train/N51zJ7F3mw/11.png new file mode 100644 index 0000000000000000000000000000000000000000..188264acc2aeb08aa6433d24699534918e136242 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5a8fba25aef8e98a0797f29ac3101896761e6947ec80e1c9a68226a2f9b3db5 +size 169285 diff --git a/vlm/train/N51zJ7F3mw/2.png b/vlm/train/N51zJ7F3mw/2.png new file mode 100644 index 0000000000000000000000000000000000000000..6a2a7fe6ebc90fc476ed3cc13262f651f9d846f5 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d51c13af0797529da06287eabe56f627d3c49aa94e86ff6490bc5aeb071d2e8c +size 529584 diff --git a/vlm/train/N51zJ7F3mw/3.png b/vlm/train/N51zJ7F3mw/3.png new file mode 100644 index 0000000000000000000000000000000000000000..313934a7aefef9514e560e8692ca7daad919c555 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9c91907ddf959da937bee2a196151e5f4851bc447189dca8de223894690fe5a +size 471504 diff --git a/vlm/train/N51zJ7F3mw/4.png b/vlm/train/N51zJ7F3mw/4.png new file mode 100644 index 0000000000000000000000000000000000000000..3f3545ba7af15c444110bfc7ce5c6f7748f72c86 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17e43e19e861a3d377e5ba755cb738249a39fe477dd28d3ed819eae53824c8b4 +size 707765 diff --git a/vlm/train/N51zJ7F3mw/5.png b/vlm/train/N51zJ7F3mw/5.png new file mode 100644 index 0000000000000000000000000000000000000000..f8ddc8c7c58bf3e6aa74391772dc00ab32936a1d --- /dev/null +++ b/vlm/train/N51zJ7F3mw/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d56cfe5e1cbf3460b4adceb5d17d7cdb342ca553a5305444c49687b0dc17e387 +size 510377 diff --git a/vlm/train/N51zJ7F3mw/6.png b/vlm/train/N51zJ7F3mw/6.png new file mode 100644 index 0000000000000000000000000000000000000000..a80d6f5e56080b374ea8b79c2fdd8167efe51b6d --- /dev/null +++ b/vlm/train/N51zJ7F3mw/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd376d3457c82af9e9581ae4b83f311b213a4dd4ca3ab99e150364c68617c739 +size 524531 diff --git a/vlm/train/N51zJ7F3mw/7.png b/vlm/train/N51zJ7F3mw/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f4bdde990336b205271a06fe72d030dfe9189990 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91aa3ba9f916c6946a0eabe8b2f5a38240e0f85fd2b1fe168509a3ba4e8f5fbe +size 480887 diff --git a/vlm/train/N51zJ7F3mw/8.png b/vlm/train/N51zJ7F3mw/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8f18e42390008e41df3862c0571e396e674573d2 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25ef58f329e2f77ac107ea55ab6a0ab00145c773960fae0ba9d1d169bf61817b +size 569992 diff --git a/vlm/train/N51zJ7F3mw/9.png b/vlm/train/N51zJ7F3mw/9.png new file mode 100644 index 0000000000000000000000000000000000000000..c498f0e1af5f35b7bcc284025fbd6fd782e61470 --- /dev/null +++ b/vlm/train/N51zJ7F3mw/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5769e4242526a8533488a524dc81f96ba64203448f809881f40b0e5ea72962f4 +size 519764 diff --git a/vlm/train/NSBrFgJAHg/13.png b/vlm/train/NSBrFgJAHg/13.png new file mode 100644 index 0000000000000000000000000000000000000000..f01c5086099f2bfd640e4b437ed016cd4eb56a22 --- /dev/null +++ b/vlm/train/NSBrFgJAHg/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea2d762b64ac745addaeafe4b0a7940b0e5304851b40e5acd06515b721ca4c2a +size 509478 diff --git a/vlm/train/NSBrFgJAHg/16.png b/vlm/train/NSBrFgJAHg/16.png new file mode 100644 index 0000000000000000000000000000000000000000..1c4ce0559c08841a46dbb3c4bb364624af82c05b --- /dev/null +++ b/vlm/train/NSBrFgJAHg/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05a704a055ae26bf9b11af8655d52a4cc8d02fc510d1331c2e65e5357d447010 +size 458826 diff --git a/vlm/train/NSBrFgJAHg/18.png b/vlm/train/NSBrFgJAHg/18.png new file mode 100644 index 0000000000000000000000000000000000000000..5c392c23612b37e43a177f440a7350e5a93e951b --- /dev/null +++ b/vlm/train/NSBrFgJAHg/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0d76b064ee88b88da0572f1afb8de94b7e74f2c91dcfae202563db2379ae80e +size 163378 diff --git a/vlm/train/NSBrFgJAHg/19.png b/vlm/train/NSBrFgJAHg/19.png new file mode 100644 index 0000000000000000000000000000000000000000..f74be0163446ad2a2d3baf8aeca42341e9114b73 --- /dev/null +++ b/vlm/train/NSBrFgJAHg/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65d9b4f8f92a62dba4de8b9c8fd09136cb898d25942f704ba3f40013af4f041d +size 136380 diff --git a/vlm/train/Q2iaAc-4I1v/0.png b/vlm/train/Q2iaAc-4I1v/0.png new file mode 100644 index 0000000000000000000000000000000000000000..1e9b680c6b0f8dbe270a91d94951db67c25b25cc --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45df61a38ae809db742272ca877ccee0d999745feb4446376881104c31b1f2f1 +size 483327 diff --git a/vlm/train/Q2iaAc-4I1v/1.png b/vlm/train/Q2iaAc-4I1v/1.png new file mode 100644 index 0000000000000000000000000000000000000000..c62d499fbd91696e5d62fb27adf7fb56b9fac8d5 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d61e321041bb42dfc950158bc5bc44d69ba7181fb77b7a990c4fa3fd4438e90b +size 580619 diff --git a/vlm/train/Q2iaAc-4I1v/10.png b/vlm/train/Q2iaAc-4I1v/10.png new file mode 100644 index 0000000000000000000000000000000000000000..508f35a1e805fad4d1ee3cf39320800996bfbe10 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88add7194cdc89c8fe4d37b07939edb84d53ebc2c9aa5ec70b871ba1dd4a92b5 +size 314853 diff --git a/vlm/train/Q2iaAc-4I1v/11.png b/vlm/train/Q2iaAc-4I1v/11.png new file mode 100644 index 0000000000000000000000000000000000000000..e9b720e8cadf09927000110158517fe16c0a818e --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6df44a1465e30bb46a43d9ddde1a2f827474d48f63cc1b8076109ce3847bb2e9 +size 472984 diff --git a/vlm/train/Q2iaAc-4I1v/12.png b/vlm/train/Q2iaAc-4I1v/12.png new file mode 100644 index 0000000000000000000000000000000000000000..97928d79d38daa14ecb6922ac2f3da7a8cd097b0 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:920c7da5ce6c394ddb30a07e85f0992b0fd5cdef289d4bb601cd1409a4ac0424 +size 570607 diff --git a/vlm/train/Q2iaAc-4I1v/13.png b/vlm/train/Q2iaAc-4I1v/13.png new file mode 100644 index 0000000000000000000000000000000000000000..123634380b10064cdbbb220258c08b8fa604151a --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cd436cf724813822eecfb885321cb301bef023ea6d5ac65257e6b479ea893d0 +size 504352 diff --git a/vlm/train/Q2iaAc-4I1v/14.png b/vlm/train/Q2iaAc-4I1v/14.png new file mode 100644 index 0000000000000000000000000000000000000000..2d6fcf4c4802fbc54561c11f9b925f0266c7a034 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:717b60d6712267c3cac44a23dae05064553dd8f626f37ad6c1214fd3ad27480c +size 123811 diff --git a/vlm/train/Q2iaAc-4I1v/2.png b/vlm/train/Q2iaAc-4I1v/2.png new file mode 100644 index 0000000000000000000000000000000000000000..7952edd39306e782876ee4283155aaae1e933a63 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d548e839726bd37ce4f72f71c6e08a9652ee99f44293ea444157e005cb331a31 +size 421152 diff --git a/vlm/train/Q2iaAc-4I1v/3.png b/vlm/train/Q2iaAc-4I1v/3.png new file mode 100644 index 0000000000000000000000000000000000000000..bbef9b3fd5bd1e002fe400c14555e93a607ec2f1 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e501fc8bf3ee153fd2b8c109fb51e2868b8704602b0ad9498c11ddfbf06d2c24 +size 467554 diff --git a/vlm/train/Q2iaAc-4I1v/4.png b/vlm/train/Q2iaAc-4I1v/4.png new file mode 100644 index 0000000000000000000000000000000000000000..6e9d7027fcce8ed3ba689ecb930a6f187230c40c --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0276c3e5c61b006edeb93718025b5b98406af9a73ecba249f224ea5f84b87c7 +size 553159 diff --git a/vlm/train/Q2iaAc-4I1v/5.png b/vlm/train/Q2iaAc-4I1v/5.png new file mode 100644 index 0000000000000000000000000000000000000000..385eb093ec262a0eb1e916b52aa502d781c188ad --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:718baec36464acc3245aace2d62274b901b3b4444c723b95323477b9168db930 +size 557111 diff --git a/vlm/train/Q2iaAc-4I1v/6.png b/vlm/train/Q2iaAc-4I1v/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b84a3b77056e095339bed1f7de6700809a466890 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bb04a585c3bd649aea88e87f110b1a7b52c9cc28c761b7b19a204729cb39858 +size 717802 diff --git a/vlm/train/Q2iaAc-4I1v/7.png b/vlm/train/Q2iaAc-4I1v/7.png new file mode 100644 index 0000000000000000000000000000000000000000..4f53297ad6f15e0671fb695c3f9a15e3068c4697 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4fbb44aec487d7a138fa2b0501aaac111cef378d92c255e8f560473e7a63dee6 +size 625773 diff --git a/vlm/train/Q2iaAc-4I1v/8.png b/vlm/train/Q2iaAc-4I1v/8.png new file mode 100644 index 0000000000000000000000000000000000000000..1a6213c79e794036ff4f6418503bbec3ea5d002e --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfa928c2ced4bc7ab4f3c2db20cd05b34f3dd7919038689b77d29c14f42f6b18 +size 538835 diff --git a/vlm/train/Q2iaAc-4I1v/9.png b/vlm/train/Q2iaAc-4I1v/9.png new file mode 100644 index 0000000000000000000000000000000000000000..82c2a05a224d1eb2bf970c44ade9f8a783929402 --- /dev/null +++ b/vlm/train/Q2iaAc-4I1v/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdc00fdac48189a2a0a6f464c7758a894bfe682302165f78d0a2af96f5fb72b6 +size 516458 diff --git a/vlm/train/S1HlA-ZAZ/0.png b/vlm/train/S1HlA-ZAZ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..41738dc2336469470f3bf6aac59ff498cc3c0e02 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a0e4b044204ccc50960ff1e0fad478991372ea822b5868213b0fed841f1e655 +size 504723 diff --git a/vlm/train/S1HlA-ZAZ/1.png b/vlm/train/S1HlA-ZAZ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..96fffdb7a89483ec252f88edb1210190b5258a52 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dac80876dbd194a9ccf556f55e686d54c7b2087f23bea78be49d43c14e23cc72 +size 534533 diff --git a/vlm/train/S1HlA-ZAZ/10.png b/vlm/train/S1HlA-ZAZ/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2f781480ce95857c02af18a2d115b99774175160 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ce128397f558c479ee88a1685ff8767652102f0ce7c4b690cf765a71fed2b4d +size 650958 diff --git a/vlm/train/S1HlA-ZAZ/11.png b/vlm/train/S1HlA-ZAZ/11.png new file mode 100644 index 0000000000000000000000000000000000000000..142b5d583027519d6fbceede0d5c7f7965f18205 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b31db1fe83eae40ca737637ee8eb884915ec52f5e28e8d6b561f232a6a4652e +size 537307 diff --git a/vlm/train/S1HlA-ZAZ/12.png b/vlm/train/S1HlA-ZAZ/12.png new file mode 100644 index 0000000000000000000000000000000000000000..8323297f046f8b4a7bb4aba9349d57a775381712 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:762166a0b88759bd15cf96592d7aeafd025676b9f08ced1a5ef805e98e70077d +size 292128 diff --git a/vlm/train/S1HlA-ZAZ/13.png b/vlm/train/S1HlA-ZAZ/13.png new file mode 100644 index 0000000000000000000000000000000000000000..2b392fbb106cac10a788f59d6ef686fcf94762c0 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4f90044d6a9012a05b4eb0dfa065cfb039536790faec5eca4901f7e8b0fbcfb +size 378970 diff --git a/vlm/train/S1HlA-ZAZ/14.png b/vlm/train/S1HlA-ZAZ/14.png new file mode 100644 index 0000000000000000000000000000000000000000..bb3c9518403aad96cee47f87e03543997eed6cc4 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e532933f68b23417ec56ff886745dcce71a7da407bee74d19f607d3b312d76e +size 296740 diff --git a/vlm/train/S1HlA-ZAZ/15.png b/vlm/train/S1HlA-ZAZ/15.png new file mode 100644 index 0000000000000000000000000000000000000000..14ab36c08469f19fe6110836445b03159437dd74 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a09329a5e4b120d0c58c8d1b0af1c8caa16e290f22999ff932ed85e3fa742a8 +size 105411 diff --git a/vlm/train/S1HlA-ZAZ/2.png b/vlm/train/S1HlA-ZAZ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..dc02edb90c3da48475aeb254709d7b3082aa5c1c --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29bb61c0602e698d9da8bc48a9a5281c68e087943f2df135bf48086eb8d35502 +size 493500 diff --git a/vlm/train/S1HlA-ZAZ/3.png b/vlm/train/S1HlA-ZAZ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..062d392e8d10c1becfcd58d8c7a07f7a7e6f2094 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13a1bb24213c5d5ba57dae688ad3d7075cdb4225b2de8cc365bb5689c21fa68c +size 505574 diff --git a/vlm/train/S1HlA-ZAZ/4.png b/vlm/train/S1HlA-ZAZ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..a17130f5d41c62ecfe102d4ace0686b4e94e4e42 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42ccb68763a4a59f24d93b2633e10abab74e4bc41c2d17c13d0f9a23d33a16ea +size 598042 diff --git a/vlm/train/S1HlA-ZAZ/5.png b/vlm/train/S1HlA-ZAZ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..99101c5c433749e4f5e0fe88e24fa8a316c53225 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e48f601c4631282cefa9029f01080479e6e503a34ea1155a8c7a9ff34279fa8b +size 574870 diff --git a/vlm/train/S1HlA-ZAZ/6.png b/vlm/train/S1HlA-ZAZ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..34e4a1a17bd74829430d05bc7130c941cbbac272 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de97c09c09363598fcd91204a64c2b45ddf3ca3ba2d9565c162e692909371f8f +size 811889 diff --git a/vlm/train/S1HlA-ZAZ/7.png b/vlm/train/S1HlA-ZAZ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f1151d66da3c6c3385a46204cc6c6b5da9feaf3e --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb9c280deb2a5fe3d3337802744366d21ea6cd941f78c3d97fe862ddec74260d +size 561887 diff --git a/vlm/train/S1HlA-ZAZ/8.png b/vlm/train/S1HlA-ZAZ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..efd4c02781c835dafed8f8e21387dae901502716 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c7e18ccc2d03e12da8c0732b697a81a35efebaf6125ced5f4c9f1df1e645c1 +size 538250 diff --git a/vlm/train/S1HlA-ZAZ/9.png b/vlm/train/S1HlA-ZAZ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e0f150b3405a4fa2a211ad64716eb937558c9567 --- /dev/null +++ b/vlm/train/S1HlA-ZAZ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0143cf9f49869d3ce29b2ada31254a164fd16542802dbcceccbefc9a43c96ba2 +size 580114 diff --git a/vlm/train/S1gUsoR9YX/0.png b/vlm/train/S1gUsoR9YX/0.png new file mode 100644 index 0000000000000000000000000000000000000000..e556509892738671e2430bab81d753fd86125919 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a5d96e5f5be23f0b3e9484d435963efab8b1662e70c4f0ca8eb7e9ca0096f4b +size 433430 diff --git a/vlm/train/S1gUsoR9YX/1.png b/vlm/train/S1gUsoR9YX/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7917954834bee38e9f3331df8a5c721081c05dcf --- /dev/null +++ b/vlm/train/S1gUsoR9YX/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81a784f948b711ed5c49964ae0c63b4d213f8eb7c6319a5f6e5f66dc92af9bd6 +size 549444 diff --git a/vlm/train/S1gUsoR9YX/10.png b/vlm/train/S1gUsoR9YX/10.png new file mode 100644 index 0000000000000000000000000000000000000000..bb83e30436908b1a5f29e8008f5b22515a46c63d --- /dev/null +++ b/vlm/train/S1gUsoR9YX/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca37655963d0c730de4f8d5ab0dadab322a0fbf9190004befc990157275d728c +size 543014 diff --git a/vlm/train/S1gUsoR9YX/11.png b/vlm/train/S1gUsoR9YX/11.png new file mode 100644 index 0000000000000000000000000000000000000000..5b915bda44e3645825ed648bbb7562b05663d213 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ad102d50bcf5ffdceccca0d0eb34a0a29c0bb883df6cb61e2dd1be9166e1bf2 +size 561941 diff --git a/vlm/train/S1gUsoR9YX/12.png b/vlm/train/S1gUsoR9YX/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a38e1beb0b05bbb2bb37dd35f65c2f8c595b9ae6 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f305653af34fd1a8f6a9bc091d81fdcf5c21ce9355ac7bb4cf56c48586301f93 +size 84113 diff --git a/vlm/train/S1gUsoR9YX/13.png b/vlm/train/S1gUsoR9YX/13.png new file mode 100644 index 0000000000000000000000000000000000000000..3b571950136061289bf8c40f334c5533df76472b --- /dev/null +++ b/vlm/train/S1gUsoR9YX/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68449cbb41fc0bfb4f8af588f9dbee4571b170a6545e02105fca9f0a3b0a1573 +size 366815 diff --git a/vlm/train/S1gUsoR9YX/14.png b/vlm/train/S1gUsoR9YX/14.png new file mode 100644 index 0000000000000000000000000000000000000000..c5788f32780fbc3e0bef111b7c40c4d980304516 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6331aac2fbebd476cb4b54fc25c0ff6ac67260375738fd27b6e27360e880b97 +size 392726 diff --git a/vlm/train/S1gUsoR9YX/2.png b/vlm/train/S1gUsoR9YX/2.png new file mode 100644 index 0000000000000000000000000000000000000000..e7871624e47f05354998d883475f877c842f14dc --- /dev/null +++ b/vlm/train/S1gUsoR9YX/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89d5d6ca576f9eef202a218630ade1ea6cc90eec2c55e976b40eb497a058f4bc +size 460370 diff --git a/vlm/train/S1gUsoR9YX/3.png b/vlm/train/S1gUsoR9YX/3.png new file mode 100644 index 0000000000000000000000000000000000000000..76f348012bb296ba3548cf9a3cd8d113bf4e3752 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38cc173f6b7449d9e0c1425087994ecfdcaea0a52bd2b5fa84768af348b0cadc +size 450668 diff --git a/vlm/train/S1gUsoR9YX/4.png b/vlm/train/S1gUsoR9YX/4.png new file mode 100644 index 0000000000000000000000000000000000000000..eabddc88b077d08743073b091fc43b5ff96379f4 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fbe0deb67781c3667a13227a1582884835259cfa2d25d38b59fd6a0bd75c0e5 +size 526471 diff --git a/vlm/train/S1gUsoR9YX/5.png b/vlm/train/S1gUsoR9YX/5.png new file mode 100644 index 0000000000000000000000000000000000000000..f315a0715b46e75873305bd7a6317baf561213a1 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:efd42a7b25e73f499e91dca0fa9fc45b13673a6091a773ddfd7393a6023afb90 +size 482324 diff --git a/vlm/train/S1gUsoR9YX/6.png b/vlm/train/S1gUsoR9YX/6.png new file mode 100644 index 0000000000000000000000000000000000000000..6c10b33eb33a0d84144b4ab9e182a0c9dd0a37db --- /dev/null +++ b/vlm/train/S1gUsoR9YX/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59079b23c557c581947d4f803f00aa18f11a27ab69d0526d2775cd3faa273e7c +size 411387 diff --git a/vlm/train/S1gUsoR9YX/7.png b/vlm/train/S1gUsoR9YX/7.png new file mode 100644 index 0000000000000000000000000000000000000000..8d321407c37d15aeb75a96d348e310f0c95df075 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea9638b3c99212872e9d2151522e28b6437ee632810f2b9a32b7ebdc3be4602e +size 467858 diff --git a/vlm/train/S1gUsoR9YX/8.png b/vlm/train/S1gUsoR9YX/8.png new file mode 100644 index 0000000000000000000000000000000000000000..27ef3d6e091dbb6bc933487d37745e8153901521 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ae6928f4452ae627f2e2abae0c1990709465407977fd9db670540500fce015f +size 408288 diff --git a/vlm/train/S1gUsoR9YX/9.png b/vlm/train/S1gUsoR9YX/9.png new file mode 100644 index 0000000000000000000000000000000000000000..1c4f1b6e17e909f85cf9e2f1422de8eb1a183056 --- /dev/null +++ b/vlm/train/S1gUsoR9YX/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:212a660ec1b90395b7502436ed864db527ec165c3e9c73f906a9fce96aad290e +size 550956 diff --git a/vlm/train/SJQO7UJCW/0.png b/vlm/train/SJQO7UJCW/0.png new file mode 100644 index 0000000000000000000000000000000000000000..ce4a5490ea10d1f2d88bf30789e4f4e03618aa92 --- /dev/null +++ b/vlm/train/SJQO7UJCW/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc7fcb7d7b13d2ff1f519c4c5e37db6fe04d655bb3298b7f2288b6f4e73edfda +size 508817 diff --git a/vlm/train/SJQO7UJCW/1.png b/vlm/train/SJQO7UJCW/1.png new file mode 100644 index 0000000000000000000000000000000000000000..075f076ff8b58921b08212f103b3a7575ea518ec --- /dev/null +++ b/vlm/train/SJQO7UJCW/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a783636cecc88b1fd8e5bab8a415cd0b932bb9f11391f6b07465b1e50cb96098 +size 605490 diff --git a/vlm/train/SJQO7UJCW/10.png b/vlm/train/SJQO7UJCW/10.png new file mode 100644 index 0000000000000000000000000000000000000000..282dba93d8904d6d4e20fd2dc33d375dac9b65e0 --- /dev/null +++ b/vlm/train/SJQO7UJCW/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae46d732aef2baeb64fc82c28c64dc87b0b2775c6e35290f5e77d7f0c7af5c7b +size 209688 diff --git a/vlm/train/SJQO7UJCW/11.png b/vlm/train/SJQO7UJCW/11.png new file mode 100644 index 0000000000000000000000000000000000000000..a234b7b8d670d7403689052b0b02409fb18f2a0e --- /dev/null +++ b/vlm/train/SJQO7UJCW/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d393c27728fc0ce5a64cd5fe7ad7dce5b3563983e5068f2a0ee8732743fb1b31 +size 926865 diff --git a/vlm/train/SJQO7UJCW/12.png b/vlm/train/SJQO7UJCW/12.png new file mode 100644 index 0000000000000000000000000000000000000000..eaef3e6106e96ff77804d0919330e53ed5803142 --- /dev/null +++ b/vlm/train/SJQO7UJCW/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e87739e3819fdd2a9f742bcfbd7b6dc49d5f1eca1599ba671891ce162e15dc7b +size 836318 diff --git a/vlm/train/SJQO7UJCW/13.png b/vlm/train/SJQO7UJCW/13.png new file mode 100644 index 0000000000000000000000000000000000000000..29f831950007d691493a594e2042d17feff20ebc --- /dev/null +++ b/vlm/train/SJQO7UJCW/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ad538c85e1d9e1fd1c83ea290a601a08950898d889b000fdf47eab0c309bb31 +size 716005 diff --git a/vlm/train/SJQO7UJCW/2.png b/vlm/train/SJQO7UJCW/2.png new file mode 100644 index 0000000000000000000000000000000000000000..7a19e6b6fffe6088992d4ed6feb6b3865a75a6e9 --- /dev/null +++ b/vlm/train/SJQO7UJCW/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:010671b29b38d2d3529efc69689abf1ee33826d357a945f29539b38dc94e0b17 +size 566000 diff --git a/vlm/train/SJQO7UJCW/3.png b/vlm/train/SJQO7UJCW/3.png new file mode 100644 index 0000000000000000000000000000000000000000..daa7e0b654ec627353d8ec7e9e277975db055961 --- /dev/null +++ b/vlm/train/SJQO7UJCW/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da09d55d919dd21a81c8448730fb0eea9283692368d55a36c19733050faf8f49 +size 495050 diff --git a/vlm/train/SJQO7UJCW/4.png b/vlm/train/SJQO7UJCW/4.png new file mode 100644 index 0000000000000000000000000000000000000000..dca6236586bb80813f58860f609be2a2d3ef2fae --- /dev/null +++ b/vlm/train/SJQO7UJCW/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b43d97d39cd19f337d16f83c1f3aa775fbb2058c453e7d2c21da0e317ebf53c +size 537324 diff --git a/vlm/train/SJQO7UJCW/5.png b/vlm/train/SJQO7UJCW/5.png new file mode 100644 index 0000000000000000000000000000000000000000..7d6cb2d55e19f0adbe1fdcd15ab9c258812bbf8f --- /dev/null +++ b/vlm/train/SJQO7UJCW/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a18d0015e7632b07da678348a48aec0fa67a97677244d009bd4c24741d89666 +size 595698 diff --git a/vlm/train/SJQO7UJCW/6.png b/vlm/train/SJQO7UJCW/6.png new file mode 100644 index 0000000000000000000000000000000000000000..33fbd6a72b8d0d27ad78620d77bf451557b0db3c --- /dev/null +++ b/vlm/train/SJQO7UJCW/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee3549ad7d166da9bed0a9ca0ed0842de6bd5cffdba8838e00a1e46e9dfcc428 +size 622950 diff --git a/vlm/train/SJQO7UJCW/7.png b/vlm/train/SJQO7UJCW/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ea2577d484a546de57ebe2c4727816fd8b351bc6 --- /dev/null +++ b/vlm/train/SJQO7UJCW/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21e51d94f4de51ef16e83b35a96e0f9dc009e0471e39c9e2d4aebca88c9aa741 +size 530634 diff --git a/vlm/train/SJQO7UJCW/8.png b/vlm/train/SJQO7UJCW/8.png new file mode 100644 index 0000000000000000000000000000000000000000..ef4239d046759bdc32b2a73c64bfb2d5fa493a4a --- /dev/null +++ b/vlm/train/SJQO7UJCW/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9271eacfc3cd4b74c84540aefedbf077f361d0eb5a2f024e4420085b6b5676cc +size 485843 diff --git a/vlm/train/SJQO7UJCW/9.png b/vlm/train/SJQO7UJCW/9.png new file mode 100644 index 0000000000000000000000000000000000000000..731c3f78d3b6244e28d39df8cd177c73a8ebd93e --- /dev/null +++ b/vlm/train/SJQO7UJCW/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3279fb1de7ab443c47ac7093c6c990d44ba30ce814ed5dacddddcc215789bd0e +size 513573 diff --git a/vlm/train/SJe5P6EYvS/0.png b/vlm/train/SJe5P6EYvS/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d505c4b7cc7d6d88297f823edbfcd5f03e891362 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:881ef01494a3bb1e1c465b139c1ab1fd10d5f0027c0f9688caf1eb0255c0072c +size 525734 diff --git a/vlm/train/SJe5P6EYvS/1.png b/vlm/train/SJe5P6EYvS/1.png new file mode 100644 index 0000000000000000000000000000000000000000..e651a778380613955168d7f351717d15572fba6c --- /dev/null +++ b/vlm/train/SJe5P6EYvS/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:424a7a730805da6fc764dc64cd9173589a294c80eacacc99bf2740240c9b828c +size 452056 diff --git a/vlm/train/SJe5P6EYvS/10.png b/vlm/train/SJe5P6EYvS/10.png new file mode 100644 index 0000000000000000000000000000000000000000..db2789eb6e0af33eb18be83bba2fb5761e1febbc --- /dev/null +++ b/vlm/train/SJe5P6EYvS/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5bdaca11e30dea533a1cc7237ea90031b97593439a72fc36cc509635bcbb083 +size 219487 diff --git a/vlm/train/SJe5P6EYvS/11.png b/vlm/train/SJe5P6EYvS/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ecd09cdd3dae7893c2a13b83587141246a2f483f --- /dev/null +++ b/vlm/train/SJe5P6EYvS/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f384f0a73bd533f3e84c99352a6dc2a790638fe2541981cd5fd9ace8cce5c763 +size 311143 diff --git a/vlm/train/SJe5P6EYvS/12.png b/vlm/train/SJe5P6EYvS/12.png new file mode 100644 index 0000000000000000000000000000000000000000..1afe708fb02122d3c33192dc7e5b431cd963221d --- /dev/null +++ b/vlm/train/SJe5P6EYvS/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:faf3efb5960ac7920d7ba4006c35be60fdc99e3c5aacf49913b3886dfd51ce49 +size 1198587 diff --git a/vlm/train/SJe5P6EYvS/2.png b/vlm/train/SJe5P6EYvS/2.png new file mode 100644 index 0000000000000000000000000000000000000000..bf470ee8ff01d760949999e82afcb308d5c63d80 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e8ea32d6a74b1d2f912d4d5d706412bc50eb7863b62f6b1932d7e012cbc92b3 +size 534848 diff --git a/vlm/train/SJe5P6EYvS/3.png b/vlm/train/SJe5P6EYvS/3.png new file mode 100644 index 0000000000000000000000000000000000000000..3943363ac7744d68a68378910c06036d91ddcb21 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55db8d7ac8e76a0901c1cc62f9ea46744a80e143f0fdd56dcdb2fc2cc2a23fb +size 557320 diff --git a/vlm/train/SJe5P6EYvS/4.png b/vlm/train/SJe5P6EYvS/4.png new file mode 100644 index 0000000000000000000000000000000000000000..c011e0ebb5f5a528cdfd711cabcaf5ff8e5733d8 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd7c5350d54cecf29bacb95bbc1a4fce31dbf5dce18aa71e8c42051525de3b05 +size 516776 diff --git a/vlm/train/SJe5P6EYvS/5.png b/vlm/train/SJe5P6EYvS/5.png new file mode 100644 index 0000000000000000000000000000000000000000..be3ebe0941eb0d5f55d51a02ad75b097123b86d1 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11816d96bc947152dbee856e2223fcf0564436730e4fdf66b7c1d802a6ffb24e +size 447157 diff --git a/vlm/train/SJe5P6EYvS/6.png b/vlm/train/SJe5P6EYvS/6.png new file mode 100644 index 0000000000000000000000000000000000000000..5b35abc56a6acb5abd9c33d98e0022d55d1b6815 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c788fa66b1dda168262223cb6631b7e8e4e4cf879072ea3fa09968707842b54d +size 524443 diff --git a/vlm/train/SJe5P6EYvS/7.png b/vlm/train/SJe5P6EYvS/7.png new file mode 100644 index 0000000000000000000000000000000000000000..4b309ea90e095572589082b9cf37541688257fe0 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:913f1328286158dd5a97f9f2971ddfb37c6db423d674cdc201e277574fc4498b +size 573598 diff --git a/vlm/train/SJe5P6EYvS/8.png b/vlm/train/SJe5P6EYvS/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5e3e7c3b924177946bb7b0a757ffb2e5e6586e0e --- /dev/null +++ b/vlm/train/SJe5P6EYvS/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52ae1d390761c54dc1c5dd03201726e1c17f6e68a03e4ef3974a6753cc6baecb +size 532518 diff --git a/vlm/train/SJe5P6EYvS/9.png b/vlm/train/SJe5P6EYvS/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d1f73cb3483059fde57ed0ce71a849d558adc8f7 --- /dev/null +++ b/vlm/train/SJe5P6EYvS/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73de3ab84b491d05819c4576693118b454f73f3ec269bbbd91b90238df7150c4 +size 537535 diff --git a/vlm/train/SkNQeiRpb/0.png b/vlm/train/SkNQeiRpb/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d58a889b5a169942fc2c74fade0b2cca0c396cad --- /dev/null +++ b/vlm/train/SkNQeiRpb/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eea4287a26e845e57800fba7d906a1574e16eb6d13a6283d6de365e87658de7b +size 438284 diff --git a/vlm/train/SkNQeiRpb/1.png b/vlm/train/SkNQeiRpb/1.png new file mode 100644 index 0000000000000000000000000000000000000000..7ce150a801d030a6a22cb5ba148c28d018b88af3 --- /dev/null +++ b/vlm/train/SkNQeiRpb/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78eba23acefe8472698d87fc567e0a150c746ab3c3e6633828e20f9e9ac2d763 +size 542202 diff --git a/vlm/train/SkNQeiRpb/2.png b/vlm/train/SkNQeiRpb/2.png new file mode 100644 index 0000000000000000000000000000000000000000..64ca4bc81b93a594ae286479ca91ae919ece3d8a --- /dev/null +++ b/vlm/train/SkNQeiRpb/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d46e0bf01e1e8b16bde324af7dfd7f09fd5d4468a26c79daff183ca45f257ce5 +size 465857 diff --git a/vlm/train/SkNQeiRpb/3.png b/vlm/train/SkNQeiRpb/3.png new file mode 100644 index 0000000000000000000000000000000000000000..ce0065b71974ee8f2eadcb45b0660f3aa0f7de18 --- /dev/null +++ b/vlm/train/SkNQeiRpb/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5bc2efc553c2e9f05fd411bd1261e2ae11ee9152675a911b819da0e98117a24 +size 321092 diff --git a/vlm/train/SkNQeiRpb/4.png b/vlm/train/SkNQeiRpb/4.png new file mode 100644 index 0000000000000000000000000000000000000000..8acd086de2cde669bdc453ce703296473a16fc8f --- /dev/null +++ b/vlm/train/SkNQeiRpb/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49ae5277b558a715f449dd8c62be7c200a01abc8f8302d92c64c69262f0b5b4e +size 369807 diff --git a/vlm/train/SkNQeiRpb/5.png b/vlm/train/SkNQeiRpb/5.png new file mode 100644 index 0000000000000000000000000000000000000000..bb7454a00c927cf527199dc175e90edafef011b9 --- /dev/null +++ b/vlm/train/SkNQeiRpb/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fd454bf03ed32c34ddfc5e62bceb2842c5baa9555555bded966cbeb45290cf6 +size 378683 diff --git a/vlm/train/SkNQeiRpb/6.png b/vlm/train/SkNQeiRpb/6.png new file mode 100644 index 0000000000000000000000000000000000000000..45363333cc0af6ed6d47c779af616f09d9045a67 --- /dev/null +++ b/vlm/train/SkNQeiRpb/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e58eb97de831ac95c4c72234e55124556341870dc38d26a24db6813492f6bb57 +size 387918 diff --git a/vlm/train/SkNQeiRpb/7.png b/vlm/train/SkNQeiRpb/7.png new file mode 100644 index 0000000000000000000000000000000000000000..04e89f902aa704e54a580190b6ee7cf43bc78de3 --- /dev/null +++ b/vlm/train/SkNQeiRpb/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de6ec1497de5004e2ec01bc4dd5ead52b03755813a9ba70647909d0344918f5b +size 464983 diff --git a/vlm/train/SkNQeiRpb/8.png b/vlm/train/SkNQeiRpb/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5b476457cc5571fbb961716c40b493f47cff1c55 --- /dev/null +++ b/vlm/train/SkNQeiRpb/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be2c6a72fffdf4162866e771bfb5e77201325368cfdc07b5623970cbd01a806e +size 526209 diff --git a/vlm/train/Skvd-myR-/13.png b/vlm/train/Skvd-myR-/13.png new file mode 100644 index 0000000000000000000000000000000000000000..2697744f72f3f22086a3bc322f14516faa91a0bf --- /dev/null +++ b/vlm/train/Skvd-myR-/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:360c22a4ec6b905bc9316504ea0655cd09d68204a14f7219b3fa7c5c394e628c +size 549564 diff --git a/vlm/train/Sy4tzwqxe/0.png b/vlm/train/Sy4tzwqxe/0.png new file mode 100644 index 0000000000000000000000000000000000000000..c0afd968d2f1502407c412a0efd32133636f95b5 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9d8ab884804764fc189cb2400071bf9bbd155d6eba119f33411dd8c0c499c0a +size 497358 diff --git a/vlm/train/Sy4tzwqxe/1.png b/vlm/train/Sy4tzwqxe/1.png new file mode 100644 index 0000000000000000000000000000000000000000..97d391b9b2ea24dfac2039fcd0c27c240e4be44d --- /dev/null +++ b/vlm/train/Sy4tzwqxe/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91b3e030eece8fd80e2b2b2c1a2d4dd491ba22f07494383a8984808c0d7fed04 +size 608524 diff --git a/vlm/train/Sy4tzwqxe/10.png b/vlm/train/Sy4tzwqxe/10.png new file mode 100644 index 0000000000000000000000000000000000000000..451d7fd08af4ddc15452e341f7e07c56d4f5ca8a --- /dev/null +++ b/vlm/train/Sy4tzwqxe/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34a347f94903f8346d0e88954b68eb8e43a0594f8a9c6015e1213347408e4c84 +size 483869 diff --git a/vlm/train/Sy4tzwqxe/11.png b/vlm/train/Sy4tzwqxe/11.png new file mode 100644 index 0000000000000000000000000000000000000000..065d9c777e51c28b87623dc5e23118d2d79344c0 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f160004c847e0b5775c6b31563623888d48b33bf35dd49680118a5fd3a55c9c9 +size 221799 diff --git a/vlm/train/Sy4tzwqxe/2.png b/vlm/train/Sy4tzwqxe/2.png new file mode 100644 index 0000000000000000000000000000000000000000..6ba71ffdffcc007df0e260f93d54e167a7d3b1d7 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41a0e18f5cd09139d8536512c80723093f2b3e25d9869daaa52dc7a6d8e2cf1d +size 540369 diff --git a/vlm/train/Sy4tzwqxe/3.png b/vlm/train/Sy4tzwqxe/3.png new file mode 100644 index 0000000000000000000000000000000000000000..2903058e238ed30dcd72c333064a71bd12e92544 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b6966905b9a442edd4893feed0ba30ffc6255f21ec757cf8ce9b55e3232888e +size 475112 diff --git a/vlm/train/Sy4tzwqxe/4.png b/vlm/train/Sy4tzwqxe/4.png new file mode 100644 index 0000000000000000000000000000000000000000..c641417c6549abc275eff735252d0e06db5b0779 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2df0e66a2c6f097e3a1745892c159a50347bd4490538256db07bf64e89de2fe7 +size 492755 diff --git a/vlm/train/Sy4tzwqxe/5.png b/vlm/train/Sy4tzwqxe/5.png new file mode 100644 index 0000000000000000000000000000000000000000..a998754a1d4b69c32b43328c0aa41bd7ea5cbafd --- /dev/null +++ b/vlm/train/Sy4tzwqxe/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36f0d0c9297e221d705b47f926a20823134fa54d47fd45133d693c5aaba41bdf +size 508211 diff --git a/vlm/train/Sy4tzwqxe/6.png b/vlm/train/Sy4tzwqxe/6.png new file mode 100644 index 0000000000000000000000000000000000000000..7431a5009ad62e840d3af7781dc35b3534eaaae0 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf9f9bd4f20d4b65df003cbc931ffe7e89797564257734847705d4ff936041ac +size 559240 diff --git a/vlm/train/Sy4tzwqxe/7.png b/vlm/train/Sy4tzwqxe/7.png new file mode 100644 index 0000000000000000000000000000000000000000..04094331c2ec644fc4f64c74ced6baadab73d163 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:958176928418d3cfe1ad01defc8086649620b04f2730432fb40694d282b4c543 +size 490004 diff --git a/vlm/train/Sy4tzwqxe/8.png b/vlm/train/Sy4tzwqxe/8.png new file mode 100644 index 0000000000000000000000000000000000000000..64ecfc85df7ed114a73e6b1146fc417fa543a8e4 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35624d0375c8ec1db986a4dfa16369616c7c6331ec8a8717396ca1e57e977b90 +size 537488 diff --git a/vlm/train/Sy4tzwqxe/9.png b/vlm/train/Sy4tzwqxe/9.png new file mode 100644 index 0000000000000000000000000000000000000000..aa3be6a1381cf7d008b507d730ffaab0b336e766 --- /dev/null +++ b/vlm/train/Sy4tzwqxe/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5542a94c9ead00bb726e5bc28d506515d4a369d33310cd67aac75709f381573e +size 447231 diff --git a/vlm/train/SyzVb3CcFX/0.png b/vlm/train/SyzVb3CcFX/0.png new file mode 100644 index 0000000000000000000000000000000000000000..53d488c785b81e99fb03b276785c7e33ce6b6118 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:386f2476e8b739581bd85643046237201bc1366bc3bb98679baa8db60296f84c +size 494562 diff --git a/vlm/train/SyzVb3CcFX/1.png b/vlm/train/SyzVb3CcFX/1.png new file mode 100644 index 0000000000000000000000000000000000000000..5346194b283c350e970debbaa5a4488e72db6c12 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:400d69cce1e137e0781d6d7ff22cd20c629957980b2e3357174055606da46cea +size 617068 diff --git a/vlm/train/SyzVb3CcFX/10.png b/vlm/train/SyzVb3CcFX/10.png new file mode 100644 index 0000000000000000000000000000000000000000..52afd9bb62a4647ec2f61a44cf8d3d17a6541d05 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7477309b8aae1dac3034a4ea1b676feef3652b17339b6dc337758d9244dfa8bf +size 554964 diff --git a/vlm/train/SyzVb3CcFX/11.png b/vlm/train/SyzVb3CcFX/11.png new file mode 100644 index 0000000000000000000000000000000000000000..6987fa3266276105e4b5a0666f5812794be25f38 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:903997aa8843519811fb3688e355f60c5c028aca13694848c878bd3360871318 +size 427849 diff --git a/vlm/train/SyzVb3CcFX/12.png b/vlm/train/SyzVb3CcFX/12.png new file mode 100644 index 0000000000000000000000000000000000000000..8ef838ec9e6f76afafc8dd23172793e236c4bdaa --- /dev/null +++ b/vlm/train/SyzVb3CcFX/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8e66db0e0a524a9c2481143e7a9c969d0c833e27a2a2238a86c315ec13fcc9a +size 503991 diff --git a/vlm/train/SyzVb3CcFX/13.png b/vlm/train/SyzVb3CcFX/13.png new file mode 100644 index 0000000000000000000000000000000000000000..fc511c479ba902dcb26edf5ff0121e150bfebeb2 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17314c1d02a9ec1c14f2028a75c2361d0c650f17217ca4fb1b27bd07555f9941 +size 542527 diff --git a/vlm/train/SyzVb3CcFX/14.png b/vlm/train/SyzVb3CcFX/14.png new file mode 100644 index 0000000000000000000000000000000000000000..e67d009936ab4f04446172d06fb6543f1fd17b97 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e5079ea7c0fdd41359bb38a09b1fde4e980a2dbd54c8a9a399e1d26a6dae6b9 +size 292538 diff --git a/vlm/train/SyzVb3CcFX/15.png b/vlm/train/SyzVb3CcFX/15.png new file mode 100644 index 0000000000000000000000000000000000000000..75a28cffb1a86a2afb532a05985682804394b9d7 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db2884a44e5bafc47ee08700ae158143af73100c5aa0d6064d0cd97ac8ca4e73 +size 362455 diff --git a/vlm/train/SyzVb3CcFX/16.png b/vlm/train/SyzVb3CcFX/16.png new file mode 100644 index 0000000000000000000000000000000000000000..ea83446f4eacfd62ad46a3d9b849e0bff1c8346a --- /dev/null +++ b/vlm/train/SyzVb3CcFX/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:928c2fef2a9724706cb487bd3634d1364bf362cc8d13ce4e59206b2780978cce +size 352591 diff --git a/vlm/train/SyzVb3CcFX/17.png b/vlm/train/SyzVb3CcFX/17.png new file mode 100644 index 0000000000000000000000000000000000000000..1ff65d6ef70510a4eca891b5e0bbc1ad8b40d1b2 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3baa45babd2c145ae75b753e8641b6bef38191e074c86728adf59b7bab3c49d6 +size 1224249 diff --git a/vlm/train/SyzVb3CcFX/18.png b/vlm/train/SyzVb3CcFX/18.png new file mode 100644 index 0000000000000000000000000000000000000000..5fa18a78776d07433ec6a13331e46c3c1dc695ec --- /dev/null +++ b/vlm/train/SyzVb3CcFX/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cccbfc32ffe40cd02fba8670b44b8ebf8d21eba7eaf41f2d6e2dd5942bb41d78 +size 1009574 diff --git a/vlm/train/SyzVb3CcFX/19.png b/vlm/train/SyzVb3CcFX/19.png new file mode 100644 index 0000000000000000000000000000000000000000..c5fc90ffc903e04a8841b32b36e0325689699028 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7af6c7d1a5043b18743e4436d4392180f90b04f0400362d959524ab5f36b78be +size 1079204 diff --git a/vlm/train/SyzVb3CcFX/2.png b/vlm/train/SyzVb3CcFX/2.png new file mode 100644 index 0000000000000000000000000000000000000000..a3f07d58215c699e177703e5d4e799998f8a78d2 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6a5ed6c2414f522ca3bf6940e876fe01512a61b7f7fddf893a64bd39356c3fb +size 551455 diff --git a/vlm/train/SyzVb3CcFX/3.png b/vlm/train/SyzVb3CcFX/3.png new file mode 100644 index 0000000000000000000000000000000000000000..7aec7f9b0bd69fafe2901750130fa39b43156732 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12e6cd991c2725b086d69c8989f653e7b73029790f6618666318528b118063ae +size 512431 diff --git a/vlm/train/SyzVb3CcFX/4.png b/vlm/train/SyzVb3CcFX/4.png new file mode 100644 index 0000000000000000000000000000000000000000..360247f960bfe2c178f0c9b3bd5a6a3d6bbc14a5 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3e64d60d051940964a5deb9d415e859d42aa45715d25ea7ae5d05365990e419 +size 514838 diff --git a/vlm/train/SyzVb3CcFX/5.png b/vlm/train/SyzVb3CcFX/5.png new file mode 100644 index 0000000000000000000000000000000000000000..e03443e248da9bd9cc34721953fa5ca650e3b73e --- /dev/null +++ b/vlm/train/SyzVb3CcFX/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ab40bd970f4607093f4230b254f3aaa2b7c8eecb280c6e3dc78ad746ffd5edd +size 679197 diff --git a/vlm/train/SyzVb3CcFX/6.png b/vlm/train/SyzVb3CcFX/6.png new file mode 100644 index 0000000000000000000000000000000000000000..41ad37fc1fa85982c58c46a832b4c5a3d0aa9f3e --- /dev/null +++ b/vlm/train/SyzVb3CcFX/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02d3ebdcefecf7c0b3394b3853e7c6b82950568576701fcf8f4e9d21dde9e1fe +size 841560 diff --git a/vlm/train/SyzVb3CcFX/7.png b/vlm/train/SyzVb3CcFX/7.png new file mode 100644 index 0000000000000000000000000000000000000000..a881bcadda4a88131df5e352dd99bcceca961aa5 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b054e7eaaccfe6687f8966b3e4bd57eb17eb1967426c69e4e0b738f3039de62a +size 898577 diff --git a/vlm/train/SyzVb3CcFX/8.png b/vlm/train/SyzVb3CcFX/8.png new file mode 100644 index 0000000000000000000000000000000000000000..031888fc29d64625d341b0e2affdb41262114d96 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05a417fd1a15d93c7a39ab5013445deec3d4a9d12f298393ede723afeb74977f +size 582545 diff --git a/vlm/train/SyzVb3CcFX/9.png b/vlm/train/SyzVb3CcFX/9.png new file mode 100644 index 0000000000000000000000000000000000000000..77933a395007440bef3b337f274d389366f797b3 --- /dev/null +++ b/vlm/train/SyzVb3CcFX/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c52cd42949f327efe03f71a21ad1a8dbe75ee6c10bc23b7ba8635a3712e9c45a +size 527659 diff --git a/vlm/train/Uj7pF-D-YvT/0.png b/vlm/train/Uj7pF-D-YvT/0.png new file mode 100644 index 0000000000000000000000000000000000000000..4937bd4cd9d3a5c8b9c75acfd5814e84b888dd02 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c5f7c44ff35bd4cbd89e24875a86e992ea0e4ae51fb8485f8c0287ceecf2e98 +size 390268 diff --git a/vlm/train/Uj7pF-D-YvT/1.png b/vlm/train/Uj7pF-D-YvT/1.png new file mode 100644 index 0000000000000000000000000000000000000000..6efbe53bc6ab92f4c1c61298f90b4e4e404ca5c2 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52d30558a04404531c0f6e40e8be8cf9a6f4d304970e4aab5a79fcaba1923250 +size 571728 diff --git a/vlm/train/Uj7pF-D-YvT/10.png b/vlm/train/Uj7pF-D-YvT/10.png new file mode 100644 index 0000000000000000000000000000000000000000..5bde8975319166d42f2293ea28eb87b58d3917c0 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fc83bebe9ee835b6ea1ca9178fea1b2a866a5fcd3e097b8d512512d508439720 +size 606448 diff --git a/vlm/train/Uj7pF-D-YvT/11.png b/vlm/train/Uj7pF-D-YvT/11.png new file mode 100644 index 0000000000000000000000000000000000000000..803477db0f55ab2eb158302d0a9f44d284fa83d2 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e3273599206542d6f3bfe29b8a60b3b1f97fc21f2b0db76a40d6eb7fcc3b9ca +size 445769 diff --git a/vlm/train/Uj7pF-D-YvT/2.png b/vlm/train/Uj7pF-D-YvT/2.png new file mode 100644 index 0000000000000000000000000000000000000000..51d3ae1a58216b20587353c5afbff927fca8edb9 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f5a552a86e91049fe3891715f70fc6b649bbbf0da3c1a7eeb029e5288733c15 +size 523637 diff --git a/vlm/train/Uj7pF-D-YvT/3.png b/vlm/train/Uj7pF-D-YvT/3.png new file mode 100644 index 0000000000000000000000000000000000000000..f25bd8ba5f7cf9fd5680a7b4882c2a9607a6c3c0 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c59cb261816775b075c2b3b8a9223d8fa6f27eb166bcfe918373551506db8a40 +size 575423 diff --git a/vlm/train/Uj7pF-D-YvT/4.png b/vlm/train/Uj7pF-D-YvT/4.png new file mode 100644 index 0000000000000000000000000000000000000000..cd58f8d8c24b96cb4758ec3d11c79b9232ac659a --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61e598339788a35a340a491824ff1c3991d1d02460c178de4c4fbb2f42e53cb8 +size 686210 diff --git a/vlm/train/Uj7pF-D-YvT/5.png b/vlm/train/Uj7pF-D-YvT/5.png new file mode 100644 index 0000000000000000000000000000000000000000..3abb54e3320760d4bac688317331c72080379602 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9147f8ae5765bffeb7a6910d5ca404868cbac21a60473ecc9253adec5a15a09e +size 607703 diff --git a/vlm/train/Uj7pF-D-YvT/6.png b/vlm/train/Uj7pF-D-YvT/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b7b72aefaa14c2b186855d60cbffcf27a31c02f4 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d304bc34d8751a27dc1ecdd8006e59cd1de76e650fed8acf6d01a99873468ad +size 589883 diff --git a/vlm/train/Uj7pF-D-YvT/7.png b/vlm/train/Uj7pF-D-YvT/7.png new file mode 100644 index 0000000000000000000000000000000000000000..c8764ddda3907890f9ee54e04bbe24b8af06ff9a --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c44375cb7972bc44e6ec38fb101e3328212909f17ec091a91417d56a0a07a82 +size 556341 diff --git a/vlm/train/Uj7pF-D-YvT/8.png b/vlm/train/Uj7pF-D-YvT/8.png new file mode 100644 index 0000000000000000000000000000000000000000..009d9f3e329cbe0b56e88ba7902ed6ace242af3a --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:229ab00c8615e1fce57691af0aaf93218edc41c34c05baa8cee4da8ea3d9430e +size 638868 diff --git a/vlm/train/Uj7pF-D-YvT/9.png b/vlm/train/Uj7pF-D-YvT/9.png new file mode 100644 index 0000000000000000000000000000000000000000..c69c0d124e55da5b0fc9d66eed12ff35ad60de39 --- /dev/null +++ b/vlm/train/Uj7pF-D-YvT/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8ef50262aefd5e00a9c23aad279a2e0d0e919579c0cf8ad41758100a637dded +size 555792 diff --git a/vlm/train/Uwh-v1HSw-x/0.png b/vlm/train/Uwh-v1HSw-x/0.png new file mode 100644 index 0000000000000000000000000000000000000000..8bba6a25f3ed4148001e694e84e1b52ad2a63334 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e4758d26f0210c404720fa18146526a5de0513dbe1e46468f7a498897634ddd +size 438730 diff --git a/vlm/train/Uwh-v1HSw-x/1.png b/vlm/train/Uwh-v1HSw-x/1.png new file mode 100644 index 0000000000000000000000000000000000000000..2090fa88537890a8f28cf23662e391397f3e8a7e --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a99222bf4e918f5f9a71814ce08ba8952ddb5a98ade2b16f386558655a0f8f7 +size 588163 diff --git a/vlm/train/Uwh-v1HSw-x/10.png b/vlm/train/Uwh-v1HSw-x/10.png new file mode 100644 index 0000000000000000000000000000000000000000..60a351fdc26b7b2eb6cf94543b0fb84fac9cf21b --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d777844ad7602193e9ad531d72adb445f319093d5168d8a5e81ddde01689b801 +size 536562 diff --git a/vlm/train/Uwh-v1HSw-x/11.png b/vlm/train/Uwh-v1HSw-x/11.png new file mode 100644 index 0000000000000000000000000000000000000000..64e6cf2d3de6a3418b82cbf6881e42e3329e46a9 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:388458aea51e9367699863d2e761745b4e7ae0d08e81aa869345f379f37461ab +size 566004 diff --git a/vlm/train/Uwh-v1HSw-x/12.png b/vlm/train/Uwh-v1HSw-x/12.png new file mode 100644 index 0000000000000000000000000000000000000000..48efa83364cd186e9c95081fd499cb26475f76d4 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17fde18a20ca09efab1c20bba1c58379a14077d96236c2dbb89e4238bdde4b91 +size 542953 diff --git a/vlm/train/Uwh-v1HSw-x/2.png b/vlm/train/Uwh-v1HSw-x/2.png new file mode 100644 index 0000000000000000000000000000000000000000..38af6f8012b700eaa20fb46266172559ed3023d6 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9284142589041f81252923897971a8bfabfec13200ff9a0cd3761233a09c18e2 +size 538991 diff --git a/vlm/train/Uwh-v1HSw-x/3.png b/vlm/train/Uwh-v1HSw-x/3.png new file mode 100644 index 0000000000000000000000000000000000000000..88bb45f625de225638cd3bd55d1c324cb4d2b8a0 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:068c53a35cc0181eec05f845503056953b381cf2671f31a841be5258fc44ea28 +size 592143 diff --git a/vlm/train/Uwh-v1HSw-x/4.png b/vlm/train/Uwh-v1HSw-x/4.png new file mode 100644 index 0000000000000000000000000000000000000000..b41158f0fe5c468e7b036b77c9cd498bc2993c00 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5af586355b3e247b48176451dae1405b2951ae5730870e3c865a021b34a585e5 +size 550177 diff --git a/vlm/train/Uwh-v1HSw-x/5.png b/vlm/train/Uwh-v1HSw-x/5.png new file mode 100644 index 0000000000000000000000000000000000000000..2bc0fcbeb288b5590357c483c091e4f5c1064c27 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e0593da62ec5726baf3b547900103beaee790dbfb406aee6374517b0b6a3dbc +size 616829 diff --git a/vlm/train/Uwh-v1HSw-x/6.png b/vlm/train/Uwh-v1HSw-x/6.png new file mode 100644 index 0000000000000000000000000000000000000000..f3d8001f03643da47d3343e0fb56aec747f9e47b --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1dc0b027e1f0a40c4ccff323e7a6b3a9e5976c85be6a27146ffa073bc4fbaa2 +size 486183 diff --git a/vlm/train/Uwh-v1HSw-x/7.png b/vlm/train/Uwh-v1HSw-x/7.png new file mode 100644 index 0000000000000000000000000000000000000000..fed80befe2227948c0316b148b1e446ecd088ad1 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e30f4ecacb9486ad5a5628a56ca4dc42e7eb2c770f1ecce4667e52066612ffb9 +size 590351 diff --git a/vlm/train/Uwh-v1HSw-x/8.png b/vlm/train/Uwh-v1HSw-x/8.png new file mode 100644 index 0000000000000000000000000000000000000000..a51747c59bd2b00cd07070fd0772bcc6b3542cea --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8673fcda27fab5263f13b79931f4fa030c54cefcfd9c99daec6b7db445d578aa +size 527767 diff --git a/vlm/train/Uwh-v1HSw-x/9.png b/vlm/train/Uwh-v1HSw-x/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e7f1c0f88eb66684e20f19ddb500384f2a7c6221 --- /dev/null +++ b/vlm/train/Uwh-v1HSw-x/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fde6820882387351c401ad1df25c77e3eb9c1309ceeea85a7c651da8584501b +size 513936 diff --git a/vlm/train/a5KvtsZ14ev/0.png b/vlm/train/a5KvtsZ14ev/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6226e57433869cc0889c51a0653196c1fecd9c59 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c4f06e3db0bfaa27fbc3f76eccfdad5c21d2610c88952fc5e62a6490b3567eb +size 510434 diff --git a/vlm/train/a5KvtsZ14ev/1.png b/vlm/train/a5KvtsZ14ev/1.png new file mode 100644 index 0000000000000000000000000000000000000000..96231f0c5790f3e97bef9a2ff1120950ced4e2d3 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e48f43e48f95d3aed7088c0c777f89411cc6ba448c7591f9065d0ffbae925920 +size 614494 diff --git a/vlm/train/a5KvtsZ14ev/10.png b/vlm/train/a5KvtsZ14ev/10.png new file mode 100644 index 0000000000000000000000000000000000000000..eb23d880fcea6f73e5965e55db1a4a3402eeb134 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6413fdf2757f5678fa486c4cadc1a58c66a5187cf4f602dcfd8ecb4920141d8 +size 521331 diff --git a/vlm/train/a5KvtsZ14ev/11.png b/vlm/train/a5KvtsZ14ev/11.png new file mode 100644 index 0000000000000000000000000000000000000000..63b410830fca7002dbdb043dfcbb10513c7fb96d --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea86f20b9952a6fc0165aee18384dfa9f35938099b0fbea46b8e539e89d6762e +size 538279 diff --git a/vlm/train/a5KvtsZ14ev/12.png b/vlm/train/a5KvtsZ14ev/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b68a3fd21ab1251c250cbafc96a537a58cda1af1 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54015ec35b8d5102282beccd708d6ba1e35c6bb77946987a35939ca241fb8c40 +size 294182 diff --git a/vlm/train/a5KvtsZ14ev/2.png b/vlm/train/a5KvtsZ14ev/2.png new file mode 100644 index 0000000000000000000000000000000000000000..41117555c345b8dc24955021c5084f65b90e35bd --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7df7ff12863eaa755c65c4130684b46742151ebc3bbb8da53e9cb7564ff6cfcb +size 533841 diff --git a/vlm/train/a5KvtsZ14ev/3.png b/vlm/train/a5KvtsZ14ev/3.png new file mode 100644 index 0000000000000000000000000000000000000000..6f43da92e342754cc740337cf411b430b01d85ec --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71f4b29b34048285fa0aa73f62830f2ac2a0079e2e349da8e7344414bfa9934a +size 505103 diff --git a/vlm/train/a5KvtsZ14ev/4.png b/vlm/train/a5KvtsZ14ev/4.png new file mode 100644 index 0000000000000000000000000000000000000000..9e882153603e3215bb3508b24dbfaf7565745a95 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1fefebfc0ab59326e2e7b006b53511959b1d94378a2d92375d458fc1076804ef +size 534316 diff --git a/vlm/train/a5KvtsZ14ev/5.png b/vlm/train/a5KvtsZ14ev/5.png new file mode 100644 index 0000000000000000000000000000000000000000..36ca4022846077f64cc43a1395e7218b69cd2c1c --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75ffa319b1309af5e09db4ec38cbcb281868c763358febb2dfe1cbdce8eba689 +size 578257 diff --git a/vlm/train/a5KvtsZ14ev/6.png b/vlm/train/a5KvtsZ14ev/6.png new file mode 100644 index 0000000000000000000000000000000000000000..48360bbc6afdce82f39021c98e5dbf263d1b62cd --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62a76ff9ebd4143030647e62d997d2eea2569452569f2524ab3042dde115fb60 +size 540138 diff --git a/vlm/train/a5KvtsZ14ev/7.png b/vlm/train/a5KvtsZ14ev/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e5ba7092b1d0eb447f34de4e3fe6075889c2e357 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:261710c8e935c9d1c7e921bbf21f62f408585b2f52de658577a55135031f6654 +size 596949 diff --git a/vlm/train/a5KvtsZ14ev/8.png b/vlm/train/a5KvtsZ14ev/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5d7bcbf30e532394238e7c388071da77155ceeab --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92d246ab60395e015ce8f4f359ba7c212f4fc3718fcaa4fd857f86e61c98e5c8 +size 394976 diff --git a/vlm/train/a5KvtsZ14ev/9.png b/vlm/train/a5KvtsZ14ev/9.png new file mode 100644 index 0000000000000000000000000000000000000000..91c7212df92063ed0ac7b87f4ae7f823aa9efaf1 --- /dev/null +++ b/vlm/train/a5KvtsZ14ev/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:490e8944c0a16d44327435d135b38cf7dbee21ebfb2bede55df8ed8b25e94e7c +size 564127 diff --git a/vlm/train/chPj_I5KMHG/0.png b/vlm/train/chPj_I5KMHG/0.png new file mode 100644 index 0000000000000000000000000000000000000000..281bb99bdd3cfa4343385cf1d2a62daef3b6c078 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84bdb4c0a3f22697cdd4a9d62eaa3cd1e5d47c2c9fdfb2747673db337ff00605 +size 448933 diff --git a/vlm/train/chPj_I5KMHG/1.png b/vlm/train/chPj_I5KMHG/1.png new file mode 100644 index 0000000000000000000000000000000000000000..fec67f24428cc42e00e3bef5118e669c2b899040 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09eac8dc7fba1ca750d03ee0e2ec49f1453acc49308c1e03fd0950aae48621fe +size 583493 diff --git a/vlm/train/chPj_I5KMHG/10.png b/vlm/train/chPj_I5KMHG/10.png new file mode 100644 index 0000000000000000000000000000000000000000..0072d1748f76e204bccffd037b7c20a59c6fab8c --- /dev/null +++ b/vlm/train/chPj_I5KMHG/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:716a77e4b4a4e1b015adaabcf82285fbd05e8278e23c0b1037f8e2b8871491f7 +size 542303 diff --git a/vlm/train/chPj_I5KMHG/11.png b/vlm/train/chPj_I5KMHG/11.png new file mode 100644 index 0000000000000000000000000000000000000000..c4935af1b51fe7ebfb2bee748bd69c113fe0131d --- /dev/null +++ b/vlm/train/chPj_I5KMHG/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6cdaec9113ed0f9246eac8a72fd0d14c2132c318a7ee95d719a9e52b19cb2180 +size 526755 diff --git a/vlm/train/chPj_I5KMHG/12.png b/vlm/train/chPj_I5KMHG/12.png new file mode 100644 index 0000000000000000000000000000000000000000..18eb528dba30b14cf034520c808f63f10084f669 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4299db583c196cd8b6c378f099eb5baff0bb6ca59d1a82059c84fe7e77d8112e +size 155743 diff --git a/vlm/train/chPj_I5KMHG/13.png b/vlm/train/chPj_I5KMHG/13.png new file mode 100644 index 0000000000000000000000000000000000000000..52397a8a6b311e4bbe830e321332db477f2bef78 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba7c6e304688123ea8d3ec8d4c212ddef3d992bc48aacede2ee14050cc9f12df +size 445214 diff --git a/vlm/train/chPj_I5KMHG/14.png b/vlm/train/chPj_I5KMHG/14.png new file mode 100644 index 0000000000000000000000000000000000000000..fb9a96ac7fb35df2ac64b8be17837d6e178ff88a --- /dev/null +++ b/vlm/train/chPj_I5KMHG/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2726331426cb0e02e06b4e7361206a49db79d06d60ecf414aff5ec4d34a9cc0 +size 522083 diff --git a/vlm/train/chPj_I5KMHG/15.png b/vlm/train/chPj_I5KMHG/15.png new file mode 100644 index 0000000000000000000000000000000000000000..e9b7f5c4ada2a01373067e2936593cfa771d0c32 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:720f90f2d633790f083a7fc12b935ce0bfd57d1ad7bb22a60775b531ce551312 +size 453024 diff --git a/vlm/train/chPj_I5KMHG/16.png b/vlm/train/chPj_I5KMHG/16.png new file mode 100644 index 0000000000000000000000000000000000000000..0468ce78cb507ae84cea8cfde678c2c2a13b6bce --- /dev/null +++ b/vlm/train/chPj_I5KMHG/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25a542073ea81347520257a12f142302d06d502261cdad5e64fe551ec21b3631 +size 522086 diff --git a/vlm/train/chPj_I5KMHG/17.png b/vlm/train/chPj_I5KMHG/17.png new file mode 100644 index 0000000000000000000000000000000000000000..eb91df1524731ac90fbc52979106366a783532cc --- /dev/null +++ b/vlm/train/chPj_I5KMHG/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eb2cf79c21a63f60f9b2c8d938712bccabf506c3b1bdca0260a11dc0e9f7370 +size 493488 diff --git a/vlm/train/chPj_I5KMHG/18.png b/vlm/train/chPj_I5KMHG/18.png new file mode 100644 index 0000000000000000000000000000000000000000..4e061cab513fdcff5675cc87280fb579d875797e --- /dev/null +++ b/vlm/train/chPj_I5KMHG/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51477062cd8092dd3ca02aed66cd2268f076aa32f4875bc84361099767bc844e +size 475831 diff --git a/vlm/train/chPj_I5KMHG/19.png b/vlm/train/chPj_I5KMHG/19.png new file mode 100644 index 0000000000000000000000000000000000000000..597d69bdba2cd8bdd09ab47e536056ad1f14ec03 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa3f53a35e9057be8c80a4030b41debb88772bdd52dce1bea6920473382c61df +size 463042 diff --git a/vlm/train/chPj_I5KMHG/2.png b/vlm/train/chPj_I5KMHG/2.png new file mode 100644 index 0000000000000000000000000000000000000000..7a8388a04b7168af47656905c75ea8ae52a6117b --- /dev/null +++ b/vlm/train/chPj_I5KMHG/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c15f3877f424388450f4123e9ebf87ab92fa4dfb96a26e942ba6570c18e519d +size 550868 diff --git a/vlm/train/chPj_I5KMHG/20.png b/vlm/train/chPj_I5KMHG/20.png new file mode 100644 index 0000000000000000000000000000000000000000..7a9a91fb4bb2b0393ac846f08d5d93e35645762b --- /dev/null +++ b/vlm/train/chPj_I5KMHG/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df1f78ceb7fda17d73f36beda8590899f84947d4617a4a8016ed32135492b877 +size 580085 diff --git a/vlm/train/chPj_I5KMHG/3.png b/vlm/train/chPj_I5KMHG/3.png new file mode 100644 index 0000000000000000000000000000000000000000..97cde16a25d8eb6e2ec5f5ff9be7cee6b1b70da9 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f65fe6aced8feabe31400b5ef3aaa0cbe8a11c1e5041fb21ac132c0970d25e7 +size 605574 diff --git a/vlm/train/chPj_I5KMHG/4.png b/vlm/train/chPj_I5KMHG/4.png new file mode 100644 index 0000000000000000000000000000000000000000..9abefe82408ca91843dd2b96ab8a8fd3c8661702 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d05f1a7344736c4c5277a95b913594b58e0ca92115a26980f13e84d75ef1751 +size 637094 diff --git a/vlm/train/chPj_I5KMHG/5.png b/vlm/train/chPj_I5KMHG/5.png new file mode 100644 index 0000000000000000000000000000000000000000..19d17870c1f39b9d1aedced526b42e2882441537 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8601e94e9d6b8e686e7b2f1a086b7a03ac885afe1b93f577f361d0e47c6e2127 +size 544154 diff --git a/vlm/train/chPj_I5KMHG/6.png b/vlm/train/chPj_I5KMHG/6.png new file mode 100644 index 0000000000000000000000000000000000000000..db89c2c9c25e7b838af1f511b061028e48a510d5 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb12a24ccb55dce33a35c0a8ec54158ce2e960b7691b241d5d172f94aa887a09 +size 621160 diff --git a/vlm/train/chPj_I5KMHG/7.png b/vlm/train/chPj_I5KMHG/7.png new file mode 100644 index 0000000000000000000000000000000000000000..30cc49976ec34e1844cbdd79a4b7db154916cbfb --- /dev/null +++ b/vlm/train/chPj_I5KMHG/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5974e8e3974bd3713fbcc327512f535b254c269f098a942fdaceb93a5da8ae4 +size 571106 diff --git a/vlm/train/chPj_I5KMHG/8.png b/vlm/train/chPj_I5KMHG/8.png new file mode 100644 index 0000000000000000000000000000000000000000..ce71e6245f531abb7a67e875a0c87b815594f786 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88c35a6616ba43089a0dea18ae00f678c21117eda873fc8138b1748e56e7eaf1 +size 579562 diff --git a/vlm/train/chPj_I5KMHG/9.png b/vlm/train/chPj_I5KMHG/9.png new file mode 100644 index 0000000000000000000000000000000000000000..af1a9c65aa17b70e7ef9d230059ba87c700796f0 --- /dev/null +++ b/vlm/train/chPj_I5KMHG/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a3c9bbe2506a136c80a170fd207f68a47cfa03e2b8e48641329e59f3a82543f +size 114947 diff --git a/vlm/train/l2UWXn5iBQI/0.png b/vlm/train/l2UWXn5iBQI/0.png new file mode 100644 index 0000000000000000000000000000000000000000..901b76e9d1d06ab96f99683ceb41423a3f912f52 --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92b72852fa6fc90d8c5f58a47ac0b80da44719eecbcc1ce28927c8531a7f4884 +size 451092 diff --git a/vlm/train/l2UWXn5iBQI/1.png b/vlm/train/l2UWXn5iBQI/1.png new file mode 100644 index 0000000000000000000000000000000000000000..3416074f791d77e868653503d54db48a8dfec3e9 --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67dd9cd71e4d51fe89167fa8b470ed00317c4f4319b7793c1224b2e533321fe0 +size 613022 diff --git a/vlm/train/l2UWXn5iBQI/10.png b/vlm/train/l2UWXn5iBQI/10.png new file mode 100644 index 0000000000000000000000000000000000000000..4f4bf7846e95c9e0bfe80417d40ab6e7df5df20e --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e565816d33a3bb50e22b800cde37fcc830cd5c4a096600177f698effa94d24e +size 663947 diff --git a/vlm/train/l2UWXn5iBQI/11.png b/vlm/train/l2UWXn5iBQI/11.png new file mode 100644 index 0000000000000000000000000000000000000000..2985c5bb5ab353b400fc272fc2b341e3dee5cefb --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:656195fe7af74835be9ba7d894154110fc3e03cbc6cea5def60ec37c2e50315c +size 702588 diff --git a/vlm/train/l2UWXn5iBQI/12.png b/vlm/train/l2UWXn5iBQI/12.png new file mode 100644 index 0000000000000000000000000000000000000000..a555f1e45576e077a55c638d2339d82d229021dc --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57c91dda5ec4eaf8b3d1ffeedf9a649732e22b4d2067ceaede1a004784409264 +size 431755 diff --git a/vlm/train/l2UWXn5iBQI/2.png b/vlm/train/l2UWXn5iBQI/2.png new file mode 100644 index 0000000000000000000000000000000000000000..ef766521bafbe017bc7a8a95e4e80d79a85fec14 --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94d0f098f1be38caa5f83357c9844476e5ea34a412a072093e634e3b80a4f4ca +size 577251 diff --git a/vlm/train/l2UWXn5iBQI/3.png b/vlm/train/l2UWXn5iBQI/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e00575f51d9bcc6811231e5d31f1b3efc2c663da --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16c91552c170f21c887a16b3b058f710fbdf4db304865a6174d2fac89a2c0e30 +size 579130 diff --git a/vlm/train/l2UWXn5iBQI/5.png b/vlm/train/l2UWXn5iBQI/5.png new file mode 100644 index 0000000000000000000000000000000000000000..3da8477aced2464ad9d5532bc380fa4d5ad153ad --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c680d17093c1ae98db479b5f099133ac524cbf03879bcf6f962de8cdebc1b4e2 +size 570811 diff --git a/vlm/train/l2UWXn5iBQI/6.png b/vlm/train/l2UWXn5iBQI/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ef9cd227429cdac7ac31ff1b14049ea98c65b91d --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f31186c61571ce947464def0cc13d3562c12a54581547d55235295c23dd4026 +size 586077 diff --git a/vlm/train/l2UWXn5iBQI/7.png b/vlm/train/l2UWXn5iBQI/7.png new file mode 100644 index 0000000000000000000000000000000000000000..a86d86b769f66af76cf91329b273f8e4d2580541 --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52d80e7997723d0cbd60edfec30dc5b173f95dc8078d237ee363c0552fa901ff +size 596000 diff --git a/vlm/train/l2UWXn5iBQI/8.png b/vlm/train/l2UWXn5iBQI/8.png new file mode 100644 index 0000000000000000000000000000000000000000..2d643a5dd5a40d50a3b4ed67b7118fbec55aec57 --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f518db86f0d546814aeed3fabfba3554dd05b5fda525b71ed60f7e5d0a9021d +size 631202 diff --git a/vlm/train/l2UWXn5iBQI/9.png b/vlm/train/l2UWXn5iBQI/9.png new file mode 100644 index 0000000000000000000000000000000000000000..819e13b2b5d922798e703a03a9595678842a0e85 --- /dev/null +++ b/vlm/train/l2UWXn5iBQI/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c6c193028b2e2d7c9dfca7f43197d717388dcb59439273ffed2b3d18f6e1aa36 +size 494256 diff --git a/vlm/train/r-oRRT-ElX/0.png b/vlm/train/r-oRRT-ElX/0.png new file mode 100644 index 0000000000000000000000000000000000000000..026e6d6dc7c553f89d487ec34cbee47512ab006b --- /dev/null +++ b/vlm/train/r-oRRT-ElX/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb93860f5c063dbce1770b1e49a8097cda4b6ff3f921d02fcedeb21bdeaf3432 +size 482836 diff --git a/vlm/train/r-oRRT-ElX/1.png b/vlm/train/r-oRRT-ElX/1.png new file mode 100644 index 0000000000000000000000000000000000000000..1d12d7a200a14bf7775beeca438e57fae1079dba --- /dev/null +++ b/vlm/train/r-oRRT-ElX/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41b2f92f418b9da4599e389eb2cbbba0ab1cff8a25213f8ed5c591047e074b5d +size 583660 diff --git a/vlm/train/r-oRRT-ElX/10.png b/vlm/train/r-oRRT-ElX/10.png new file mode 100644 index 0000000000000000000000000000000000000000..fc547363f3e95f0894f7e9811110b0b59ec12c8c --- /dev/null +++ b/vlm/train/r-oRRT-ElX/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e23cc5ec613d77843c1940bf2ad37701c39261ad7bf66253cb334726e93800f +size 574561 diff --git a/vlm/train/r-oRRT-ElX/11.png b/vlm/train/r-oRRT-ElX/11.png new file mode 100644 index 0000000000000000000000000000000000000000..f6672884687b177fe2134205274e1a573bbc3246 --- /dev/null +++ b/vlm/train/r-oRRT-ElX/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:00608493d0af8b360752380785109f14f2fffd91b86f40bdb16de3dcffe93055 +size 695788 diff --git a/vlm/train/r-oRRT-ElX/12.png b/vlm/train/r-oRRT-ElX/12.png new file mode 100644 index 0000000000000000000000000000000000000000..54a0dad8bf16393a2539a9f47096cdda566733f9 --- /dev/null +++ b/vlm/train/r-oRRT-ElX/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fc408635c3a8b27fabe869e31b11e7872d0730d8c5bb6335e873d9e58de27e5 +size 689082 diff --git a/vlm/train/r-oRRT-ElX/13.png b/vlm/train/r-oRRT-ElX/13.png new file mode 100644 index 0000000000000000000000000000000000000000..0695221e5a15b87f11f5d0830bad24a648391209 --- /dev/null +++ b/vlm/train/r-oRRT-ElX/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:856f4d3596dbb1ffdcc0c70a0ace4928fb333093ec79edeb9452a9fc5e34062f +size 673531 diff --git a/vlm/train/r-oRRT-ElX/2.png b/vlm/train/r-oRRT-ElX/2.png new file mode 100644 index 0000000000000000000000000000000000000000..6e2bc9868fcef2142d2c790ad6e09cf8fbeab22d --- /dev/null +++ b/vlm/train/r-oRRT-ElX/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b30edaad14aceb39e0237fc0381e98659270d029d29a7ba92562dd0d83c8f1a +size 595193 diff --git a/vlm/train/r-oRRT-ElX/3.png b/vlm/train/r-oRRT-ElX/3.png new file mode 100644 index 0000000000000000000000000000000000000000..296f05ca423b3b1834a1f83752c789699154214e --- /dev/null +++ b/vlm/train/r-oRRT-ElX/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:834b9aea4585113c86a8eaff3ad1af1c4c4dd5e8e1dd37b15d2b235353053f7f +size 591285 diff --git a/vlm/train/r-oRRT-ElX/4.png b/vlm/train/r-oRRT-ElX/4.png new file mode 100644 index 0000000000000000000000000000000000000000..287ee84274ff0c826a345ad181aee891a702b8a2 --- /dev/null +++ b/vlm/train/r-oRRT-ElX/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff3e8bdde6ed216a55b5ad78d42e6ab95d1c26a71e7898121b3bfa598b711cee +size 599563 diff --git a/vlm/train/r-oRRT-ElX/5.png b/vlm/train/r-oRRT-ElX/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8e2d997e594bc8814de5c851a4221044a4acba41 --- /dev/null +++ b/vlm/train/r-oRRT-ElX/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:524916877044d652f10aec8240b5362ebe617c99b3b8ca7e4baa044318b6de53 +size 621873 diff --git a/vlm/train/r-oRRT-ElX/6.png b/vlm/train/r-oRRT-ElX/6.png new file mode 100644 index 0000000000000000000000000000000000000000..635f326daed39f3e5c4ff579c046b8ab92fb6f7f --- /dev/null +++ b/vlm/train/r-oRRT-ElX/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:153c88130157e845bcdc325fa90c49793b0ea67e582547b33c04bcc0cd1f2fa2 +size 589912 diff --git a/vlm/train/r-oRRT-ElX/7.png b/vlm/train/r-oRRT-ElX/7.png new file mode 100644 index 0000000000000000000000000000000000000000..4e84e13449a18484a76805f53f78a8e5fc4dac3f --- /dev/null +++ b/vlm/train/r-oRRT-ElX/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb909f7659cc25e63d084c262c860307b1437fdeacc03e478b7c42051dc978ab +size 582763 diff --git a/vlm/train/r-oRRT-ElX/8.png b/vlm/train/r-oRRT-ElX/8.png new file mode 100644 index 0000000000000000000000000000000000000000..760da06b92d88d4d1ed54628a98355fafeb7a1f2 --- /dev/null +++ b/vlm/train/r-oRRT-ElX/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b71f2e006545760f88cdf46230434fd1e804a2959b6c6efa81311c55d01c244c +size 694585 diff --git a/vlm/train/r-oRRT-ElX/9.png b/vlm/train/r-oRRT-ElX/9.png new file mode 100644 index 0000000000000000000000000000000000000000..7e0cf4a63a4b02b68a3a304795b7c8d2ecc126fc --- /dev/null +++ b/vlm/train/r-oRRT-ElX/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:312aabdf5b7b6800069c49df88cd9e49b1348b3d0689df59a4de6a70f0efa3e4 +size 607197 diff --git a/vlm/train/r1YqWz-R-/0.png b/vlm/train/r1YqWz-R-/0.png new file mode 100644 index 0000000000000000000000000000000000000000..6b86326d3e71d30e5c19b52402033d1b84337918 --- /dev/null +++ b/vlm/train/r1YqWz-R-/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5771d77cce9820a479e88d07d5c458cab88352b56fe03a67344b245c96005d92 +size 507571 diff --git a/vlm/train/r1YqWz-R-/1.png b/vlm/train/r1YqWz-R-/1.png new file mode 100644 index 0000000000000000000000000000000000000000..725ee50aa291715a675e9cbe84d906cb0d18d622 --- /dev/null +++ b/vlm/train/r1YqWz-R-/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e7337a59d3d21d8f5bc024aa63d7805b33ac4b10b53c569b6975633815b16b5 +size 617395 diff --git a/vlm/train/r1YqWz-R-/10.png b/vlm/train/r1YqWz-R-/10.png new file mode 100644 index 0000000000000000000000000000000000000000..be3a519cf3de08aa99a0a9b38b1d19e14e918a8c --- /dev/null +++ b/vlm/train/r1YqWz-R-/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:601d5da04027bce324b7962678980adc80f2e69f742ad66caeb539fe1a4ca9ac +size 223546 diff --git a/vlm/train/r1YqWz-R-/11.png b/vlm/train/r1YqWz-R-/11.png new file mode 100644 index 0000000000000000000000000000000000000000..d46d8904e690ea20aa9922ec33dd8fb324dd6f27 --- /dev/null +++ b/vlm/train/r1YqWz-R-/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97f32b36dc79bed110b8f1b9917d0e55d0ff077e06393d6ed1db247c4ab40bde +size 404530 diff --git a/vlm/train/r1YqWz-R-/12.png b/vlm/train/r1YqWz-R-/12.png new file mode 100644 index 0000000000000000000000000000000000000000..136db8832ac15a69eeaf206bdfc2818db4b3f481 --- /dev/null +++ b/vlm/train/r1YqWz-R-/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed3681f3b91efe34337bd5a9f97e99151d1141ff3fe99c716ea70f88cf3b1e6f +size 328826 diff --git a/vlm/train/r1YqWz-R-/13.png b/vlm/train/r1YqWz-R-/13.png new file mode 100644 index 0000000000000000000000000000000000000000..56e4f194d0061ae66fca21eeaaa5b8218dea926c --- /dev/null +++ b/vlm/train/r1YqWz-R-/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79d3da1332e43cffbc111287bb5a45d326da213eaa6c0656aa9325bff6044113 +size 317572 diff --git a/vlm/train/r1YqWz-R-/2.png b/vlm/train/r1YqWz-R-/2.png new file mode 100644 index 0000000000000000000000000000000000000000..a619a9f5c1ce7374bfd4e19aeb8053e476c6569c --- /dev/null +++ b/vlm/train/r1YqWz-R-/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d0cb808d39f3e10d56bb6ae397ea41b7b9ed66e001f10c29abda20bc7c30315 +size 467075 diff --git a/vlm/train/r1YqWz-R-/3.png b/vlm/train/r1YqWz-R-/3.png new file mode 100644 index 0000000000000000000000000000000000000000..c5304c51d06293c08822f0b580235cf69492c9e6 --- /dev/null +++ b/vlm/train/r1YqWz-R-/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b44ca9fb327f18a1f218c8d11fecb8b71116d02a4709818296ccb9489b097d3 +size 527906 diff --git a/vlm/train/r1YqWz-R-/4.png b/vlm/train/r1YqWz-R-/4.png new file mode 100644 index 0000000000000000000000000000000000000000..65ed4e3d51176cc892db90edd63de0e92abcce7d --- /dev/null +++ b/vlm/train/r1YqWz-R-/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dec2dfda87f6a7a74d56d57197bf43be48774acdee105a6eb402080697fce8f9 +size 370798 diff --git a/vlm/train/r1YqWz-R-/5.png b/vlm/train/r1YqWz-R-/5.png new file mode 100644 index 0000000000000000000000000000000000000000..2a07a1af21dd254b42716b24e99b7926ce1b8c82 --- /dev/null +++ b/vlm/train/r1YqWz-R-/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebead43a99408835e677842031e2966d99a62f62143d3c0f780b1224ac1d8744 +size 553684 diff --git a/vlm/train/r1YqWz-R-/6.png b/vlm/train/r1YqWz-R-/6.png new file mode 100644 index 0000000000000000000000000000000000000000..6e0180196098240f50bcb6f708ba73f479ceb98b --- /dev/null +++ b/vlm/train/r1YqWz-R-/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d31c4419a0dd44fb5f20063c8c0b76f1c559acb433abcc95f645d8c1e7635c29 +size 506515 diff --git a/vlm/train/r1YqWz-R-/7.png b/vlm/train/r1YqWz-R-/7.png new file mode 100644 index 0000000000000000000000000000000000000000..779b562f96ad2b53bbf46fe5c63fe1434550db03 --- /dev/null +++ b/vlm/train/r1YqWz-R-/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eca5575d87e2ef8a66525aaf0d01757930a7347e636fd04712d0549b6255c547 +size 506028 diff --git a/vlm/train/r1YqWz-R-/8.png b/vlm/train/r1YqWz-R-/8.png new file mode 100644 index 0000000000000000000000000000000000000000..7823c98ce4c87a50af5e9fce9c80ed119cb78ace --- /dev/null +++ b/vlm/train/r1YqWz-R-/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d3936e1ff49d07e1d33f51a5e2de1139e829b7ba6dca34e95c377466eda541b +size 440938 diff --git a/vlm/train/r1YqWz-R-/9.png b/vlm/train/r1YqWz-R-/9.png new file mode 100644 index 0000000000000000000000000000000000000000..d897b94cf0b7fbe6789f94634d92ee6d5e30c127 --- /dev/null +++ b/vlm/train/r1YqWz-R-/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9124b763794feef28e22ffd28aef304807a008df8499ff744eb83cf9108734dc +size 518957 diff --git a/vlm/train/r1lPleBFvH/0.png b/vlm/train/r1lPleBFvH/0.png new file mode 100644 index 0000000000000000000000000000000000000000..74cf8858dc9ab0f7ba081f64e80deaca5a579951 --- /dev/null +++ b/vlm/train/r1lPleBFvH/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adda5cae19c494e03e921ff32c4966be6de264998e2e639dfc181abe1e283d9a +size 485662 diff --git a/vlm/train/r1lPleBFvH/1.png b/vlm/train/r1lPleBFvH/1.png new file mode 100644 index 0000000000000000000000000000000000000000..442a3977a5687831de2f447ddf49d585593c15d1 --- /dev/null +++ b/vlm/train/r1lPleBFvH/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a81b02fb960f60d3f0758379e1a8de472f4b3b0cc2553e8e1bbab98053ce4db +size 568698 diff --git a/vlm/train/r1lPleBFvH/10.png b/vlm/train/r1lPleBFvH/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d49e254f07cebc60d4c3c70322d69f27777861bf --- /dev/null +++ b/vlm/train/r1lPleBFvH/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9720aa808ac89043c38851390a7f546ccb6f8e9516f21ef50ea8d1c4ca0c679 +size 118900 diff --git a/vlm/train/r1lPleBFvH/11.png b/vlm/train/r1lPleBFvH/11.png new file mode 100644 index 0000000000000000000000000000000000000000..537be42ffcad73792932936625ac0483aed7afba --- /dev/null +++ b/vlm/train/r1lPleBFvH/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed94b3906996f2f528fd54f3806caf456260416883fc8ca5828e65044552a283 +size 523977 diff --git a/vlm/train/r1lPleBFvH/12.png b/vlm/train/r1lPleBFvH/12.png new file mode 100644 index 0000000000000000000000000000000000000000..ac0cd417cfa4bfbc7de2b41fb2d6307acfcd0276 --- /dev/null +++ b/vlm/train/r1lPleBFvH/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf0ed22498d59bea37cdb35fb6d9d483682fba418c83b212c47929e96406b165 +size 469708 diff --git a/vlm/train/r1lPleBFvH/13.png b/vlm/train/r1lPleBFvH/13.png new file mode 100644 index 0000000000000000000000000000000000000000..62fb641c7eb1d55df6e0cdc141d15397a5325a57 --- /dev/null +++ b/vlm/train/r1lPleBFvH/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62a537a59269cf0c263a42182236775d03a8ac4fd94935a128cf7bf671e16007 +size 389868 diff --git a/vlm/train/r1lPleBFvH/14.png b/vlm/train/r1lPleBFvH/14.png new file mode 100644 index 0000000000000000000000000000000000000000..86abe8834d0e63c4ad3ec98bb64032888cb4dc58 --- /dev/null +++ b/vlm/train/r1lPleBFvH/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcf1f1476c73dc606d112d73cf799732c789cc2b2d38753feb6c86a25940d4e9 +size 324259 diff --git a/vlm/train/r1lPleBFvH/15.png b/vlm/train/r1lPleBFvH/15.png new file mode 100644 index 0000000000000000000000000000000000000000..4d46401c62bab93ea5f9313c5b029323c4bef240 --- /dev/null +++ b/vlm/train/r1lPleBFvH/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e34be9fbe002413fce976cfcd77c356e9f9258e8811468777ca8026107e9a16 +size 190478 diff --git a/vlm/train/r1lPleBFvH/2.png b/vlm/train/r1lPleBFvH/2.png new file mode 100644 index 0000000000000000000000000000000000000000..0d1050d660bf05a9b056c8179865aba782f26512 --- /dev/null +++ b/vlm/train/r1lPleBFvH/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d1d59eeb475b6fab11e212cc856864b9e349a06052988f61fe6198af2511fa7 +size 526500 diff --git a/vlm/train/r1lPleBFvH/3.png b/vlm/train/r1lPleBFvH/3.png new file mode 100644 index 0000000000000000000000000000000000000000..aa149e2e47760e9ffb6f102f78c6ab675abd6940 --- /dev/null +++ b/vlm/train/r1lPleBFvH/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5397d598479dab392459dd96ed78973450435eb5e9fb3ff7acdd414a4f50c3ba +size 563765 diff --git a/vlm/train/r1lPleBFvH/4.png b/vlm/train/r1lPleBFvH/4.png new file mode 100644 index 0000000000000000000000000000000000000000..4dd9b813e41e4cd649902a86cd7c70acd30dd202 --- /dev/null +++ b/vlm/train/r1lPleBFvH/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12c930893ef32c28fb040b38c78d2f4d355f8a21f5457ab7d57a8ff79bf78972 +size 477456 diff --git a/vlm/train/r1lPleBFvH/5.png b/vlm/train/r1lPleBFvH/5.png new file mode 100644 index 0000000000000000000000000000000000000000..68d0294e4b2710dce4736e0d581d9e95fd542d51 --- /dev/null +++ b/vlm/train/r1lPleBFvH/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:536851efc613f799f20f054a974280ad558890e2128265851ba5d0d76959cfa7 +size 548511 diff --git a/vlm/train/r1lPleBFvH/6.png b/vlm/train/r1lPleBFvH/6.png new file mode 100644 index 0000000000000000000000000000000000000000..7fb3119a96eed917bb4291eff04bcf243ebd42f7 --- /dev/null +++ b/vlm/train/r1lPleBFvH/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8ebdc7fd4963830828716f8f64ba2ed9f3429b90fc10852a988f72c916d15b4 +size 753608 diff --git a/vlm/train/r1lPleBFvH/7.png b/vlm/train/r1lPleBFvH/7.png new file mode 100644 index 0000000000000000000000000000000000000000..47c8164ac6492017311c6aff0bc7138000e6791f --- /dev/null +++ b/vlm/train/r1lPleBFvH/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ba642f1e412057a66855a67381bff3d0761d3a1cde67dea125594175a7f5cf4 +size 506111 diff --git a/vlm/train/r1lPleBFvH/8.png b/vlm/train/r1lPleBFvH/8.png new file mode 100644 index 0000000000000000000000000000000000000000..3beba7a84d0f4cc698031c1e8b8d017688096119 --- /dev/null +++ b/vlm/train/r1lPleBFvH/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd1286bc05127aefdf9d24c0adc2c3384ede286447903c97edb1b333b6809c04 +size 552612 diff --git a/vlm/train/r1lPleBFvH/9.png b/vlm/train/r1lPleBFvH/9.png new file mode 100644 index 0000000000000000000000000000000000000000..295721dfb87a532b3bc7f7782627912138763ab2 --- /dev/null +++ b/vlm/train/r1lPleBFvH/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3746c8bcd71c9ce211909af4eed7e88528f278f8863c473851d9ba99ac22dc1f +size 554331 diff --git a/vlm/train/rkecl1rtwB/0.png b/vlm/train/rkecl1rtwB/0.png new file mode 100644 index 0000000000000000000000000000000000000000..85701e4a4343bea1a0c1b86b52f4889d744fde82 --- /dev/null +++ b/vlm/train/rkecl1rtwB/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8bfcf82800e6f620be14627b93e1b4a722ded0f69f80194aa8b71ddc852d6b6d +size 537821 diff --git a/vlm/train/rkecl1rtwB/1.png b/vlm/train/rkecl1rtwB/1.png new file mode 100644 index 0000000000000000000000000000000000000000..442c6b70473c7fa1f133115022da2b3d5ff6e40c --- /dev/null +++ b/vlm/train/rkecl1rtwB/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d5e2e5468b28b325de7396166c4b6219d082919a513da79f06553018bd289163 +size 572370 diff --git a/vlm/train/rkecl1rtwB/10.png b/vlm/train/rkecl1rtwB/10.png new file mode 100644 index 0000000000000000000000000000000000000000..f93cbbd84a4ebc02c9648126b07b78bfa22fd5a7 --- /dev/null +++ b/vlm/train/rkecl1rtwB/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dfa2c62669c451b9590b05cede43326a59a8d0e4f17b2f1d39439780ee116a7 +size 257092 diff --git a/vlm/train/rkecl1rtwB/11.png b/vlm/train/rkecl1rtwB/11.png new file mode 100644 index 0000000000000000000000000000000000000000..16a88990c17d918a6e61700ea4a8f9052fbb5fdf --- /dev/null +++ b/vlm/train/rkecl1rtwB/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b8c5fe63e709fc938854837402fbcc75200b28513b00a1c503ec22815e90425 +size 333499 diff --git a/vlm/train/rkecl1rtwB/12.png b/vlm/train/rkecl1rtwB/12.png new file mode 100644 index 0000000000000000000000000000000000000000..9d859bc19d538d826e546b6c4c7b92fc453f2b5e --- /dev/null +++ b/vlm/train/rkecl1rtwB/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5f535fd90ed4804b97b9760db2e5fdda35b48ca27e1d1e8ce0eeff667ff1114 +size 381353 diff --git a/vlm/train/rkecl1rtwB/13.png b/vlm/train/rkecl1rtwB/13.png new file mode 100644 index 0000000000000000000000000000000000000000..8948ae28d1bded83ad0aee656946ea612bd3b756 --- /dev/null +++ b/vlm/train/rkecl1rtwB/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15ced6d8cf073a5dd81cfe931f11d74ad7d417451035b3b6dd80b5021b0110a6 +size 456411 diff --git a/vlm/train/rkecl1rtwB/14.png b/vlm/train/rkecl1rtwB/14.png new file mode 100644 index 0000000000000000000000000000000000000000..7e20e47b6c1af7c5bab7ce97fc31b26df7354571 --- /dev/null +++ b/vlm/train/rkecl1rtwB/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f735519e8edb79b6f9a1b673745642c189ce29a248ebd6d4bc8b8672b07dc80b +size 518497 diff --git a/vlm/train/rkecl1rtwB/15.png b/vlm/train/rkecl1rtwB/15.png new file mode 100644 index 0000000000000000000000000000000000000000..c28967b3243d65a9aca6c821c75a40ff626d4a01 --- /dev/null +++ b/vlm/train/rkecl1rtwB/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e4d941be24ecb91d00c5bf64a5e2a0408e47c1210964e02e804a5cf525122b8 +size 601062 diff --git a/vlm/train/rkecl1rtwB/16.png b/vlm/train/rkecl1rtwB/16.png new file mode 100644 index 0000000000000000000000000000000000000000..1bc50cd7f21a32a0def5a842c819ca49b2e6ff09 --- /dev/null +++ b/vlm/train/rkecl1rtwB/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1b5af8677d5cbbc5672753dafc9c398392a8267b0070a6683f1d603030c9b96 +size 389004 diff --git a/vlm/train/rkecl1rtwB/2.png b/vlm/train/rkecl1rtwB/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c8cea3adc81b35209481ba9f8acef5af22dab93f --- /dev/null +++ b/vlm/train/rkecl1rtwB/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49fa77f9b09e02c59db7f6a625a746dd1846b92afbc44d5dd93ac6aca469891b +size 537130 diff --git a/vlm/train/rkecl1rtwB/3.png b/vlm/train/rkecl1rtwB/3.png new file mode 100644 index 0000000000000000000000000000000000000000..8870ce180f43fde9c377a53c10738640e1755649 --- /dev/null +++ b/vlm/train/rkecl1rtwB/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6f660b4e5c25563a0a4ff54084eb0df7909ced8cc493ff533a060fea34b8b28c +size 539030 diff --git a/vlm/train/rkecl1rtwB/4.png b/vlm/train/rkecl1rtwB/4.png new file mode 100644 index 0000000000000000000000000000000000000000..16735edc349f217c5ed6ea8f2a4b60d81404559d --- /dev/null +++ b/vlm/train/rkecl1rtwB/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:141ba671894b9ec121e281202b24e60e216587a34bb2c63d2c7bbefb08d5bb5f +size 502185 diff --git a/vlm/train/rkecl1rtwB/5.png b/vlm/train/rkecl1rtwB/5.png new file mode 100644 index 0000000000000000000000000000000000000000..9a8d4d180afb62e367bbf246d8df7f2d4bf53a17 --- /dev/null +++ b/vlm/train/rkecl1rtwB/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea18363aeafd2fb949337487debe268e7e21920daf353623106221a6a6bf78ab +size 618605 diff --git a/vlm/train/rkecl1rtwB/6.png b/vlm/train/rkecl1rtwB/6.png new file mode 100644 index 0000000000000000000000000000000000000000..bfef71fc466e0e683ec2a7221cb14ca51b1b2bdc --- /dev/null +++ b/vlm/train/rkecl1rtwB/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:816148eaf2d7fccae255810a15226ce9f3f0dbd72a5986ea13962f6022972f7b +size 664755 diff --git a/vlm/train/rkecl1rtwB/7.png b/vlm/train/rkecl1rtwB/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ae5ea49e670d7cd47289016af98b4e76deef77e2 --- /dev/null +++ b/vlm/train/rkecl1rtwB/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:97263eee82a23da21e7fb15242b5b80afb59fa0810dd99c8346eebf2800bc0d0 +size 564845 diff --git a/vlm/train/rkecl1rtwB/8.png b/vlm/train/rkecl1rtwB/8.png new file mode 100644 index 0000000000000000000000000000000000000000..d4f19bb346db6a9b5e0514fcacb37556d90e7ed8 --- /dev/null +++ b/vlm/train/rkecl1rtwB/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d8c81dcaea8cfd15abfe2f347421833f99c8aaeb2e3a804c0331c86b6d59f08 +size 506787 diff --git a/vlm/train/rkecl1rtwB/9.png b/vlm/train/rkecl1rtwB/9.png new file mode 100644 index 0000000000000000000000000000000000000000..8e4a82b5b802aa394528657233e8998aa9c3b149 --- /dev/null +++ b/vlm/train/rkecl1rtwB/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47361a04303a10958b88e67ccb1d1cb66a531c26fc9b117a667b0ae9c1cc91f0 +size 339709 diff --git a/vlm/train/rkl03ySYDH/0.png b/vlm/train/rkl03ySYDH/0.png new file mode 100644 index 0000000000000000000000000000000000000000..06270f7472d8cd0d9791e124c78d6711b4a1b729 --- /dev/null +++ b/vlm/train/rkl03ySYDH/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d5457ba1561c8417204706b5e7fb309455b70c0424b4497af9de744113c540f +size 495878 diff --git a/vlm/train/rkl03ySYDH/1.png b/vlm/train/rkl03ySYDH/1.png new file mode 100644 index 0000000000000000000000000000000000000000..6ec344f6f34b47f118a524e077161904320308a5 --- /dev/null +++ b/vlm/train/rkl03ySYDH/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c73685e501917fe9f0a888945c9847d3d0f9a0ea05fd1ac9e0241dd0686593f +size 541874 diff --git a/vlm/train/rkl03ySYDH/10.png b/vlm/train/rkl03ySYDH/10.png new file mode 100644 index 0000000000000000000000000000000000000000..4aafa8ba3110f119cb93320a50ae8eba861ecaaf --- /dev/null +++ b/vlm/train/rkl03ySYDH/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d1b2e386ff03165c6ea4c734063f4dca2b0086fde7eb52bf1c01979aef071b8 +size 373309 diff --git a/vlm/train/rkl03ySYDH/11.png b/vlm/train/rkl03ySYDH/11.png new file mode 100644 index 0000000000000000000000000000000000000000..9e485ad119ede2da77b72e22e34948840a855b0c --- /dev/null +++ b/vlm/train/rkl03ySYDH/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad8a015dc761e1a32931fbbe10e51fbb03fa4c9e57ecc9297533307b8546e172 +size 623390 diff --git a/vlm/train/rkl03ySYDH/12.png b/vlm/train/rkl03ySYDH/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b7e9014f235b0910eaa864c0861d6d8d7bdc330d --- /dev/null +++ b/vlm/train/rkl03ySYDH/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca4ea7b2f312b838d2ec8895402a2b6e415bcb3a403b2ea269f787cb5d452769 +size 673509 diff --git a/vlm/train/rkl03ySYDH/13.png b/vlm/train/rkl03ySYDH/13.png new file mode 100644 index 0000000000000000000000000000000000000000..be0fd7456c305f8aec9d8b3927aa68492a258b2e --- /dev/null +++ b/vlm/train/rkl03ySYDH/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e2b09e8fe83cd940fc6435bf67d9ff22513a03e4db14db3dd0c7353e11bc870 +size 911024 diff --git a/vlm/train/rkl03ySYDH/14.png b/vlm/train/rkl03ySYDH/14.png new file mode 100644 index 0000000000000000000000000000000000000000..830fecb98435f38cb70cb547f5806e1cb27a93d8 --- /dev/null +++ b/vlm/train/rkl03ySYDH/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a54a4565c3f38d8fe298cf987b5e1051d124862b0f779170e5d0eedf8399c52d +size 411197 diff --git a/vlm/train/rkl03ySYDH/15.png b/vlm/train/rkl03ySYDH/15.png new file mode 100644 index 0000000000000000000000000000000000000000..ae68a31a660577876a4151fd10be7824d3391cb8 --- /dev/null +++ b/vlm/train/rkl03ySYDH/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5965052c2724a2ea8a12754ce7d1737b7f79a45ef5e926ef070b82d91b396b6 +size 359357 diff --git a/vlm/train/rkl03ySYDH/16.png b/vlm/train/rkl03ySYDH/16.png new file mode 100644 index 0000000000000000000000000000000000000000..9a2e35407b5b360c8e4ef9c5c49a1962e88f5917 --- /dev/null +++ b/vlm/train/rkl03ySYDH/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3e5ff13309d82dddeabdccb06e90bc5d593e8423035e6d8f174c643952e555d +size 293832 diff --git a/vlm/train/rkl03ySYDH/17.png b/vlm/train/rkl03ySYDH/17.png new file mode 100644 index 0000000000000000000000000000000000000000..029bd7b0795351409febadff83dee46e9f37ea31 --- /dev/null +++ b/vlm/train/rkl03ySYDH/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:12ddd3175e7c01f22ba6114c48047d6ba2ddfca1a67ccc2a3eb2cf5178d7b3bc +size 391166 diff --git a/vlm/train/rkl03ySYDH/18.png b/vlm/train/rkl03ySYDH/18.png new file mode 100644 index 0000000000000000000000000000000000000000..1052ffaead10a1abe68023eeeb4f9d0d604220a4 --- /dev/null +++ b/vlm/train/rkl03ySYDH/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a517b528841affea596836951e47c6e194a8002c29ca6c237a30d360de923938 +size 379499 diff --git a/vlm/train/rkl03ySYDH/19.png b/vlm/train/rkl03ySYDH/19.png new file mode 100644 index 0000000000000000000000000000000000000000..a1cad2308274ec7f2487ef76109ddfd8c01035c1 --- /dev/null +++ b/vlm/train/rkl03ySYDH/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5680b14974e71ed4e77def6f436c52722a675da695c02cf399a85acae8c997b +size 297545 diff --git a/vlm/train/rkl03ySYDH/2.png b/vlm/train/rkl03ySYDH/2.png new file mode 100644 index 0000000000000000000000000000000000000000..e40bebdf6703f850665fb95a97197f0777302d8e --- /dev/null +++ b/vlm/train/rkl03ySYDH/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:339985e628481a3717e8aee2a663cf61523e3229d10c85fa5c6e0542a8b5a3af +size 544440 diff --git a/vlm/train/rkl03ySYDH/20.png b/vlm/train/rkl03ySYDH/20.png new file mode 100644 index 0000000000000000000000000000000000000000..7365b14a1ac4c6b60101fa8d2f1e39b9ac3d7f39 --- /dev/null +++ b/vlm/train/rkl03ySYDH/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:90ac7a7ab24bc214dbff885bf94b324931dfa6247d4b51245c90d4574d423b83 +size 302879 diff --git a/vlm/train/rkl03ySYDH/21.png b/vlm/train/rkl03ySYDH/21.png new file mode 100644 index 0000000000000000000000000000000000000000..01311e95bc82f741858f9f64798de26a5ef6ae28 --- /dev/null +++ b/vlm/train/rkl03ySYDH/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5cd335d46772619e367d2c8e113635a340fc16c5af02aea1a7261fb9ae048eba +size 344700 diff --git a/vlm/train/rkl03ySYDH/3.png b/vlm/train/rkl03ySYDH/3.png new file mode 100644 index 0000000000000000000000000000000000000000..6518988d28144610871586402b30e7ac88d4de87 --- /dev/null +++ b/vlm/train/rkl03ySYDH/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2d5242a335aa49af8497edd455cf2ce6771a40960eff5904d180e98d580de97 +size 472001 diff --git a/vlm/train/rkl03ySYDH/4.png b/vlm/train/rkl03ySYDH/4.png new file mode 100644 index 0000000000000000000000000000000000000000..f7649aa2a3dc82eef28b071f39942e8e54c70c2b --- /dev/null +++ b/vlm/train/rkl03ySYDH/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a64ac6e0124a75ab56a13674afe3ed79990b0b11ed735c56071c95176190a58 +size 618568 diff --git a/vlm/train/rkl03ySYDH/5.png b/vlm/train/rkl03ySYDH/5.png new file mode 100644 index 0000000000000000000000000000000000000000..83fbb57ffd3a52672103c72060835124bd9edff6 --- /dev/null +++ b/vlm/train/rkl03ySYDH/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:386b50a559a9cbb90a41e8e9e512fd0803dfee90bb4d2dc4a81f101594032487 +size 727278 diff --git a/vlm/train/rkl03ySYDH/6.png b/vlm/train/rkl03ySYDH/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ff03b13e63b66efe13d474212abbfd1b3ad7d4bf --- /dev/null +++ b/vlm/train/rkl03ySYDH/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9123554cca532138e2681ff78093407343c88509f7749954f2ce8fc94477116b +size 706828 diff --git a/vlm/train/rkl03ySYDH/7.png b/vlm/train/rkl03ySYDH/7.png new file mode 100644 index 0000000000000000000000000000000000000000..ec990715683bc148b48231d7dcd84d064f5e8d29 --- /dev/null +++ b/vlm/train/rkl03ySYDH/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8649f109a83116efdac1b6b30b836fd2ac32689443e5f7eb969f84d5785ff20c +size 595901 diff --git a/vlm/train/rkl03ySYDH/8.png b/vlm/train/rkl03ySYDH/8.png new file mode 100644 index 0000000000000000000000000000000000000000..53be71bcdf69c5489888d4e77189675e85a10695 --- /dev/null +++ b/vlm/train/rkl03ySYDH/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a2062285f3b69f4cba9e76068f37da320a31326247e9ad59d225a65796d409c +size 555638 diff --git a/vlm/train/rkl03ySYDH/9.png b/vlm/train/rkl03ySYDH/9.png new file mode 100644 index 0000000000000000000000000000000000000000..a8769c6581286e8b848312625e5f5ce7341a2fba --- /dev/null +++ b/vlm/train/rkl03ySYDH/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eeefad1669ea63c39586a0aeba5988d6597e1f00a4255abc3044ad3ee18c4526 +size 483472 diff --git a/vlm/train/rklaWn0qK7/0.png b/vlm/train/rklaWn0qK7/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d9e6959e4b9a04c9609e8ab0613a3b6eff4c361c --- /dev/null +++ b/vlm/train/rklaWn0qK7/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:449a204e95733f012efa1ab27f2c00470bc7815ed1a5ebc6c0240053e059e9ee +size 460129 diff --git a/vlm/train/rklaWn0qK7/1.png b/vlm/train/rklaWn0qK7/1.png new file mode 100644 index 0000000000000000000000000000000000000000..2fe16804dac25181d25ff8d5e5436ae66bd03270 --- /dev/null +++ b/vlm/train/rklaWn0qK7/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c49081e6e580e998d2dd84d96ee8b82b81603089bfa560067d49694953543683 +size 529731 diff --git a/vlm/train/rklaWn0qK7/10.png b/vlm/train/rklaWn0qK7/10.png new file mode 100644 index 0000000000000000000000000000000000000000..8ecbb11c8e0bc76a1e696325b4a3537b3bf0b497 --- /dev/null +++ b/vlm/train/rklaWn0qK7/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1913cfd2c32a006a702a71b211113925d01028d27e1fa07f091fe4e47a457580 +size 541752 diff --git a/vlm/train/rklaWn0qK7/11.png b/vlm/train/rklaWn0qK7/11.png new file mode 100644 index 0000000000000000000000000000000000000000..185c600d47a8f8e5ac57928ba8b5a57de7b6a240 --- /dev/null +++ b/vlm/train/rklaWn0qK7/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2ffe4d6424d916db1a3d03ec0b05834ccc872ee7ac5954af565ecd864935d1d +size 191537 diff --git a/vlm/train/rklaWn0qK7/12.png b/vlm/train/rklaWn0qK7/12.png new file mode 100644 index 0000000000000000000000000000000000000000..b14df811020f218868fe63c54b9987e2c603298e --- /dev/null +++ b/vlm/train/rklaWn0qK7/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5a764d735d444ed09d79fe30d95ec5124377797fc91cb094b1a47ef9f4079cf +size 418226 diff --git a/vlm/train/rklaWn0qK7/13.png b/vlm/train/rklaWn0qK7/13.png new file mode 100644 index 0000000000000000000000000000000000000000..7895978fbcb167b578cf6602a7f81be25cb0dda5 --- /dev/null +++ b/vlm/train/rklaWn0qK7/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96f61b758622d47c142d9a610c8d6af4ad8b72574fffc4a0c5f28d1049ed7072 +size 164708 diff --git a/vlm/train/rklaWn0qK7/2.png b/vlm/train/rklaWn0qK7/2.png new file mode 100644 index 0000000000000000000000000000000000000000..cfda99b6451e817ae59d38e7659ffbc0691c7866 --- /dev/null +++ b/vlm/train/rklaWn0qK7/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dd390abaf9504381726dca24ddad064cbe80d23d0430e7d79946b8261a76c73 +size 430230 diff --git a/vlm/train/rklaWn0qK7/3.png b/vlm/train/rklaWn0qK7/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b0f6b228f264412faab3a32df2f919318c6b6c09 --- /dev/null +++ b/vlm/train/rklaWn0qK7/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05ad78f8e73aebb61302efdd13d8dbfa5e1a5e6919130c738abb1df92751e859 +size 464951 diff --git a/vlm/train/rklaWn0qK7/4.png b/vlm/train/rklaWn0qK7/4.png new file mode 100644 index 0000000000000000000000000000000000000000..1cfb521bfdc0732f7c3b022355705b03de6c98e3 --- /dev/null +++ b/vlm/train/rklaWn0qK7/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:393d04c0427c3781fa155f3a0a78488f73f8928602bf765b62aca745856ed466 +size 552423 diff --git a/vlm/train/rklaWn0qK7/5.png b/vlm/train/rklaWn0qK7/5.png new file mode 100644 index 0000000000000000000000000000000000000000..75fe71a62428d73f394061e32d9d48c62e0aba31 --- /dev/null +++ b/vlm/train/rklaWn0qK7/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8af27d5e58034b8759e6695b7443f5c9f1230d4b56472a851b46edca6fee52f0 +size 509715 diff --git a/vlm/train/rklaWn0qK7/6.png b/vlm/train/rklaWn0qK7/6.png new file mode 100644 index 0000000000000000000000000000000000000000..81cd4b249c8b1929f8895184c68d9270804e4900 --- /dev/null +++ b/vlm/train/rklaWn0qK7/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d7a2642dc6536b38c526d4f876276204e79312660bc7e411cda9e8908d0f268 +size 498401 diff --git a/vlm/train/rklaWn0qK7/7.png b/vlm/train/rklaWn0qK7/7.png new file mode 100644 index 0000000000000000000000000000000000000000..3fedec86957e05d73151f493f21c547bed259f3f --- /dev/null +++ b/vlm/train/rklaWn0qK7/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b641f1fc7cb4bbcb7e0102ff8aaa7e43f885610120657b2a2ed909710d2eabf +size 512318 diff --git a/vlm/train/rklaWn0qK7/8.png b/vlm/train/rklaWn0qK7/8.png new file mode 100644 index 0000000000000000000000000000000000000000..7b8830de01b1bc3a8babae73dcaba91a605be481 --- /dev/null +++ b/vlm/train/rklaWn0qK7/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d53f793b0ead997238a0c89b762727c0fc74c0848a4d731f1fde80a8778040c4 +size 437632 diff --git a/vlm/train/rklaWn0qK7/9.png b/vlm/train/rklaWn0qK7/9.png new file mode 100644 index 0000000000000000000000000000000000000000..ec1552ba3ea39ba5cdc35d68ee913d8018483783 --- /dev/null +++ b/vlm/train/rklaWn0qK7/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f1bccd6122758203d978be3f514a0d09398b8a48113eebba21452b77b24a2a6 +size 336148 diff --git a/vlm/train/ryG6xZ-RZ/0.png b/vlm/train/ryG6xZ-RZ/0.png new file mode 100644 index 0000000000000000000000000000000000000000..d0f69a7b751f7499a64adf9c5d1f5310ba08c885 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcf37bdb123dee182c32a8e3270c613964baf51134e3fcabf269685e8e3bd7f1 +size 465330 diff --git a/vlm/train/ryG6xZ-RZ/1.png b/vlm/train/ryG6xZ-RZ/1.png new file mode 100644 index 0000000000000000000000000000000000000000..a9a2e979debc32881ce8cfd30f40ca61e80c9a08 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e9816feffb319d24e4b2cd77f1c54badf2c9b0985ceb92a84733c65d6370d15 +size 476082 diff --git a/vlm/train/ryG6xZ-RZ/2.png b/vlm/train/ryG6xZ-RZ/2.png new file mode 100644 index 0000000000000000000000000000000000000000..d0d22885dfaaf357dc2815e8e81a8be606aea60f --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26c6a51a036455c26cecc6219340ad3b8de73f1d3e013af86d06a4fd3beb2b2e +size 502669 diff --git a/vlm/train/ryG6xZ-RZ/3.png b/vlm/train/ryG6xZ-RZ/3.png new file mode 100644 index 0000000000000000000000000000000000000000..b4da7cc610932f8d8e9e6a0dbb64561697826909 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6848bfa82f5286d3b96e7b3f6ca5625f55466f79b1b4e77a509a8e274ff2b3cb +size 519312 diff --git a/vlm/train/ryG6xZ-RZ/4.png b/vlm/train/ryG6xZ-RZ/4.png new file mode 100644 index 0000000000000000000000000000000000000000..c32db020c33832dcdba75ca062418f4500017568 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f0239ed1f9a6610ba5fa3f7a5a132eba64241e6b1edd33238e8ceca42662709 +size 461636 diff --git a/vlm/train/ryG6xZ-RZ/5.png b/vlm/train/ryG6xZ-RZ/5.png new file mode 100644 index 0000000000000000000000000000000000000000..57fb0aae72febf9f49be8b497a0a72576b68df40 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:176bf781a8e690f055637f49a7fcc8392323ad8c4b666ca70d3d3247463ed5a2 +size 557954 diff --git a/vlm/train/ryG6xZ-RZ/6.png b/vlm/train/ryG6xZ-RZ/6.png new file mode 100644 index 0000000000000000000000000000000000000000..d8643b111bb1cfabf9ac412dd6e43df4d9f3a886 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2741c17f45cdaa1ae505cae8f85304a0ef35163499af6949842fb5e24f5f4db4 +size 603225 diff --git a/vlm/train/ryG6xZ-RZ/7.png b/vlm/train/ryG6xZ-RZ/7.png new file mode 100644 index 0000000000000000000000000000000000000000..bec1134b657372bbb8f23666f8cd019ea597b949 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1390dff57f0c52b995a60860b7d5e9d0030e413bcce4727af5e58f32f08d859e +size 441696 diff --git a/vlm/train/ryG6xZ-RZ/8.png b/vlm/train/ryG6xZ-RZ/8.png new file mode 100644 index 0000000000000000000000000000000000000000..8f01f2f1b67fad6cfcec59002d4810b742b83d9f --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54f1789bbcf7e5b7f3a449a97c9760c1b57d6b20e5f8926feca9d739c0f2e7e2 +size 613444 diff --git a/vlm/train/ryG6xZ-RZ/9.png b/vlm/train/ryG6xZ-RZ/9.png new file mode 100644 index 0000000000000000000000000000000000000000..0f769a530481aa47d50a7cd7ba3f342cbc645f99 --- /dev/null +++ b/vlm/train/ryG6xZ-RZ/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bed70cdab7f3fd0c40ec47b6ea41c3c0aa40539bb26a7c5ffd7fd51bb9288c57 +size 367608 diff --git a/vlm/train/ryenvpEKDr/0.png b/vlm/train/ryenvpEKDr/0.png new file mode 100644 index 0000000000000000000000000000000000000000..08d99d5844e83d350882822e39921db5549a8a3c --- /dev/null +++ b/vlm/train/ryenvpEKDr/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ee5b4b700057588d43b9ab1ef1dce56e95723213ce873beb299a1b91564a168 +size 493403 diff --git a/vlm/train/ryenvpEKDr/1.png b/vlm/train/ryenvpEKDr/1.png new file mode 100644 index 0000000000000000000000000000000000000000..3fa457908e8df2c6df4dfe1119cf457c3c6206cd --- /dev/null +++ b/vlm/train/ryenvpEKDr/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4daddaa6d3cadb80a326d9bab51620c308f8d8bbf380a8d13b47fdd4bd1faefa +size 565311 diff --git a/vlm/train/ryenvpEKDr/10.png b/vlm/train/ryenvpEKDr/10.png new file mode 100644 index 0000000000000000000000000000000000000000..d8c1a38b56a5b83722745aee315821c8c7b7078d --- /dev/null +++ b/vlm/train/ryenvpEKDr/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1162010fbdfb1af2bd1bd03339feb747f3e9887bfb76fb1c8a18a940a4ce0a2c +size 568596 diff --git a/vlm/train/ryenvpEKDr/11.png b/vlm/train/ryenvpEKDr/11.png new file mode 100644 index 0000000000000000000000000000000000000000..ab380f4bc6e776bdf48c33a499eaa5bd907ec2d3 --- /dev/null +++ b/vlm/train/ryenvpEKDr/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8d01df654edaa681e5aa254d7d9fb63fe648f3ff647005b573c711199a306a63 +size 577337 diff --git a/vlm/train/ryenvpEKDr/12.png b/vlm/train/ryenvpEKDr/12.png new file mode 100644 index 0000000000000000000000000000000000000000..9b56d1a56b52ecd1d7a4ec4f822aeca17a8b935d --- /dev/null +++ b/vlm/train/ryenvpEKDr/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38b89b9393cdf4cafc1d7c17745e8bcde0b616049e01b67c7855e082142fe892 +size 176091 diff --git a/vlm/train/ryenvpEKDr/13.png b/vlm/train/ryenvpEKDr/13.png new file mode 100644 index 0000000000000000000000000000000000000000..807c4904f15bc6f90121b40a2aebc0135e9ef967 --- /dev/null +++ b/vlm/train/ryenvpEKDr/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c909cff247dec09df2ddf8ffc9de641294de63306a8b4b6735125d88782b369d +size 514777 diff --git a/vlm/train/ryenvpEKDr/14.png b/vlm/train/ryenvpEKDr/14.png new file mode 100644 index 0000000000000000000000000000000000000000..d8364ada3e2e3848c0d4bc6ac1bb16102fca5ded --- /dev/null +++ b/vlm/train/ryenvpEKDr/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10bd73826910e823346b4fcac3661f936bcd707584930557fcb2c1181aacf014 +size 259249 diff --git a/vlm/train/ryenvpEKDr/15.png b/vlm/train/ryenvpEKDr/15.png new file mode 100644 index 0000000000000000000000000000000000000000..e00366d228b8e3069ad66136cedbba3dac5b3e34 --- /dev/null +++ b/vlm/train/ryenvpEKDr/15.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95ec8c85d84002516920a561b3cac74d871c6314297e3507a75e0ba0807c2366 +size 203747 diff --git a/vlm/train/ryenvpEKDr/16.png b/vlm/train/ryenvpEKDr/16.png new file mode 100644 index 0000000000000000000000000000000000000000..654527e6ca136e30ef3e7151a2047f942432f0b8 --- /dev/null +++ b/vlm/train/ryenvpEKDr/16.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:819e204116f181702b6dfb9a71bd985af9814e75f79b75e7cb2e74bb67a0f375 +size 350962 diff --git a/vlm/train/ryenvpEKDr/17.png b/vlm/train/ryenvpEKDr/17.png new file mode 100644 index 0000000000000000000000000000000000000000..ddddcf7ab11977b1c4486c54d07f2d9314b167ff --- /dev/null +++ b/vlm/train/ryenvpEKDr/17.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:461674c8a23e5e03b0a5e5227f3129f7fb4aee18ce4b4759c6d4d00db4b6c35d +size 696001 diff --git a/vlm/train/ryenvpEKDr/18.png b/vlm/train/ryenvpEKDr/18.png new file mode 100644 index 0000000000000000000000000000000000000000..7753e75259e0262343d037a8f6fd18e260bdbcae --- /dev/null +++ b/vlm/train/ryenvpEKDr/18.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c4bc3c2a6af3bfbfb5efd026eb7248c07a5f47023d1cedbde9e9b4e722e3d50f +size 710890 diff --git a/vlm/train/ryenvpEKDr/19.png b/vlm/train/ryenvpEKDr/19.png new file mode 100644 index 0000000000000000000000000000000000000000..ec7192997e1ec4791e041987aca3d59af7c278e2 --- /dev/null +++ b/vlm/train/ryenvpEKDr/19.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98a0d1b7870fac96c3fbe21a776899390082a3d0934fb73cc4c4a3ec1a4e1fd0 +size 735370 diff --git a/vlm/train/ryenvpEKDr/2.png b/vlm/train/ryenvpEKDr/2.png new file mode 100644 index 0000000000000000000000000000000000000000..c6f05d0ebd31c5e67797a4e21b26b5852a00dd4a --- /dev/null +++ b/vlm/train/ryenvpEKDr/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c4d84f1dd6ebc5ac2031b90f49f81e3fc5f63873bdaf16dcbe37ceee72e7c6d +size 483542 diff --git a/vlm/train/ryenvpEKDr/20.png b/vlm/train/ryenvpEKDr/20.png new file mode 100644 index 0000000000000000000000000000000000000000..2ab4edfc8f3d49e95900e54bcec456128247f378 --- /dev/null +++ b/vlm/train/ryenvpEKDr/20.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b396f31554b86bc4a9844379324a04afa7f0a22417b41db4ab8b50aa61ce2a2 +size 374419 diff --git a/vlm/train/ryenvpEKDr/21.png b/vlm/train/ryenvpEKDr/21.png new file mode 100644 index 0000000000000000000000000000000000000000..6c81c28861ad3c566ae89e0903311bea5d75ba83 --- /dev/null +++ b/vlm/train/ryenvpEKDr/21.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1a882fe7426c214466df71c1383e262f60009f0fb78aaf247b1a4202183a5620 +size 318173 diff --git a/vlm/train/ryenvpEKDr/22.png b/vlm/train/ryenvpEKDr/22.png new file mode 100644 index 0000000000000000000000000000000000000000..53971fc6686f4c31a4020fe705e1fc06d78939fd --- /dev/null +++ b/vlm/train/ryenvpEKDr/22.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:83212d6833d69f44b4093e1c7f81eb50e9c84225c4d30cc1d57a3826568e3ffc +size 255450 diff --git a/vlm/train/ryenvpEKDr/23.png b/vlm/train/ryenvpEKDr/23.png new file mode 100644 index 0000000000000000000000000000000000000000..7ce89cfc6c0a60cd446c09a2d18e490f0b2efb59 --- /dev/null +++ b/vlm/train/ryenvpEKDr/23.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4f4e8610102abbe1d7e529a24f8997e34f36a1bae794546972f5cf7f7dc59a7 +size 228499 diff --git a/vlm/train/ryenvpEKDr/24.png b/vlm/train/ryenvpEKDr/24.png new file mode 100644 index 0000000000000000000000000000000000000000..b88026ac443a419f858bf0cff8a4f820e5675ffb --- /dev/null +++ b/vlm/train/ryenvpEKDr/24.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4b71870a0cd261b94a7e1e836e4005836c2cd80286c3c8440d6eed2ddb57e96a +size 218984 diff --git a/vlm/train/ryenvpEKDr/25.png b/vlm/train/ryenvpEKDr/25.png new file mode 100644 index 0000000000000000000000000000000000000000..ebdac0dd51186a739a606bb353b5876e99d905be --- /dev/null +++ b/vlm/train/ryenvpEKDr/25.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e9b60421fc19d27a382b9cd0d5548c3f305e8d2180e1e1ee79499b65ee93ff8 +size 252579 diff --git a/vlm/train/ryenvpEKDr/26.png b/vlm/train/ryenvpEKDr/26.png new file mode 100644 index 0000000000000000000000000000000000000000..54457b658e2acdbf941a5d80f7dda079b0d04d71 --- /dev/null +++ b/vlm/train/ryenvpEKDr/26.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09c2346a2c8981183691c34db63a1d740e6b81164d2bf1e17624261aeca89e51 +size 258304 diff --git a/vlm/train/ryenvpEKDr/27.png b/vlm/train/ryenvpEKDr/27.png new file mode 100644 index 0000000000000000000000000000000000000000..8a921d1dc70a7d41bfeccdea8e7258224ec50887 --- /dev/null +++ b/vlm/train/ryenvpEKDr/27.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c321a948c164308cf126d573160236b3804139df7ce99d267dfa7ee52001da75 +size 241553 diff --git a/vlm/train/ryenvpEKDr/28.png b/vlm/train/ryenvpEKDr/28.png new file mode 100644 index 0000000000000000000000000000000000000000..f851ac466e397b59d59c72f9b288ad0713e28031 --- /dev/null +++ b/vlm/train/ryenvpEKDr/28.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93f92b0b16255f85a213758e815e5bfa75ac7125d62ca1227b9828230e006795 +size 252353 diff --git a/vlm/train/ryenvpEKDr/29.png b/vlm/train/ryenvpEKDr/29.png new file mode 100644 index 0000000000000000000000000000000000000000..8497605bbc74f73e7f4da9e76f46730fe128d8aa --- /dev/null +++ b/vlm/train/ryenvpEKDr/29.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:79b9ea18bfccca4e0fa81a3d85dc6f17d43c2d95afcc6f95610196fdacb149d2 +size 266131 diff --git a/vlm/train/ryenvpEKDr/3.png b/vlm/train/ryenvpEKDr/3.png new file mode 100644 index 0000000000000000000000000000000000000000..e7cf3a4f29c6ec3e31b2f1a2596236205712ddb6 --- /dev/null +++ b/vlm/train/ryenvpEKDr/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf18969403fa2cf9791ae4907000cca7a058134d09765826e42aad7c0365c1bc +size 654791 diff --git a/vlm/train/ryenvpEKDr/4.png b/vlm/train/ryenvpEKDr/4.png new file mode 100644 index 0000000000000000000000000000000000000000..a4830160edfb6502306f61f3b5aac06ca1f8b697 --- /dev/null +++ b/vlm/train/ryenvpEKDr/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50b63359d7c3431df71e0727be8d81184dfd8f9314b221bbd8a0724d6141edac +size 475092 diff --git a/vlm/train/ryenvpEKDr/5.png b/vlm/train/ryenvpEKDr/5.png new file mode 100644 index 0000000000000000000000000000000000000000..8fdafdbb750d4c6706655728bd77c77b86e66b7d --- /dev/null +++ b/vlm/train/ryenvpEKDr/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b7b974e4fabe9866f25a6e99c89ef7aeeb0090d7b3a5ec27597d59b4335826be +size 490483 diff --git a/vlm/train/ryenvpEKDr/6.png b/vlm/train/ryenvpEKDr/6.png new file mode 100644 index 0000000000000000000000000000000000000000..3cdc6914b652e848e41d7c8b2dd5688d8ce8b6aa --- /dev/null +++ b/vlm/train/ryenvpEKDr/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e44b3eefedb2a882350455ee32cd216c4b7d8aca002659727c77cdd2471e9825 +size 506289 diff --git a/vlm/train/ryenvpEKDr/7.png b/vlm/train/ryenvpEKDr/7.png new file mode 100644 index 0000000000000000000000000000000000000000..f5e7d36dc6ab15d146a27daeb733831ddc7e1122 --- /dev/null +++ b/vlm/train/ryenvpEKDr/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:745bbd3db67aec09043fa9e11ca3587b10dbc29f9c182ed0ddda34aeff5e0819 +size 542520 diff --git a/vlm/train/ryenvpEKDr/8.png b/vlm/train/ryenvpEKDr/8.png new file mode 100644 index 0000000000000000000000000000000000000000..22f9d85af52c9c2883635f6934f4fd114fd0e7a4 --- /dev/null +++ b/vlm/train/ryenvpEKDr/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2027467a30ead8276e31aab4f0a9eca694f8f002605a6b437457031ce269c58b +size 530331 diff --git a/vlm/train/ryenvpEKDr/9.png b/vlm/train/ryenvpEKDr/9.png new file mode 100644 index 0000000000000000000000000000000000000000..da0917b4022b71e85d61ac6bf13f6188d0bad746 --- /dev/null +++ b/vlm/train/ryenvpEKDr/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a756d8d0b6db1943d8ec43ce7b366bfe227dc8fdd9ae88ebde39c715f23b2e5 +size 520799 diff --git a/vlm/train/v8b3e5jN66j/0.png b/vlm/train/v8b3e5jN66j/0.png new file mode 100644 index 0000000000000000000000000000000000000000..9fe1ad008c926a0d84538af3a5f99f9523ab040f --- /dev/null +++ b/vlm/train/v8b3e5jN66j/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8fb7f00ca0e53e4b8f81c4949b366289d92c8db0347b1f8e3d103a0e600b98c1 +size 517302 diff --git a/vlm/train/v8b3e5jN66j/1.png b/vlm/train/v8b3e5jN66j/1.png new file mode 100644 index 0000000000000000000000000000000000000000..1297f6ada68a11428449d8d7230f4d5c1ec03857 --- /dev/null +++ b/vlm/train/v8b3e5jN66j/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64c2b3c12276a88246cc7ced34ba82444f186fa78eb6f1400dd1b24a141fc558 +size 564622 diff --git a/vlm/train/v8b3e5jN66j/10.png b/vlm/train/v8b3e5jN66j/10.png new file mode 100644 index 0000000000000000000000000000000000000000..268aa66d7fba6e9397a4afc5164dd637357688ea --- /dev/null +++ b/vlm/train/v8b3e5jN66j/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c0030ef4a5e305242ded5d7a1a1699e9d6a67774b39af1976503db1fb9b5fdc3 +size 580347 diff --git a/vlm/train/v8b3e5jN66j/11.png b/vlm/train/v8b3e5jN66j/11.png new file mode 100644 index 0000000000000000000000000000000000000000..13c493ab88280523c3b12fc4c478ddde9994057c --- /dev/null +++ b/vlm/train/v8b3e5jN66j/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b59e36884787ff2337067d66fa3bd8f52fec6316f36304de3eceaa0bdb2424c6 +size 501164 diff --git a/vlm/train/v8b3e5jN66j/2.png b/vlm/train/v8b3e5jN66j/2.png new file mode 100644 index 0000000000000000000000000000000000000000..0e5a0088aec7c0f3319b07f68085248370584ac9 --- /dev/null +++ b/vlm/train/v8b3e5jN66j/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1efb776022654379f11b73bee6d19bdfa55520e86f3e5f5410839cd9f826a37b +size 577398 diff --git a/vlm/train/v8b3e5jN66j/3.png b/vlm/train/v8b3e5jN66j/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0a57628f56a2c5f88d129d7844c0d8f5db3578fd --- /dev/null +++ b/vlm/train/v8b3e5jN66j/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf4a7e78ab71aa813cd54f31e637b40d707420e33500be16ce778afe67a4990f +size 592238 diff --git a/vlm/train/v8b3e5jN66j/4.png b/vlm/train/v8b3e5jN66j/4.png new file mode 100644 index 0000000000000000000000000000000000000000..61a4a62b53f6894db3e0e376ede18255825904d7 --- /dev/null +++ b/vlm/train/v8b3e5jN66j/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9207fca6d3749257a68240c45a57d14fc172a9cc6d6df3ef0db81d82da68953 +size 627218 diff --git a/vlm/train/v8b3e5jN66j/5.png b/vlm/train/v8b3e5jN66j/5.png new file mode 100644 index 0000000000000000000000000000000000000000..0d1563f45ba7e8f670087a4f1cc12df89919305b --- /dev/null +++ b/vlm/train/v8b3e5jN66j/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da45f7c749a972d75e4c9ca33656ebe1f349d2ff904bc2b4f44c6bd799519453 +size 616998 diff --git a/vlm/train/v8b3e5jN66j/6.png b/vlm/train/v8b3e5jN66j/6.png new file mode 100644 index 0000000000000000000000000000000000000000..294e8be3da816b55710cc6b310e2f83d4abc9c94 --- /dev/null +++ b/vlm/train/v8b3e5jN66j/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d10601f10291c932ed54b1e3a8d0c374cd7688b01aeae07be9e82e0da60a8d40 +size 538667 diff --git a/vlm/train/v8b3e5jN66j/7.png b/vlm/train/v8b3e5jN66j/7.png new file mode 100644 index 0000000000000000000000000000000000000000..4ed3730d9d788e39990994727f9aabddb925860c --- /dev/null +++ b/vlm/train/v8b3e5jN66j/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9159a7b1746fd46e2b53bcbae7df06319ed7ce158e9d815244200aa4e839cf5f +size 660116 diff --git a/vlm/train/v8b3e5jN66j/8.png b/vlm/train/v8b3e5jN66j/8.png new file mode 100644 index 0000000000000000000000000000000000000000..c679e1dd32ba0926877f98b1f289719a808a7e4f --- /dev/null +++ b/vlm/train/v8b3e5jN66j/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:030042b50a2ae61e1795a6bc1c971dcbb73d740bdb86a7ed48d3d50de1b902a5 +size 580945 diff --git a/vlm/train/v8b3e5jN66j/9.png b/vlm/train/v8b3e5jN66j/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e8301dbb3284554e93880941ad53953904642584 --- /dev/null +++ b/vlm/train/v8b3e5jN66j/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:634a1ec6d644ebaa7a6162441dcb6d0fc9475ca47112f141e2cf9de572c6b58b +size 528010 diff --git a/vlm/train/xN3XX6pKSD5/0.png b/vlm/train/xN3XX6pKSD5/0.png new file mode 100644 index 0000000000000000000000000000000000000000..48551f488566e8232e2fb5a7aaa99a225598128e --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78a1382c6044ee694d5f6ad04c209363288c025354e2c8295e7c8a959b3c5d24 +size 441107 diff --git a/vlm/train/xN3XX6pKSD5/1.png b/vlm/train/xN3XX6pKSD5/1.png new file mode 100644 index 0000000000000000000000000000000000000000..852d4a5d1c723615260d2242b217daa3ce5c0e5a --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46bc25781082a1de6c67c271bbf818d685482048a3cb007361b916dd68f3c560 +size 611531 diff --git a/vlm/train/xN3XX6pKSD5/10.png b/vlm/train/xN3XX6pKSD5/10.png new file mode 100644 index 0000000000000000000000000000000000000000..9bb20abf231abe3811b54101ce1d8a62c9c2fbc2 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21d436c8f4b96d620487417f6eedbf39da24e7cab24f7898c1a6bf23c55a2ee5 +size 571098 diff --git a/vlm/train/xN3XX6pKSD5/11.png b/vlm/train/xN3XX6pKSD5/11.png new file mode 100644 index 0000000000000000000000000000000000000000..9c87f0a3fa0ccea9b769b38f9b6e3c2b2990f763 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:051860a8ade3d088f1f54006449bcb6ca8ac58428d260ee93fcf814da67eb506 +size 575229 diff --git a/vlm/train/xN3XX6pKSD5/12.png b/vlm/train/xN3XX6pKSD5/12.png new file mode 100644 index 0000000000000000000000000000000000000000..050d692889b3863e004aff16e06872cdd3945492 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/12.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:137b96c0cfcbeeeb084522950419ebe14b477e7091fe2d79b740bd471a81571e +size 572058 diff --git a/vlm/train/xN3XX6pKSD5/13.png b/vlm/train/xN3XX6pKSD5/13.png new file mode 100644 index 0000000000000000000000000000000000000000..b14c297f4efdbc30309b0bcb8b2781671396f784 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/13.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3c823132d9ddde88033bc51368d00547124ed970c9b18d3ddaa686c86374d2a +size 69379 diff --git a/vlm/train/xN3XX6pKSD5/14.png b/vlm/train/xN3XX6pKSD5/14.png new file mode 100644 index 0000000000000000000000000000000000000000..028531ca413da33032ca13a4894fe1a53a300658 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/14.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87a5ef8eec972ca22c947e37a01414a19728f5f07a1221a3e738c110a77ebadf +size 457887 diff --git a/vlm/train/xN3XX6pKSD5/2.png b/vlm/train/xN3XX6pKSD5/2.png new file mode 100644 index 0000000000000000000000000000000000000000..94fcd1feb9aa9919f9f5927d5d4c59ccd491e45e --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:129b8ac884ab5c1b8471839c8c49a3f1f97b4882a9cde769ad71bb92069cd449 +size 692071 diff --git a/vlm/train/xN3XX6pKSD5/3.png b/vlm/train/xN3XX6pKSD5/3.png new file mode 100644 index 0000000000000000000000000000000000000000..0543f3038e2510d1f1f0ea26278afc820c06ee48 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:673933f0677d60c53ad9795f1765ef533927ad7e621564a243396ece75aa3c14 +size 622402 diff --git a/vlm/train/xN3XX6pKSD5/4.png b/vlm/train/xN3XX6pKSD5/4.png new file mode 100644 index 0000000000000000000000000000000000000000..490c4a79ce51500bfeb1dcb51eedce45b465110d --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13ccff2b8d3800c7b7a034d9ff81de9e908cfea27eef29552e33d606a6651fff +size 663803 diff --git a/vlm/train/xN3XX6pKSD5/5.png b/vlm/train/xN3XX6pKSD5/5.png new file mode 100644 index 0000000000000000000000000000000000000000..28518951a065f17ffea40b53a4e4a2770c456b05 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6dde2976df4b53f020baad325f23b19fcea7ecad46cdc3889b101a16e392107a +size 529946 diff --git a/vlm/train/xN3XX6pKSD5/6.png b/vlm/train/xN3XX6pKSD5/6.png new file mode 100644 index 0000000000000000000000000000000000000000..ab1933a939585d56067bc45f568d78123290d677 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd6b135d51d135d8e6ab040b65b7903d5e68b409d35a2df876c8594be123cf28 +size 1182501 diff --git a/vlm/train/xN3XX6pKSD5/7.png b/vlm/train/xN3XX6pKSD5/7.png new file mode 100644 index 0000000000000000000000000000000000000000..e0415d04ce854f9973f9bdde3df0c5436bc76584 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ecf23904e1dce979832ef75e4a72b29377093bc88e1b743021b35a0a31e6aef +size 627052 diff --git a/vlm/train/xN3XX6pKSD5/8.png b/vlm/train/xN3XX6pKSD5/8.png new file mode 100644 index 0000000000000000000000000000000000000000..5f342c3a495adc4de77ee3605d6914150b2c24fe --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b19b0422643cb3c577eb0c95ab29e662148fe3176e3574d5ba49aeb7af216de1 +size 1051087 diff --git a/vlm/train/xN3XX6pKSD5/9.png b/vlm/train/xN3XX6pKSD5/9.png new file mode 100644 index 0000000000000000000000000000000000000000..5ce6df78e24b57d346b694ad776138b2d49ad9d0 --- /dev/null +++ b/vlm/train/xN3XX6pKSD5/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bb7e315ee2e1de9f1634b55e3a88bf5af899ddbbc4950b694cef3c8157b1590 +size 666310 diff --git a/vlm/train/z-l1kpDXs88/0.png b/vlm/train/z-l1kpDXs88/0.png new file mode 100644 index 0000000000000000000000000000000000000000..ab033ce17a4c0072d2aef3a473d72b3dbf1a2fa5 --- /dev/null +++ b/vlm/train/z-l1kpDXs88/0.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2b0cfbba503a32b3acc77ebf46753c01c5bd294bf01e561d8a425047f2c275f +size 427773 diff --git a/vlm/train/z-l1kpDXs88/1.png b/vlm/train/z-l1kpDXs88/1.png new file mode 100644 index 0000000000000000000000000000000000000000..5f7b8e664918c6be7ba005284ca3ddfe752c34ec --- /dev/null +++ b/vlm/train/z-l1kpDXs88/1.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:204ba8de44d87f6f30b94d76e25b65cb0465a11c73cd64be655e409642aeb326 +size 422969 diff --git a/vlm/train/z-l1kpDXs88/10.png b/vlm/train/z-l1kpDXs88/10.png new file mode 100644 index 0000000000000000000000000000000000000000..2f512fa2dc0116177f5be5055b28569bf52418f5 --- /dev/null +++ b/vlm/train/z-l1kpDXs88/10.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd3e16982d0a6b13231903961875585989c4289561c40898f343370d644d3bbc +size 577781 diff --git a/vlm/train/z-l1kpDXs88/11.png b/vlm/train/z-l1kpDXs88/11.png new file mode 100644 index 0000000000000000000000000000000000000000..0f441ab31d2889845b57c014b4567f9ea281f0fe --- /dev/null +++ b/vlm/train/z-l1kpDXs88/11.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d60b592740be271b13b4556a7ec4deba3e6b90338e6e663fc0ca266d59204b06 +size 638488 diff --git a/vlm/train/z-l1kpDXs88/2.png b/vlm/train/z-l1kpDXs88/2.png new file mode 100644 index 0000000000000000000000000000000000000000..371760296c38ce637fb5a397000eae5e62f383aa --- /dev/null +++ b/vlm/train/z-l1kpDXs88/2.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36c2c5b92e6363d1259d01de413f15f71393be8d1839e68860008e66dbb2e935 +size 474058 diff --git a/vlm/train/z-l1kpDXs88/3.png b/vlm/train/z-l1kpDXs88/3.png new file mode 100644 index 0000000000000000000000000000000000000000..9bafdad626af669bb519628ce24b71d160abbc80 --- /dev/null +++ b/vlm/train/z-l1kpDXs88/3.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:75b87485e72dcd228297af39d5ea0e2423fb5dcd55ce469b8d2fa0a344d527b8 +size 450230 diff --git a/vlm/train/z-l1kpDXs88/4.png b/vlm/train/z-l1kpDXs88/4.png new file mode 100644 index 0000000000000000000000000000000000000000..168728b440520bc6ac3e0d16a2f99b4e3906652c --- /dev/null +++ b/vlm/train/z-l1kpDXs88/4.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e91ecbe73673f869db86f5752a1836d7ed6eb89897c778a046b61f10ddc27259 +size 541248 diff --git a/vlm/train/z-l1kpDXs88/5.png b/vlm/train/z-l1kpDXs88/5.png new file mode 100644 index 0000000000000000000000000000000000000000..549a9658301d700ef9a7784b8b31c28b542a7084 --- /dev/null +++ b/vlm/train/z-l1kpDXs88/5.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae504b9a4622d5785ce2ebd1c43b095754e1588e652d5e483aba3cd4fcb3ed48 +size 447002 diff --git a/vlm/train/z-l1kpDXs88/6.png b/vlm/train/z-l1kpDXs88/6.png new file mode 100644 index 0000000000000000000000000000000000000000..b132258d166bf7dd8d7fcb3019eddcbd6c48c6c3 --- /dev/null +++ b/vlm/train/z-l1kpDXs88/6.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdb2f4f446743acac31f2b200e32eec5187a066aec241ca7ad3bf25979573d19 +size 520017 diff --git a/vlm/train/z-l1kpDXs88/7.png b/vlm/train/z-l1kpDXs88/7.png new file mode 100644 index 0000000000000000000000000000000000000000..d8222105b003949f08d1dd1ffa487787d0e1f00c --- /dev/null +++ b/vlm/train/z-l1kpDXs88/7.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5330d3f2c1a27dd50379b264cccc82ea17bddea5c23ba6e4110522496d47fe27 +size 472933 diff --git a/vlm/train/z-l1kpDXs88/8.png b/vlm/train/z-l1kpDXs88/8.png new file mode 100644 index 0000000000000000000000000000000000000000..4bb3ccb4a9aca638fb07f4f6a148464df74067af --- /dev/null +++ b/vlm/train/z-l1kpDXs88/8.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5352089486e6fcd4ecea4d78aae04510c494d0c49eb61a119548507d8ce9d496 +size 541516 diff --git a/vlm/train/z-l1kpDXs88/9.png b/vlm/train/z-l1kpDXs88/9.png new file mode 100644 index 0000000000000000000000000000000000000000..e934c11c67a2a18c12adcbcf9471dfa4fa369829 --- /dev/null +++ b/vlm/train/z-l1kpDXs88/9.png @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e1bf4af1425b68ab662b629f3f74df62a6d3f7a42d88c389502f0ceb96c1b6f2 +size 518410